diff --git a/.agents/issue-index.md b/.agents/issue-index.md
index 1cea113a2..194e076e0 100644
--- a/.agents/issue-index.md
+++ b/.agents/issue-index.md
@@ -692,6 +692,7 @@ rather than merged. `scripts/check-agent-record.py` gates both.
 | [#1843](https://github.com/mudler/vllm.cpp/issues/1843) | `FIX-FP8-PLAN-CAPTURE-1843` | The fp8 cuBLASLt lane still queries `cublasLtMatmulAlgoGetHeuristic` inside CUDA-graph capture because `VT_FP8_PLAN_CACHE` ships OFF (`fp8_plan_cache.h:49-59` @ `364f2a898`), so a captured decode on an fp8-tower model dies on CUDA 13.3 even with PR #1741 -- measured on `dgx:gpu0` (GB10, staged CUDA 13.3.73): #1741 alone fails on the fp8 lane, `VT_FP8_PLAN_CACHE=1` alone fails on the bf16-TN lane, both together pass the graphed 35B gate token-exact on all three arms. The fix is the default flip #1741's spec owed, with the same not-a-performance-knob polarity argument its `gemm_plan_cache.h` records. Claimed by row `FIX-FP8-PLAN-CAPTURE-1843` ([spec](specs/fix-fp8-plan-capture.md)) | bug |
 | [#1849](https://github.com/mudler/vllm.cpp/issues/1849) | `SPEC-DFLASH2` | **The DFlash2 draft step costs a flat ~23 ms at EVERY K, and the two levers #1849 names resolve differently once read from the records.** Lever A (quantize the shared head) is ALREADY LANDED for the measured subject: `r0b0tlab/Qwen3.8-27B-NVFP4-MTP-sm121` stores `lm_head` as W4A16_NVFP4 g16 (header-verified 2026-08-21, quantization-matrix `QUANT-QWEN38-27B-NVFP4-ARM`), upstream computes with it packed through `lm_head.quant_method.apply`, and both our reads have been packed since #1628 — so the head traffic is ~2×0.72 GB not 2×2.54, the draft-phase floor re-derives to ~9 ms, and the unattributed residual GROWS to ~13-14 ms. Lever B (launch/sync trim) is counted in code at one replay + ~10 launches + ~76 B up / 64 B down + one sync — well under 0.5 ms, so the residual sits INSIDE kernels and needs on-box attribution. W9 lands `VT_SPEC_TRACE=2` (the `[spec-phase-dev]` pre/fwd/select/walk split) as the instrument, and borrow-first loading for the draft's shared bf16 embed+head (~5.1 GB host on the bf16 arm, ~2.5 GB on the r0b0tlab arm; memory only, no step-time claim). The bf16-target arm's 2×2.54 GB head reads are upstream's own serving dtype and stand as a recorded ceiling. Wave spec [dflash2-draft-fixed-cost.md](specs/dflash2-draft-fixed-cost.md); the K-ladder rerun, the `ncu`/`nsys` attribution and any step delta are owed there, operator-run | perf |
 | [#1844](https://github.com/mudler/vllm.cpp/issues/1844) | `ENG-MM-INPUT-PIPELINE` | **`scripts/mm/tower_skip_rss.sh` killed every measured leg mid-load, because `run_arm`'s `/health` poll was answered by the PREVIOUS leg's server.** First real run (`thor:gpu0`, worker `rc-worker-kk96r`, `d60692c8`): checkpoint staged and verified (29 files, 8887294190 B), both binaries built sha256-identical (`78d582e4...`), live target query green on both build dirs -- and then **five 0-byte `.time` files** and `VOID` on both pairs. `warmup` reached `listening on http://0.0.0.0:18607` (1286 B log); all four measured legs stopped at `loading model from ...` (363/345 B) inside one minute. TWO defects, one shape. (a) `$PORT` is fixed and the poll starts immediately, so a stale listener answers it and the leg is ready before it has read a tensor. (b) `kill "$pid"` signals `/usr/bin/time`, which installs no handler: the timer dies before writing its `-o` file and the server is reparented to init and KEEPS THE PORT -- which is what was answering. Measured: `/usr/bin/time -v -o f sleep 100 & kill $!` leaves `f` at 0 B and `sleep` alive with ppid 1; signalling the CHILD leaves `f` at 752 B with a `Maximum resident set size` line. **Nothing could catch it**: `test_tower_skip_rss_report.py` was 60/60 green over finished files, and `run_arm`, the poll and the teardown ran only under a lease -- the residual [#1819](https://github.com/mudler/vllm.cpp/issues/1819) recorded verbatim, now observed. FIXED IN FLOW: a leg refuses to start into an occupied port; readiness requires the banner in the leg's OWN log (its stdout, which no other server can write) before `/health`; SIGTERM goes to the SERVER so the timer survives to write; the port must stop accepting before the next leg; each of the three waits is bounded and each bound REFUSES; and a leg whose `.time` carries no `Maximum resident set size` line fails AT that leg rather than as VOID four legs later. GATED: `TOWER_SKIP_RSS_SOURCE_ONLY=1` sources the harness for its functions alone and `tests/scripts/test_tower_skip_rss_arm.py` drives `run_arm` against a fake server on a scratch port -- stale listener, absent banner, death during load, the happy path's non-empty `.time`, five legs in the declared order, and both halves restored as mutations. RED-first: 11 of 14 cases fail against the pre-fix `run_arm`, the stale-listener case reporting `LEG default OK` with a 0-byte `.time` | bug |
+| [#1846](https://github.com/mudler/vllm.cpp/issues/1846) | `MODEL-MM-dots3-note-dots3-note-for-causal-lm` | **The released `dots3-note-prev` shard index declares `indexer_rope_layout: "leading"` and `indexer_rope_converted_from: "tail"` in its `metadata` block, and NOTHING reads either key** — `git grep indexer_rope_layout` over vLLM `origin/main` returns nothing. Measured at W2 while reading the whole index. It is the publisher stating how the DSA indexer's `wq_b`/`wk` are laid out along the 128-wide index head, and it agrees with what upstream's code does anyway: `DeepseekV2Indexer` rotates `[..., :rope_dim]` and leaves `[..., rope_dim:]` (`deepseek_v2.py:805,:814`, `rope_dim` 64 of `index_head_dim` 128), which is a LEADING slice. NOT spec §4 trap 2: that one is about which PAIRS rope rotates (GPT-J vs NeoX), this one about which HALF of the head it rotates, and both are numerically silent on a row spec §6.4 says has no oracle. W2 pins both values in an assertion so a re-published checkpoint cannot flip the layout silently; W2 consumes neither, because W2 writes no maths. W3 owes the slice | feature |
 | [#1857](https://github.com/mudler/vllm.cpp/issues/1857) | `SPEC-DFLASH2` | **The q>1 DFlash2 verify rides the PREFILL flash lane (`is_prefill = num_tokens > num_reqs`, `fa2_decode` hard-requires `num_tokens == num_reqs`), costing +9 ms/step from q=2 to q=9 on the #1574 K-ladder — the last attributed gap against SGLang (27.60 vs 25.07 tok/s at equal acceptance, ~109 vs ~122 ms/step, both paying the same ~20 ms draft).** W10 mirrors upstream's spec-as-decode: the reorder-threshold policy `1 + (parallel_drafting ? 2 : 1) * K` (`backend.py:718-736` @ `b389ac2946`, identical at the pin) classifies the runner's already-verified uniform verify length onto the decode class, the classification travels `CommonAttentionMetadata -> PagedAttentionArgs`, and a new ADDITIVE d256 launcher serves it with the exact presentation upstream `mha_fwd_kvcache` uses at seqlen_q>1 — batched split-KV, bottom-right causal against `seqused_k` (the draft mask with no new mask code), `set_params_splitkv` heuristic. The shipped q==1 arms and every unclassified batch stay dispatch-identical; `VT_FA2_SPEC_DECODE=0` restores the prefill route for a same-binary A/B. Wave spec [dflash2-spec-as-decode.md](specs/dflash2-spec-as-decode.md); the GPU step-time delta (the −8-9 ms claim), the GPU token gates and the first CUDA compile are owed there, operator-run | perf |
 | [#1853](https://github.com/mudler/vllm.cpp/issues/1853) | `LTX25-DIT-ATTN-FLASH` | **`PENDING` on a `dgx:gpu0` lease: the arithmetic-perturbation reference render that would make [#1743](https://github.com/mudler/vllm.cpp/issues/1743)'s criterion RELATIVE.** §11 of [`ltx25-dit-attn-flash.md`](specs/ltx25-dit-attn-flash.md) relocates the pixel verdict onto **correspondence** and **incoherence**, which discriminate a degraded render from a separated trajectory. It deliberately does NOT answer the other half of #1743: is the swap's divergence no worse than this pipeline's own divergence under an arithmetic perturbation of comparable size. That needs one further arm - the **naive** path at `768x448/49f`, seed `20260820`, on §10.7's pinned binary and checkpoints, with a bounded `+/-1` bf16 ULP dither injected at the DiT attention output at the `8.6e-05` to `3.7e-04` per-element flip rate §10.2 derives - after which `D(flash, naive) <= D(dither, naive)` is a bound with NO chosen constant. **No lease was authorised for #1743, so this is PENDING and not skipped.** The cross-build `baseline-20260820` vs `naive` figure (mean \|delta\| **9.452407**, LARGER than the swap's **6.414156**) is NOT that control and is not used as one: the binary lineage differs, so every other commit between `a50c57d69` and `3e2961ef0` sits inside it, which §10.8 already records. NOT FIXED IN FLOW: it needs a GPU lease this work does not have. Owned by row `LTX25-DIT-ATTN-FLASH` and listed under `## Owed` | bug |
 | [#1854](https://github.com/mudler/vllm.cpp/issues/1854) | `LTX25-DIT-ATTN-FLASH` | **NOT GATEABLE in this tree, declared rather than proxied: nothing asks whether an LTX-2.5 render is GOOD, only whether two renders are the SAME.** Raised while relocating #1743's criterion. The RELATIVE form of the question IS now answered and gated - the coherence checks of §11.3 assert that neither arm is systematically sharper, blockier, quieter or less mobile than the other, at `K <= 0.5` where a one-directional degradation gives `K = 1` exactly. The ABSOLUTE form is not answered: **prompt adherence needs a vision-language model**, which §10.8 already refuses to approximate ("a check for 'is this a golden retriever shaking off water' is a model, not a threshold"), and **artefact-freedom needs an absolute reference render** from an oracle that runs this pipeline, which `.agents/oracles/` does not have. `scripts/ltx25-render-compare.py` therefore computes an **absolute quality panel per arm** - 8-grid and 32-grid blockiness ratios, clipped-pixel fraction, mean sharpness - prints it, records it in the JSON and **checks none of it**, saying so in its own output, rather than inventing a threshold that means nothing without a reference. NOT FIXED IN FLOW and deliberately not: a proxy for perceptual quality is the `a-shape-valid-gate-passes-a-wrong-artefact` failure. Owned by row `LTX25-DIT-ATTN-FLASH` and listed under `## Owed` in §11.5 | bug |
diff --git a/.agents/model-matrix.md b/.agents/model-matrix.md
index f1fafff5f..294d16930 100644
--- a/.agents/model-matrix.md
+++ b/.agents/model-matrix.md
@@ -119,7 +119,7 @@ Engaged architectures (the 53 non-`INVENTORIED` rows):
 | 🚧 | `KimiLinearForCausalLM` | Kimi-Linear-48B-A3B | **PAGED-INCREMENTAL DECODE LANDS the 5× speed win (2026-08-07, §19, `row/KIMI-PAGED-INCREMENTAL` #113):** the §18 real lever (e) BUILT — `KimiDecodeCache` + `ForwardPrefillIncremental`/`ForwardDecodeStepIncremental` (`kimi_linear_device.cpp`): prefill-once (KDA recurrent+conv state carried via `vt::KdaGatedDeltaRule` state in/out + `vt::CausalConv1dFwd` tap-carry; NoPE-MLA latent-KV cached) + recurrent decode-step, MIRRORING vLLM `kimi_gdn_linear_attn._forward` (prefill=`chunk_kda_with_fused_gate` output_final_state / decode=`fused_recurrent_kda` initial_state, `vllm-src` `a4e3cb4`; divergences: host state vs paged slot cache, materialized-MHA MLA vs paged-FA2 — named residuals). CPU byte-exact state-carry gate `test_kimi_linear_forward` **15/15·875** (NEW case l: carried decode == fresh full-recompute byte-identical + greedy-identical). Full 48.9B GB10 (single-load/config, flock, drop_caches, min-avail 18-21 GiB, no reboot, §12 golden md5 `bfa5bdbf`): recompute 122/128 @ 4.23 tok/s (reproduces #111); incremental+recurrence 120/128 @ 16.63; **incremental+chunk-prefill 122/128 @ 18.87/19.03 tok/s (2 runs) — token-IDENTICAL to recompute (Gate A PASS, p7 `got` byte-exact) = 4.5× over recompute, 0.90× of vLLM ~21 (the 5× decode gap 0.20×→0.90×)**. Gate B STRICT NOT reached (122/128): chunk-prefill in the RIGHT vehicle reproduces recompute EXACTLY, does NOT close p7 — REFUTES the #111 "p7 in the right vehicle → STRICT" hypothesis; p7 intrinsic (§13/§14 f32-vs-bf16 near-tie at a comma). Decode decomposition (nsys, ours, 99 steps, same-tool): **~90% is the SAME cuBLAS `internal::gemvx::kernel` vLLM calls (batch-1 GEMV-parity)**, KdaScanKernel 2.3%, MoE glue 2.3%, CastBf16 3%; chunk kernels 20 inst = prefill only (prefill=chunk/decode=recurrent IN VIVO). Killing O(n²) ALONE reaches parity-class; no lever load-bearing beyond it — residual = ~15% host-orchestration idle + 3% CastBf16 (a bf16 residual stream = the ONE lever ALSO closing the p7-STRICT near-tie) + paged-FA2 MLA decode. vLLM-live-nsys@0.82 NOT run (box-safety: 95-98 GiB reservation + nsys below the 15 GiB LIFE-CRITICAL floor). `--incremental` opt-in; `VT_KIMI_DEVICE_KDA`/`_CHUNK` STAY OFF (122/128 ≠ STRICT). Row STAYS 🚧. **chunk_kda PREFILL PHASE-2 MEASURED — op CORRECT, chunk-EVERY-STEP REGRESSES 122→102 (2026-08-07, §18, `row/KIMI-CHUNK-KDA-P2` #111):** the `chunk_kda` prefill family regenerated + vendored for ALL 6 arches (reproducible — only new `kda_*`+MANIFEST; GDN cubins byte-identical; drift GREEN) + wired through the new op `vt::KdaChunkPrefill` (the 6-cubin `_chunk_kda_fwd_with_cumulative_g`; `cuda_gdn.cu.o` -Werror clean; RED-first unit `test_ops_kda_chunk_prefill` **2/2·4** on GB10 — chunk-vs-recurrence mean_abs **4.68e-5**, wrong-gate **72×**; GDN untouched 66/66·4242). Full 48.9B GB10 gate (flock, min-avail 21 GiB, no reboot): control device-KDA reproduces **122/128, 4.24 tok/s** EXACTLY; **+chunk-prefill (`VT_KIMI_DEVICE_KDA_CHUNK`) REGRESSES to 102/128, 4.08 tok/s** (p3 16→3, p6 16→11). Root cause: the island's O(n²) recompute applies chunk EVERY decode step over the growing sequence — NOT vLLM's prefill=chunk/decode=recurrent split — so it coin-flips near-ties the recurrence-every-step (control) doesn't (the recurrence matches vLLM's DECODE; chunk only matches its PREFILL). vLLM speed arm (§12 recipe, util 0.82, triton MoE, eager, single-seq; min-avail 15 GiB, no reboot): **~21 tok/s median** 16-token aggregate (25.3 cold-discarded; TTFT not isolable in 0.25.0) vs ours **4.24** (recurrence) / **4.08** (chunk) STEADY decode → **ours/vLLM ≈ 0.20** (vLLM ~5× faster on decode — the O(n²)-recompute vs paged-incremental distance, = the coupled STRICT+speed lever). `VT_KIMI_DEVICE_KDA_CHUNK` STAYS OFF (a regression isn't a flip); device-KDA (122, OFF) still best. The op + regen are the validated prefill half of the named real lever (e) paged-incremental decode (chunk-prefill ONCE + recurrent-decode over PERSISTENT state — kills the O(n²); the STRICT + speed lever, coupled). Row STAYS 🚧. **DEVICE-KDA GB10 122/128 + 4.24 tok/s (§15, #104); device NoPE-MLA lever MEASURED-NEGATIVE (2026-08-07, §16, `row/KIMI-STRICT-CLOSE` #107):** the per-channel-decay device recurrence `vt::KdaGatedDeltaRule` moves 106→**122/128** (p0-p6 16/16; sole p7 pos-6 comma near-tie) AND **1.35→4.24 tok/s (3.1×)** — vLLM's ACTUAL f32-on-bf16 arithmetic, beats §14's host-precision 120. The §15 residual (d) was attempted in device-COMPUTE form: `VT_KIMI_DEVICE_MLA` routes the 7 NoPE-MLA layers' softmax core through `vt::Attention` (pad-V: value zero-padded qk_nope+qk_rope=192 vs v=128, `out[:,:,:v]` byte-exact). CPU RED-first gate GREEN (`test_kimi_linear_forward` **14/14·825**, pad-V==f64 ref rtol 3e-3; perturbation fails 108). Full 48.9B GB10 gate (single-load, flock, min-avail 21 GiB, no reboot): control device-KDA reproduces **122/128, 4.24 tok/s** EXACTLY; **+device-MLA REGRESSES to 109/128 AND 3.89 tok/s** — `vt::Attention`'s f32 online-softmax is the right math but a DIFFERENT reduction order than vLLM's FA2, so it coin-flips near-ties (breaks p3 16→3 into §14's `163586×` repeat) and the per-(t,h) build slows the O(n²) recompute. `VT_KIMI_DEVICE_MLA` STAYS OFF, kept as a documented-MEASURED-NEGATIVE A/B knob (§14 `ISLAND_F32ACC` precedent). MLA dims VERIFIED from the real config (nah=32, qk_nope=128, qk_rope=64, v=128, kv_lora=512, q_lora=None; 7 full-attn/20 KDA). Both device knobs default OFF (122 ≠ STRICT, K=3-deterministic golden). STRICT residual, sharpened: needs vLLM's ACTUAL kernels — (c) chunk_kda prefill family (Triton-AOT regen for sm_121a) + (d) paged FA2 `mla::ForwardMlaAttentionBlock` (NOT the vt::Attention approximation) + (e) paged-incremental decode (needs a decode/paged-attn op, query_len≠key_len; kills the O(n²)) — each a substantial multi-kernel brick (§16). Row STAYS 🚧. **FULL-MODEL GB10 e2e RUNS — NEAR-TIE 106/128 (2026-08-06, `row/MODEL-KIMI-LINEAR-BF16`):** the bf16-resident path CLEARS the f32-loader block — the full 48.9B model now runs e2e on one GB10. dgx CUDA build (`-Werror` clean, 14 GDN AOT symbols nm-linked, `test_kimi_linear_forward` 13/13·656 in the CUDA binary); `kimi-linear-gen --gpu` greedy-decodes the §12 8-prompt battery x16 vs `greedy_ids.npy`. MEMORY: load 117.6s, host RSS PEAK **1.7 GiB** (stage-then-ReleaseHost), device peak 98.5 GiB, min-avail **21.6 GiB** (above the 15 GiB floor, matches the ~25 GiB pool-math headroom), NO OOM/reboot. TOKEN gate **NEAR-TIE 106/128 (82.8%)** — prompts 0,1,3,4,5,6 are 16/16 token-exact, p2/p7 diverge at punctuation/word near-ties; 96 consecutive exact tokens across 6 prompts prove the WIRING (a wiring bug can't). Root cause (honest): the f32 residual stream + host-f64 islands are MORE precise than vLLM's bf16 device kernels, so they flip the argmax where vLLM's deterministic bf16 top-1 has a small margin. STRICT path = the named W7-speed residuals (device GDN/MLA islands -> bf16 stream matching vLLM's rounding). 1.59 tok/s (recompute+island rate). `VT_KIMI_DEVICE_COMPUTE` STAYS OFF (parity-enablers: near-tie != token-exact). Row STAYS 🚧. **bf16-RESIDENT loader/forward IMPLEMENTED + CPU-gated (2026-08-06, `row/MODEL-KIMI-LINEAR-BF16`):** the §13 design is coded — `LoadKimiLinearResidentBf16Weights`/`StageKimiResidentBf16`/`BuildKimiResidentFromHost` (`kimi_linear_weights.cpp`; `LoadBf16Direct` -> `OwnedTensor`, per-tensor stage-to-`d_dev` + `ReleaseHost`, tiny vectors host f32), `KimiLinearResidentWeights` (`kimi_linear.h`), bf16 device forward `DeviceForwardBodyBf16` + `Gemm Bf16` cast-act at ~20 GEMM sites with the two host-fallback islands EXTRACTED+shared (`kimi_linear_device.cpp`), `ForwardDevice` resident-path dispatch (`kimi_linear.cpp`), and the `kimi-linear-gen` e2e harness. CPU **13/13·656** (12/12·614 f32 path UNTOUCHED + NEW tiny-config bf16-vs-f32 gate). PENDING: dgx CUDA build + full-model GB10 e2e vs the STRICT golden. Row STAYS 🚧. **bf16-RESIDENT brick POOL-MATH+DESIGN (2026-08-06, `row/MODEL-KIMI-LINEAR-BF16`):** pool math CLOSES (91.5 GiB bf16 device-resident + ~2.4 GiB act/norms/ctx ≈ 94 GiB, ~25 GiB headroom); design grounded §13 (Laguna `GemmBf16` cast-act + `OwnedTensor::d_dev`, `LoadBf16Direct`, f32 `MaterializeHost` kept for the unit gate). Impl (loader/forward rewrite + gate + e2e) pending. Row STAYS 🚧. **§8 GOLDEN CAPTURED — STRICT (2026-08-06, `row/MODEL-KIMI-LINEAR-E2E`):** the §8 SACRED oracle golden is captured on GB10 (0.25.0-stage, util 0.82, moe=triton, min 15 GiB avail, NO reboot), **8/8 prompts DETERMINISTIC over K=3 → STRICT gate**, committed at `tests/parity/goldens/kimi_linear_greedy/`. Full our-engine e2e BLOCKED on OUR f32 loader (materializes ~183 GiB > 119 pool), the bf16-residency residual; row STAYS 🚧. **W7 GPU-VERIFY (2026-08-06, branch `row/MODEL-KIMI-LINEAR-GPU`):** the device compute runs **12/12·614 GREEN on GB10 sm_121a CUDA build**, BOTH arms (`VT_KIMI_DEVICE_COMPUTE=1` + host-ref); prod stack (CUTLASS-NVFP4 GEMM + FA2 ENABLED + Triton-AOT GDN, 14 cubins nm-verified); f32 device==W2 ref, no divergence, no DeepSeek-class trap. Oracle gateability re-confirmed (0.25.0-stage registers `KimiLinearForCausalLM`). e2e §8 SACRED golden STILL disk-blocked (91.5 GiB checkpoint absent, dgx root 100% full, 34G free). Row STAYS 🚧. **W7 DBuf-resident device COMPUTE landed, CPU-gated** (`CLAIM-KIMI-LINEAR-W7`): the real device compute (`ForwardDeviceCompute`, `kimi_linear_device.cpp`) composes the whole 27-layer KDA/NoPE-MLA + 256-expert-MoE hybrid over pooled f32 `DBuf`s through the SHARED `vt::` ops (embed/`FusedChain` add+RMSNorm/`MatmulBT` projections/`CausalConv1dFwd` convs/`L2Norm`/`RmsNormGated`/`MoeRouterTopK` sigmoid-`noaux_tc`/`MoeSiluMul`/`MoeCombine`/lm_head), returning DEVICE-RESIDENT logits; 2 documented HOST-FALLBACK islands (the KDA per-k-channel gated-delta recurrence + its exp/softplus decay gate — `vt::GdnDecode` carries only a per-HEAD scalar decay; the NoPE-MLA softmax core — the paged `mla::ForwardMlaAttentionBlock` device path is born-on-runner) are the W7-speed residuals. CPU-gated vs the W2 host reference (the CPU backend runs the SAME `vt::` dispatch): `test_kimi_linear_forward` **12/12·614** (per-op KDA/NoPE-MLA/MoE/dense device==ref within f32-accumulation tolerance; the whole `ForwardDeviceCompute` == ref logits + greedy-token-identical + device-resident). Runner opt-in via `VT_KIMI_DEVICE_COMPUTE=1` (default OFF keeps the CPU-verified W6 host-ref compose). GPU numerics (bf16 activations, GDN Triton-AOT cubins, paged het-KV, grouped-MoE slabs) + the e2e SACRED golden stay a NAMED pending (box down) — row STAYS 🚧. ON TOP OF **W6 DEVICE forward SEAM** (`CLAIM-KIMI-LINEAR-W6`): the born-on-the-runner `ForwardDevice` (the DEFAULT `gather_logits` runner path) no longer refuses — it composes the `[rows,vocab]` logits via the CPU reference and hands them back DEVICE-RESIDENT (a pooled `DBuf`, wrapped like deepseek_v2 `WrapDeviceLogits`; `on_device()==true` on CPU+CUDA) so the on-GPU sampler consumes them with NO host download. Kimi-Linear now ROUTES device-resident (`check-runner-routing-consistency` reclassifies it, refuse-skipped stubs 2→1, NO allowlist; `check-fusion-consistency` green); `test_kimi_linear_forward` **7/7·300** (adds the `ForwardDevice`==host-ref device-resident gate). The DBuf-resident device COMPUTE (KDA via the GDN family, NoPE-MLA via `mla::ForwardMlaAttentionBlock`, DeepSeek-V2 grouped-MoE over the paged het-KV; full plan in `kimi_linear.cpp`) is the GPU-verify-pending W7 residual. ON TOP OF **W2-W6 CPU REFERENCE forward** (`CLAIM-KIMI-LINEAR-W2`): the real host `KimiLinearModel::Forward` composes the whole 27-layer hybrid from the landed primitives (KDA layer via `vllm::kimi_kda` refs + the gated-delta recurrence; NoPE-MLA materialized-MHA ref; sigmoid `noaux_tc` MoE + shared expert; dense SwiGLU); loader now materializes host float weights; `test_kimi_linear_forward` 6/6·246 (per-op gates + finite whole forward + greedy decode). ON TOP OF **W1 scaffolding** (registry + `ParseKimiLinearParams` 20 KDA + 7 NoPE-MLA + index-verified name-map + het-KV spec). e2e-gateable (FITS one GB10, 0.77× pool). RESIDUAL = the DEVICE born-on-runner forward (KDA kernel/absorbed-MLA/grouped-MoE slabs) + the W0/W7 e2e SACRED golden. **ROW 7 RUNNER FOLD (2026-08-07, §21, #122): Kimi decode now runs THROUGH `ModelRegistry::Forward` on the runner's paged state (B1 config synthesis, B2 KDA-paged via `KdaChunkPrefill`/`KdaGatedDeltaRule` over `gdn_state`, B3 paged NoPE-MLA with `mla::ForwardMlaAttentionBlock` default-ON, B4 `vllm_complete_tokens` ABI v13 + thin-client example). GB10: engine==CLI 128/128 byte-identical, vs golden 122/128 (the intrinsic near-tie profile), SACRED post-fold green (35B 315/315, 27B 235/235); server stream 19.0 tok/s wall vs vLLM ~21 (~0.90×) = the open speed residual.** Row stays `ACTIVE` on the speed thread | `MODEL-TEXT-kimi-linear-kimi-linear-for-causal-lm` |
 | 📋 | `MuseGlimmerForConditionalGeneration` | Muse Glimmer 30B (Meta, agentic multimodal; iRoPE + gated attn + perception encoder + DFlash) | **W0-W7 ON `row/MODEL-MUSE-GLIMMER` (PR #279, NOT merged): 52-layer text forward, 50-layer perception encoder, mm wiring, ATEM reasoning + tool parsers — the forward no longer refuses.** The evidence behind that is NARROW and says so: text agrees with a torch transcription of #51655 and with HF's own `muse_glimmer` on REAL 30B tensors at **reduced depth 4/52 only** (5 prefill argmax positions, no generated tokens); full depth never ran, multi-step decode is untested, the perception encoder has NO reference check, nothing has run through the server, and the ATEM parsers' channel scoping does not work at server defaults (the seam has no `adjust_request` dispatch site and `skip_special_tokens` defaults true, so the framing is stripped first). The lifecycle token stays `SPIKE`/📋 deliberately: advancing it owes `docs/STATUS.md` and `.agents/NOW.md` rows, and both are byte-exactly at their shrink-only ratchet/budget, so the advance belongs to the landing commit that can pay for them. Released 2026-08-08, **BEYOND-PIN**: no Muse code at `555967922` nor on vLLM `main`; the only upstream impl is the OPEN PR [#51655](https://github.com/vllm-project/vllm/pull/51655) head `075d645af` (approved, 3/20 CI red), ported from on explicit developer direction as a tracked exception (porting-inventory §9 deviation 16). **No oracle ⇒ no speed axis is claimable**; correctness gates against the HF reference. Text tower fully traced (Gemma-style sandwich norms w/ baked +1 offset + split eps, Llama-4-style iRoPE NoPE/full vs RoPE/sliding, weightless pre-RoPE QK-norm, ~3.87 query pre-scale w/ dual config schema, Qwen3.5-style attn output gate reading the LAYER INPUT); 50-layer perception encoder (linear patchify, bilinear pos-emb interp, width-first 2D RoPE, block-windowed attn, pixel-shuffle merge) reuses the Qwen3-VL vision seam; DFlash is recognition + `is_neox_style` threading on the existing speculator, not a new drafter | `MODEL-MM-muse-glimmer-muse-glimmer-for-conditional-generation` |
 | 📋 | `KimiK3ForConditionalGeneration` | Kimi K3 (2.8T MoE + MoonViT-V2, DERIVE-AND-SHIP) | **W2/W5 CPU scaffolding landed** (registry stub + nested text/vision/quant config descent + text-backbone structural name-map + REFUSE-by-name forward + MXFP4-refuse loader; clean CPU build, scaffold gate 6/6). text backbone IS `KimiLinearForCausalLM` (KDA+MLA+MoE hybrid, HEAVY reuse); **does NOT fit GB10 (~1.56 TB MXFP4, ~12×)** and NOT in the pinned oracle ⇒ no on-box golden — DERIVED, proxy-gated on Kimi-Linear-48B; forward + MXFP4 + KDA delta + MoonViT-V2 not implemented (NOT-YET-BUILDABLE) | `MODEL-MM-kimi-k3-kimi-k3-for-conditional-generation` |
-| 📋 | `Dots3NoteForCausalLM` | dots3-note (280B-A16B multimodal MoE: DeepSeek-V3.2 text tower + sliding-window MLA, MoE ViT, `dots` audio encoder) | **W1 LANDED (config + registry): the arch RESOLVES, parses the REAL released `config.json`, and accounts 1614/1614 tensors on a committed slice of the released shard index; load, GGUF and the forward each REFUSE BY NAME.** All six §4 config traps are gated RED-first (evidence table in spec §4.1), which matters more here than on any other row because spec §6.4 records that NO oracle runs on hardware we own, so no token gate exists downstream. W1 also RESOLVED §1.4 from the checkpoint: exactly ONE nextn layer (`model.layers.46.*`, 18 tensors) with the SLIDING geometry and a DENSE MLP, and no `shared_head.head.weight` (`has_own_lm_head=False`); and CORRECTED spec §4 item 6 ([#1804](https://github.com/mudler/vllm.cpp/issues/1804)) — BOTH MLA ropes are GPT-J (`deepseek_v2.py:1093-1098` for the full class), the geometries differ on the THETA (5e4 vs 8e7), not the layout. Memory format recorded: `mlp.gate.e_score_correction_bias` ships F32 in an otherwise BF16 tower. Row stays `SPIKE` deliberately — making an arch resolve is not porting a model, and the §8.1 heading restructure `ACTIVE` requires belongs to the brick where the forward stops refusing. W0 scope, unchanged below. Upstream subclasses DeepSeek (`Dots3NoteModel(DeepseekV32Model)`, `Dots3NoteMoE(DeepseekV2MoE)`), so our gated MLA + DSA indexer + `noaux_tc` router + Qwen3-VL vision + Voxtral audio carry most of it; net-new is windowed MLA over 33 of 46 layers with a second latent geometry (576 vs 1088 rows), the padded/heterogeneous MLA KV spec, the headwise attention gate, the MoE ViT and the `dots` audio stem. **BEYOND-PIN** (vLLM `main` only, `#51255`, still being patched) and **oracle-memory-infeasible**: ~576 GB bf16 / ~290 GB fp8 against a 119-122 GiB ceiling on every host we own, so no performance axis is claimable and no token gate can bind until §6.4 of the spec is answered | `MODEL-MM-dots3-note-dots3-note-for-causal-lm` |
+| 📋 | `Dots3NoteForCausalLM` | dots3-note (280B-A16B multimodal MoE: DeepSeek-V3.2 text tower + sliding-window MLA, MoE ViT, `dots` audio encoder) | **W2 LANDED (the whole weight map): the complete released shard index read for real — 38006/38006 accounted, 35381 language / 2195 vision / 430 audio, every bucket asserted BY NUMBER, zero unaccounted, and the two tower files carried as NAMED W6/W7 deferral records rather than integer counters.** Headers only: 4770592 bytes over 133 safetensors headers, no tensor byte. It settled three things the W1 slice could not — the backbone has exactly FOUR distinct layer shapes so the 1/2 repeat holds, the full/sliding split derived from the shipped indexer tensors matches `config.layer_types` exactly, and the checkpoint carries 62 F32 tensors in TWO families (the 45 language `e_score_correction_bias` W1 predicted plus 17 vision `router_bias`, spec R5's shape). And it FOUND one: the index declares `indexer_rope_layout: "leading"` / `converted_from: "tail"`, which no upstream code reads ([#1846](https://github.com/mudler/vllm.cpp/issues/1846)); W3 owes the slice. **W1 LANDED (config + registry): the arch RESOLVES, parses the REAL released `config.json`, and accounts 1614/1614 tensors on a committed slice of the released shard index; load, GGUF and the forward each REFUSE BY NAME.** All six §4 config traps are gated RED-first (evidence table in spec §4.1), which matters more here than on any other row because spec §6.4 records that NO oracle runs on hardware we own, so no token gate exists downstream. W1 also RESOLVED §1.4 from the checkpoint: exactly ONE nextn layer (`model.layers.46.*`, 18 tensors) with the SLIDING geometry and a DENSE MLP, and no `shared_head.head.weight` (`has_own_lm_head=False`); and CORRECTED spec §4 item 6 ([#1804](https://github.com/mudler/vllm.cpp/issues/1804)) — BOTH MLA ropes are GPT-J (`deepseek_v2.py:1093-1098` for the full class), the geometries differ on the THETA (5e4 vs 8e7), not the layout. Memory format recorded: `mlp.gate.e_score_correction_bias` ships F32 in an otherwise BF16 tower. Row stays `SPIKE` deliberately — making an arch resolve is not porting a model, and the §8.1 heading restructure `ACTIVE` requires belongs to the brick where the forward stops refusing. W0 scope, unchanged below. Upstream subclasses DeepSeek (`Dots3NoteModel(DeepseekV32Model)`, `Dots3NoteMoE(DeepseekV2MoE)`), so our gated MLA + DSA indexer + `noaux_tc` router + Qwen3-VL vision + Voxtral audio carry most of it; net-new is windowed MLA over 33 of 46 layers with a second latent geometry (576 vs 1088 rows), the padded/heterogeneous MLA KV spec, the headwise attention gate, the MoE ViT and the `dots` audio stem. **BEYOND-PIN** (vLLM `main` only, `#51255`, still being patched) and **oracle-memory-infeasible**: ~576 GB bf16 / ~290 GB fp8 against a 119-122 GiB ceiling on every host we own, so no performance axis is claimable and no token gate can bind until §6.4 of the spec is answered | `MODEL-MM-dots3-note-dots3-note-for-causal-lm` |
 | 🚧 | `MiniMaxH3DiTModel` | MiniMax-H3 (33.1B omni-modal video+audio DiT, DERIVE-AND-SHIP) | **W1/W2 landed**: packed layout (fl2va + ref2va, fp64 position grid BIT-EXACT), latent packing, euler-ancestral eta0 scheduler, and the full DiT forward all parity-gated against the UPSTREAM vLLM-Omni modules executed at reduced dimensions (**max abs diff 1.6e-7**, 10/10 cases / 2539 assertions). NOT autoregressive (no KV cache, no sampler, no logits) and **e2e HW-BLOCKED** (~354 GB checkpoint, ~133 GB/rank on 4x B300 vs 119 GiB unified); bf16 production stream + request planning + the ComfyUI-GGUF arm also landed (535 REAL tensors resolve onto our contract, geometry from shapes alone). **HW verdict CORRECTED: quantized arms FIT (~41 GB in 119 GiB)**, so e2e + speed are reachable; encoder/VAEs/audio VAE DONE (4.2e-9 vs the checkpoint's remote code); NVFP4 layout GATED as identical to ours (speed path is loader wiring); BOTH VAE DECODERS done (audio 4.2e-9, video ViT3D 8.9e-8); video tiling + 3D-CNN encoder (conditioning only) pending; encoder TEXT tower done (1.2e-7); **serving `/v1/videos` DONE and the DEVICE-RESIDENT forward (W2b, f32) LANDED + GPU-VERIFIED on Thor sm_110 at video 1.49e-7 / audio 8.94e-8**; bf16 stream + fusion folds + the FP4 path (needs sm_121a) + a real-checkpoint run pending. **2026-08-05: the AUDIO-VAE ENCODER is ported** (DAC analysis stack + `pre_block` AttnProjection + `mean_proj`, gated stage by stage vs the checkpoint's own remote code at 2.98e-8 / 1.64e-7 / 1.86e-8) with its own checkpoint loader gated on the real 1087-tensor manifest — so **ref2va AUDIO and VIDEO+AUDIO references are now WIRED** (audio rows move by 0.51 / 0.71; a different waveform still moves them by 7.1e-4). Both VAEs are now complete in both directions. **bf16 13-SHARD RELEASE INDEXES 2026-08-07 (`row/H3-BF16-SHARDED-DIT`)**: `MiniMaxH3ShardedCheckpoint` resolves the ORIGINAL 66.3 GB release through its own `model.safetensors.index.json` (a tensor named in the index but missing from its shard throws BY NAME), `EnumerateMiniMaxH3ShardedTensors` feeds the shared shapes-only geometry parser, and `LoadMiniMaxH3DitFromShards` is the host-f32 reference loader. Gated CPU-only at 72/72/54497 (post-rebase): every tensor resolves to the shard the index named AND to the bytes written there, the derived geometry equals the single-file path field for field, and a SPARSE 13-shard release with the REAL 535 tensors at REAL shapes (66.3 GB declared, 144 KB on disk) derives the SHIPPED geometry (50/5376/56/128/14336/24/32/1x2x2/5120). **STREAMS 2026-08-07 (`row/H3-BF16-SHARDED-STREAM`)**: `StreamMiniMaxH3ShardedToDeviceBf16` uploads it one tensor at a time — a BF16 tensor bound for a bf16 device slot goes straight from the mmap with ZERO host buffer, so peak host is bounded by ONE tensor (observed `host_peak=8192`, `direct=37 converted=9`); bit-exact vs the non-streamed `StageMiniMaxH3DitWeights` reference over all 46 views with identical logits, rope.inv_freq host-resident, 73/73/55203. Spec §8.14. **bf16 TEXT ENCODER + THE CONDITIONING NUMBER 2026-08-07 (`row/H3-ENC-BF16-COND-DIFF`)**: the 14-shard 63 GB bf16 Qwen3-VL-32B encoder streams to device too (`StreamMiniMaxH3EncoderShardsToDevice`, q/k/v and gate/up fused ON DEVICE), `--encoder-only` runs the tower alone (peak ~96 -> ~49 GiB by not loading the DiT first), and the widening is gated BIT-IDENTICAL vs an f32-staged tower so the A/B cannot be confounded. MEASURED on Thor over 233 tokens: Q4_K_M vs bf16 conditioning is cos 0.99745 mean / 0.909 min, rel RMS 6.85% excluding the attention sink, median rotation 3.5 deg — same energy as a ONE-WORD prompt edit but DIFFUSE (232/233 tokens rotate vs 172/233). Whether the RENDER changes is NOT established. 75/75/55609. Spec §8.15. This UNBLOCKS the quantization-quality question; no bf16-vs-quant render or speed number is claimed. Spec §8.13. **W-FP4a LANDED (CPU) 2026-08-06 (`row/H3-FP4-SPEED`)**: the device DiT forward now routes the NVFP4 projections through the shared Marlin W4A16 dispatcher (fp4 kept packed; no new quant code), fp4-vs-bf16 wiring gate GREEN (62/62·30039). **W-FP4a GB10 leg LANDED 2026-08-06 (`row/H3-FP4-GPU-E2E`, PR #64):** on sm_121a the Marlin W4A16 path RAN for all 11 projections (`dense_gemms==11` default — VT_MARLIN_DENSE is default-ON → vLLM's own DENSE Marlin GEMM, not the grouped route; `marlin_gemms==11` under VT_MARLIN_DENSE=0; `fallback_gemms==0`), fp4-vs-bf16 BYTE-EXACT (max\|diff\|=0), and the fp4 arm is a MEMORY win not a diffusion-forward speed win (per-forward bf16/fp4 3.47× @seq64 → 0.79–0.83× @seq4224–7040; ~16 vs ~66 GB device). Real-checkpoint fp4-resident t2va e2e RUNS (real 18.75 GB NVFP4 DiT + VAEs + GGUF Qwen3-VL-32B encoder → valid mp4/wav; DiT s/step 5.45/20.0/209 s @512/768/REF-209f) but frames are a non-scene patch-grid at 12/20/50 steps → OPEN render bug (device VAE/denoise). vLLM-Omni has no quantized H3 arm (BF16-only) so any comparison is HW/loader-forced-indirect — spec §8. **ONE-SURFACE ROW 2 LANDED 2026-08-08 (`row/H3-VIDEO-ABI`)**: the whole assembly pipeline is library-owned (`vllm::multimodal::MiniMaxH3VideoEngine`, `minimax_h3_video.cpp`) behind the ABI v12 `vllm_video_*` entry points; `/v1/videos` routes through the SAME seam; `minimax_h3_gen`+`minimax_h3_mux` are thin `vllm.h` clients, frames+WAV byte-identical to the pre-fold binary on the committed fold fixture (`test_minimax_h3_video_fold` 3-arm gate + the v12 `test_capi` section); GB10 real-video re-verify via the v12 ABI = named residual. **ROW 2 DEVICE-SEAM FOLLOW-UP (#135; replaces #134):** ABI 0/1 maps once to `vt::DeviceType`; shared code dispatches through `GetBackend(device_type)`, restoring DSR 34→32 without a baseline/allowlist change; CPU compile/fold test pending in CI due shared-disk pressure **PRUNED CHECKPOINTS LOAD AND RUN 2026-08-10 (`row/H3-PRUNED`, #241, spec §8.21):** the community `pruned` variants (`unsloth/MiniMax-H3-GGUF` Q2_K..Q8_0, `lilcheaty/MiniMax-H3-NVFP4` `*_pruned_nvfp4`) are not lossily pruned but ComfyUI's AdaLN timestep-CURVE refactor (`comfy/ldm/minimax/model.py:419-432,610-615`): an `adaln_t_table` [1025, 8] lerp replaces the sinusoidal+MLP time embedder, no SiLU before the AdaLN linear, and its in_features drop 2688 -> 8, collapsing `adaln_proj` from 13.04B of 33.12B to 0.04B and the DiT to 20.11B — so a pruned Q8_0 (21.4 GB) costs what our unpruned Q4_K_M (19.9 GB) costs. Gated on the REAL 532-tensor pruned GGUF manifest (532 = 535 - 4 `time_embedder.*` + 1 table, header-only), the clamped-lerp golden, and a CONSTRUCTED-curve test proving the pruned forward is numerically identical to the unpruned one; 79/79/57299. | `MODEL-DIFFUSION-minimax-h3-mini-max-h3-dit` |
 | 🚧 | `MiniMaxMusic3ForConditionalGeneration` | MiniMax-Music3 (text-to-music: ~8.6B Qwen3 global LLM + 0.646B 8-codebook RVQ depth decoder + 2.4B fp32 flow-matching DiT + 0.054B DAC Flow-VAE) | **W0 DONE + W1 LANDED (#672, [spec](specs/minimax-music3.md)).** Diffusers oracle gateable with committed per-stage goldens (#708); modular loader in the tree, 1413/1413 assertions against the real checkpoint (#714). Geometry MEASURED from every component `config.json` plus safetensors headers by range request. First row to exercise the fallback-oracle rule (#647): absent from vLLM AND vLLM-Omni, so its primary oracle is the OPEN diffusers PR #14456 head `c6da9936` with SGLang-Omni `748a0b43` as the e2e cross-check. **W2/W3 LANDED**: the AR half computes and is gated. **W4/W5 LANDED**: the ACOUSTIC half — flow-matching DiT, `FlowMatchEulerDiscreteScheduler` with `invert_sigmas`, the CFG mix, the denoise loop's window bookkeeping and the DAC Flow-VAE vocoder over the shared `vocoder1d` primitives — gated at reduced dimensions (27 cases / 265 assertions, no checkpoint) AND at full scale against the committed capture (scheduler BIT-EXACT 22 016/22 016; waveform 88 064 samples, 0 outside a tolerance calibrated against torch reproducing its own goldens). **W6 LANDED**: registered as a `SpeechRegistry` family (`minimax-music3`, detection INSPECTS `modular_model_index.json`), `SpeechGenParams` extended ADDITIVELY with `lyrics`/`description` plus three generation controls (IndexTTS-2.5 byte-identical, 0 lines changed), the new `vllm_speech_*` C-ABI surface at v20, and `POST /v1/audio/speech` on the example server as a thin ABI client. The denoise + decode composition reproduces the capture's `waveform.npy` — 88 064 values, 0 outside tolerance, max abs-diff 4.52e-06 — and the WAV payload is BIT-EXACT against the quantization of that golden. **W7 LANDED: ONE ARM IMPLEMENTED AND VALUE-GATED, the rest refused.** The RVQ depth decoder loads from GGUF Q4_K (`audio-cpp/MiniMax-Music3-GGUF`@`c36aaeed`, `rvq_depth_decoder_q4_k.gguf`, 405 752 480 B, sha256 `4c5d41b2...c70cbdd0` — pinned because an unpinned quantized checkpoint is not reproducible): 47 tensors as 36 Q4_K + 9 BF16 + 2 F16, dequantized to bf16 through the SHARED `gguf_dequant.h` seam, lineage keyed on `audiocpp.model_spec.family` because `general.architecture` reads `audiocpp`/`mm3`/`qwen3`/`wan` for this ONE model and `wan` collides with real Wan video GGUFs. Gated 6 cases / 319 assertions against the W3 golden's 716 800 values at a bound DERIVED rather than chosen: worst per-tensor relative L2 0.0742 (bound 0.10), full-scale mean abs-diff 0.0324 / max abs-diff 0.3125 / 2.84% bit-identical. **THE GATE-DESIGN FINDING, which generalizes past this row: an upper-bound-only tolerance CANNOT distinguish a real quantized arm from a silent dequant fallback, because the fallback is CLOSER to the golden** — bf16 weights through the identical forward give mean abs-diff 0.00182, 17.8x nearer than the genuine Q4_K arm, so every plausible upper bound PASSES the failure. What catches it is a LOWER bound (5e-3, sitting between the two), gated by a positive control that actually runs the fallback and asserts the gate rejects it. Two further INDEPENDENT legs prove the quantized path was taken rather than inferring it: the RESIDENT ggml type of every tensor (loader bookkeeping — localizes a fault to a name, but a loader that lied about its tallies would pass it, which is why mutation QM2 exists) and the Q4_K LATTICE (a property of the DATA needing no loader cooperation — 0 of 524 288 32-element windows exceed 16 distinct values, against a bf16 control from the same file at 127 of 128). Second finding: the 2 F16 islands do NOT round-trip exactly (rel 3.23e-08 / 2.36e-08) because the quantizer re-encoded them BF16->F16 and F16's exponent range is NARROWER, flushing weights below ~6e-08 to zero — so the gate splits the islands, BF16 bit-exact and F16 bounded; asserting "unquantized means exact" for all 11 REDS a correct reader. 18 mutations across both layers all fire, one of them (QM4, the lineage guard) only after a COVERAGE HOLE it exposed was closed with synthetic GGUF fixtures. The 4 OTHER components' GGUF arms, the `mm3` and ComfyUI lineages, and every non-GGUF format are REFUSED BY NAME and owed. Quantized MiniMax-Music3 checkpoints DO exist — a HuggingFace survey on 2026-08-14 found 14 community repositories in 5 formats (GGUF of all five components, int8/w4a8, MLX, plus a proprietary Cortiq arm), recorded query-by-query with result counts in spec §9.1. `minimax_music3_quant.{h,cpp}` is a SEPARATE translation unit per porting-a-model.md that diagnoses 8 formats at the THREE places a quantized checkpoint announces itself (tree `.gguf`, manifest sidecars plus the dtype-only fp8/int8 cases, and `quantization_config`/MLX `quantization`) and refuses each BY NAME with the evidence, the missing piece, the supported arm, the phase and the issue. 29 cases / 125 assertions without a checkpoint plus 6 cases / 319 assertions against the pinned artifact, RED first at 8/8 probe checks, 18 of 18 mutations fire; every tolerance is DERIVED from a printed measurement rather than chosen, and no speed number is claimed. GGUF headers MEASURED by range request (56 MiB, no weight bytes): "the GGUF arm" is THREE MUTUALLY INCOMPATIBLE LINEAGES and `general.architecture` cannot separate them (`audiocpp`/`mm3`/`qwen3`/`wan`, and `wan` collides with real Wan video GGUFs), the ComfyUI lineage ships the DiT + condition encoder ONLY so it can never generate audio, and `comfy.gguf.orig_shape.*` is absent from all ten files. The 8.6B **W2's remainder LANDED (this PR): the 8.6B `Qwen3ForCausalLM` forward runs.** `Qwen3DenseModel::ForwardEmbeds` gives the landed dense path the `inputs_embeds` entry its multimodal siblings already had (`qwen3_vl.h:145,159`, `gemma4.h:210-218`, `muse_glimmer.h:369-380`), additively and proved BIT-IDENTICAL to the token-id `Forward` in logits AND in the paged KV it writes. Teacher-forced on the capture's own codes, 25 steps reproduce `frame_hiddens[:, :4096]` - 102 400 values, 9337 bit-identical, mean abs-diff 1.763e-02 - INSIDE a matched control (upstream's own model under `sdpa_kernel(MATH)`: 12 036 bit-identical, mean abs-diff 1.475e-02), with the per-step error FLAT rather than compounding. The oracle's OWN sampled semantic codes rank 2.48 on average under our guided logits where chance is 8191.5, and the gate carries its own NEGATIVE CONTROL: a one-step-shifted alignment reads mean abs-diff 0.803, 46x outside. **The e2e case at `POST /v1/audio/speech` is OBSERVED TO PASS (2026-08-15, #852)**: 200 `audio/wav`, 12332 bytes, 2 AR frames -> 6 latent frames -> 3072 samples/channel of 44100 Hz stereo, 6144 int16 samples all non-zero, 0 clipped, 2818 of 3072 positions differing between L and R; 21 assertions in 7:54, and 5 cases / 535 assertions for the whole file in 31:14. It had never been seen because the request body posted `audio_duration_s`, the FIELD name, where `ParseSpeechRequest` reads `audio_duration` - so the key was silently dropped and the family's 60 s default substituted, making every run a ~750x job (1500 AR frames not 2, 8 denoise windows not 1, 5167 vocoder latents not 6) that could never have passed its own payload assertion. The near-miss key is now REFUSED (#925). **The earlier diagnosis was wrong and the way it was wrong generalizes**: open LM shard fds do NOT locate the program counter, because `LoadBf16Direct` BORROWS the mapping and a borrowed `OwnedTensor` keeps it alive for the whole request, while the depth decoder's tensors are COPIED so its fd closes on return. A symbol-resolved profile puts the LM weight load at 180 s (1 thread, state `D`, ~92 MB/s off the NAS) in BOTH binaries, and the cost in `LinearNoBias` 42-57% / `Threadpool::Barrier` 25% / `vocoder1d::ConvTranspose1d` 88.5%; the suspected 200000 x 4096 `lm_head` transpose is 1.388 s. The GGUF k-quant ARM for the remaining components is still owed; spec §5's token-exact gate is WITHDRAWN because upstream's AR stage has no greedy path  **PARITY SWEEP + MUSIC-ONLY SERVER LANDED (#672, #953)**: upstream's whole user-facing surface was enumerated field by field (SGLang-Omni `748a0b43` `models/minimax_music3/`, diffusers `c6da9936` `modular_pipelines/minimax_music3/`) and recorded with `file:line` in spec §10.1. **`--model` is now OPTIONAL when `--speech-model` is given**, so `vllm-server --speech-model <dir>` serves `/v1/audio/speech` and registers NOTHING else — upstream's own shape (`sgl-omni serve --model MiniMaxAI/MiniMax-Music3`, `config.py:29-63`: three stages, no chat LLM), and the third instance of the task-conditional dispatch already in `server_main.cpp` for pooling and transcription-only checkpoints (vLLM `api_server.py:255-265`). ADDITIVE and proved: the ONLY case whose verdict changes is BOTH flags absent, which was an error and remains one with a message naming both options; the route table is gated in BOTH directions over a REAL SOCKET. New `examples/minimax_music3_gen` — the music family was the one generative family in this tree with NO example — a THIN C-ABI client (`vllm_speech_engine_load` + `vllm_synthesize`, `include/vllm.h` only). **Five keys upstream refuses BY NAME were SILENT here** and are now refused (#953): `temperature`/`top_p`/`top_k`/`repetition_penalty` (`request_builders.py:14-19,109-114`; the AR stage has ONE sampler, a fixed top-50 draw) and `max_new_tokens` (`request_builders.py:56-68`; upstream's length spelling in 25 Hz FRAMES, not seconds) — the #925 class exactly. **The e2e gate no longer reports a SKIP WEARING A PASS**: it read 5 cases / 5 passed with `assertions: 0` when the checkpoint was absent, the same shape that fooled this project on `test_qwen3_paged_engine`. Split into a CHECKPOINT-FREE half that runs unconditionally in CI (request contract, both ceilings, the speech-only route table over a real socket with a stub synthesizer) and the env-gated checkpoint half, whose HTTP case now runs over a REAL SOCKET against the music-only server shape rather than calling `handle_audio_speech`; a coverage-report case prints which arms ran every run. No-env arm 9 cases / 37 assertions where it was 5 / 0. **The WEIGHTS are documented** per `porting-a-model.md` §2.1: `docs/USAGE.md` carries component-by-component tables for the diffusers arm at `MiniMaxAI/MiniMax-Music3`@`fbdf52fbaaca799592917417eb05f1899f1255ec` (**28.5 GB resident, 28 517 617 303 B MEASURED**, out of a 57.4 GB repository, and why they differ), the native `.pth` arm we refuse and SGLang-Omni serves, the one implemented GGUF Q4_K artifact with its sha256, and the 14 THIRD-PARTY quantized repositories in 5 formats each marked refused. The revision is VERIFIED not copied: `condition_encoder/diffusion_pytorch_model.safetensors` on disk hashes to `83179c5e…a202c2a4d`, that revision's own LFS record. OWED and named rather than discovered: the non-`wav` response formats (no encoder vendored; note upstream DOWNMIXES TO MONO for every non-wav format, `client/audio.py:328-334`), request batching and `/v1/audio/speech/batch` (SGLang runs continuous batching at 16 with TWO engine rows per request for the CFG twin, `engine_builder.py:74-77`), and the 32 kHz delivery resample. Streaming is a PERMANENT refusal, not a debt: neither upstream arm has it (`supports_streaming_vocoder=False`). One place we are AHEAD of both arms: `guidance_scale` is a real per-request control here, where diffusers freezes it at 1.7 into the guider component (`denoise.py:180`) and SGLang exposes it only as a serve-time knob. **DEVICE ARM LANDED (#672, this PR) — PARTIAL, and the split is the point.** `SpeechModelParams` grew `device` (0 = CPU, 1 = the accelerator this build resolves; the `VideoModelParams::device` polarity, NOT `vllm_model_params.device`'s 0=auto, because every Music3 gate was taken on the CPU arm and `auto` would move every zero-filled caller off it), resolved through a NEW SHARED `multimodal::SpeechEngineDeviceType` rather than a third copy of a mapping the tree already carried twice (`minimax_h3_video.cpp:255`, `ltx2_video.cpp:706`). The engine's `vt::Queue` was a COMPILE-TIME CONSTANT (`minimax_music3_speech.cpp:492`, whose own comment said "a device arm is a queue, not a fork"); it is now that queue, built once in the constructor BEFORE any weight is read (the unified-memory recipe). `Music3LmSession` allocates its 36-layer paged KV on the queue's device, because `dense_attn::KvSlice` labels the pointer with `d.q.device` and a host `std::vector` handed to a CUDA forward is a host pointer wearing a device tensor's label. C ABI **v21**: `vllm_speech_model_params.device` + `vllm_speech_engine_device`, which reports what was GRANTED not what was requested; `--speech-device` on the server, `--device` on `minimax-music3-gen`. **WHAT MOVES is the 8.6B `Qwen3ForCausalLM` half and nothing else**: it already rides the shared `Qwen3DenseModel::ForwardEmbeds` five text registrations use, so no model was forked. **WHAT DOES NOT, named rather than implied**: the 0.646B RVQ depth decoder and the 2.4B fp32 DiT are host `std::vector<float>` scalar loops under `-ffp-contract=off` (kept reproducible against torch by W2-W5) and are OWED; the DAC Flow-VAE vocoder needs `ConvTranspose1d`, and **`vt` HAS NO SUCH OP** — `vt::Conv2d` and `vt::DepthwiseConv1d` exist but are registered for the **CPU only** and neither expresses a transposed convolution — so it is recorded as owed rather than hand-rolled outside the seam. The CPU arm is BIT-IDENTICAL: it takes the same host code with the same `std::vector` KV, and every Music3 gate is unchanged. | `MODEL-MUSIC-minimax-music3-mini-max-music3-for-conditional-generation` |
 | 📋 | `LTX2VideoTransformer3DModel` | LTX-2.5 (21.00B joint video+audio flow-matching DiT, Lightricks) | **L0 spec committed 2026-08-11 (#435, [spec](specs/ltx-2-5.md)).** Geometry MEASURED from the FP8 checkpoint's own safetensors header by HTTP range request (6124 tensors, 881,048-byte header, no payload downloaded): **21.00B** params — 48 blocks @ 386.7M = 18.560B, audio embeddings connector 2.016B, global 0.427B. The filename says `22b` and the Diffusers card says ~19B; the MEASURED count is what this row uses. Video stream 4096 (32 heads x 128), audio stream 2048 (32 heads x 64), in/out channels 128 both. Per block SIX attentions — `attn1` (video self), `attn2` (video<->text, cross 4096), `audio_attn1`, `audio_attn2` (cross 2048), and the two CROSS-MODAL `audio_to_video_attn` / `video_to_audio_attn` — which is the structural break from MiniMax-H3: H3 packs every modality into ONE sequence with per-row token tags, LTX runs TWO streams coupled by explicit cross-attention. **Per-head gated attention on every attention** (`to_gate_logits` = `Linear(query_dim, heads, bias=True)`, `attention.py:513-514`, applied AFTER the attention output at `:577`) — H3 has no analogue and getting it wrong renders plausibly-wrong rather than erroring. FFN is gelu-approximate 4096->16384->4096 with **NO bias** while `audio_ff` 2048->8192->2048 **HAS** bias, which independently confirms `ff_bias=false` / `audio_ff_bias=true` from `model_configurator.py:78-80` — checkpoint and source agreeing, not either alone. **RETRACTED 2026-08-12, was billed as a FREE WIN.** The spec claimed 2.5 sets `use_prompt_adaln_single=false` so the cross-attention K/V are timestep-free and cacheable. The shipped checkpoint DISPROVES it: it carries 12 `prompt_adaln_single`/`audio_prompt_adaln_single` tensors including a `timestep_embedder.linear_1 [4096, 256]` (256 = the sinusoidal timestep width), and `model.py:223-227` builds that module ONLY when the flag is TRUE. `transformer.py:441` was quoted as proof of no timestep term, but `:442-443` add one whenever `prompt_timestep` is not None, and the comment above them says exactly that. NO SHIPPED DEFECT: `ltx2_dit.cpp:672` refuses the cache by name when the flag is on, so the feature is correct-and-inapplicable rather than silently wrong, and stays gated bit-identical and prompt-bound for any checkpoint that does set it false. **ORACLE:** vLLM-Omni does NOT support 2.5 — `resolve_ltx_pipeline_recipe` keys on 2 / 2.3 only and RAISES otherwise (`ltx2_recipes.py:162-166`), with upstream [vllm-omni#6066](https://github.com/vllm-project/vllm-omni/issues/6066) filed 2026-08-11 — but its `DiffusersAdapterPipeline` is generic (`DiffusionPipeline.from_pretrained`, `pipeline_diffusers_adapter.py:116`), so vLLM-Omni CAN execute 2.5 via `--load-format diffusers`. Binding oracle = that adapter; immediate cross-check = Lightricks `ltx-pipelines`. **HW: FITS ONE GB10** at ~29 GB (NVFP4 DiT 18.72 + NVFP4 Gemma-4 TE 7.40 + VAEs 1.83 + upscaler 1.00) vs H3's ~41 GB. **OWED UP FRONT, not to be discovered later:** the speed axis lands `PENDING` because the adapter is a black box (`supports_step_execution=False`, `supports_request_batch=False`) and therefore NOT vLLM's production configuration, which AGENTS.md requires as the denominator; DiffVAE (`NADiffusionDecoder`, neighborhood attention) is REFUSED BY NAME until its own row rather than silently downgraded to the Conv VAE; and no render-quality claim is made from structural e2e. | `MODEL-DIFFUSION-ltx-2-5-ltx2-video-transformer-3d-model` |
@@ -451,7 +451,7 @@ Transformers compatibility is capability-driven and excluded from finite counts.
 | `MODEL-MM-kimi-audio-kimi-audio-for-conditional-generation` | `MoonshotKimiaForCausalLM` | `registry.py:449`; `vllm/model_executor/models/kimi_audio.py::KimiAudioForConditionalGeneration` | conditional generation / audio+image | MM processor; encoder/merge; audio/ASR frontend; vision encoder | ☐ required | `INVENTORIED` | none | unassigned |
 | `MODEL-MM-muse-glimmer-muse-glimmer-for-conditional-generation` | `MuseGlimmerForConditionalGeneration` | **BEYOND-PIN — NOT in `555967922`** (Muse Glimmer released 2026-08-08, after the pin, and NOT on vLLM `main` either); the only upstream implementation is the OPEN, approved-but-CI-red PR [vllm#51655](https://github.com/vllm-project/vllm/pull/51655) at head `075d645af` — `vllm/model_executor/models/muse_glimmer.py` (text attn `:1083`, decoder layer `:1212`, vision encoder `:692`, adapter `:1036`, MM processor `:262`), `vllm/transformers_utils/configs/muse_glimmer.py`, `vllm/transformers_utils/processors/muse_glimmer.py`, DFlash hooks in `qwen3_dflash.py:75-94` + `v1/spec_decode/dflash.py:83-95` | conditional generation / image+video | model loader/forward; MM processor; vision encoder + merge; video path; speculative decoding (DFlash); reasoning + tool parsers | [muse-glimmer spec](specs/muse-glimmer.md), issue [#268](https://github.com/mudler/vllm.cpp/issues/268) | `SPIKE` | **W-SPEC (2026-08-10, `row/MODEL-MUSE-GLIMMER`, records+spec only, NO code).** Traced end-to-end from PR #51655 @ `075d645af`. `meta-models/Muse-Glimmer-30B`, Apache-2.0, bf16, `model_type: muse_glimmer`; registry maps BOTH `MuseGlimmerForConditionalGeneration` and `MuseGlimmerForCausalLM` onto one class. **Text tower** L=52, H=6656, 32 q-heads / 2 kv-heads (GQA 16:1), head_dim 128, vocab 202048, 131k ctx, rope theta 5e5: Gemma2-style SANDWICH norms (input/post-attn/pre-ffn/post-ffn) computed in fp32 with a BAKED `+1` weight offset and SPLIT eps (pre-norms `rms_norm_eps`, post-norms the smaller `post_norm_eps`) (`muse_glimmer.py:1236-1247`); **iRoPE** — `no_rope_layers[i]==1` ⇒ RoPE AND sliding-window, `==0` ⇒ NoPE AND full attention, sliding×3 then full every 4th (`:1114-1116,:1167-1168`) — the ONE mechanism with no local analogue (we have no Llama-4); WEIGHTLESS QK-norm over head_dim in fp32 applied BEFORE RoPE (`:1189-1196`); post-QK-norm QUERY PRE-SCALE `scale_query_by`≈3.87 with softmax scaling left at `head_dim**-0.5` (`:1112,:1192`); per-head sigmoid ATTENTION OUTPUT GATE whose gate reads the LAYER INPUT, not the attn output (`:1203-1206`); no logit softcapping. **Two named correctness traps:** (a) the query pre-scale ships under two schemas — native raw `qk_scale_factor`≈43.784 vs modular PRE-FOLDED ≈3.87 — disambiguated upstream BY MAGNITUDE against `sqrt(head_dim)` (`:472-517`); mis-reading it scales every query by 11.3×; (b) `use_qk_norm` / `use_attn_output_gate` read as `None` (not `True`) in the modular schema and only an explicit `False` disables them (`:456-469`), so a naive `getattr(...,False)` silently drops BOTH mechanisms while still emitting plausible text. **Perception encoder** L=50, H=1536, 16 heads (head_dim 96), patch 14×14, `patch_temporal`=2, 32×32 learned pos-emb grid, interleaved window/full per `layer_types`, projector 4096→6144, image tok 200092 / video tok 200091, placeholders `<\|patch\|>`/`<\|image\|>`/`<\|video\|>`: LINEAR patchify (`conv1_linear`, NOT a conv) (`:696,:710`), BILINEAR pos-emb interpolation with per-corner validity masking and a half-pixel `+0.5/-0.5` convention (`:761-820`), 2D RoPE with **width-first** `cat([freq_w,freq_h])` (`:741-759`), block-windowed attention via a `-1`-padded permutation whose per-block valid counts become `seq_lens` (`:844-867`), pixel-shuffle downsample asserting `output_dim == hidden*merge^2` (`:822-842`), plain `LayerNorm` (not RMSNorm) at ln_pre/ln_post. **DFlash** adds NO new drafter — PR #51655 only recognises `MuseGlimmerAssistantModel` as method `dflash` and threads the TARGET's `is_neox_style` into the draft config, because a draft/target RoPE-layout mismatch is SILENT (acceptance collapses, nothing errors, output stays correct). **REUSE MAP:** sandwich norms → `gemma2.cpp`/`gemma4.cpp`; output gate → `qwen3_5.cpp`; weightless QK-norm → Qwen3 family; windowed vision tower → `qwen3_vl_vision.cpp`; gated MLP → `layers::MlpGateUpMethodBase`+`vt::MergedGemmGroup`; decode → `ModelRegistry::Forward`+`dense_attn::AttnBlock`; fusion → `vt::FusedChain`; DFlash → the existing speculator row. **HONESTY:** the pinned oracle CANNOT load `muse_glimmer` (and the checkpoint wants transformers 5.15.0.dev0 vs the pin's 5.14.1), so there is NO gateable oracle and **every performance axis is an OPEN GAP by construction** — no parity/throughput claim may be made from this row until #51655 merges and the pin advances. Correctness gates against the HF reference instead, with per-mechanism RED-first mutation carrying the weight the missing oracle would have. Weights (~60 GiB bf16) NOT downloaded; GB10 fit needs a quantized arm. W-plan W0-W7 in the spec. **W0 CPU SCAFFOLDING LANDED (2026-08-10, `CLAIM-MUSE-GLIMMER-W0`, CPU-only, no weights, no GPU):** additive TUs register BOTH `MuseGlimmerForCausalLM` and `MuseGlimmerForConditionalGeneration` onto ONE factory (mirroring registry.py @ #51655); `ParseMuseGlimmerParams` descends the canonical NESTED layout AND normalizes the older FLAT layout (configs/muse_glimmer.py:186-305 — without it a flat config silently deserializes to an ALL-DEFAULT text config, ignoring every checkpoint value with no error); `NormalizeMuseGlimmerWeightName` ports the `hf_to_vllm_mapper` (:1389-1425) for BOTH checkpoint conventions; `EnumerateMuseGlimmerTensors` is the structural name map, deliberately OMITTING the three WEIGHTLESS modules (`embed_norm` :1286, per-head `qk_norm` :1121, `perception_emb_norm` :1470) that ship no tensor; forward REFUSES-by-name. Gate `tests/vllm/models/test_muse_glimmer_scaffold.cpp` **11/11 cases, 73/73 assertions**, clean CPU `-Werror` build. **RED-FIRST MUTATION-PROVEN, all four named traps** (each mutated in-tree, gate went RED, tree restored byte-identical): (1) treating the native raw `qk_scale_factor` as pre-folded → 3 RED (the 11.3x query blow-up); (2) defaulting the absent `use_qk_norm`/`use_attn_output_gate` to FALSE → 4 RED; (3) counting the iRoPE mask FORWARD instead of backward from the last layer → 5 RED; (4) applying the legacy-guac sandwich-norm renames in the wrong order (which SWAPS post-attention with pre-feedforward) → 1 RED. KV-cache spec is a documented W0 placeholder (one full-attention group; the real sliding/full split rides the Gemma-4 per-layer seam at W1). **W1-W7 FOLLOWED ON THE SAME BRANCH (`row/MODEL-MUSE-GLIMMER`, PR #279, NOT merged), and this paragraph's "config/name-map/registry only, NO forward" is superseded:** W1 the 52-layer text forward (sandwich norms w/ split eps, iRoPE, weightless QK-norm before RoPE, query pre-scale, attn output gate, SwiGLU, untied lm_head, output_multiplier before the soft-cap) gated vs an independent fp32 transcription of `075d645af` plus a property test per mechanism (`test_muse_glimmer_text`, 21 cases / 487 assertions, and the SAME binary re-run under `VT_FUSED_CHAIN_ADOPT=0` so the non-FusedChain fallback arm is gated too); W3 the 50-layer perception encoder (`test_muse_glimmer_vision`); W4 the mm wiring — tower → adapter → projection → `perception_emb_norm` → masked scatter onto the image/video placeholder rows, with the released 30B's 1436 tensors accounted 1436/1436 (`test_muse_glimmer_wiring`, 9/10316); W7 the ATEM reasoning + tool parsers (34 cases / 251 assertions). **REAL-WEIGHT EVIDENCE, and its exact limits:** on real 30B tensors at **reduced depth 4 of 52** our forward's 5 prefill argmax positions are identical to our torch transcription AND to HF's own `muse_glimmer` (`exportable-muse` @`a9e337e8`); max abs diff 0.0889745, cosine 0.999981. NOT established, and not to be implied: the full-depth 52-layer arm NEVER RAN; these are prefill argmax positions, NOT generated tokens, so multi-step decode and the sliding window across steps are untested; the perception encoder has NO reference check of any kind (the wiring gate proves reachability and placement, not that an image produces the right tokens); nothing has run end to end through the server; and the ATEM parsers' channel scoping DOES NOT WORK at server defaults, because the ToolParser/ReasoningParser seam has no `adjust_request` dispatch site while `skip_special_tokens` defaults true and the checkpoint marks `<\|start\|>`/`<\|message\|>`/`<\|eom\|>`/`<\|eot\|>` special (open gap, spec §6.7). **NO vLLM SPEED AXIS ON ANY DIMENSION** — unchanged and unchangeable until #51655 merges and the pin advances. A SECONDARY, quant-matched llama.cpp bar now EXISTS and is measured (#333, 2026-08-11, `row/MUSE-BENCH-2`, spec §14): same 16.76 GB Q4_K_M file both engines on an idle GB10 CPU, 128-token prefill **tie at 0.997x** (0.898x at 20 threads), decode **0.232x**, 512-token prefill 0.170x, peak RSS **1.92x MORE** (30.29 vs 15.74 GiB, the §10.2 dequantized qkv/lm_head/embed). It is a secondary reference and is NEVER the denominator; the vLLM cell stays an open gap by construction. Row token stays `SPIKE` deliberately: advancing it owes `docs/STATUS.md` and `.agents/NOW.md` rows and both surfaces sit byte-exactly at their shrink-only ratchet/budget, so the advance belongs to the landing commit | `CLAIM-MUSE-GLIMMER-SPEC`, `CLAIM-MUSE-GLIMMER-W0`, `CLAIM-MUSE-GLIMMER-W1`, `CLAIM-MUSE-GLIMMER-W3`, `CLAIM-MUSE-GLIMMER-W4`, `CLAIM-MUSE-GLIMMER-W7`, `CLAIM-MUSE-GLIMMER-FIX` |
 | `MODEL-MM-kimi-k3-kimi-k3-for-conditional-generation` | `KimiK3ForConditionalGeneration` | **BEYOND-PIN — NOT in `555967922`** (K3 released 2026-07-27, after the pin); closest registered = its literal text backbone `KimiLinearForCausalLM` (`registry.py:140`; `vllm/model_executor/models/kimi_linear.py`) + the K2.5 vision wrapper `kimi_k25.py:290` / tower `kimi_k25_vit.py` | conditional generation / image (text-first) | model loader/forward; FusedMoE/grouped GEMM; GDN/linear-attention state (KDA); MLA/latent KV; MXFP4 compressed-tensors quant; MM processor + MoonViT-V2 encoder/merge | [kimi-k3 spike](specs/kimi-k3.md) | `SPIKE` | **W0 SCOPE (2026-07-28, `CLAIM-KIMI-K3-SCOPE`, DERIVE-AND-SHIP, records-only).** From the HF `config.json` (fetch-derived): `architectures:["KimiK3ForConditionalGeneration"]`, `text_config.architectures:["KimiLinearForCausalLM"]` — the text backbone IS the pinned Kimi-Linear hybrid, MASSIVELY scaled: **H=7168, L=93 (69 KDA + 24 MLA full-attn), 896 experts / top-16 / 2 shared, `moe_intermediate_size=3072`**; MLA geometry `kv_lora=512`/`q_lora=1536`/`qk_nope=128`/`qk_rope=64` (= our landed DeepSeek-V3 dims); KDA `head_dim=128`/`num_heads=96`/`short_conv=4`/`gate_lower_bound=-5.0`; quant **`mxfp4-pack-quantized` (compressed-tensors, group 32, e8m0) + MXFP8 acts (QAT)**; vision **MoonViT-V2** (~401M, patch 14, 27L). **HEAVY REUSE** — GDN (KDA's parent, `cuda_gdn.cu`/`gdn_attn.cpp`), DeepSeek MLA (`deepseek_v2.cpp`/`mla_attention.*`, exact geometry), DeepSeek-style MoE (`qwen3_moe.cpp`/`cuda_moe.cu`, scale to 896), and the Qwen3.6-35B GDN-hybrid-MoE model skeleton (`qwen3_5_moe.cpp`) are the structural twins; Kimi-K2 tokenizer/tool parser (`parser/kimi_k2.cpp`) reused. **NET-NEW:** the KDA kernel delta (per-channel `[H,D]` low-rank decay `f_a_proj`/`f_b_proj` + sigmoid-gated output norm + 3 q/k/v convs — already scoped on the Kimi-Linear row), **MXFP4** (we have NVFP4 group-16, not MXFP4 group-32/e8m0), **AttnRes** (report-only, UNCONFIRMED — not in config.json nor pinned `kimi_linear.py`), and the **MoonViT-V2 tower**. **HW-fit: DOES NOT FIT GB10** — 2.8T MXFP4 ≈ **1.56 TB ≈ ~12× over the 119 GiB pool**; no small K3 exists. **DERIVE-AND-SHIP** (no on-box golden, like the beyond-vLLM CUDA bricks): (a) REAL proxy gate of KDA+MLA+MoE on the FITTING `Kimi-Linear-48B-A3B` (~89–91 GiB) vs the pinned oracle, (b) build-verify + structural review for the K3 scale-up. The pinned oracle has NO `kimi_k3` ⇒ even HW-rich users need a pin advance to oracle-gate K3 itself. CORRECTS the 2026-07-25 sweep note ("loads as `DeepseekV3ForCausalLM`" — true for K2, NOT K3). W-plan W1-W8 in the spec. **W2/W5 CPU SCAFFOLDING LANDED (2026-07-28, `CLAIM-KIMI-K3-W2-W5`, DERIVED+BUILD-VERIFIED, NOT pushed):** additive registry TU registers `KimiK3ForConditionalGeneration` (info: text-gen + `is_hybrid` + `has_inner_state` + `supports_multimodal`); config-descent `ParseKimiK3Params` reads the nested `text_config` (KimiLinear KDA+MLA+MoE scalars — note upstream key `num_experts_per_token`), `vision_config` (MoonViT-V2 PARTIAL), `quantization_config` (MXFP4 detect) grounded in `configs/kimi_linear.py:11-148`; pure `EnumerateKimiK3TextBackboneTensors` is the 93-layer KDA/MLA + 896-expert MoE structural name-map grounded 1:1 in `kimi_linear.py:104-378,460-554` + `kimi_gdn_linear_attn.py:102-226` (per-layer KDA vs MLA-with-qLoRA vs MoE-vs-dense branching); forward REFUSE-by-name (`VT_CHECK(false)`, mirrors `deepseek_v4.cpp`); loader REFUSES MXFP4 (a real K3 checkpoint's dtype) deferring to the shared DeepSeek-V4 MXFP4 row. **NOT-YET-BUILDABLE (correctly deferred):** MXFP4 materialization (→ `CLAIM-DEEPSEEK-V4-*` / quantization-matrix MXFP4), the KDA kernel delta (→ Kimi-Linear row `MODEL-TEXT-kimi-linear-*`), MoonViT-V2 vision (W7), K3 multimodal-wrapper weight prefix (post-pin). Code: `src/vllm/model_executor/models/kimi_k3{,_registry,_weights}.cpp` + `include/vllm/model_executor/models/kimi_k3.h`. Test: `tests/vllm/models/test_kimi_k3_scaffold.cpp` (6/6, 63 assertions — registry-resolve + config-descent + split logic + enumeration + reject + MXFP4-refuse). Clean CPU build (`-DVLLM_CPP_CUDA=OFF`). Row stays `SPIKE` (no on-box e2e; forward not implemented); registration `src/vllm/model_executor/models/kimi_k3_registry.cpp:126`; test `tests/vllm/models/test_kimi_k3_scaffold.cpp:123` | `CLAIM-KIMI-K3-SCOPE`, `CLAIM-KIMI-K3-W2-W5` |
-| `MODEL-MM-dots3-note-dots3-note-for-causal-lm` | `Dots3NoteForCausalLM` | **BEYOND-PIN — NOT in `555967922`** (the checkout at the pin carries only `dots_ocr.py`, verified); present on vLLM `main` as the platform-split package `vllm/models/dots3_note/{common,nvidia}/` (15 files, ~5.7k LoC) added by [vllm#51255](https://github.com/vllm-project/vllm/pull/51255) @ `9035151d6`, last touched `170592a93` (2026-08-13, [vllm#52172](https://github.com/vllm-project/vllm/pull/52172)); `registry.py:381` (`_MULTIMODAL_MODELS`) -> `("vllm.models.dots3_note", "Dots3NoteForCausalLM")` (re-derived at `c205726108df54bb6fbf15b19e725a4a3add2b18`; W0 read `:375`); config `transformers_utils/configs/dots3_note.py:7` `Dots3NoteConfig(DeepseekV3Config)`; text tower `nvidia/model.py` (`Dots3NoteMoE:76`, `Dots3NotePaddedMLAAttention:204`, `Dots3NoteFullAttention:219`, `Dots3NoteSlidingAttention:329`, `Dots3NoteDecoderLayer:481`, `Dots3NoteModel:549`, shared body `_forward_note_mla:135`); SWA-MLA stack `nvidia/attention.py` (`_gather_swa_kv_kernel:48`, `_apply_swa_score_mask_kernel:118`, `Dots3NoteMLAMetadataBuilder:306`, `Dots3NoteTritonMLAImpl:438`, `Dots3NotePaddedSparseImpl:689`); vision `nvidia/vision.py:508` + `nvidia/vision_moe.py`; audio `nvidia/audio_encoder.py:437`; MM front end `nvidia/multimodal.py:49`; processor `common/processor.py` + `common/video.py` | conditional generation / text+image+video+audio understanding | model loader/forward; MLA/latent KV; **sliding-window MLA (new)**; heterogeneous/padded MLA KV spec (new); DSA lightning indexer; ungrouped `noaux_tc` FusedMoE 256/top-8; MM processor + MoE ViT encoder/merge; video path; `dots` Whisper-variant audio tower; blockwise-FP8 + owed GGUF k-quant arms; MTP | [dots3-note spec](specs/dots3-note.md), issue [#699](https://github.com/mudler/vllm.cpp/issues/699) | `SPIKE` | **W1 LANDED 2026-08-23 (`row/MODEL-MM-dots3-note-W1`, config + registry, CPU-only, no GPU, no checkpoint downloaded).** Additive TU `src/vllm/model_executor/models/dots3_note_registry.cpp` with ONE `REGISTER_VLLM_MODEL` and zero edit to a shared array; `dots3_note.{h,cpp}` carry `Dots3NoteParams`, the on-disk name map and a forward that REFUSES BY NAME. `Dots3NoteMTPModel` is deliberately NOT registered and stays `INVENTORIED` (W10): a speculator that cannot propose makes the engine accept a config it then dies on mid-run. Gate `test_dots3_note_scaffold` **19 cases / 3876 assertions** (15/3694 as first landed; the #1805 review added four). **ALL SIX §4 CONFIG TRAPS RED-FIRST**, captured on a RED arm that COMPILED and ran (7/15 cases red, 16 assertions) before the correct values existed: `n_group` 8→1 and `topk_group` 4→1 (DeepseekV3Config `:168-169` would regroup the router at every MoE layer, and a grouped config is now REFUSED by name); `indexer_rope_interleave` false→true (`deepseek_v2.py:1148` consumes it as `is_neox_style = not …`, so V3.2's absent-key default rotates split-half NeoX); `num_nextn_predict_layers` 0→1, whose RED knock-on was 19 UNCLAIMED tensors ("first: model.layers.46.eh_proj.weight"); the two `apply_mla_qkv_lora_rescale` scalars, which DISAGREE between the geometries (kv `sqrt(10)` full vs `sqrt(5)` sliding); and `swa_rope_theta` 5e4 against the full layers' 8e7. **§4 item 6 CORRECTED in place** ([#1804](https://github.com/mudler/vllm.cpp/issues/1804)): `is_neox_style=False` is NOT sliding-only — `Dots3NoteFullAttention` inherits the same hard-coded value from `deepseek_v2.py:1093-1098`, so BOTH MLA ropes are GPT-J and only the theta differs. **§1.4 RESOLVED by the checkpoint**: exactly ONE nextn layer, carrying the SLIDING attention set (no `indexer.*`, `q_b_proj` [16384,1024]) and a DENSE MLP; `model.mtp.embed_tokens.weight` present and `shared_head.head.weight` absent, matching `has_own_embed_tokens=True`/`has_own_lm_head=False` (`mtp.py:141-142`). Upstream cannot answer the sliding half — `config.layer_types[layer_idx]` (`model.py:503`) has no entry at index 46 — so the checkpoint is the authority and W10 owes the reconciliation. **Tensor accounting 1614/1614 both ways** over a committed HEADERS-ONLY slice of `model.safetensors.index.json` @ `1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b` (root + layer 0 dense/full + layer 1 MoE/full + layer 2 MoE/sliding + nextn 46), out of the 38006 the checkpoint ships; the 2625 vision/audio tensors are NAMED W6/W7 deferrals in the loader's classifier, never silent drops. W2 owes the whole index. Shapes verified against the released safetensors headers (no tensor byte read): full kv row 576 vs sliding 1088, headwise `g_proj` width = head count, indexer 64x128. KV spec emits ONE MLA group at the PADDED 1088 row (`model.py:283`, `:213`), not the full layers' 576. **Memory format**: `mlp.gate.e_score_correction_bias` is F32 in an otherwise BF16 tower. The forward refusal is driven through the REAL model the factory returns — a fabricated `LoadedModel` stub is UB the moment the handle opens ([#730](https://github.com/mudler/vllm.cpp/issues/730)/[#775](https://github.com/mudler/vllm.cpp/issues/775)) — with a separate foreign-model case for the checked-downcast guarantee alone. Config parsing lives in the model TU, NOT `hf_config.cpp` as spec §3.2 item 6 proposed: `hf_config.cpp` is the shared container reader and AGENTS.md forbids a surface every PR must write (spec §4.2); it needs no edit for this checkpoint. Upstream read at vLLM `origin/main` `c205726108df54bb6fbf15b19e725a4a3add2b18` — still BEYOND-PIN. Row stays `SPIKE`: the §8.1 heading restructure `ACTIVE` requires belongs to the brick where the forward stops refusing. NO number is claimable on any axis while spec §6.4 option B holds. **THREE REVIEW FINDINGS REPAIRED IN THE SAME PR ([#1805](https://github.com/mudler/vllm.cpp/pull/1805)).** (F1) The config readers SUBSTITUTED A DEFAULT for an absent or wrong-typed key, which is the §4 item 4 trap and the one item W1 shipped ungated: deleting `apply_mla_qkv_lora_rescale` and `swa_rope_theta` made the parse SUCCEED with all four LoRA scales at 1.0 and 33 of 46 layers at 1e4 instead of 5e4, and neither key is one of the four setdefaults — upstream raises `AttributeError`. Measured blast radius: **26 of 36 required keys parsed clean when deleted, all 36 when wrong-typed**. Every read is now REQUIRED (refuse by name) or OPTIONAL-with-a-named-upstream-default (type-strict), a wrapped `text_config` layout refuses, and §4.3 records the deliberate stricter-than-upstream call for the DeepseekV3Config-inherited subset. (F2) The tower-deferral subcase asserted only `REQUIRE_NOTHROW`, so reclassifying all 2625 tower tensors as LANGUAGE stayed GREEN; `acc.vision`/`acc.audio`/`acc.language` are now asserted by count — the mutation reports `acc.vision=0, acc.language=35386` against 35381. (F3) `docs/FEATURES.md` listed this arch as BOTH registered and "not registered", and its 36/40 counts were stale at 37/41. Also repaired: §4.1/§4.3 trap numbering unified on §4's own 1-6, `model.py` anchors re-derived (`:283` not `:315`, `__init__` `:222-308`/`:332-460`), §1.1's rope row and §1.4 corrected, the `rope_is_neox_style` struct default flipped to the never-correct value so deleting the resolution step fails loudly, and the CMake comment orphaning undone. **W0 SCOPE (2026-08-14, `CLAIM-MODEL-DOTS3-NOTE-W0`, records-only — no code, no build, no download, no GPU).** `dots-studio/dots3-note-prev`, 280B total / 16B active, `model_type: dots3_note`. **Measured** from `config.json` + the HF API: L=46, H=5120, `intermediate_size` 13824, vocab 152064, 524288 positions, `rope_theta` 8e7; MoE 256 routed + 1 shared, top-8, `moe_intermediate_size` 1536, `first_k_dense_replace` 1, sigmoid/`noaux_tc`. **Attention is hybrid with TWO geometries:** 13 `full_attention` layers at indices 0,1,5,9,...,45 (128 heads, `kv_lora_rank` 512, `qk_nope` 128, DSA indexer `index_n_heads` 64 / `index_head_dim` 128 / `index_topk` 2048) and 33 `sliding_attention` layers (window **513**, 64 heads, `swa_kv_lora_rank` **1024**, `swa_qk_nope_head_dim` **192**, own `rope_theta` 5e4, `is_neox_style=False`, no indexer) => latent rows 576 vs 1088, reconciled upstream by padding the physical row (`Dots3NotePaddedMLAAttention.get_kv_cache_spec`). Vision = MoE ViT, 42 layers, `embed_dim` 1536, patch 14, `spatial_merge_size` 2, `use_qk_norm`, `pre_pixel_shuffle`, patch-merger adapter 1536->5120, **pyramid** `pyramid_num_routed` (dense to layer 24, then 4->64). Audio = `encoder_type: "dots"` Whisper variant, `d_model` 1280 / 32 layers / 128 mel / `max_source_positions` **6000** / **swiglu**, conv2d stem, RoPE (partial 0.5), RMSNorm, 60 s chunking, adapter 1280->5120. **HEAVY REUSE:** upstream literally subclasses DeepSeek — `Dots3NoteModel(DeepseekV32Model)`, `Dots3NoteDecoderLayer(DeepseekV32DecoderLayer)`, `Dots3NoteMoE(DeepseekV2MoE)`, `Dots3NoteFullAttention(DeepseekV2MLAAttention)` — so our gated MLA (`deepseek_v2.cpp`, SACRED 8/8), DSA indexer (`deepseek_v4_dsa.cpp`), `noaux_tc` router, 256-expert grouped MoE, `qwen3_vl_vision.cpp` (image+video STRICT 32/32) and `voxtral.cpp`/`whisper_audio.cpp` (audio 16/16) carry most of it. **NET-NEW:** windowed MLA (our every MLA path is full-attention — `deepseek_v4_registry.cpp` sets `is_hybrid=false`), the padded/heterogeneous MLA KV spec, the headwise sigmoid attention gate + the extra `k_rope_only_layernorm` over the 64-dim rope-only k slice + the two `apply_mla_qkv_lora_rescale` scalars, the MoE ViT (all our ViTs are dense), the `dots` audio stem, and the GGUF converter (llama.cpp has no `dots3_note` => no quant-matched llama.cpp bar). **FOUR SILENT CONFIG TRAPS** the checkpoint's `config.json` does NOT carry, all set in `Dots3NoteConfig.__init__:12-25`: `n_group=1`/`topk_group=1` (upstream: *"Do not inherit DeepSeek-V3's 8-group/4-group router defaults ... A different grouping changes the selected experts at every MoE layer"* — our router is gated at V3's grouped dims), `indexer_rope_interleave=True` (GPT-J adjacent pairs, where V3.2 and therefore our indexer default to split-half NeoX), and `num_nextn_predict_layers=1`. A fifth, in-JSON: `apply_mla_qkv_lora_rescale` => `sqrt(hidden/lora_rank)` after each layernorm. **HW-fit: DOES NOT FIT ANY HOST WE OWN.** Measured via the HF API: 131 language shards + `model-vision.safetensors` 13.7 GB + `model-audio.safetensors` 1.77 GB = **~576 GB bf16**; the `-fp8` sibling ~290 GB; upstream's own recipe is `--tensor-parallel-size 8` on H100s. GB10 is ~119 GiB and Thor ~122 GiB RAM / 123 GiB free disk, so **the vLLM oracle cannot run this model here at any published precision** and there is no smaller `dots3-note` in the `dots-studio` org — the same wall [deepseek-v4-flash](specs/deepseek-v4-flash.md) hit at 156.7 GiB, three times worse. Our own arm is plausible at <=2 bpw (~75-90 GiB, the ds4flash IQ2_XXS pattern) but an arm with no oracle beside it yields NO parity number, so **every performance axis is an open gap by construction**. Beyond-pin on top of that: gating needs a pin advance, and upstream is still moving (`#52172` landed the day before this scope). Thor `192.168.68.23` is the developer-designated e2e CUDA host (2026-08-14) and probed the same day — 14 cores, 122 GB RAM, 123 GiB free, but **no nvcc, no cmake, no ninja, no venv**, and `nvidia-smi` refuses under non-interactive ssh — so W0.5 provisions it; it can carry our low-bit arm, the sm_110 axis and every unit gate, but not a 290 GB oracle. **DECISION 2026-08-15: option B** (developer) — port brick by brick against independent references, ship our own low-bit arm on Thor, and carry the e2e parity gate as an OPEN GAP. Consequence recorded on the row, not just in the spec: **no performance number for this model is claimable on any axis** while B holds, because there is no denominator; and a brick gated against an in-test reference proves two implementations AGREE, never that either matches vLLM. The owed gate lives in the spec's `## Owed`. W0.5 (provision Thor) is in flight and is NOT dots3-specific — Thor is the only non-GB10 CUDA host, and it runs an immutable OS with `/` READ-ONLY, so the recipe goes in environment.md. W-plan W0-W11 in the spec | `CLAIM-MODEL-DOTS3-NOTE-W0` |
+| `MODEL-MM-dots3-note-dots3-note-for-causal-lm` | `Dots3NoteForCausalLM` | **BEYOND-PIN — NOT in `555967922`** (the checkout at the pin carries only `dots_ocr.py`, verified); present on vLLM `main` as the platform-split package `vllm/models/dots3_note/{common,nvidia}/` (15 files, ~5.7k LoC) added by [vllm#51255](https://github.com/vllm-project/vllm/pull/51255) @ `9035151d6`, last touched `170592a93` (2026-08-13, [vllm#52172](https://github.com/vllm-project/vllm/pull/52172)); `registry.py:381` (`_MULTIMODAL_MODELS`) -> `("vllm.models.dots3_note", "Dots3NoteForCausalLM")` (re-derived at `c205726108df54bb6fbf15b19e725a4a3add2b18`; W0 read `:375`); config `transformers_utils/configs/dots3_note.py:7` `Dots3NoteConfig(DeepseekV3Config)`; text tower `nvidia/model.py` (`Dots3NoteMoE:76`, `Dots3NotePaddedMLAAttention:204`, `Dots3NoteFullAttention:219`, `Dots3NoteSlidingAttention:329`, `Dots3NoteDecoderLayer:481`, `Dots3NoteModel:549`, shared body `_forward_note_mla:135`); SWA-MLA stack `nvidia/attention.py` (`_gather_swa_kv_kernel:48`, `_apply_swa_score_mask_kernel:118`, `Dots3NoteMLAMetadataBuilder:306`, `Dots3NoteTritonMLAImpl:438`, `Dots3NotePaddedSparseImpl:689`); vision `nvidia/vision.py:508` + `nvidia/vision_moe.py`; audio `nvidia/audio_encoder.py:437`; MM front end `nvidia/multimodal.py:49`; processor `common/processor.py` + `common/video.py` | conditional generation / text+image+video+audio understanding | model loader/forward; MLA/latent KV; **sliding-window MLA (new)**; heterogeneous/padded MLA KV spec (new); DSA lightning indexer; ungrouped `noaux_tc` FusedMoE 256/top-8; MM processor + MoE ViT encoder/merge; video path; `dots` Whisper-variant audio tower; blockwise-FP8 + owed GGUF k-quant arms; MTP | [dots3-note spec](specs/dots3-note.md), issue [#699](https://github.com/mudler/vllm.cpp/issues/699) | `SPIKE` | **W2 LANDED 2026-08-24 (`row/MODEL-MM-dots3-note-W2`, the WHOLE weight map, CPU-only, no GPU, no tensor byte downloaded).** The complete released `model.safetensors.index.json` @ `1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b` (sha256 `95a364b4…`) joined to the safetensors HEADER of all 133 files it names — two HTTP Range requests each, **4770592 header bytes and ZERO tensor bytes** against a 576886825984-byte checkpoint. **Accounting 38006/38006: 35381 language / 2195 vision / 430 audio, zero unaccounted, zero missing, zero duplicated, zero invented**, every bucket asserted BY NUMBER and the whole 38006-name set driven through `ModelRegistry::Resolve(...).factory->load_weights` as well as through the classifier. Gate `test_dots3_note_scaffold` **26 cases / 110821 assertions** (19/3876 at W1). The two towers stopped being integers: `Dots3NoteDeferredTowers()` is a table of records — prefix, the ONE file each tower ships in, the brick that owes it (W6 vision, W7 audio) and what it is — the classifier dispatches on it, and the load refusal prints it, so an unknown tensor is distinguishable from a deferred one. **THREE facts the W1 slice could not reach.** (1) The backbone has EXACTLY FOUR distinct layer shapes — `{0}` dense+full 19 tensors, the 12 full+MoE `{1,5,…,45}` at 789, the 33 sliding+MoE at 784, `{46}` at 18 — so W1's "the remaining 42 layers repeat 1/2 exactly" is now measured and no layer breaks the pattern. (2) The full/sliding split derived from the SHIPPED `indexer.wk` tensors equals `config.layer_types` exactly (13 full, `q_b_proj` [24576,1024] against [16384,1024]), which is two independent released artifacts agreeing. (3) The checkpoint carries **62 F32 tensors in TWO families, not one**: the 45 language `mlp.gate.e_score_correction_bias` W1 predicted, and **17 `vision_encoder.blocks.{25..41}.mlp.router_bias`** the language-only slice could not see, whose widths are the pyramid's own 4,8,…,64,64 — fp32 learned router statistics inside a BF16 tower, and the census closes at 37944 BF16 + 62 F32 with no third dtype. **NOT spec R5**, which an earlier draft of this cell claimed: R5/§2.4 is the FP32 dynamic ACTIVATION SCALES in `note_vision_fused_moe_fp8`, and the bf16 checkpoint ships no scale tensors at all — R5 stays wholly owed by W6. Audio is BF16 throughout, and §1.2's vision pyramid is now checkpoint-measured rather than config-inferred. **A FOURTH is a finding, not a confirmation** ([#1846](https://github.com/mudler/vllm.cpp/issues/1846)): the index's `metadata` block declares `indexer_rope_layout: "leading"` and `indexer_rope_converted_from: "tail"`, and `git grep indexer_rope_layout` over vLLM `origin/main` returns NOTHING — upstream reads neither. It agrees with what upstream's code does anyway (`DeepseekV2Indexer` rotates `[..., :rope_dim]`, `deepseek_v2.py:805,:814`), and it is NOT §4 trap 2: trap 2 is which PAIRS rope rotates, this is which HALF of the 128-wide head it rotates. W2 pins both values in an assertion and consumes neither; W3 owes the slice. Evidence, the fetch recipe, the tower inventories W6/W7 inherit and the 15-row mutation table (every row RED with `compile_err=0` printed beside it, one of them a REPLACEMENT for a guard that came back GREEN and was deleted) are spec §4.4. Row stays `SPIKE`: reading a weight map is not porting a model. **W1 LANDED 2026-08-23 (`row/MODEL-MM-dots3-note-W1`, config + registry, CPU-only, no GPU, no checkpoint downloaded).** Additive TU `src/vllm/model_executor/models/dots3_note_registry.cpp` with ONE `REGISTER_VLLM_MODEL` and zero edit to a shared array; `dots3_note.{h,cpp}` carry `Dots3NoteParams`, the on-disk name map and a forward that REFUSES BY NAME. `Dots3NoteMTPModel` is deliberately NOT registered and stays `INVENTORIED` (W10): a speculator that cannot propose makes the engine accept a config it then dies on mid-run. Gate `test_dots3_note_scaffold` **19 cases / 3876 assertions** (15/3694 as first landed; the #1805 review added four). **ALL SIX §4 CONFIG TRAPS RED-FIRST**, captured on a RED arm that COMPILED and ran (7/15 cases red, 16 assertions) before the correct values existed: `n_group` 8→1 and `topk_group` 4→1 (DeepseekV3Config `:168-169` would regroup the router at every MoE layer, and a grouped config is now REFUSED by name); `indexer_rope_interleave` false→true (`deepseek_v2.py:1148` consumes it as `is_neox_style = not …`, so V3.2's absent-key default rotates split-half NeoX); `num_nextn_predict_layers` 0→1, whose RED knock-on was 19 UNCLAIMED tensors ("first: model.layers.46.eh_proj.weight"); the two `apply_mla_qkv_lora_rescale` scalars, which DISAGREE between the geometries (kv `sqrt(10)` full vs `sqrt(5)` sliding); and `swa_rope_theta` 5e4 against the full layers' 8e7. **§4 item 6 CORRECTED in place** ([#1804](https://github.com/mudler/vllm.cpp/issues/1804)): `is_neox_style=False` is NOT sliding-only — `Dots3NoteFullAttention` inherits the same hard-coded value from `deepseek_v2.py:1093-1098`, so BOTH MLA ropes are GPT-J and only the theta differs. **§1.4 RESOLVED by the checkpoint**: exactly ONE nextn layer, carrying the SLIDING attention set (no `indexer.*`, `q_b_proj` [16384,1024]) and a DENSE MLP; `model.mtp.embed_tokens.weight` present and `shared_head.head.weight` absent, matching `has_own_embed_tokens=True`/`has_own_lm_head=False` (`mtp.py:141-142`). Upstream cannot answer the sliding half — `config.layer_types[layer_idx]` (`model.py:503`) has no entry at index 46 — so the checkpoint is the authority and W10 owes the reconciliation. **Tensor accounting 1614/1614 both ways** over a committed HEADERS-ONLY slice of `model.safetensors.index.json` @ `1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b` (root + layer 0 dense/full + layer 1 MoE/full + layer 2 MoE/sliding + nextn 46), out of the 38006 the checkpoint ships; the 2625 vision/audio tensors are NAMED W6/W7 deferrals in the loader's classifier, never silent drops. W2 owes the whole index. Shapes verified against the released safetensors headers (no tensor byte read): full kv row 576 vs sliding 1088, headwise `g_proj` width = head count, indexer 64x128. KV spec emits ONE MLA group at the PADDED 1088 row (`model.py:283`, `:213`), not the full layers' 576. **Memory format**: `mlp.gate.e_score_correction_bias` is F32 in an otherwise BF16 tower. The forward refusal is driven through the REAL model the factory returns — a fabricated `LoadedModel` stub is UB the moment the handle opens ([#730](https://github.com/mudler/vllm.cpp/issues/730)/[#775](https://github.com/mudler/vllm.cpp/issues/775)) — with a separate foreign-model case for the checked-downcast guarantee alone. Config parsing lives in the model TU, NOT `hf_config.cpp` as spec §3.2 item 6 proposed: `hf_config.cpp` is the shared container reader and AGENTS.md forbids a surface every PR must write (spec §4.2); it needs no edit for this checkpoint. Upstream read at vLLM `origin/main` `c205726108df54bb6fbf15b19e725a4a3add2b18` — still BEYOND-PIN. Row stays `SPIKE`: the §8.1 heading restructure `ACTIVE` requires belongs to the brick where the forward stops refusing. NO number is claimable on any axis while spec §6.4 option B holds. **THREE REVIEW FINDINGS REPAIRED IN THE SAME PR ([#1805](https://github.com/mudler/vllm.cpp/pull/1805)).** (F1) The config readers SUBSTITUTED A DEFAULT for an absent or wrong-typed key, which is the §4 item 4 trap and the one item W1 shipped ungated: deleting `apply_mla_qkv_lora_rescale` and `swa_rope_theta` made the parse SUCCEED with all four LoRA scales at 1.0 and 33 of 46 layers at 1e4 instead of 5e4, and neither key is one of the four setdefaults — upstream raises `AttributeError`. Measured blast radius: **26 of 36 required keys parsed clean when deleted, all 36 when wrong-typed**. Every read is now REQUIRED (refuse by name) or OPTIONAL-with-a-named-upstream-default (type-strict), a wrapped `text_config` layout refuses, and §4.3 records the deliberate stricter-than-upstream call for the DeepseekV3Config-inherited subset. (F2) The tower-deferral subcase asserted only `REQUIRE_NOTHROW`, so reclassifying all 2625 tower tensors as LANGUAGE stayed GREEN; `acc.vision`/`acc.audio`/`acc.language` are now asserted by count — the mutation reports `acc.vision=0, acc.language=35386` against 35381. (F3) `docs/FEATURES.md` listed this arch as BOTH registered and "not registered", and its 36/40 counts were stale at 37/41. Also repaired: §4.1/§4.3 trap numbering unified on §4's own 1-6, `model.py` anchors re-derived (`:283` not `:315`, `__init__` `:222-308`/`:332-460`), §1.1's rope row and §1.4 corrected, the `rope_is_neox_style` struct default flipped to the never-correct value so deleting the resolution step fails loudly, and the CMake comment orphaning undone. **W0 SCOPE (2026-08-14, `CLAIM-MODEL-DOTS3-NOTE-W0`, records-only — no code, no build, no download, no GPU).** `dots-studio/dots3-note-prev`, 280B total / 16B active, `model_type: dots3_note`. **Measured** from `config.json` + the HF API: L=46, H=5120, `intermediate_size` 13824, vocab 152064, 524288 positions, `rope_theta` 8e7; MoE 256 routed + 1 shared, top-8, `moe_intermediate_size` 1536, `first_k_dense_replace` 1, sigmoid/`noaux_tc`. **Attention is hybrid with TWO geometries:** 13 `full_attention` layers at indices 0,1,5,9,...,45 (128 heads, `kv_lora_rank` 512, `qk_nope` 128, DSA indexer `index_n_heads` 64 / `index_head_dim` 128 / `index_topk` 2048) and 33 `sliding_attention` layers (window **513**, 64 heads, `swa_kv_lora_rank` **1024**, `swa_qk_nope_head_dim` **192**, own `rope_theta` 5e4, `is_neox_style=False`, no indexer) => latent rows 576 vs 1088, reconciled upstream by padding the physical row (`Dots3NotePaddedMLAAttention.get_kv_cache_spec`). Vision = MoE ViT, 42 layers, `embed_dim` 1536, patch 14, `spatial_merge_size` 2, `use_qk_norm`, `pre_pixel_shuffle`, patch-merger adapter 1536->5120, **pyramid** `pyramid_num_routed` (dense to layer 24, then 4->64). Audio = `encoder_type: "dots"` Whisper variant, `d_model` 1280 / 32 layers / 128 mel / `max_source_positions` **6000** / **swiglu**, conv2d stem, RoPE (partial 0.5), RMSNorm, 60 s chunking, adapter 1280->5120. **HEAVY REUSE:** upstream literally subclasses DeepSeek — `Dots3NoteModel(DeepseekV32Model)`, `Dots3NoteDecoderLayer(DeepseekV32DecoderLayer)`, `Dots3NoteMoE(DeepseekV2MoE)`, `Dots3NoteFullAttention(DeepseekV2MLAAttention)` — so our gated MLA (`deepseek_v2.cpp`, SACRED 8/8), DSA indexer (`deepseek_v4_dsa.cpp`), `noaux_tc` router, 256-expert grouped MoE, `qwen3_vl_vision.cpp` (image+video STRICT 32/32) and `voxtral.cpp`/`whisper_audio.cpp` (audio 16/16) carry most of it. **NET-NEW:** windowed MLA (our every MLA path is full-attention — `deepseek_v4_registry.cpp` sets `is_hybrid=false`), the padded/heterogeneous MLA KV spec, the headwise sigmoid attention gate + the extra `k_rope_only_layernorm` over the 64-dim rope-only k slice + the two `apply_mla_qkv_lora_rescale` scalars, the MoE ViT (all our ViTs are dense), the `dots` audio stem, and the GGUF converter (llama.cpp has no `dots3_note` => no quant-matched llama.cpp bar). **FOUR SILENT CONFIG TRAPS** the checkpoint's `config.json` does NOT carry, all set in `Dots3NoteConfig.__init__:12-25`: `n_group=1`/`topk_group=1` (upstream: *"Do not inherit DeepSeek-V3's 8-group/4-group router defaults ... A different grouping changes the selected experts at every MoE layer"* — our router is gated at V3's grouped dims), `indexer_rope_interleave=True` (GPT-J adjacent pairs, where V3.2 and therefore our indexer default to split-half NeoX), and `num_nextn_predict_layers=1`. A fifth, in-JSON: `apply_mla_qkv_lora_rescale` => `sqrt(hidden/lora_rank)` after each layernorm. **HW-fit: DOES NOT FIT ANY HOST WE OWN.** Measured via the HF API: 131 language shards + `model-vision.safetensors` 13.7 GB + `model-audio.safetensors` 1.77 GB = **~576 GB bf16**; the `-fp8` sibling ~290 GB; upstream's own recipe is `--tensor-parallel-size 8` on H100s. GB10 is ~119 GiB and Thor ~122 GiB RAM / 123 GiB free disk, so **the vLLM oracle cannot run this model here at any published precision** and there is no smaller `dots3-note` in the `dots-studio` org — the same wall [deepseek-v4-flash](specs/deepseek-v4-flash.md) hit at 156.7 GiB, three times worse. Our own arm is plausible at <=2 bpw (~75-90 GiB, the ds4flash IQ2_XXS pattern) but an arm with no oracle beside it yields NO parity number, so **every performance axis is an open gap by construction**. Beyond-pin on top of that: gating needs a pin advance, and upstream is still moving (`#52172` landed the day before this scope). Thor `192.168.68.23` is the developer-designated e2e CUDA host (2026-08-14) and probed the same day — 14 cores, 122 GB RAM, 123 GiB free, but **no nvcc, no cmake, no ninja, no venv**, and `nvidia-smi` refuses under non-interactive ssh — so W0.5 provisions it; it can carry our low-bit arm, the sm_110 axis and every unit gate, but not a 290 GB oracle. **DECISION 2026-08-15: option B** (developer) — port brick by brick against independent references, ship our own low-bit arm on Thor, and carry the e2e parity gate as an OPEN GAP. Consequence recorded on the row, not just in the spec: **no performance number for this model is claimable on any axis** while B holds, because there is no denominator; and a brick gated against an in-test reference proves two implementations AGREE, never that either matches vLLM. The owed gate lives in the spec's `## Owed`. W0.5 (provision Thor) is in flight and is NOT dots3-specific — Thor is the only non-GB10 CUDA host, and it runs an immutable OS with `/` READ-ONLY, so the recipe goes in environment.md. W-plan W0-W11 in the spec | `CLAIM-MODEL-DOTS3-NOTE-W0` |
 | `MODEL-DIFFUSION-minimax-h3-mini-max-h3-dit` | `MiniMaxH3DiTModel` | **BEYOND-PIN AND OUT-OF-REPO** — not in `555967922` (H3 released after the pin) and not in the vLLM repository at all: it lives in `vllm-project/vllm-omni`, `vllm_omni/diffusion/models/minimax_h3/` (`minimax_h3_transformer.py`, `packed_sequence.py`, `packed_tokens.py`, `scheduling_minimax_h3_euler_ancestral.py`, `denoise_loop.py`, `vae.py`, `encoder.py`, `pipeline_minimax_h3.py`) | diffusion generation / video + audio (text/image/video/audio in) | flow-matching denoise loop; packed varlen NON-CAUSAL attention; AdaLN modulation; 3D MM-RoPE; video VAE + audio VAE (checkpoint REMOTE CODE); Qwen3-VL-derived encoder; MP4 muxing; Ulysses sequence parallelism | [minimax-h3 spike](specs/minimax-h3.md) | `PARTIAL` | **W0-W2 LANDED (2026-08-03, `CLAIM-MINIMAX-H3-W0-W2`, DERIVE-AND-SHIP).** H3 is NOT an autoregressive LLM: it is a CFG-distilled joint video+audio DIFFUSION transformer (50 blocks, H=5376, 56 MHA heads x 128, SwiGLU 14336, AdaLN 6x3xH, 3D RoPE rotating 96 of 128 dims, video row width 96, audio latent 32) forwarded ONCE PER STEP of a 50-step flow-matching loop — no KV cache, no sampler, no logits, so the SACRED token-exact methodology does not apply. **HW VERDICT: e2e is IMPOSSIBLE on this project's hardware** — ~354 GB checkpoint (DiT 66.3 GB + Qwen3-VL-derived encoder 51.5 GB + video VAE ~10 GB + audio VAE ~0.6 GB), upstream validates on **4x NVIDIA B300 at ~133 GB peak per rank**, vs ONE GB10 with 119 GiB UNIFIED memory; CPU offload does not help because the pool IS host RAM. **WHAT IS GATED (and it is exact):** upstream's modules are pure Python, so they are executed at REDUCED DIMENSIONS on CPU as the oracle (`scripts/gen-minimax-h3-goldens.py` imports them by file path and freezes their outputs; both sides rebuild weights/inputs from an identical FNV-1a + splitmix64 stream, so no weight byte is checked in). Results: fl2va + ref2va packed layouts EXACT including the **fp64 position grid BIT-EXACT** (it feeds RoPE — the port reproduces numpy's `linspace(endpoint=False)` evaluation order and upstream's deliberately-split pairwise vs sequential span summations, `packed_sequence.py:101-113`); patchify/unpatchify/audio pack EXACT + round-trip; scheduler EXACT; **full DiT forward max abs diff 1.6e-7 (video) / 1.5e-7 (audio)** — f32 round-off. **REUSE:** the packed varlen non-causal attention routes through the SHARED `vt::DFlashBlockAttention(causal=false)` (its per-document bidirectional contract IS upstream's varlen FA call) and every projection through `vt::MatmulBT` — NO new kernel was added. **NOT-YET-BUILT (honest):** device-resident/bf16 forward + fusion folds (W2b), H3-Encoder on our existing Qwen3-VL tower (W3), the two VAEs — which are **checkpoint REMOTE CODE** under `trust_remote_code` and must be reimplemented in C++, not adapted (W4/W5) — pipeline/tasks (W6), `/v1/videos` + MP4 muxing, which needs a NEW dependency decision (W7), and USP multi-GPU (W8). No speed number is claimed; upstream reports the DiT at 88% of request latency. Code: [minimax_h3.h](../include/vllm/model_executor/models/minimax_h3.h#L1-L333), [minimax_h3.cpp](../src/vllm/model_executor/models/minimax_h3.cpp#L410-L640) (`MiniMaxH3DitForward`), [minimax_h3_packing.cpp](../src/vllm/model_executor/models/minimax_h3_packing.cpp#L259-L400) (`BuildMiniMaxH3PackedSequence`). Test: [test_minimax_h3.cpp](../tests/vllm/models/test_minimax_h3.cpp#L376-L470) (DiT forward parity; 10/10 cases / 2539 assertions, clean CPU build) + generator [gen-minimax-h3-goldens.py](../scripts/gen-minimax-h3-goldens.py#L1-L60); [ledger](parity-ledger.md#L889). **W6A+W9 LANDED (2026-08-03, `CLAIM-MINIMAX-H3-W6A-W9`) + HW VERDICT CORRECTED.** (a) the **bf16 PRODUCTION dtype policy** now runs (upstream's cast points with the fp32 islands preserved; gated vs a bf16 upstream golden at max abs diff 2.4e-3); (b) **request planning** — 17n+5 frame snapping, video/audio latent shapes, the rectified-flow time-shift sigma schedule, canvas resolution and t2va/fl2va/ref2va dispatch — ported and EXACT vs `time_request.py` + `pipeline_minimax_h3.py:121-122,207-222,374-434`; (c) the **ComfyUI-GGUF arm**: the name map is the IDENTITY and every one of the **535 real tensors** of `MiniMax-H3-FL2VA-Q3_K_M.gguf` resolves onto our contract, with the geometry derived from SHAPES ALONE equal to the shipped config (gated on a manifest read from the file's own header by range request — no payload downloaded). Two shape rules recorded: GGUF `ne` is reversed vs torch, and `comfy.gguf.orig_shape.<name>` overrides it where ComfyUI reshaped a tensor for quant-block alignment (the 50 AdaLN projections: logical [96768, 2688], 2688 not a multiple of the 256-element Q3_K block). **★ HARDWARE VERDICT CORRECTED — the earlier 'e2e is IMPOSSIBLE on this hardware' was WRONG** because it reasoned from the bf16 release alone: quantized H3 checkpoints exist and FIT (GGUF DiT Q3_K_M 15.6 GB + Qwen3-VL encoder Q4_K_M 14.6 GB + VAEs ~11 GB ~= **41 GB** in a 119 GiB pool; `lilcheaty/MiniMax-H3-NVFP4` likewise). So e2e AND a speed comparison are REACHABLE; NVFP4 is the likely speed path (sm_121 native FP4 tensor cores + our tuned NVFP4 stack). **W5 LANDED (audio VAE)**: H3's VAEs are checkpoint REMOTE CODE under `trust_remote_code`, so a no-Python engine must REIMPLEMENT them — the DAC-lineage BigVGAN audio decoder (weight-norm materialization (w = g*v divided by the row norm), anti-aliased SnakeBeta with kaiser-sinc up/down resampling, replicate padding, final clamp) is ported and gated against the checkpoint's OWN modules at **max abs diff 4.2e-9**. The VIDEO VAE (`klvae.py` ~48 KB + CNN/ViT + tiling) is the largest remaining brick. **W10 GROUNDED + W4 SCOPED from REAL manifests** (safetensors headers captured by range request, no payload downloaded): the NVFP4 checkpoint's 1051 tensors are textbook compressed-tensors NVFP4 (U8 packed 2-per-byte + E4M3 `weight_scale` at group 16 + F32 scalar `weight_scale_2`; 258 quantized projections; fp32/bf16 islands left unquantized; names identical to our contract) — i.e. EXACTLY the layout our tuned NVFP4 stack already consumes, so W10 is loader wiring, not a new quant scheme. The video VAE's 560 tensors show its ENCODER is the 3D CNN (rank-5 Conv3d) while its DECODER — the half generation needs — is a 36-block TRANSFORMER (to_qkv/to_out, ff.w1/w2, 2 norms + 2 learned residual scales per block, x_embedder/mask_token/register_tokens/proj_out), materially smaller than klvae.py's 48 KB suggested. **W4 BLOCK LANDED**: the video-VAE decoder's repeated `TransformerBlock` is ported and gated at **6.0e-8** against the checkpoint's OWN remote code — RMSNorm + per-head RMS qk-norm (no affine) + full attention + gated-SiLU FF + LEARNED PER-CHANNEL residual scales, and critically the PER-HEAD-INTERLEAVED qkv layout ([head][q,k,v], NOT the DiT's [q_all,k_all,v_all]) that would otherwise produce a plausible-but-wrong image. **W4 DECODER DONE**: the FULL ViT3D video-VAE decoder — pack, x_embedder, register/cls tokens, 3D RoPE (RotaryEmbeddingND, length-normalized ids, angle scale 2pi, tiled freq blocks), the 36-block stack, LayerNorm norm_out, proj_out and unpatchify — is ported and gated at **8.9e-8** against the checkpoint's own `ViT3DDecoder`, at its real hyperparameters (36 layers, 32 heads x 64, rope_theta 100, rope_dim_ratio 0.75). BOTH VAE decoders are now done (audio 4.2e-9). **W3 TEXT TOWER DONE**: the H3-Encoder's truncated Qwen3-VL text tower — gated at **1.2e-7** vs upstream — with all three H3 deltas exercised: layer truncation (min(num_hidden_layers, 50)), the UNNORMALIZED layer-49 output (NO final RMSNorm, unlike stock Qwen3-VL — applying one silently shifts every conditioning vector), and DeepStack visual injection into the first N layers; plus interleaved M-RoPE, fused QKV, per-head q/k RMSNorm, causal GQA and the gated-SiLU MLP. **W6 t2va ASSEMBLED — the WHOLE PATH COMPOSES**: `MiniMaxH3GenerateT2va` wires packed layout -> rectified-flow sigma schedules -> the multi-step denoise loop of DiT forwards -> unpatchify + audio unpack -> per-channel denormalize -> BOTH VAE decoders, producing correctly-shaped, finite frames and a stereo waveform in [-1, 1]; gated by a structural end-to-end test at reduced dimensions with random weights (NOT a quality result). Assembling it also caught a real gap: the audio decode needed the checkpoint's `dec_in_proj` (Conv1d k=1, vae_latent_channels -> num_mels) ahead of BigVGAN. **W9 GGUF ARM DONE**: `LoadMiniMaxH3DitFromGguf` materializes the DiT from a ComfyUI-format GGUF — dequantizing every tensor through the SHARED GGUF dequant entry point (so the Q2_K/Q3_K/Q4_K families the H3 GGUFs use are covered by the same code every other GGUF model uses), recovering the geometry from shapes alone, and binding the forward's views with missing tensors throwing BY NAME rather than reading as zeros; gated by a synthetic-file load-AND-RUN test (a real DiT forward executes off the loaded weights). **W10 LOADER DONE**: `LoadMiniMaxH3DitFromNvfp4` materializes the DiT from an NVFP4 compressed-tensors checkpoint — the U8-packed [out, in/2] weight plus its E4M3 group-16 `weight_scale` and F32 scalar `weight_scale_2` go through the project's EXISTING NVFP4 dequant (no new quant code), sidecars are excluded from the model tensor set, and the logical [out, in] shape is recovered; gated by a synthetic-file load-AND-RUN test. BOTH quantized loaders are now done. **W3 VISION BLOCK DONE** (6.0e-8): the repeated unit of the encoder's Qwen3-VL vision tower, which differs from the text tower in every way that matters numerically — LayerNorm WITH BIAS (not RMSNorm), a [q_all, k_all, v_all] qkv layout (not the video VAE ViT's per-head interleave), fp32 rotary, NON-CAUSAL attention segmented by `cu_seqlens` (the test asserts a perturbation in one packed image leaves the other's outputs BIT-IDENTICAL), and the TANH-approximate GELU. **W3 ENCODER COMPLETE**: the FULL vision tower also lands — Conv3d patch embed (kernel == stride, so a linear over the flattened patch), BILINEAR resampling of the learned position grid into spatial-merge order, the 2D rotary table, per-frame `cu_seqlens`, and both merger flavours (the final merger norms the PRE-shuffle width while the DeepStack mergers norm the POST-shuffle width, and both use exact-erf GELU unlike the block MLP's tanh approximation), gated over a RAGGED two-image batch. Only the MM processor remains on the encoder. **CONDITION-NOISE augmentation DONE** (fl2va/ref2va): the noised-anchor mix plus its ROW ACCOUNTING — each visual condition draws noise of length `target_latent_t + imgvid_cond_num_frames` and slices the PREFIX matching its own latent_t, every condition restarts the SAME seed (so concatenating and drawing once would differ for multi-reference requests), and rows advance by that condition's own patchified count. Gated EXACT with the noise supplied, so the comparison isolates the accounting from torch's RNG. **REFERENCE-VIDEO geometry + FRAME SCHEDULE DONE** (the pure-math half of `reference_video.py`): the canvas pipeline (aspect clamp -> 768 short edge -> max-pixel rescale -> nearest multiple of 32) and the 24-to-2 FPS frame resample with per-temporal-patch block timestamps, both EXACT. NOTE the rest of that module (probe, transcode, frame extraction, audio decode) shells out to ffmpeg and is blocked on the SAME external dependency decision as `/v1/videos` MP4 muxing — one decision unlocks reference-video INPUT decode and generated-video OUTPUT encode together. **VIDEO VAE TILING DONE**: the tile plan (smallest tile count whose MINIMUM overlaps still cover the axis, leftover slack distributed in whole `vae_ratio` units ROUND-ROBIN across the seams) plus the linear seam cross-fade, both EXACT. Shipped config tile 256 / overlap 64 / vae_ratio 16 (= prod(space_down), the 'f16' in f16t4). Getting the slack distribution wrong shifts every tile after the first and surfaces as seam artifacts rather than an error. **PRESENTATION TOKEN TAGS DONE** — the fl2va vision-span override the denoise loop requires callers to have applied: a vision block is `vision_start + pad*count + vision_end` and the WHOLE block, MARKERS INCLUDED, is tagged VIDEO; tagging only the pads would leave two markers as TEXT and shift every AdaLN modulation index after them. Gated EXACT, with the test proving each VIDEO run is a whole vision block. **VAE 3D-CNN ENCODER PRIMITIVES DONE**: causal Conv3d (all temporal padding on the LEFT so a frame never sees the future, `reflect` spatial padding), GroupNorm3D (32 groups, eps 1e-6, statistics spanning TIME as well as space) and ResnetBlock3D, gated EXACT — with CAUSALITY proven directly on the bare convolution (a change to the last frame provably cannot reach earlier frames). **Downsample3D DONE** too: the strided inter-level conv, whose subtlety is the ASYMMETRIC pre-pad — one pixel on the RIGHT of W and the BOTTOM of H before a stride-2 conv with padding (1,0,0); padding symmetrically instead shifts the whole sampling lattice by half a pixel, which is a silent wrong latent rather than an error. Only the EncoderFCN3D level-loop assembly remains on the VAE encoder. **VIDEO VAE COMPLETE — encoder AND decoder**: the whole 3D-CNN encoder level loop (conv_in -> per level [ResnetBlock3D x N then Downsample3D or a 1x1x1 channel match] -> GroupNorm -> SiLU -> conv_out) is gated EXACT. **MM PROCESSOR = REUSE, gated**: H3's `FL2VA/processor` is a stock `Qwen3VLProcessor`, so the multimodal front end is this project's EXISTING Qwen3-VL processor rather than a new port; H3's own config is parsed and driven through it (patch 16 / temporal 2 / merge 2, **0.5 normalization rather than CLIP statistics**, a 32-pixel grid, the 768x1344 default canvas proven an IDENTITY under smart_resize, and VIDEO bounds deliberately looser than the image ones). **With this every PORTABLE piece of the lane is done**; what remains is one dependency decision (ffmpeg media I/O + MP4 muxing) and GPU-blocked work (the device-resident FP4 forward and any speed number). **WAV OUTPUT** added: the decoded stereo waveform serializes to RIFF/WAVE 16-bit PCM, converting the VAE's CHANNEL-MAJOR layout to INTERLEAVED (getting that backwards yields audio that plays but with the channels time-smeared) and clamping rather than wrapping. Deliberately dependency-free, and required under EITHER outcome of the open MP4/muxer decision. **VIDEO OUTPUT PATH DONE**: PPM frame serialization (planar [C,T,H,W] -> row-major interleaved RGB, [-1,1] -> [0,255] clamped) plus the MP4 mux argv (h264/yuv420p + AAC, `-shortest`, `+faststart`). The built argv was RUN through real ffmpeg 6.1.1 and produced a VALID MP4 (ffprobe: h264 yuv420p video + AAC stereo at 32 kHz). The library never spawns a process — `src/vllm/` has no subprocess precedent — so it builds the artifacts and the command while the example/server layer invokes it. **`/v1/videos` API LOGIC DONE**: the request contract (H3 defaults — 50 steps, flow shift 12 video / 3 audio; both the vLLM-Omni `extra_params` nesting and a flat top-level spelling accepted; malformed input rejected with a reason rather than silently defaulted) and the job store (queued -> running -> succeeded/failed, illegal transitions throw, unknown ids reported so the route can 404, status JSON omits fields that do not apply, and concurrent creation is thread-safe). Remaining is mechanical glue: route registration + runner injection, with the ffmpeg call in `examples/` per the ratified boundary. Test: 34/34 + video-api 4/4 (9233 + 63 assertions). OPEN: there is no vllm-omni parity PIN — the upstream-sync protocol covers only the vLLM repo | `CLAIM-MINIMAX-H3-W0-W2`, `CLAIM-MINIMAX-H3-W6A-W9` |
 | `MODEL-MM-moss-tts-moss-tts-delay-talker-for-generation` | `MossTTSDelayModel` | **OUT-OF-REPO** — not in `555967922` and not in the vLLM repository at all (absent from every dictionary of `registry.py` on `main` as well): it is registered by the separate `vllm-project/vllm-omni` repository, `vllm_omni/model_executor/models/registry.py:350-354` @ `bbe6ccc512a404a2df8c977ea29003002f2683e8` → module `vllm_omni/model_executor/models/moss_tts/modeling_moss_tts_talker.py`, class `MossTTSDelayTalkerForGeneration`, with the shared stage-1 codec decoder at `registry.py:368-372` → module `moss_tts/modeling_moss_tts_codec.py`, class `MossTTSCodecDecoder`. Module and class are given as separate fields rather than in the compact colon-pair form, as the `MiniMaxH3DiTModel` row does: that form marks a target inside the PINNED vLLM registry inventory, and this target is in another repository entirely. Documented at `vllm-omni` `docs/models/supported_models.md:78`. Official recipes: `OpenMOSS-Team/MOSS-TTS`, `MOSS-TTSD-v1.0`, `MOSS-SoundEffect`, `MOSS-VoiceGenerator` — four of the 157 recipes at `vllm-project/recipes` `86c7777a`, all four declaring `architectures: ["MossTTSDelayModel"]`. Tracked by [#610](https://github.com/mudler/vllm.cpp/issues/610) | conditional generation / text+audio in, AUDIO OUT (TTS) | MM processor; audio detokenizer / codec decoder; delay-pattern multi-codebook decoding; reference-audio speaker encoder; streaming audio output. **We serve no audio-generation output modality at all today** | ☐ required | `INVENTORIED` | none. There is no `vllm-omni` parity PIN — the upstream-sync protocol covers only the vLLM repository — so an oracle for this row has to be established before it can advance | unassigned |
 | `MODEL-MM-moss-tts-moss-tts-realtime-talker-for-generation` | `MossTTSRealtime` | **OUT-OF-REPO** — not in `555967922`, and absent from `registry.py` on vLLM `main`; registered by `vllm-project/vllm-omni` at `vllm_omni/model_executor/models/registry.py:356-360` @ `bbe6ccc512a404a2df8c977ea29003002f2683e8` → module `vllm_omni/model_executor/models/moss_tts/modeling_moss_tts_talker.py`, class `MossTTSRealtimeTalkerForGeneration` (separate module and class fields rather than the compact colon-pair form, per the out-of-repo convention) — the SAME module as `MossTTSDelayModel` but a different class, so the two are separate targets rather than aliases. Documented at `vllm-omni` `docs/models/supported_models.md:79`. Official recipe: `OpenMOSS-Team/MOSS-TTS-Realtime` (1.7B). Tracked by [#610](https://github.com/mudler/vllm.cpp/issues/610) | conditional generation / text+audio in, AUDIO OUT (realtime/streaming TTS) | MM processor; audio detokenizer / codec decoder; realtime full-duplex streaming; incremental audio decode | ☐ required | `INVENTORIED` | none. Same missing-`vllm-omni`-pin blocker as the other omni rows | unassigned |
diff --git a/.agents/porting-inventory.md b/.agents/porting-inventory.md
index 8707f0c1a..ecf89b309 100644
--- a/.agents/porting-inventory.md
+++ b/.agents/porting-inventory.md
@@ -1741,6 +1741,17 @@ Examples: `examples/cli` ✅ (C-API client), `examples/server` ✅ (OpenAI serve
     `file:line` W1 cites was read at `origin/main` =
     `c205726108df54bb6fbf15b19e725a4a3add2b18`, and the anchors that decide
     correctness are named in `src/vllm/model_executor/models/dots3_note.h`.
+    **W2 (2026-08-24) re-read its own anchors at `origin/main` =
+    `185cada36b`**, which is 20 days of upstream later, and cites two: the
+    tower prefixes in `nvidia/multimodal.py`'s `hf_to_vllm_mapper` (`:70-78`,
+    the two prefixes at `:75-76`; `vision_encoder.` -> `visual.` and
+    `audio_encoder.` -> `audio_tower.`. First written here as `:53-62`, which is
+    W1's `c205726108` line and an INHERITED anchor rather than the re-read one
+    this paragraph claims — corrected in place, review F4 on
+    [#1847](https://github.com/mudler/vllm.cpp/pull/1847)) and
+    the DSA indexer's rope slice in
+    `vllm/model_executor/models/deepseek_v2.py` (`:805`, `:814`). Point (b)
+    below is why it re-read rather than inheriting W1's revision.
     Same shape as deviations 16 and 17, taken for the same reason: at the pin
     the architecture does not exist, so there is nothing there to mirror. It
     is recorded here, and argued for in the commit that introduced it,
diff --git a/.agents/specs/dots3-note.md b/.agents/specs/dots3-note.md
index 4ee1155c3..ee188442e 100644
--- a/.agents/specs/dots3-note.md
+++ b/.agents/specs/dots3-note.md
@@ -16,11 +16,12 @@ parallelism for Dots3 NOTE"). **NOT present at our parity pin.**
 fleet device **`thor:gpu0` through an `rc` lease and never by `ssh`** — the host
 address is recorded in `environment.md` to identify the box, not as a way into
 it. §6.3 records what that host can and cannot carry for this model, measured.
-**Status:** W1 — config + registry landed (§7 W1, evidence §4.1). The arch
-RESOLVES and parses; load, GGUF and the forward all REFUSE BY NAME. No GPU was
-used and no tensor byte of the checkpoint was downloaded: the two committed
-fixtures are the released `config.json` and a headers-only projection of the
-shard index. The row stays `SPIKE`.
+**Status:** W2 — the whole weight map landed (§7 W2, evidence §4.4), on top of
+W1's config + registry (§4.1). The arch RESOLVES, parses, and accounts for
+38006/38006 of the released checkpoint's tensors; load, GGUF and the forward all
+REFUSE BY NAME. No GPU was used and no tensor byte of the checkpoint was
+downloaded: the committed fixtures are the released `config.json` and a
+headers-only projection of the complete shard index. The row stays `SPIKE`.
 
 ---
 
@@ -465,6 +466,265 @@ language tower is BF16 except one family. `mlp.gate.e_score_correction_bias`
 ships **F32**. A loader that assumed one dtype for the checkpoint would misread
 it, and no token gate could see the difference.
 
+### 4.4 W2 read the whole index, and three things the slice could not say
+
+**LANDED at W2** (`row/MODEL-MM-dots3-note-W2`, same TU and same test file as
+W1, upstream re-read at vLLM `origin/main` `185cada36b`). CPU-only. No GPU
+lease was taken and none was needed.
+
+**What was fetched, exactly.** The complete
+`model.safetensors.index.json` of `dots-studio/dots3-note-prev` at revision
+`1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b` (3436982 bytes, sha256
+`95a364b468a93ccad6adcb9c3aa110cb7a1411c2575c334c39022f9f84d456e1`), then the
+safetensors HEADER of every one of the 133 files it names — two HTTP Range
+requests each, 8 bytes for the header length and then the header JSON,
+**4770592 header bytes in total and not one tensor byte**. The checkpoint is
+576886825984 bytes and was never downloaded. The committed fixture is
+`tests/vllm/models/fixtures/dots3_note_prev/index_full.json`; it collapses ONE
+index, the routed-expert index, to `{E}` with the member count beside it, and
+expands back to exactly 38006 names. Every backbone layer, vision block and
+audio layer is a separate entry on purpose: W2 exists to MEASURE that the
+layers repeat, and a fixture that collapsed them would assume the answer.
+
+The released `config.json` was re-fetched at the same revision and is
+byte-identical to the committed fixture, sha256
+`99b7de680dd456111c36efb8749f8ae7177328e97b65a3e39a6700cbc1173833`.
+
+**The gate, met.** `test_dots3_note_scaffold` — **26 cases / 110821
+assertions**, CPU-only, no GPU, no checkpoint (19/3876 at W1). The accounting
+reads **38006 / 38006**: 35381 language, 2195 vision, 430 audio, zero
+unaccounted, zero missing, zero duplicated, zero invented. All three buckets are
+asserted BY NUMBER in every case that touches them, and the whole 38006-name set
+is driven through `ModelRegistry::Resolve(...).factory->load_weights` as well as
+through the classifier, so the map is proved reachable and not merely correct.
+
+**One thing changed shape rather than only growing.** W1 classified the towers
+with two prefix literals and two integer counters. A counter cannot say whether
+2625 weights are deferred on purpose or lost, so `Dots3NoteDeferredTowers()`
+is now a table of records — prefix, the one file the tower ships in, the brick
+that owes it, and what it is — and `AccountDots3NoteTensors` dispatches on that
+table. The load refusal prints it, so an unknown tensor is distinguishable from
+a deferred one in the message a user gets.
+
+#### The three facts the slice could not reach
+
+1. **The backbone has exactly FOUR distinct layer shapes, and no layer breaks
+   the pattern.** Grouping all 47 layers by their (suffix, dtype, shape) set
+   with the expert index collapsed gives `{0}` (dense MLP + full attention, 19
+   tensors), the 12 full+MoE layers `{1, 5, 9, ... 45}` (789 each), the 33
+   sliding+MoE layers (784 each) and `{46}` (18). W1 recorded "the remaining 42
+   backbone layers repeat layers 1/2 exactly" as a claim about a checkpoint
+   nobody had read. It holds. A fifth class would have meant the port was
+   reading some layer with the wrong map.
+
+2. **The full/sliding split derived from the WEIGHTS matches
+   `config.layer_types` exactly.** The DSA indexer ships only on the full class,
+   so the shipped `self_attn.indexer.wk.weight` names give the schedule
+   independently of the config: `{0, 1, 5, 9, 13, 17, 21, 25, 29, 33, 37, 41,
+   45}`, 13 layers, and `q_b_proj` is [24576, 1024] on every one of them against
+   [16384, 1024] everywhere else. Two independent released artifacts agreeing is
+   a stronger statement than either alone, and §1.1's table is now measured
+   rather than transcribed.
+
+3. **The checkpoint carries 62 F32 tensors in TWO families, not one.** W1 saw
+   one and predicted the family. The language tower's 45 are
+   `mlp.gate.e_score_correction_bias`, one per MoE layer, layers 1 to 45 — as
+   predicted. The other 17 are **`vision_encoder.blocks.{25..41}.mlp.router_bias`**,
+   which W1's language-only slice could not see at all. Their widths are the
+   pyramid's own routed-expert counts, 4, 8, 12 … 64, 64, which is §1.2's
+   `pyramid_num_routed` confirmed from the weights. The vision MoE keeps its
+   router statistics in fp32 inside an otherwise BF16 tower, and a loader that
+   resolved one dtype for the checkpoint would read them wrong with no shape
+   change and no error. Audio is BF16 throughout, and the census closes: 37944
+   BF16 + 62 F32 = 38006, with no third dtype anywhere.
+
+   **This is NOT R5, and an earlier draft of this section said it was.** R5 and
+   §2.4 are about the FP32 **dynamic activation scales** inside
+   `note_vision_fused_moe_fp8` — a quantized-path memory format. The bf16
+   checkpoint carries no scale tensors at all; every one of its 38006 entries is
+   a named parameter. `router_bias` is a learned fp32 parameter, which is a real
+   and useful finding and a different one. **R5 stays entirely owed by W6**, and
+   a W6 implementer must not read this row as confirming it.
+
+#### And a fourth, which is a finding rather than a confirmation
+
+**The released index declares an indexer RoPE layout that nothing reads.** Its
+`metadata` block carries `"indexer_rope_layout": "leading"` and
+`"indexer_rope_converted_from": "tail"` beside `total_size`. `git grep
+indexer_rope_layout` over vLLM `origin/main` returns nothing, so upstream never
+reads either key: it is the publisher stating how the DSA indexer's `wq_b` and
+`wk` are laid out along the 128-wide index head, and saying the published
+weights were re-ordered to get there.
+
+It agrees with what upstream's code does anyway. `DeepseekV2Indexer` splits both
+`q` and `k` as `[..., : rope_dim]` for the rotated half and `[..., rope_dim :]`
+for the rest (`deepseek_v2.py:805`, `:814`, `rope_dim` = 64 of `index_head_dim`
+= 128), which is a LEADING slice. **This is not §4 trap 2.** Trap 2 is about
+which PAIRS the rope rotates, GPT-J against NeoX. This is about which HALF of
+the head is rotated at all. Both are numerically silent, they are independent,
+and §6.4 says this row has no oracle that could catch either.
+
+W2 pins both values in an assertion so a re-published checkpoint cannot flip the
+layout silently, and consumes neither, because W2 writes no maths.
+[#1846](https://github.com/mudler/vllm.cpp/issues/1846) owns it and W3 owes the
+slice.
+
+#### What the two towers actually ship, for W6 and W7
+
+W2 does not port either tower and writes no maths for them. It does read their
+whole tensor list, because that is what "named deferral" has to mean, and four
+of those facts are worth stating where the brick that owes them will look.
+
+**Vision (2195 tensors, `model-vision.safetensors`).** 42 blocks, each with
+`norm_1`, `norm_2`, `attn.{qkv, proj, q_norm, k_norm}` and an MLP. `qkv` is
+[4608, 1536] — one fused projection, no bias — and `q_norm`/`k_norm` are [64],
+so `use_qk_norm` acts per head at head_dim 64 over 24 heads. Blocks 0 to 24 are
+DENSE with `mlp.{fc1, fc2, fc3}` at [4224, 1536] / [1536, 4224] / [4224, 1536],
+a three-tensor SwiGLU rather than a gate/up/down triple. Blocks 25 to 41 are
+MoE, with `mlp.experts.{E}.{fc1, fc2, fc3}` at the `moe_intermediate_size` of
+2112 and a router that is `mlp.gate_weight` + `mlp.router_bias` — NOT the
+`mlp.gate.weight` + `mlp.gate.e_score_correction_bias` spelling the language
+tower uses. Outside the blocks: `patch_embed.proj` [1536, 3, 14, 14] with a
+bias, `patch_embed.norm`, `post_trunk_norm`, and the patch-merger adapter
+`adapter.{ln_q, mlp.0, mlp.2}` folding 4x1536 = 6144 to 5120.
+
+**Audio (430 tensors, `model-audio.safetensors`).** 32 encoder layers of
+`self_attn.{q_proj, k_proj, v_proj, out_proj}` at [1280, 1280], with a bias on
+q, v and out and NONE on k — Whisper's own convention. `fc1` is [10240, 1280]
+against `fc2` [1280, 5120], so the SwiGLU gate and up are packed into one
+tensor at twice the 5120 `encoder_ffn_dim`. The stem is
+`conv2d1` [480, 1, 3, 3], `conv2d2` and `conv2d3` [480, 480, 3, 3], then
+`conv_out` [1280, 7680] = 16 x the 480 `downsample_hidden_size`. The adapter is
+`audio_adapter.proj.{0, 1, 3}`, 1280 to 5120.
+
+**There is NO learned positional embedding in the audio tower**, and that is
+checkpoint and upstream agreeing rather than an absence to explain: at
+`nvidia/audio_encoder.py:507-519` `DotsSpeechEncoder` sets
+`self.embed_positions = None` when `use_rope` is true, and the released
+`audio_config` sets it true. W7 must not go looking for one.
+
+#### The mutation table
+
+Every mutation was applied to the tracked source — or, for the `F` rows, to the
+committed fixture — rebuilt, run, and reverted, with the tree verified
+byte-for-byte afterwards. **The compiler exit status is printed beside each
+row**, because a mutation that fails to build reads as a passing test and this
+project has been bitten by that repeatedly. Every row compiled.
+`cases`/`assertions` are what `doctest` reported failing.
+
+| id | mutation | compiler exit | result | cases | assertions | first failing case |
+|---|---|---:|---|---:|---:|---|
+| M1 | the towers are counted as LANGUAGE | 0 | RED | 2 | 9 | W2: all 38006 tensors … are claimed |
+| M2 | the audio tower is dropped from the deferral table | 0 | RED | 4 | 12 | W2: all 38006 tensors … are claimed |
+| M3 | the vision deferral names the WRONG brick (W7) | 0 | RED | 2 | 2 | W2: the two tower files are NAMED W6/W7 deferrals |
+| M4 | the vision deferral names the WRONG file | 0 | RED | 1 | 3 | W2: the two tower files are NAMED W6/W7 deferrals |
+| M5 | the nextn layer is emitted with the FULL attention set | 0 | RED | 5 | 10 | enumeration: all 1614 tensors of the released slice |
+| M6 | every backbone layer is treated as MoE (`first_k_dense_replace` ignored) | 0 | RED | 7 | 19 | config: the REAL released config.json parses |
+| M7 | every backbone layer is treated as FULL attention | 0 | RED | 8 | 86 | config: the REAL released config.json parses |
+| M8 | the headwise gate `g_proj` is dropped from the name map | 0 | RED | 6 | 16 | enumeration: all 1614 tensors of the released slice |
+| M9 | `k_rope_only_layernorm` is dropped from the name map | 0 | RED | 6 | 15 | enumeration: all 1614 tensors of the released slice |
+| M10 | the MoE shared expert is dropped from the name map | 0 | RED | 5 | 13 | enumeration: all 1614 tensors of the released slice |
+| M11 | one routed expert per MoE layer is dropped (255, not 256) | 0 | RED | 5 | 13 | enumeration: all 1614 tensors of the released slice |
+| M12b | a VISION tensor is added to the language name map | 0 | RED | 5 | 14 | enumeration: all 1614 tensors of the released slice |
+| F1 | FIXTURE: one vision `router_bias` is re-typed BF16 | 0 | RED | 1 | 2 | W2: the memory format of the WHOLE checkpoint |
+| F2 | FIXTURE: the indexer rope layout reads `tail` | 0 | RED | 1 | 1 | W2: the released index states an indexer RoPE layout |
+| F3 | FIXTURE: one language tensor is moved into the vision tower file | 0 | RED | 1 | 1 | W2: the two tower files are NAMED W6/W7 deferrals |
+
+**A sixteenth row came from the fresh review, not from W2.** R8 deleted the
+PRODUCTION CALL SITE of `AccountDots3NoteTensors` in `LoadDots3NoteWeights` —
+the one thing W2's own table never mutated, because W2 wrote the call — and it
+came back RED. So the map is reached through the registry rather than only
+exercised by helpers, and the reachability claim is measured rather than
+asserted. The same review re-derived every number in this section against the
+live release independently, including the 266 Range requests, the two fixture
+hashes, all four bucket counts and the whole tower inventory below.
+
+M1 is the row this brick exists for. It is the W1 review's M15 at full scale:
+fold the towers into the language count and every "nothing was left over"
+assertion stays green while 2625 weights go unloaded. It fires.
+
+**One mutation came back GREEN, and the CODE changed rather than the table.**
+W2 first wrote "a name cannot be both loaded and deferred" as a runtime
+`VT_CHECK` inside `AccountDots3NoteTensors`. Deleting it left the whole gate
+passing, because no config can make `EnumerateDots3NoteTensors` emit a
+tower-prefixed name: every name it emits is `model.`- or `lm_head`-prefixed by
+construction. That is production code no input reaches, which is the shape
+AGENTS.md's reachability rule names, and the honest answer to a green mutation
+is to remove what the gate cannot see rather than to keep it and note it. The
+invariant is real, so the tower case asserts it over the real map, and **M12b
+replaces the deleted row by injecting the defect the guard was meant to catch**
+— a `vision_encoder.` name added to the map — which takes the gate red. The
+guarantee is kept; the unreachable copy of it is not.
+
+**M11 is a re-run, and the first attempt had a cause this spec got WRONG.** In
+the batch it exited 135 with no parseable `doctest` summary. W2 wrote that up as
+disk pressure — the box read 92% full at that moment — and **that was the wrong
+cause**, corrected in place because `main` is never rewritten. The
+[#1847](https://github.com/mudler/vllm.cpp/pull/1847) review found the real one
+and reproduced it 3/3 **at 47 GB free and 61 GB RAM available**: `TempConfig`
+and `TempCheckpoint` built their `/tmp` paths from a **per-process**
+`static int counter`, so two concurrent runs of the same binary shared one
+directory — both were watched sharing `/tmp/dots3_note_cfg_8`. Each constructor
+rewrites a file the other has mmapped through `SafetensorsFile::Open` and each
+destructor `remove_all()`s the other's, which is SIGBUS, exit 135, and a
+block-buffered `doctest` summary lost with the process.
+
+**That failure mode reads as NO RESULT, not as a failure**, which is why it is
+worth more than the mutation row it corrupted: under §6.4 option B this row has
+no oracle, this file is its only instrument, and a second agent building on the
+same box is routine here. Both paths are now process-unique. The identical shape
+in at least `test_laguna_nvfp4_loader`, `test_kimi_linear_scaffold`,
+`test_loader_unaligned_offsets`, `test_ltx2_lora` and
+`test_minimax_h3_video_fold` is [#1860](https://github.com/mudler/vllm.cpp/issues/1860),
+not this row.
+
+Re-run alone the same mutation compiled clean and took the gate red on 5 cases /
+13 assertions, naming `model.layers.1.mlp.experts.255.down_proj.weight` and 135
+unaccounted tensors. A crash is not a red test, so the row carries the reading
+that has a summary behind it.
+
+**The fix carries its own RED-before pair, measured here rather than inherited
+from the review.** Two concurrent runs of the same binary, same box, same
+minute:
+
+| arm | compiler exit | run A | run B | `Status: SUCCESS` printed |
+|---|---:|---|---|---|
+| `UniqueTempDir` (fixed) | 0 | exit 0 | exit 0 | both |
+| `static int counter` (RED) | 0 | exit 1 | **exit 135, `Bus error (core dumped)`** | **neither** |
+
+Taken at **45 GB free and 34 GB of free RAM**, which settles the cause: the
+first write-up blamed disk, and the crash reproduces with plenty of both. The
+RED arm needed one extra edit to COMPILE — reverting the two call sites leaves
+`UniqueTempDir` unused and `-Werror=unused-function` fails the build — and a
+mutation that fails to build reads as a passing test, so `[[maybe_unused]]` was
+added to the RED arm and its `compile_err=0` is recorded above beside the
+result. Note what the RED row does NOT say: run A "failed" with exit 1 and run B
+printed nothing at all. **Neither process printed a summary**, which is the
+whole hazard — the mode this defect produces is *no result*, and no result reads
+like a run that has not finished.
+
+**Two more pieces of production code went the same way as M12, on the same
+argument.** `Dots3NoteAccounting::deferred()` had no production caller — its
+only three were in the test, two lines below assertions that already read
+`acc.vision == 2195` and `acc.audio == 430` directly — and its second
+`VT_CHECK` was unreachable by exactly M12's reasoning, so it is deleted rather
+than staged (review F2). And the classifier's `else ++acc.audio` would have
+counted a hypothetical THIRD registered tower as audio: the table decided
+language-versus-deferred correctly and then inflated the wrong bucket. It now
+dispatches on the table INDEX and reports a counter-less tower as UNACCOUNTED,
+so the load refuses naming it instead of miscounting (review F3). That branch is
+unreachable while the table has two entries, and it is written as a safe
+degradation rather than as a guard this gate can prove.
+
+**What the gate costs.** 26 cases, CPU-only, and at `-DCMAKE_BUILD_TYPE=Debug
+-DCMAKE_CXX_FLAGS_DEBUG=-O0` it runs in about 170 s, of which roughly 135 s
+predates W2. Three cases dominate and all three build a whole-tower synthetic
+safetensors and drive it through the registry; the assertion count is not the
+cost. W2 halved its own share by loading once and reading the message instead of
+running a 38006-tensor load per `CHECK_THROWS_WITH_AS`, and by reporting one
+assertion per defect class with the first offender named rather than one per
+tensor — which also stops a single classifier defect printing 2197 lines.
+
 ## 5. Gates
 
 **Correctness first, and the gate form is chosen by measurement, not in advance**
@@ -786,18 +1046,24 @@ dispatchable in order, under the constraints that answer imposes.
   the same reasoning `MODEL-MM-muse-glimmer-*` records for staying `SPIKE` with a
   whole text forward landed. `docs/USAGE.md` owes the checkpoint table when a
   capability becomes reachable; `docs/FEATURES.md` carries the arch row now.
-- **W2 — weight map.** `model.safetensors.index.json` read for real, in full:
-  the 42 backbone layers W1's committed slice does not cover, and the two tower
-  files (`model-vision.safetensors`, `model-audio.safetensors`) that W1
-  classifies as named W6/W7 deferrals. The full/sliding split, `g_proj`,
-  `k_rope_only_layernorm`, the indexer tensors, the 256-expert w13/w2 mapping
-  and the nextn tail are ALREADY gated over the slice at W1, and **§1.4 is
-  resolved** (§4.1) rather than owed here. Gate: name-map checker over all
-  38006 tensors, no unclaimed tensor.
+- **W2 — weight map. DONE** (`row/MODEL-MM-dots3-note-W2`,
+  [#699](https://github.com/mudler/vllm.cpp/issues/699)).
+  `model.safetensors.index.json` read for real, in full: the 42 backbone layers
+  W1's committed slice does not cover, and the two tower files
+  (`model-vision.safetensors`, `model-audio.safetensors`) that W1 classifies as
+  named W6/W7 deferrals. **Gate met: 38006/38006, buckets 35381 language /
+  2195 vision / 430 audio, zero unaccounted, zero invented, zero duplicated.**
+  §4.4 carries the evidence, the mutation table and the three things the slice
+  could not see. §1.2's vision pyramid and §1.4 are now checkpoint-measured
+  rather than config-inferred.
 - **W3 — full-attention layer.** `_forward_note_mla` over our DeepSeek MLA:
   lora rescales, `k_rope_only_layernorm`, headwise gate, DSA indexer at
   `indexer_rope_interleave=True`. Gate: independent double-precision reference,
-  RED-first, mutation-proved.
+  RED-first, mutation-proved. **W2 added one obligation here**
+  ([#1846](https://github.com/mudler/vllm.cpp/issues/1846)): the released index
+  declares `indexer_rope_layout: "leading"`, so the indexer's rope slice is the
+  LEADING 64 of the 128-wide head, and W3 asserts that rather than inheriting
+  it silently. See §4.4.
 - **W4 — sliding-window MLA.** The §2.3 stack: windowed metadata, the gather, the
   score mask, and the padded/heterogeneous KV spec. The largest brick; likely
   splits further once W3 lands.
@@ -958,13 +1224,30 @@ porting a model, and the §8.1 heading restructure that `ACTIVE` requires belong
 to the brick where the forward stops refusing. `MODEL-MM-muse-glimmer-*` records
 the same reasoning with a whole text forward landed.
 
-**Next dispatchable: W2 — the weight map.** The whole
-`model.safetensors.index.json` rather than W1's four-layer slice: all 38006
-tensors, the 42 backbone layers W1's slice does not cover, and the two tower
-files (`model-vision.safetensors`, `model-audio.safetensors`) that W1 classifies
-as named W6/W7 deferrals. W1's `EnumerateDots3NoteTensors` and
-`AccountDots3NoteTensors` are the seam it extends; the shapes for the slice are
-already committed, so W2's new work is the towers and the live re-verification.
-W10 additionally owes one reconciliation W1 could not make: upstream's
+**W2 — DONE 2026-08-24.** The whole `model.safetensors.index.json` rather than
+W1's four-layer slice: **38006/38006 accounted**, 35381 language / 2195 vision /
+430 audio, every bucket asserted by number, zero unaccounted. Headers only —
+4770592 bytes over the 133 shard headers, no tensor byte, no GPU. The two tower
+files are now NAMED DEFERRAL RECORDS rather than integer counters, and the load
+refusal prints the table. §4.4 carries the evidence, the fetch recipe and the
+mutation table.
+
+**W2 settled three things W1 could only claim, and found a fourth.** The
+backbone has exactly four distinct layer shapes, so the 1/2 repeat holds and no
+layer breaks it. The full/sliding split read off the shipped indexer tensors
+equals `config.layer_types` exactly. The checkpoint carries 62 F32 tensors in
+TWO families — the 45 language `e_score_correction_bias` W1 predicted, plus 17
+`vision_encoder.blocks.{25..41}.mlp.router_bias` its language-only slice could
+not see, which is spec R5's shape. And the index declares
+`indexer_rope_layout: "leading"` / `indexer_rope_converted_from: "tail"`, which
+NO upstream code reads: [#1846](https://github.com/mudler/vllm.cpp/issues/1846),
+owed by W3.
+
+W10 still owes one reconciliation neither W1 nor W2 could make: upstream's
 `config.layer_types[layer_idx]` has no entry at the nextn index, so the
 checkpoint — not `model.py:503` — is what says that block is sliding.
+
+**Next dispatchable: W3 — the full-attention layer.** `_forward_note_mla` over
+our DeepSeek MLA, gated against an independent double-precision reference under
+option B. It inherits two things from W2: the indexer rope slice above, and a
+weight map that no longer has to be guessed at.
diff --git a/docs/FEATURES.md b/docs/FEATURES.md
index 29cf0610d..6d6685f81 100644
--- a/docs/FEATURES.md
+++ b/docs/FEATURES.md
@@ -148,7 +148,7 @@ speed-pending, which [BENCHMARKS.md](BENCHMARKS.md) tracks.
 | `LagunaForCausalLM` | poolside/Laguna-S-2.1-NVFP4, GGUF-Q4_K, Laguna-XS | byte-exact near-tie (distributional vs vLLM) | vLLM parity+ 1.03x, default on, via the `laguna-gen` CLI; the registered engine forward VT_CHECKs non-bf16 (`ARCH-ONE-SURFACE` fold) |
 | `KimiLinearForCausalLM` | Kimi-Linear-48B-A3B (KDA + NoPE-MLA + MoE) | **Folded onto the shared paged runner (ROW 7 §21, #122): engine==CLI 128/128 byte-identical; vs golden 122/128 (the intrinsic near-tie profile); FA2 paged MLA default-ON; SACRED post-fold green** | Served via `vllm_engine_load` + `vllm_complete_tokens` (ABI v13); server 19.0 tok/s wall vs vLLM ~21 (~0.90×), speed residual open |
 | `KimiK3ForConditionalGeneration` | Kimi-K3 (2.8T MoE) | scaffold: registry+config+enumeration gated, forward refuses | HW-infeasible (~1.56 TB); no run |
-| `Dots3NoteForCausalLM` | `dots-studio/dots3-note-prev` @`1e1e7b0c` (280B-A16B multimodal MoE, ~576 GB bf16; the `-fp8` sibling is ~290 GB). Headers only — no tensor byte downloaded | W1 scaffold: registry + config gated off the REAL released `config.json`, with one assertion per §4 config trap (ungrouped 1/1 router, GPT-J indexer RoPE, one nextn layer, the two LoRA rescales, the sliding theta); name map accounted 1614/1614 over a committed slice of the released index; load, GGUF and forward all REFUSE BY NAME | **No oracle, on any host we own** (~290 GB fp8 against a 122 GiB ceiling), so NO number is claimable on any axis and the e2e gate is an open gap by construction ([spec](../.agents/specs/dots3-note.md) §6.4, #699) |
+| `Dots3NoteForCausalLM` | `dots-studio/dots3-note-prev` @`1e1e7b0c` (280B-A16B multimodal MoE, ~576 GB bf16; the `-fp8` sibling is ~290 GB). Headers only — no tensor byte downloaded | W1+W2 scaffold: registry + config gated off the REAL released `config.json`, with one assertion per §4 config trap (ungrouped 1/1 router, GPT-J indexer RoPE, one nextn layer, the two LoRA rescales, the sliding theta); name map accounted **38006/38006** over the WHOLE released index — 35381 language, 2195 vision, 430 audio, with the two tower files carried as named W6/W7 deferrals rather than dropped; load, GGUF and forward all REFUSE BY NAME | **No oracle, on any host we own** (~290 GB fp8 against a 122 GiB ceiling), so NO number is claimable on any axis and the e2e gate is an open gap by construction ([spec](../.agents/specs/dots3-note.md) §6.4, #699) |
 | `NemotronHForCausalLM` | Nemotron-3.5-Lightning-30B-A3B-NVFP4 (`nvidia` @`29f2d174`) | config+enumeration+KV-shape gated; hybrid forward COMPUTES; loader materializes 18487/18487 as SHIPPED; A3 e2e gate 96/96 `STRICT PASS` on GB10 at `0ea5d249f` (#1221); NO run against current `main` | **PAGED (#810 A2-P): K/V go to the runner's pages; conv+SSM rows carry at the metadata's state indices.** G-SAFE: `num_reqs <= 1`. Device `lm_head` (A2-Q2b), UNMEASURED. Owed: FP8 mamba (A2-Q1), MTP, GGUF |
 | `MuseGlimmerForCausalLM` | real tensors, **bf16 depth 4/52 only**: 5 prefill argmax positions match a torch transcription of vllm#51655 and HF. GGUF full depth generates coherently (#347, #359) but is **NOT token-exact** | text forward + loader vs an fp32 reference, per-mechanism property tests, scaffold 11/11, GGUF gate 17/17. An ABSENT config key now takes the architecture's constant (#412): GGUF post-norms ran at 1e-5, not 1e-8 | no vLLM denominator (pin cannot load it); SECONDARY llama.cpp, same GGUF, GB10 CPU: prefill tie **0.997x**, decode 0.232x, RSS 1.92x (#333) |
 | `MuseGlimmerForConditionalGeneration` | vision: **no reference run of any kind**; enumeration gated vs the released 30B index (1436/1436). Image/video need bf16 safetensors: `mmproj-kquant.gguf` is refused by name | perception encoder loaded and wired, so an image or video prompt runs; `perception_emb_norm` now armed by default (#405). Reachability plus placeholder scatter only, no image or video correctness | not measurable; anchored to open vllm#51655 |
diff --git a/src/vllm/model_executor/models/dots3_note.cpp b/src/vllm/model_executor/models/dots3_note.cpp
index 855c0e5ee..def014360 100644
--- a/src/vllm/model_executor/models/dots3_note.cpp
+++ b/src/vllm/model_executor/models/dots3_note.cpp
@@ -523,6 +523,33 @@ std::vector<Dots3NoteTensor> EnumerateDots3NoteTensors(
                                    /*include_nextn=*/true);
 }
 
+const std::vector<Dots3NoteDeferredTower>& Dots3NoteDeferredTowers() {
+  // The prefixes are upstream's own, read from the hf_to_vllm_mapper at
+  // `nvidia/multimodal.py:70-78` (the two prefixes at `:75-76`):
+  // "vision_encoder." -> "visual." and "audio_encoder." -> "audio_tower.".
+  // RE-DERIVED at vLLM `origin/main` = `185cada36b`, which is where W2 read it;
+  // the same mapper sits at `:54-62` at W1's `c205726108`, and citing that from
+  // here was an inherited anchor rather than a re-read one (review F4 on
+  // #1847). The FILE beside each one is the
+  // released checkpoint's, from `model.safetensors.index.json`'s weight_map at
+  // revision 1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b: each tower ships whole
+  // in one standalone file rather than across the 131 numbered language shards.
+  static const std::vector<Dots3NoteDeferredTower> kTowers{
+      {"vision_encoder.", "model-vision.safetensors", "W6",
+       "the MoE ViT vision tower (nvidia/vision.py, nvidia/vision_moe.py)"},
+      {"audio_encoder.", "model-audio.safetensors", "W7",
+       "the `dots` Whisper-variant audio tower (nvidia/audio_encoder.py)"},
+  };
+  return kTowers;
+}
+
+const Dots3NoteDeferredTower* Dots3NoteDeferralFor(const std::string& name) {
+  for (const Dots3NoteDeferredTower& t : Dots3NoteDeferredTowers()) {
+    if (name.rfind(t.prefix, 0) == 0) return &t;
+  }
+  return nullptr;
+}
+
 Dots3NoteAccounting AccountDots3NoteTensors(
     const Dots3NoteParams& p, const std::vector<std::string>& present,
     const std::vector<int64_t>& expected_layers) {
@@ -534,6 +561,16 @@ Dots3NoteAccounting AccountDots3NoteTensors(
   for (const Dots3NoteTensor& t : claimed) {
     VT_CHECK(!t.consumer.empty(),
              "dots3-note: enumerated " + t.name + " with no named consumer");
+    // NO RUNTIME GUARD HERE, deliberately. A name cannot be both a language
+    // weight and a deferred tower weight — the language branch below wins, so
+    // the tower count would silently drop while the total still read 100%
+    // accounted. W2 first wrote that invariant as a VT_CHECK on this line and
+    // MEASURED it dead: no config makes `EnumerateDots3NoteTensors` emit a
+    // `vision_encoder.` or `audio_encoder.` name, every name it emits is
+    // `model.`- or `lm_head`-prefixed by construction, and deleting the check
+    // left the whole gate green (spec §4.4, mutation M12). The invariant is
+    // real, so it is asserted over the real map in the test instead, where
+    // adding such a name to this function is what fires it.
     if (!claimed_names.insert(t.name).second) acc.duplicated.push_back(t.name);
   }
 
@@ -542,16 +579,31 @@ Dots3NoteAccounting AccountDots3NoteTensors(
     if (on_disk.count(name) == 0) acc.missing.push_back(name);
   }
 
-  const auto starts_with = [](const std::string& s, const char* prefix) {
-    return s.rfind(prefix, 0) == 0;
-  };
   for (const std::string& name : present) {
     if (claimed_names.count(name) != 0) {
       ++acc.language;
-    } else if (starts_with(name, "vision_encoder.")) {
-      ++acc.vision;  // W6, named deferral
-    } else if (starts_with(name, "audio_encoder.")) {
-      ++acc.audio;  // W7, named deferral
+      continue;
+    }
+    // NOT an else-branch on a prefix literal: the deferral TABLE decides, so a
+    // tower this port forgot to register cannot quietly pass as language.
+    const Dots3NoteDeferredTower* tower = Dots3NoteDeferralFor(name);
+    if (tower == nullptr) {
+      acc.unaccounted.push_back(name);
+      continue;
+    }
+    // Dispatch on the table INDEX. `else ++acc.audio` would count a THIRD
+    // registered tower as audio — the table would decide language-versus-
+    // deferred correctly and then silently inflate the wrong bucket (review F3
+    // on #1847). A tower with no counter is reported UNACCOUNTED instead, so
+    // the load refuses naming it, and the refusal prints the table beside it so
+    // a reader can see that it IS registered and only the counter is missing.
+    // The branch is unreachable while the table has two entries, and it is a
+    // safe degradation rather than a guard this gate can prove.
+    const size_t which = static_cast<size_t>(tower - Dots3NoteDeferredTowers().data());
+    if (which == 0) {
+      ++acc.vision;
+    } else if (which == 1) {
+      ++acc.audio;
     } else {
       acc.unaccounted.push_back(name);
     }
@@ -588,11 +640,17 @@ Dots3NoteWeights LoadDots3NoteWeights(const std::vector<SafetensorsFile>& shards
                std::to_string(w.accounting.missing.size()) +
                " enumerated tensors are absent) — a weight nobody loads reads "
                "as zeros");
+  std::string towers;
+  for (const Dots3NoteDeferredTower& t : Dots3NoteDeferredTowers()) {
+    if (!towers.empty()) towers += ", ";
+    towers += std::string(t.prefix) + "* (" + t.brick + ")";
+  }
   VT_CHECK(w.accounting.unaccounted.empty(),
            "dots3-note: no consumer claims " + w.accounting.unaccounted.front() +
                " (" + std::to_string(w.accounting.unaccounted.size()) +
-               " unaccounted tensors) — see .agents/specs/dots3-note.md W2 and "
-               "issue #699");
+               " unaccounted tensors), and it is not one of the DEFERRED "
+               "towers " + towers +
+               " — see .agents/specs/dots3-note.md and issue #699");
 
   // W2 owns the materialization. Returning an UNMATERIALIZED model rather than
   // throwing is deliberate: the accounting above is a real production result
diff --git a/src/vllm/model_executor/models/dots3_note.h b/src/vllm/model_executor/models/dots3_note.h
index 5769386e4..899255509 100644
--- a/src/vllm/model_executor/models/dots3_note.h
+++ b/src/vllm/model_executor/models/dots3_note.h
@@ -218,22 +218,51 @@ struct Dots3NoteTensor {
 // tensors), routed experts unstacked, `model.` prefix. Ordered: root, backbone
 // layers ascending, then the nextn tail.
 //
-// `layers` selects WHICH backbone layers to enumerate. W1 gates a single-layer
-// slice per class rather than the whole 38006-tensor checkpoint; W2 owns the
-// full map plus the two tower files (`model-vision.safetensors`,
-// `model-audio.safetensors`), which this function deliberately does NOT claim.
+// `layers` selects WHICH backbone layers to enumerate. Pass every backbone
+// layer for a production load; a slice is for a focused test. This function
+// claims the LANGUAGE tower only — the two tower files are named deferrals,
+// see `Dots3NoteDeferredTowers()` below.
 std::vector<Dots3NoteTensor> EnumerateDots3NoteTensors(
     const Dots3NoteParams& params, const std::vector<int64_t>& layers,
     bool include_root, bool include_nextn);
 
-// Convenience: every backbone layer, the root tensors and the nextn tail.
+// Convenience: every backbone layer, the root tensors and the nextn tail. Over
+// the released checkpoint this is exactly 35381 names.
 std::vector<Dots3NoteTensor> EnumerateDots3NoteTensors(
     const Dots3NoteParams& params);
 
+// A tower the released checkpoint ships that this port does NOT load yet, and
+// the brick that owes it. This table is the difference between a DEFERRAL and a
+// SILENT DROP: a tensor matched here is refused a language-tower consumer on
+// purpose, by a record that names the brick, the file it ships in and what it
+// is. A tensor matched by nothing lands in `unaccounted` and refuses the load.
+//
+// The counters alone could not carry that meaning. Folding the towers into the
+// language count leaves every "100% accounted" assertion green while 2625
+// weights go unloaded — the exact mutation the W1 review found unguarded
+// (#1805, M15), and W2 is the scale at which it would have mattered.
+struct Dots3NoteDeferredTower {
+  const char* prefix;  // the on-disk name prefix, e.g. "vision_encoder."
+  const char* file;    // the ONE shard file every one of them ships in
+  const char* brick;   // the phase of `.agents/specs/dots3-note.md` §7 that owes it
+  const char* what;    // what it is, for the message a reader gets
+};
+
+// The complete deferral table, in the order a report should print it.
+const std::vector<Dots3NoteDeferredTower>& Dots3NoteDeferredTowers();
+
+// The deferral that claims `name`, or nullptr when no deferral does. A nullptr
+// for a name the language map does not claim either is an UNACCOUNTED tensor.
+const Dots3NoteDeferredTower* Dots3NoteDeferralFor(const std::string& name);
+
 // What an accounting pass over a checkpoint's tensor NAMES found. Every name on
 // disk lands in exactly one bucket, and `unaccounted` must be empty: a tensor
 // nobody claims is a silently dropped weight, which reads as zeros and renders.
 // The two tower buckets are NAMED deferrals (W6 vision, W7 audio), not silence.
+//
+// Over the whole released `dots-studio/dots3-note-prev` index the three buckets
+// are 35381 / 2195 / 430 = 38006. Assert them BY NUMBER: "nothing was left
+// over" is also true of a classifier that claims the towers as language.
 struct Dots3NoteAccounting {
   int64_t language = 0;  // claimed by a named language-tower consumer
   int64_t vision = 0;    // `vision_encoder.*`, deferred to W6
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 8c2d69e3a..dd981266d 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -894,14 +894,17 @@ target_include_directories(test_muse_glimmer_gguf PRIVATE
 # .agents/specs/kimi-linear.md.
 vllm_cpp_add_test(test_kimi_linear_scaffold vllm/models/test_kimi_linear_scaffold.cpp)
 
-# dots3-note W1 STRUCTURAL gate (#699, .agents/specs/dots3-note.md §4 / §7 W1):
+# dots3-note W1+W2 STRUCTURAL gate (#699, .agents/specs/dots3-note.md §4 / §7):
 # registry-resolve + config descent off the REAL released config.json, with one
 # RED-first assertion per §4 CONFIG TRAP (n_group/topk_group=1,
 # indexer_rope_interleave=True, num_nextn_predict_layers=1, the two lora
 # rescales, the sliding rope's own theta and the GPT-J layout both ropes share),
-# the on-disk name map accounted BOTH WAYS over a committed slice of the
-# released shard index (1614 tensors: the root, one layer of every class, the
-# nextn tail), the padded 1088-wide MLA row, and the load/GGUF/forward refusals
+# the on-disk name map accounted BOTH WAYS over the WHOLE released shard index
+# (W2: all 38006 tensors -- 35381 language, 2195 vision, 430 audio -- from a
+# committed headers-only projection of the index joined to all 133 safetensors
+# headers, plus W1's 1614-tensor slice kept as a cross-check), the two tower
+# files as NAMED W6/W7 deferrals asserted by count, the padded 1088-wide MLA
+# row, and the load/GGUF/forward refusals
 # — the last driven through the REAL model the factory returns, never a
 # fabricated LoadedModel (#730/#784). Spec §6.4 records that NO oracle for this
 # model runs on hardware this project owns, so these unit assertions are the
diff --git a/tests/vllm/models/fixtures/dots3_note_prev/index_full.json b/tests/vllm/models/fixtures/dots3_note_prev/index_full.json
new file mode 100644
index 000000000..3194a1fce
--- /dev/null
+++ b/tests/vllm/models/fixtures/dots3_note_prev/index_full.json
@@ -0,0 +1,1824 @@
+{
+ "_provenance": "The COMPLETE released model.safetensors.index.json of dots-studio/dots3-note-prev @ 1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b, joined to the safetensors HEADER of all 133 shard files it names. HEADERS ONLY: each file was read with two HTTP Range requests (8 bytes for the header length, then the header JSON), 4770592 header bytes in total across the 133 files, and NOT ONE TENSOR BYTE -- the checkpoint is ~576 GB and was never downloaded. This is the WHOLE checkpoint, not a slice: all 38006 tensors, including the 2195 in model-vision.safetensors and the 430 in model-audio.safetensors. ONE index is collapsed, the routed-expert index, to {E} with `count` beside it; every backbone layer, vision block and audio layer is its own entry, because W2 exists to MEASURE that the layers repeat and a fixture that collapsed them would assume the answer.",
+ "_schema": "tensors[name] = [count, dtype, shape, file]. `file` is the shard the tensors ship in, or \"*\" when a {E} family spans more than one shard.",
+ "repo": "dots-studio/dots3-note-prev",
+ "revision": "1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b",
+ "index_json_sha256": "95a364b468a93ccad6adcb9c3aa110cb7a1411c2575c334c39022f9f84d456e1",
+ "index_metadata": {"total_size": 576886825984, "indexer_rope_layout": "leading", "indexer_rope_converted_from": "tail"},
+ "checkpoint_total_tensors": 38006,
+ "checkpoint_total_size_bytes": 576886825984,
+ "shard_file_count": 133,
+ "tower_files": {"vision_encoder.": "model-vision.safetensors", "audio_encoder.": "model-audio.safetensors"},
+ "bucket_totals": {"language": 35381, "vision": 2195, "audio": 430},
+ "total_families": 1808,
+ "tensors": {
+  "audio_encoder.audio_adapter.proj.0.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.audio_adapter.proj.0.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.audio_adapter.proj.1.bias": [1, "BF16", [5120], "model-audio.safetensors"],
+  "audio_encoder.audio_adapter.proj.1.weight": [1, "BF16", [5120, 1280], "model-audio.safetensors"],
+  "audio_encoder.audio_adapter.proj.3.bias": [1, "BF16", [5120], "model-audio.safetensors"],
+  "audio_encoder.audio_adapter.proj.3.weight": [1, "BF16", [5120, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.conv2d1.bias": [1, "BF16", [480], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.conv2d1.weight": [1, "BF16", [480, 1, 3, 3], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.conv2d2.bias": [1, "BF16", [480], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.conv2d2.weight": [1, "BF16", [480, 480, 3, 3], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.conv2d3.bias": [1, "BF16", [480], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.conv2d3.weight": [1, "BF16", [480, 480, 3, 3], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.conv_out.weight": [1, "BF16", [1280, 7680], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.0.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.1.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.10.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.11.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.12.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.13.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.14.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.15.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.16.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.17.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.18.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.19.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.2.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.20.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.21.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.22.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.23.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.24.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.25.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.26.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.27.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.28.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.29.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.3.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.30.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.31.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.4.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.5.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.6.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.7.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.8.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.fc1.bias": [1, "BF16", [10240], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.fc1.weight": [1, "BF16", [10240, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.fc2.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.fc2.weight": [1, "BF16", [1280, 5120], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.final_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.self_attn.k_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.self_attn.out_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.self_attn.out_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.self_attn.q_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.self_attn.q_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.self_attn.v_proj.bias": [1, "BF16", [1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.self_attn.v_proj.weight": [1, "BF16", [1280, 1280], "model-audio.safetensors"],
+  "audio_encoder.dots_encoder.speech_encoder.layers.9.self_attn_layer_norm.weight": [1, "BF16", [1280], "model-audio.safetensors"],
+  "lm_head.weight": [1, "BF16", [152064, 5120], "model-00069-of-00131.safetensors"],
+  "model.embed_tokens.weight": [1, "BF16", [152064, 5120], "model-00083-of-00131.safetensors"],
+  "model.layers.0.input_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.0.mlp.down_proj.weight": [1, "BF16", [5120, 13824], "model-00086-of-00131.safetensors"],
+  "model.layers.0.mlp.gate_proj.weight": [1, "BF16", [13824, 5120], "model-00086-of-00131.safetensors"],
+  "model.layers.0.mlp.up_proj.weight": [1, "BF16", [13824, 5120], "model-00086-of-00131.safetensors"],
+  "model.layers.0.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00064-of-00131.safetensors"],
+  "model.layers.0.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00049-of-00131.safetensors"],
+  "model.layers.0.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00082-of-00131.safetensors"],
+  "model.layers.0.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00082-of-00131.safetensors"],
+  "model.layers.0.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00058-of-00131.safetensors"],
+  "model.layers.0.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00048-of-00131.safetensors"],
+  "model.layers.0.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00005-of-00131.safetensors"],
+  "model.layers.0.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00082-of-00131.safetensors"],
+  "model.layers.0.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00081-of-00131.safetensors"],
+  "model.layers.0.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00035-of-00131.safetensors"],
+  "model.layers.0.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00104-of-00131.safetensors"],
+  "model.layers.0.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00084-of-00131.safetensors"],
+  "model.layers.0.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00077-of-00131.safetensors"],
+  "model.layers.0.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00022-of-00131.safetensors"],
+  "model.layers.0.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00089-of-00131.safetensors"],
+  "model.layers.1.input_layernorm.weight": [1, "BF16", [5120], "model-00065-of-00131.safetensors"],
+  "model.layers.1.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.1.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.1.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.1.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00081-of-00131.safetensors"],
+  "model.layers.1.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00037-of-00131.safetensors"],
+  "model.layers.1.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00006-of-00131.safetensors"],
+  "model.layers.1.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00131-of-00131.safetensors"],
+  "model.layers.1.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00131-of-00131.safetensors"],
+  "model.layers.1.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00066-of-00131.safetensors"],
+  "model.layers.1.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00053-of-00131.safetensors"],
+  "model.layers.1.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00082-of-00131.safetensors"],
+  "model.layers.1.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00082-of-00131.safetensors"],
+  "model.layers.1.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00059-of-00131.safetensors"],
+  "model.layers.1.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00051-of-00131.safetensors"],
+  "model.layers.1.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00062-of-00131.safetensors"],
+  "model.layers.1.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00082-of-00131.safetensors"],
+  "model.layers.1.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00081-of-00131.safetensors"],
+  "model.layers.1.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00036-of-00131.safetensors"],
+  "model.layers.1.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00114-of-00131.safetensors"],
+  "model.layers.1.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00084-of-00131.safetensors"],
+  "model.layers.1.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00078-of-00131.safetensors"],
+  "model.layers.1.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00024-of-00131.safetensors"],
+  "model.layers.1.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00089-of-00131.safetensors"],
+  "model.layers.10.input_layernorm.weight": [1, "BF16", [5120], "model-00125-of-00131.safetensors"],
+  "model.layers.10.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.10.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.10.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.10.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00129-of-00131.safetensors"],
+  "model.layers.10.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00123-of-00131.safetensors"],
+  "model.layers.10.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00117-of-00131.safetensors"],
+  "model.layers.10.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00117-of-00131.safetensors"],
+  "model.layers.10.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00117-of-00131.safetensors"],
+  "model.layers.10.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00125-of-00131.safetensors"],
+  "model.layers.10.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00125-of-00131.safetensors"],
+  "model.layers.10.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00131-of-00131.safetensors"],
+  "model.layers.10.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00127-of-00131.safetensors"],
+  "model.layers.10.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00116-of-00131.safetensors"],
+  "model.layers.10.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00006-of-00131.safetensors"],
+  "model.layers.10.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00003-of-00131.safetensors"],
+  "model.layers.10.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00128-of-00131.safetensors"],
+  "model.layers.10.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00118-of-00131.safetensors"],
+  "model.layers.10.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00008-of-00131.safetensors"],
+  "model.layers.11.input_layernorm.weight": [1, "BF16", [5120], "model-00126-of-00131.safetensors"],
+  "model.layers.11.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.11.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.11.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.11.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00129-of-00131.safetensors"],
+  "model.layers.11.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00123-of-00131.safetensors"],
+  "model.layers.11.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00117-of-00131.safetensors"],
+  "model.layers.11.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00117-of-00131.safetensors"],
+  "model.layers.11.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00117-of-00131.safetensors"],
+  "model.layers.11.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00126-of-00131.safetensors"],
+  "model.layers.11.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00125-of-00131.safetensors"],
+  "model.layers.11.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00131-of-00131.safetensors"],
+  "model.layers.11.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00128-of-00131.safetensors"],
+  "model.layers.11.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00116-of-00131.safetensors"],
+  "model.layers.11.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00006-of-00131.safetensors"],
+  "model.layers.11.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00003-of-00131.safetensors"],
+  "model.layers.11.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00128-of-00131.safetensors"],
+  "model.layers.11.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00119-of-00131.safetensors"],
+  "model.layers.11.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00008-of-00131.safetensors"],
+  "model.layers.12.input_layernorm.weight": [1, "BF16", [5120], "model-00018-of-00131.safetensors"],
+  "model.layers.12.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.12.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.12.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.12.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00021-of-00131.safetensors"],
+  "model.layers.12.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00016-of-00131.safetensors"],
+  "model.layers.12.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00011-of-00131.safetensors"],
+  "model.layers.12.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00010-of-00131.safetensors"],
+  "model.layers.12.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00010-of-00131.safetensors"],
+  "model.layers.12.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00018-of-00131.safetensors"],
+  "model.layers.12.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00018-of-00131.safetensors"],
+  "model.layers.12.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00023-of-00131.safetensors"],
+  "model.layers.12.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00020-of-00131.safetensors"],
+  "model.layers.12.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00010-of-00131.safetensors"],
+  "model.layers.12.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00030-of-00131.safetensors"],
+  "model.layers.12.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00026-of-00131.safetensors"],
+  "model.layers.12.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00020-of-00131.safetensors"],
+  "model.layers.12.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00012-of-00131.safetensors"],
+  "model.layers.12.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00032-of-00131.safetensors"],
+  "model.layers.13.input_layernorm.weight": [1, "BF16", [5120], "model-00019-of-00131.safetensors"],
+  "model.layers.13.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.13.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.13.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.13.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00021-of-00131.safetensors"],
+  "model.layers.13.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00016-of-00131.safetensors"],
+  "model.layers.13.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00011-of-00131.safetensors"],
+  "model.layers.13.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00011-of-00131.safetensors"],
+  "model.layers.13.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00011-of-00131.safetensors"],
+  "model.layers.13.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00019-of-00131.safetensors"],
+  "model.layers.13.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00017-of-00131.safetensors"],
+  "model.layers.13.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00022-of-00131.safetensors"],
+  "model.layers.13.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00023-of-00131.safetensors"],
+  "model.layers.13.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00018-of-00131.safetensors"],
+  "model.layers.13.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00017-of-00131.safetensors"],
+  "model.layers.13.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00010-of-00131.safetensors"],
+  "model.layers.13.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00023-of-00131.safetensors"],
+  "model.layers.13.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00022-of-00131.safetensors"],
+  "model.layers.13.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00016-of-00131.safetensors"],
+  "model.layers.13.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00032-of-00131.safetensors"],
+  "model.layers.13.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00025-of-00131.safetensors"],
+  "model.layers.13.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00020-of-00131.safetensors"],
+  "model.layers.13.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00012-of-00131.safetensors"],
+  "model.layers.13.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00029-of-00131.safetensors"],
+  "model.layers.14.input_layernorm.weight": [1, "BF16", [5120], "model-00019-of-00131.safetensors"],
+  "model.layers.14.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.14.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.14.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.14.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00022-of-00131.safetensors"],
+  "model.layers.14.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00016-of-00131.safetensors"],
+  "model.layers.14.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00013-of-00131.safetensors"],
+  "model.layers.14.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00013-of-00131.safetensors"],
+  "model.layers.14.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00013-of-00131.safetensors"],
+  "model.layers.14.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00019-of-00131.safetensors"],
+  "model.layers.14.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00018-of-00131.safetensors"],
+  "model.layers.14.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00023-of-00131.safetensors"],
+  "model.layers.14.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00020-of-00131.safetensors"],
+  "model.layers.14.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00010-of-00131.safetensors"],
+  "model.layers.14.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00030-of-00131.safetensors"],
+  "model.layers.14.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00026-of-00131.safetensors"],
+  "model.layers.14.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00020-of-00131.safetensors"],
+  "model.layers.14.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00013-of-00131.safetensors"],
+  "model.layers.14.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00032-of-00131.safetensors"],
+  "model.layers.15.input_layernorm.weight": [1, "BF16", [5120], "model-00019-of-00131.safetensors"],
+  "model.layers.15.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.15.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.15.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.15.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00022-of-00131.safetensors"],
+  "model.layers.15.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00017-of-00131.safetensors"],
+  "model.layers.15.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00015-of-00131.safetensors"],
+  "model.layers.15.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00014-of-00131.safetensors"],
+  "model.layers.15.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00014-of-00131.safetensors"],
+  "model.layers.15.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00019-of-00131.safetensors"],
+  "model.layers.15.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00018-of-00131.safetensors"],
+  "model.layers.15.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00023-of-00131.safetensors"],
+  "model.layers.15.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00020-of-00131.safetensors"],
+  "model.layers.15.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00011-of-00131.safetensors"],
+  "model.layers.15.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00030-of-00131.safetensors"],
+  "model.layers.15.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00027-of-00131.safetensors"],
+  "model.layers.15.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00021-of-00131.safetensors"],
+  "model.layers.15.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00014-of-00131.safetensors"],
+  "model.layers.15.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00033-of-00131.safetensors"],
+  "model.layers.16.input_layernorm.weight": [1, "BF16", [5120], "model-00043-of-00131.safetensors"],
+  "model.layers.16.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.16.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.16.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.16.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00047-of-00131.safetensors"],
+  "model.layers.16.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00041-of-00131.safetensors"],
+  "model.layers.16.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00036-of-00131.safetensors"],
+  "model.layers.16.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00035-of-00131.safetensors"],
+  "model.layers.16.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00035-of-00131.safetensors"],
+  "model.layers.16.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00043-of-00131.safetensors"],
+  "model.layers.16.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00043-of-00131.safetensors"],
+  "model.layers.16.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00048-of-00131.safetensors"],
+  "model.layers.16.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00045-of-00131.safetensors"],
+  "model.layers.16.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00035-of-00131.safetensors"],
+  "model.layers.16.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00055-of-00131.safetensors"],
+  "model.layers.16.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00052-of-00131.safetensors"],
+  "model.layers.16.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00045-of-00131.safetensors"],
+  "model.layers.16.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00037-of-00131.safetensors"],
+  "model.layers.16.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00057-of-00131.safetensors"],
+  "model.layers.17.input_layernorm.weight": [1, "BF16", [5120], "model-00044-of-00131.safetensors"],
+  "model.layers.17.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.17.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.17.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.17.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00047-of-00131.safetensors"],
+  "model.layers.17.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00041-of-00131.safetensors"],
+  "model.layers.17.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00036-of-00131.safetensors"],
+  "model.layers.17.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00036-of-00131.safetensors"],
+  "model.layers.17.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00036-of-00131.safetensors"],
+  "model.layers.17.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00044-of-00131.safetensors"],
+  "model.layers.17.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00042-of-00131.safetensors"],
+  "model.layers.17.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00048-of-00131.safetensors"],
+  "model.layers.17.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00048-of-00131.safetensors"],
+  "model.layers.17.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00043-of-00131.safetensors"],
+  "model.layers.17.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00042-of-00131.safetensors"],
+  "model.layers.17.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00035-of-00131.safetensors"],
+  "model.layers.17.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00048-of-00131.safetensors"],
+  "model.layers.17.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00047-of-00131.safetensors"],
+  "model.layers.17.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00041-of-00131.safetensors"],
+  "model.layers.17.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00057-of-00131.safetensors"],
+  "model.layers.17.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00050-of-00131.safetensors"],
+  "model.layers.17.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00045-of-00131.safetensors"],
+  "model.layers.17.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00037-of-00131.safetensors"],
+  "model.layers.17.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00054-of-00131.safetensors"],
+  "model.layers.18.input_layernorm.weight": [1, "BF16", [5120], "model-00044-of-00131.safetensors"],
+  "model.layers.18.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.18.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.18.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.18.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00047-of-00131.safetensors"],
+  "model.layers.18.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00041-of-00131.safetensors"],
+  "model.layers.18.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00038-of-00131.safetensors"],
+  "model.layers.18.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00038-of-00131.safetensors"],
+  "model.layers.18.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00038-of-00131.safetensors"],
+  "model.layers.18.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00044-of-00131.safetensors"],
+  "model.layers.18.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00043-of-00131.safetensors"],
+  "model.layers.18.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00048-of-00131.safetensors"],
+  "model.layers.18.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00045-of-00131.safetensors"],
+  "model.layers.18.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00035-of-00131.safetensors"],
+  "model.layers.18.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00055-of-00131.safetensors"],
+  "model.layers.18.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00052-of-00131.safetensors"],
+  "model.layers.18.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00046-of-00131.safetensors"],
+  "model.layers.18.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00038-of-00131.safetensors"],
+  "model.layers.18.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00057-of-00131.safetensors"],
+  "model.layers.19.input_layernorm.weight": [1, "BF16", [5120], "model-00044-of-00131.safetensors"],
+  "model.layers.19.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.19.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.19.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.19.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00047-of-00131.safetensors"],
+  "model.layers.19.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00042-of-00131.safetensors"],
+  "model.layers.19.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00040-of-00131.safetensors"],
+  "model.layers.19.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00039-of-00131.safetensors"],
+  "model.layers.19.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00039-of-00131.safetensors"],
+  "model.layers.19.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00044-of-00131.safetensors"],
+  "model.layers.19.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00043-of-00131.safetensors"],
+  "model.layers.19.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00048-of-00131.safetensors"],
+  "model.layers.19.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00046-of-00131.safetensors"],
+  "model.layers.19.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00036-of-00131.safetensors"],
+  "model.layers.19.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00055-of-00131.safetensors"],
+  "model.layers.19.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00053-of-00131.safetensors"],
+  "model.layers.19.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00046-of-00131.safetensors"],
+  "model.layers.19.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00039-of-00131.safetensors"],
+  "model.layers.19.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00059-of-00131.safetensors"],
+  "model.layers.2.input_layernorm.weight": [1, "BF16", [5120], "model-00066-of-00131.safetensors"],
+  "model.layers.2.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.2.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.2.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.2.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00081-of-00131.safetensors"],
+  "model.layers.2.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00037-of-00131.safetensors"],
+  "model.layers.2.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00091-of-00131.safetensors"],
+  "model.layers.2.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00013-of-00131.safetensors"],
+  "model.layers.2.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00013-of-00131.safetensors"],
+  "model.layers.2.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00067-of-00131.safetensors"],
+  "model.layers.2.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00060-of-00131.safetensors"],
+  "model.layers.2.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00082-of-00131.safetensors"],
+  "model.layers.2.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00079-of-00131.safetensors"],
+  "model.layers.2.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00008-of-00131.safetensors"],
+  "model.layers.2.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00090-of-00131.safetensors"],
+  "model.layers.2.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00087-of-00131.safetensors"],
+  "model.layers.2.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00079-of-00131.safetensors"],
+  "model.layers.2.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00026-of-00131.safetensors"],
+  "model.layers.2.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00126-of-00131.safetensors"],
+  "model.layers.20.input_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.20.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.20.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.20.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.20.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00063-of-00131.safetensors"],
+  "model.layers.20.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00060-of-00131.safetensors"],
+  "model.layers.20.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00078-of-00131.safetensors"],
+  "model.layers.20.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00078-of-00131.safetensors"],
+  "model.layers.20.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00078-of-00131.safetensors"],
+  "model.layers.20.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.20.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00064-of-00131.safetensors"],
+  "model.layers.20.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00063-of-00131.safetensors"],
+  "model.layers.20.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00063-of-00131.safetensors"],
+  "model.layers.20.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00063-of-00131.safetensors"],
+  "model.layers.20.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00075-of-00131.safetensors"],
+  "model.layers.20.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00073-of-00131.safetensors"],
+  "model.layers.20.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00064-of-00131.safetensors"],
+  "model.layers.20.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00060-of-00131.safetensors"],
+  "model.layers.20.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00076-of-00131.safetensors"],
+  "model.layers.21.input_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.21.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.21.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.21.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.21.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00063-of-00131.safetensors"],
+  "model.layers.21.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00060-of-00131.safetensors"],
+  "model.layers.21.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00078-of-00131.safetensors"],
+  "model.layers.21.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00078-of-00131.safetensors"],
+  "model.layers.21.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00078-of-00131.safetensors"],
+  "model.layers.21.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00064-of-00131.safetensors"],
+  "model.layers.21.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00063-of-00131.safetensors"],
+  "model.layers.21.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00063-of-00131.safetensors"],
+  "model.layers.21.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00063-of-00131.safetensors"],
+  "model.layers.21.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00064-of-00131.safetensors"],
+  "model.layers.21.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00062-of-00131.safetensors"],
+  "model.layers.21.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00077-of-00131.safetensors"],
+  "model.layers.21.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00064-of-00131.safetensors"],
+  "model.layers.21.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00064-of-00131.safetensors"],
+  "model.layers.21.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00064-of-00131.safetensors"],
+  "model.layers.21.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00076-of-00131.safetensors"],
+  "model.layers.21.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00069-of-00131.safetensors"],
+  "model.layers.21.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00063-of-00131.safetensors"],
+  "model.layers.21.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00060-of-00131.safetensors"],
+  "model.layers.21.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00074-of-00131.safetensors"],
+  "model.layers.22.input_layernorm.weight": [1, "BF16", [5120], "model-00062-of-00131.safetensors"],
+  "model.layers.22.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.22.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.22.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.22.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00063-of-00131.safetensors"],
+  "model.layers.22.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00061-of-00131.safetensors"],
+  "model.layers.22.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00079-of-00131.safetensors"],
+  "model.layers.22.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00079-of-00131.safetensors"],
+  "model.layers.22.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00079-of-00131.safetensors"],
+  "model.layers.22.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.22.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00062-of-00131.safetensors"],
+  "model.layers.22.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00063-of-00131.safetensors"],
+  "model.layers.22.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00064-of-00131.safetensors"],
+  "model.layers.22.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00064-of-00131.safetensors"],
+  "model.layers.22.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00075-of-00131.safetensors"],
+  "model.layers.22.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00073-of-00131.safetensors"],
+  "model.layers.22.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00063-of-00131.safetensors"],
+  "model.layers.22.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00061-of-00131.safetensors"],
+  "model.layers.22.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00076-of-00131.safetensors"],
+  "model.layers.23.input_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.23.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.23.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.23.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.23.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00063-of-00131.safetensors"],
+  "model.layers.23.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00061-of-00131.safetensors"],
+  "model.layers.23.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00080-of-00131.safetensors"],
+  "model.layers.23.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00079-of-00131.safetensors"],
+  "model.layers.23.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00079-of-00131.safetensors"],
+  "model.layers.23.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.23.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00063-of-00131.safetensors"],
+  "model.layers.23.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00064-of-00131.safetensors"],
+  "model.layers.23.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00064-of-00131.safetensors"],
+  "model.layers.23.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00062-of-00131.safetensors"],
+  "model.layers.23.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00075-of-00131.safetensors"],
+  "model.layers.23.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00074-of-00131.safetensors"],
+  "model.layers.23.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00063-of-00131.safetensors"],
+  "model.layers.23.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00061-of-00131.safetensors"],
+  "model.layers.23.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00076-of-00131.safetensors"],
+  "model.layers.24.input_layernorm.weight": [1, "BF16", [5120], "model-00068-of-00131.safetensors"],
+  "model.layers.24.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.24.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.24.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.24.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00081-of-00131.safetensors"],
+  "model.layers.24.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00039-of-00131.safetensors"],
+  "model.layers.24.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00085-of-00131.safetensors"],
+  "model.layers.24.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00085-of-00131.safetensors"],
+  "model.layers.24.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00085-of-00131.safetensors"],
+  "model.layers.24.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00070-of-00131.safetensors"],
+  "model.layers.24.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00060-of-00131.safetensors"],
+  "model.layers.24.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00083-of-00131.safetensors"],
+  "model.layers.24.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00080-of-00131.safetensors"],
+  "model.layers.24.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00015-of-00131.safetensors"],
+  "model.layers.24.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00090-of-00131.safetensors"],
+  "model.layers.24.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00087-of-00131.safetensors"],
+  "model.layers.24.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00080-of-00131.safetensors"],
+  "model.layers.24.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00026-of-00131.safetensors"],
+  "model.layers.24.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00024-of-00131.safetensors"],
+  "model.layers.25.input_layernorm.weight": [1, "BF16", [5120], "model-00070-of-00131.safetensors"],
+  "model.layers.25.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.25.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.25.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.25.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00081-of-00131.safetensors"],
+  "model.layers.25.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00043-of-00131.safetensors"],
+  "model.layers.25.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00089-of-00131.safetensors"],
+  "model.layers.25.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00125-of-00131.safetensors"],
+  "model.layers.25.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00125-of-00131.safetensors"],
+  "model.layers.25.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00070-of-00131.safetensors"],
+  "model.layers.25.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00057-of-00131.safetensors"],
+  "model.layers.25.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00082-of-00131.safetensors"],
+  "model.layers.25.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00082-of-00131.safetensors"],
+  "model.layers.25.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00061-of-00131.safetensors"],
+  "model.layers.25.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00054-of-00131.safetensors"],
+  "model.layers.25.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00105-of-00131.safetensors"],
+  "model.layers.25.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00082-of-00131.safetensors"],
+  "model.layers.25.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00082-of-00131.safetensors"],
+  "model.layers.25.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00037-of-00131.safetensors"],
+  "model.layers.25.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00086-of-00131.safetensors"],
+  "model.layers.25.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00085-of-00131.safetensors"],
+  "model.layers.25.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00080-of-00131.safetensors"],
+  "model.layers.25.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00027-of-00131.safetensors"],
+  "model.layers.25.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00089-of-00131.safetensors"],
+  "model.layers.26.input_layernorm.weight": [1, "BF16", [5120], "model-00071-of-00131.safetensors"],
+  "model.layers.26.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.26.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.26.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.26.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00082-of-00131.safetensors"],
+  "model.layers.26.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00045-of-00131.safetensors"],
+  "model.layers.26.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00123-of-00131.safetensors"],
+  "model.layers.26.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00116-of-00131.safetensors"],
+  "model.layers.26.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00116-of-00131.safetensors"],
+  "model.layers.26.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00072-of-00131.safetensors"],
+  "model.layers.26.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00061-of-00131.safetensors"],
+  "model.layers.26.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00082-of-00131.safetensors"],
+  "model.layers.26.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00081-of-00131.safetensors"],
+  "model.layers.26.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00015-of-00131.safetensors"],
+  "model.layers.26.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00090-of-00131.safetensors"],
+  "model.layers.26.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00088-of-00131.safetensors"],
+  "model.layers.26.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00081-of-00131.safetensors"],
+  "model.layers.26.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00028-of-00131.safetensors"],
+  "model.layers.26.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00093-of-00131.safetensors"],
+  "model.layers.27.input_layernorm.weight": [1, "BF16", [5120], "model-00074-of-00131.safetensors"],
+  "model.layers.27.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.27.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.27.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.27.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00082-of-00131.safetensors"],
+  "model.layers.27.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00046-of-00131.safetensors"],
+  "model.layers.27.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00003-of-00131.safetensors"],
+  "model.layers.27.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00092-of-00131.safetensors"],
+  "model.layers.27.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00092-of-00131.safetensors"],
+  "model.layers.27.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00075-of-00131.safetensors"],
+  "model.layers.27.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00062-of-00131.safetensors"],
+  "model.layers.27.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00083-of-00131.safetensors"],
+  "model.layers.27.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00081-of-00131.safetensors"],
+  "model.layers.27.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00016-of-00131.safetensors"],
+  "model.layers.27.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00037-of-00131.safetensors"],
+  "model.layers.27.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00088-of-00131.safetensors"],
+  "model.layers.27.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00081-of-00131.safetensors"],
+  "model.layers.27.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00030-of-00131.safetensors"],
+  "model.layers.27.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00094-of-00131.safetensors"],
+  "model.layers.28.input_layernorm.weight": [1, "BF16", [5120], "model-00099-of-00131.safetensors"],
+  "model.layers.28.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.28.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.28.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.28.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00102-of-00131.safetensors"],
+  "model.layers.28.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00097-of-00131.safetensors"],
+  "model.layers.28.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00092-of-00131.safetensors"],
+  "model.layers.28.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00091-of-00131.safetensors"],
+  "model.layers.28.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00091-of-00131.safetensors"],
+  "model.layers.28.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00099-of-00131.safetensors"],
+  "model.layers.28.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00099-of-00131.safetensors"],
+  "model.layers.28.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00104-of-00131.safetensors"],
+  "model.layers.28.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00101-of-00131.safetensors"],
+  "model.layers.28.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00091-of-00131.safetensors"],
+  "model.layers.28.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00111-of-00131.safetensors"],
+  "model.layers.28.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00107-of-00131.safetensors"],
+  "model.layers.28.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00101-of-00131.safetensors"],
+  "model.layers.28.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00093-of-00131.safetensors"],
+  "model.layers.28.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00113-of-00131.safetensors"],
+  "model.layers.29.input_layernorm.weight": [1, "BF16", [5120], "model-00099-of-00131.safetensors"],
+  "model.layers.29.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.29.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.29.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.29.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00103-of-00131.safetensors"],
+  "model.layers.29.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00097-of-00131.safetensors"],
+  "model.layers.29.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00092-of-00131.safetensors"],
+  "model.layers.29.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00092-of-00131.safetensors"],
+  "model.layers.29.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00092-of-00131.safetensors"],
+  "model.layers.29.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00100-of-00131.safetensors"],
+  "model.layers.29.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00098-of-00131.safetensors"],
+  "model.layers.29.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00104-of-00131.safetensors"],
+  "model.layers.29.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00104-of-00131.safetensors"],
+  "model.layers.29.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00099-of-00131.safetensors"],
+  "model.layers.29.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00098-of-00131.safetensors"],
+  "model.layers.29.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00091-of-00131.safetensors"],
+  "model.layers.29.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00104-of-00131.safetensors"],
+  "model.layers.29.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00103-of-00131.safetensors"],
+  "model.layers.29.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00097-of-00131.safetensors"],
+  "model.layers.29.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00113-of-00131.safetensors"],
+  "model.layers.29.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00106-of-00131.safetensors"],
+  "model.layers.29.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00101-of-00131.safetensors"],
+  "model.layers.29.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00093-of-00131.safetensors"],
+  "model.layers.29.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00110-of-00131.safetensors"],
+  "model.layers.3.input_layernorm.weight": [1, "BF16", [5120], "model-00076-of-00131.safetensors"],
+  "model.layers.3.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.3.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.3.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.3.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00082-of-00131.safetensors"],
+  "model.layers.3.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00047-of-00131.safetensors"],
+  "model.layers.3.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00117-of-00131.safetensors"],
+  "model.layers.3.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00116-of-00131.safetensors"],
+  "model.layers.3.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00116-of-00131.safetensors"],
+  "model.layers.3.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00076-of-00131.safetensors"],
+  "model.layers.3.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00062-of-00131.safetensors"],
+  "model.layers.3.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00083-of-00131.safetensors"],
+  "model.layers.3.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00081-of-00131.safetensors"],
+  "model.layers.3.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00019-of-00131.safetensors"],
+  "model.layers.3.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00094-of-00131.safetensors"],
+  "model.layers.3.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00088-of-00131.safetensors"],
+  "model.layers.3.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00081-of-00131.safetensors"],
+  "model.layers.3.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00034-of-00131.safetensors"],
+  "model.layers.3.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00103-of-00131.safetensors"],
+  "model.layers.30.input_layernorm.weight": [1, "BF16", [5120], "model-00100-of-00131.safetensors"],
+  "model.layers.30.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.30.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.30.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.30.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00103-of-00131.safetensors"],
+  "model.layers.30.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00097-of-00131.safetensors"],
+  "model.layers.30.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00094-of-00131.safetensors"],
+  "model.layers.30.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00093-of-00131.safetensors"],
+  "model.layers.30.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00093-of-00131.safetensors"],
+  "model.layers.30.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00100-of-00131.safetensors"],
+  "model.layers.30.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00099-of-00131.safetensors"],
+  "model.layers.30.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00104-of-00131.safetensors"],
+  "model.layers.30.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00101-of-00131.safetensors"],
+  "model.layers.30.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00091-of-00131.safetensors"],
+  "model.layers.30.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00111-of-00131.safetensors"],
+  "model.layers.30.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00108-of-00131.safetensors"],
+  "model.layers.30.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00101-of-00131.safetensors"],
+  "model.layers.30.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00093-of-00131.safetensors"],
+  "model.layers.30.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00113-of-00131.safetensors"],
+  "model.layers.31.input_layernorm.weight": [1, "BF16", [5120], "model-00100-of-00131.safetensors"],
+  "model.layers.31.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.31.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.31.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.31.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00103-of-00131.safetensors"],
+  "model.layers.31.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00097-of-00131.safetensors"],
+  "model.layers.31.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00096-of-00131.safetensors"],
+  "model.layers.31.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00096-of-00131.safetensors"],
+  "model.layers.31.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00096-of-00131.safetensors"],
+  "model.layers.31.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00100-of-00131.safetensors"],
+  "model.layers.31.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00099-of-00131.safetensors"],
+  "model.layers.31.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00104-of-00131.safetensors"],
+  "model.layers.31.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00101-of-00131.safetensors"],
+  "model.layers.31.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00092-of-00131.safetensors"],
+  "model.layers.31.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00111-of-00131.safetensors"],
+  "model.layers.31.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00108-of-00131.safetensors"],
+  "model.layers.31.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00102-of-00131.safetensors"],
+  "model.layers.31.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00094-of-00131.safetensors"],
+  "model.layers.31.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00114-of-00131.safetensors"],
+  "model.layers.32.input_layernorm.weight": [1, "BF16", [5120], "model-00126-of-00131.safetensors"],
+  "model.layers.32.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.32.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.32.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.32.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00129-of-00131.safetensors"],
+  "model.layers.32.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00124-of-00131.safetensors"],
+  "model.layers.32.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00120-of-00131.safetensors"],
+  "model.layers.32.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00119-of-00131.safetensors"],
+  "model.layers.32.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00119-of-00131.safetensors"],
+  "model.layers.32.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00126-of-00131.safetensors"],
+  "model.layers.32.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00125-of-00131.safetensors"],
+  "model.layers.32.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00131-of-00131.safetensors"],
+  "model.layers.32.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00128-of-00131.safetensors"],
+  "model.layers.32.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00117-of-00131.safetensors"],
+  "model.layers.32.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00006-of-00131.safetensors"],
+  "model.layers.32.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00004-of-00131.safetensors"],
+  "model.layers.32.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00128-of-00131.safetensors"],
+  "model.layers.32.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00119-of-00131.safetensors"],
+  "model.layers.32.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00008-of-00131.safetensors"],
+  "model.layers.33.input_layernorm.weight": [1, "BF16", [5120], "model-00126-of-00131.safetensors"],
+  "model.layers.33.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.33.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.33.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.33.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00130-of-00131.safetensors"],
+  "model.layers.33.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00124-of-00131.safetensors"],
+  "model.layers.33.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00122-of-00131.safetensors"],
+  "model.layers.33.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00121-of-00131.safetensors"],
+  "model.layers.33.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00121-of-00131.safetensors"],
+  "model.layers.33.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00126-of-00131.safetensors"],
+  "model.layers.33.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00124-of-00131.safetensors"],
+  "model.layers.33.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00131-of-00131.safetensors"],
+  "model.layers.33.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00130-of-00131.safetensors"],
+  "model.layers.33.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00125-of-00131.safetensors"],
+  "model.layers.33.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00124-of-00131.safetensors"],
+  "model.layers.33.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00116-of-00131.safetensors"],
+  "model.layers.33.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00131-of-00131.safetensors"],
+  "model.layers.33.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00130-of-00131.safetensors"],
+  "model.layers.33.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00123-of-00131.safetensors"],
+  "model.layers.33.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00009-of-00131.safetensors"],
+  "model.layers.33.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00002-of-00131.safetensors"],
+  "model.layers.33.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00128-of-00131.safetensors"],
+  "model.layers.33.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00121-of-00131.safetensors"],
+  "model.layers.33.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00005-of-00131.safetensors"],
+  "model.layers.34.input_layernorm.weight": [1, "BF16", [5120], "model-00126-of-00131.safetensors"],
+  "model.layers.34.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.34.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.34.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.34.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00130-of-00131.safetensors"],
+  "model.layers.34.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00124-of-00131.safetensors"],
+  "model.layers.34.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00117-of-00131.safetensors"],
+  "model.layers.34.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00117-of-00131.safetensors"],
+  "model.layers.34.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00117-of-00131.safetensors"],
+  "model.layers.34.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00126-of-00131.safetensors"],
+  "model.layers.34.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00125-of-00131.safetensors"],
+  "model.layers.34.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00131-of-00131.safetensors"],
+  "model.layers.34.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00128-of-00131.safetensors"],
+  "model.layers.34.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00116-of-00131.safetensors"],
+  "model.layers.34.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00007-of-00131.safetensors"],
+  "model.layers.34.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00004-of-00131.safetensors"],
+  "model.layers.34.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00128-of-00131.safetensors"],
+  "model.layers.34.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00121-of-00131.safetensors"],
+  "model.layers.34.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00009-of-00131.safetensors"],
+  "model.layers.35.input_layernorm.weight": [1, "BF16", [5120], "model-00126-of-00131.safetensors"],
+  "model.layers.35.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.35.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.35.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.35.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00130-of-00131.safetensors"],
+  "model.layers.35.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00124-of-00131.safetensors"],
+  "model.layers.35.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00117-of-00131.safetensors"],
+  "model.layers.35.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00116-of-00131.safetensors"],
+  "model.layers.35.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00116-of-00131.safetensors"],
+  "model.layers.35.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00127-of-00131.safetensors"],
+  "model.layers.35.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00125-of-00131.safetensors"],
+  "model.layers.35.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00001-of-00131.safetensors"],
+  "model.layers.35.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00128-of-00131.safetensors"],
+  "model.layers.35.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00118-of-00131.safetensors"],
+  "model.layers.35.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00007-of-00131.safetensors"],
+  "model.layers.35.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00004-of-00131.safetensors"],
+  "model.layers.35.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00129-of-00131.safetensors"],
+  "model.layers.35.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00122-of-00131.safetensors"],
+  "model.layers.35.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00009-of-00131.safetensors"],
+  "model.layers.36.input_layernorm.weight": [1, "BF16", [5120], "model-00019-of-00131.safetensors"],
+  "model.layers.36.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.36.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.36.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.36.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00022-of-00131.safetensors"],
+  "model.layers.36.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00017-of-00131.safetensors"],
+  "model.layers.36.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00011-of-00131.safetensors"],
+  "model.layers.36.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00010-of-00131.safetensors"],
+  "model.layers.36.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00010-of-00131.safetensors"],
+  "model.layers.36.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00019-of-00131.safetensors"],
+  "model.layers.36.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00018-of-00131.safetensors"],
+  "model.layers.36.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00023-of-00131.safetensors"],
+  "model.layers.36.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00021-of-00131.safetensors"],
+  "model.layers.36.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00010-of-00131.safetensors"],
+  "model.layers.36.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00031-of-00131.safetensors"],
+  "model.layers.36.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00028-of-00131.safetensors"],
+  "model.layers.36.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00021-of-00131.safetensors"],
+  "model.layers.36.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00015-of-00131.safetensors"],
+  "model.layers.36.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00033-of-00131.safetensors"],
+  "model.layers.37.input_layernorm.weight": [1, "BF16", [5120], "model-00019-of-00131.safetensors"],
+  "model.layers.37.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.37.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.37.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.37.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00022-of-00131.safetensors"],
+  "model.layers.37.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00017-of-00131.safetensors"],
+  "model.layers.37.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00011-of-00131.safetensors"],
+  "model.layers.37.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00010-of-00131.safetensors"],
+  "model.layers.37.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00010-of-00131.safetensors"],
+  "model.layers.37.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00019-of-00131.safetensors"],
+  "model.layers.37.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00017-of-00131.safetensors"],
+  "model.layers.37.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00023-of-00131.safetensors"],
+  "model.layers.37.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00023-of-00131.safetensors"],
+  "model.layers.37.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00018-of-00131.safetensors"],
+  "model.layers.37.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00017-of-00131.safetensors"],
+  "model.layers.37.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00013-of-00131.safetensors"],
+  "model.layers.37.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00023-of-00131.safetensors"],
+  "model.layers.37.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00022-of-00131.safetensors"],
+  "model.layers.37.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00016-of-00131.safetensors"],
+  "model.layers.37.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00034-of-00131.safetensors"],
+  "model.layers.37.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00026-of-00131.safetensors"],
+  "model.layers.37.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00021-of-00131.safetensors"],
+  "model.layers.37.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00015-of-00131.safetensors"],
+  "model.layers.37.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00029-of-00131.safetensors"],
+  "model.layers.38.input_layernorm.weight": [1, "BF16", [5120], "model-00020-of-00131.safetensors"],
+  "model.layers.38.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.38.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.38.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.38.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00022-of-00131.safetensors"],
+  "model.layers.38.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00017-of-00131.safetensors"],
+  "model.layers.38.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00012-of-00131.safetensors"],
+  "model.layers.38.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00012-of-00131.safetensors"],
+  "model.layers.38.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00012-of-00131.safetensors"],
+  "model.layers.38.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00020-of-00131.safetensors"],
+  "model.layers.38.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00018-of-00131.safetensors"],
+  "model.layers.38.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00023-of-00131.safetensors"],
+  "model.layers.38.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00021-of-00131.safetensors"],
+  "model.layers.38.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00011-of-00131.safetensors"],
+  "model.layers.38.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00031-of-00131.safetensors"],
+  "model.layers.38.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00028-of-00131.safetensors"],
+  "model.layers.38.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00021-of-00131.safetensors"],
+  "model.layers.38.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00016-of-00131.safetensors"],
+  "model.layers.38.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00011-of-00131.safetensors"],
+  "model.layers.39.input_layernorm.weight": [1, "BF16", [5120], "model-00020-of-00131.safetensors"],
+  "model.layers.39.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.39.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.39.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.39.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00022-of-00131.safetensors"],
+  "model.layers.39.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00017-of-00131.safetensors"],
+  "model.layers.39.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00016-of-00131.safetensors"],
+  "model.layers.39.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00015-of-00131.safetensors"],
+  "model.layers.39.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00015-of-00131.safetensors"],
+  "model.layers.39.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00020-of-00131.safetensors"],
+  "model.layers.39.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00018-of-00131.safetensors"],
+  "model.layers.39.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00023-of-00131.safetensors"],
+  "model.layers.39.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00021-of-00131.safetensors"],
+  "model.layers.39.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00012-of-00131.safetensors"],
+  "model.layers.39.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00031-of-00131.safetensors"],
+  "model.layers.39.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00028-of-00131.safetensors"],
+  "model.layers.39.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00021-of-00131.safetensors"],
+  "model.layers.39.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00016-of-00131.safetensors"],
+  "model.layers.39.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00011-of-00131.safetensors"],
+  "model.layers.4.input_layernorm.weight": [1, "BF16", [5120], "model-00100-of-00131.safetensors"],
+  "model.layers.4.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.4.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.4.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.4.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00103-of-00131.safetensors"],
+  "model.layers.4.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00097-of-00131.safetensors"],
+  "model.layers.4.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00091-of-00131.safetensors"],
+  "model.layers.4.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00091-of-00131.safetensors"],
+  "model.layers.4.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00091-of-00131.safetensors"],
+  "model.layers.4.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00100-of-00131.safetensors"],
+  "model.layers.4.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00099-of-00131.safetensors"],
+  "model.layers.4.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00104-of-00131.safetensors"],
+  "model.layers.4.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00102-of-00131.safetensors"],
+  "model.layers.4.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00091-of-00131.safetensors"],
+  "model.layers.4.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00112-of-00131.safetensors"],
+  "model.layers.4.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00109-of-00131.safetensors"],
+  "model.layers.4.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00102-of-00131.safetensors"],
+  "model.layers.4.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00095-of-00131.safetensors"],
+  "model.layers.4.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00115-of-00131.safetensors"],
+  "model.layers.40.input_layernorm.weight": [1, "BF16", [5120], "model-00044-of-00131.safetensors"],
+  "model.layers.40.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.40.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.40.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.40.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00047-of-00131.safetensors"],
+  "model.layers.40.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00042-of-00131.safetensors"],
+  "model.layers.40.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00035-of-00131.safetensors"],
+  "model.layers.40.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00035-of-00131.safetensors"],
+  "model.layers.40.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00035-of-00131.safetensors"],
+  "model.layers.40.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00044-of-00131.safetensors"],
+  "model.layers.40.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00043-of-00131.safetensors"],
+  "model.layers.40.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00048-of-00131.safetensors"],
+  "model.layers.40.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00046-of-00131.safetensors"],
+  "model.layers.40.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00035-of-00131.safetensors"],
+  "model.layers.40.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00056-of-00131.safetensors"],
+  "model.layers.40.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00053-of-00131.safetensors"],
+  "model.layers.40.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00046-of-00131.safetensors"],
+  "model.layers.40.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00040-of-00131.safetensors"],
+  "model.layers.40.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00059-of-00131.safetensors"],
+  "model.layers.41.input_layernorm.weight": [1, "BF16", [5120], "model-00044-of-00131.safetensors"],
+  "model.layers.41.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.41.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.41.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.41.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00047-of-00131.safetensors"],
+  "model.layers.41.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00042-of-00131.safetensors"],
+  "model.layers.41.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00035-of-00131.safetensors"],
+  "model.layers.41.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00035-of-00131.safetensors"],
+  "model.layers.41.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00035-of-00131.safetensors"],
+  "model.layers.41.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00044-of-00131.safetensors"],
+  "model.layers.41.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00042-of-00131.safetensors"],
+  "model.layers.41.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00048-of-00131.safetensors"],
+  "model.layers.41.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00048-of-00131.safetensors"],
+  "model.layers.41.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00043-of-00131.safetensors"],
+  "model.layers.41.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00042-of-00131.safetensors"],
+  "model.layers.41.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00038-of-00131.safetensors"],
+  "model.layers.41.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00048-of-00131.safetensors"],
+  "model.layers.41.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00047-of-00131.safetensors"],
+  "model.layers.41.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00041-of-00131.safetensors"],
+  "model.layers.41.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00059-of-00131.safetensors"],
+  "model.layers.41.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00052-of-00131.safetensors"],
+  "model.layers.41.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00046-of-00131.safetensors"],
+  "model.layers.41.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00040-of-00131.safetensors"],
+  "model.layers.41.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00054-of-00131.safetensors"],
+  "model.layers.42.input_layernorm.weight": [1, "BF16", [5120], "model-00045-of-00131.safetensors"],
+  "model.layers.42.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.42.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.42.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.42.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00048-of-00131.safetensors"],
+  "model.layers.42.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00042-of-00131.safetensors"],
+  "model.layers.42.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00038-of-00131.safetensors"],
+  "model.layers.42.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00037-of-00131.safetensors"],
+  "model.layers.42.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00037-of-00131.safetensors"],
+  "model.layers.42.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00045-of-00131.safetensors"],
+  "model.layers.42.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00043-of-00131.safetensors"],
+  "model.layers.42.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00048-of-00131.safetensors"],
+  "model.layers.42.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00046-of-00131.safetensors"],
+  "model.layers.42.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00036-of-00131.safetensors"],
+  "model.layers.42.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00057-of-00131.safetensors"],
+  "model.layers.42.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00053-of-00131.safetensors"],
+  "model.layers.42.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00046-of-00131.safetensors"],
+  "model.layers.42.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00041-of-00131.safetensors"],
+  "model.layers.42.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00036-of-00131.safetensors"],
+  "model.layers.43.input_layernorm.weight": [1, "BF16", [5120], "model-00045-of-00131.safetensors"],
+  "model.layers.43.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.43.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.43.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.43.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00047-of-00131.safetensors"],
+  "model.layers.43.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00042-of-00131.safetensors"],
+  "model.layers.43.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00041-of-00131.safetensors"],
+  "model.layers.43.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00040-of-00131.safetensors"],
+  "model.layers.43.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00040-of-00131.safetensors"],
+  "model.layers.43.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00045-of-00131.safetensors"],
+  "model.layers.43.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00043-of-00131.safetensors"],
+  "model.layers.43.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00049-of-00131.safetensors"],
+  "model.layers.43.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00046-of-00131.safetensors"],
+  "model.layers.43.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00037-of-00131.safetensors"],
+  "model.layers.43.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00057-of-00131.safetensors"],
+  "model.layers.43.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00054-of-00131.safetensors"],
+  "model.layers.43.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00046-of-00131.safetensors"],
+  "model.layers.43.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00041-of-00131.safetensors"],
+  "model.layers.43.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00036-of-00131.safetensors"],
+  "model.layers.44.input_layernorm.weight": [1, "BF16", [5120], "model-00064-of-00131.safetensors"],
+  "model.layers.44.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.44.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.44.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.44.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00064-of-00131.safetensors"],
+  "model.layers.44.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00062-of-00131.safetensors"],
+  "model.layers.44.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00080-of-00131.safetensors"],
+  "model.layers.44.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00080-of-00131.safetensors"],
+  "model.layers.44.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00080-of-00131.safetensors"],
+  "model.layers.44.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00062-of-00131.safetensors"],
+  "model.layers.44.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00063-of-00131.safetensors"],
+  "model.layers.44.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00063-of-00131.safetensors"],
+  "model.layers.44.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00064-of-00131.safetensors"],
+  "model.layers.44.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00060-of-00131.safetensors"],
+  "model.layers.44.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00075-of-00131.safetensors"],
+  "model.layers.44.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00074-of-00131.safetensors"],
+  "model.layers.44.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00063-of-00131.safetensors"],
+  "model.layers.44.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00061-of-00131.safetensors"],
+  "model.layers.44.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00077-of-00131.safetensors"],
+  "model.layers.45.input_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.45.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.45.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.45.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.45.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00062-of-00131.safetensors"],
+  "model.layers.45.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00062-of-00131.safetensors"],
+  "model.layers.45.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00061-of-00131.safetensors"],
+  "model.layers.45.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00061-of-00131.safetensors"],
+  "model.layers.45.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00061-of-00131.safetensors"],
+  "model.layers.45.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.45.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00063-of-00131.safetensors"],
+  "model.layers.45.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00064-of-00131.safetensors"],
+  "model.layers.45.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00063-of-00131.safetensors"],
+  "model.layers.45.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00063-of-00131.safetensors"],
+  "model.layers.45.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00063-of-00131.safetensors"],
+  "model.layers.45.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00077-of-00131.safetensors"],
+  "model.layers.45.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00063-of-00131.safetensors"],
+  "model.layers.45.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00063-of-00131.safetensors"],
+  "model.layers.45.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00064-of-00131.safetensors"],
+  "model.layers.45.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00077-of-00131.safetensors"],
+  "model.layers.45.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00070-of-00131.safetensors"],
+  "model.layers.45.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00064-of-00131.safetensors"],
+  "model.layers.45.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00062-of-00131.safetensors"],
+  "model.layers.45.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00075-of-00131.safetensors"],
+  "model.layers.46.eh_proj.weight": [1, "BF16", [5120, 10240], "model-00072-of-00131.safetensors"],
+  "model.layers.46.enorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.46.hnorm.weight": [1, "BF16", [5120], "model-00062-of-00131.safetensors"],
+  "model.layers.46.input_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.46.mlp.down_proj.weight": [1, "BF16", [5120, 13824], "model-00072-of-00131.safetensors"],
+  "model.layers.46.mlp.gate_proj.weight": [1, "BF16", [13824, 5120], "model-00071-of-00131.safetensors"],
+  "model.layers.46.mlp.up_proj.weight": [1, "BF16", [13824, 5120], "model-00071-of-00131.safetensors"],
+  "model.layers.46.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00063-of-00131.safetensors"],
+  "model.layers.46.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00064-of-00131.safetensors"],
+  "model.layers.46.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00064-of-00131.safetensors"],
+  "model.layers.46.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00062-of-00131.safetensors"],
+  "model.layers.46.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00063-of-00131.safetensors"],
+  "model.layers.46.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00076-of-00131.safetensors"],
+  "model.layers.46.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00074-of-00131.safetensors"],
+  "model.layers.46.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00063-of-00131.safetensors"],
+  "model.layers.46.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00063-of-00131.safetensors"],
+  "model.layers.46.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00077-of-00131.safetensors"],
+  "model.layers.46.shared_head.norm.weight": [1, "BF16", [5120], "model-00064-of-00131.safetensors"],
+  "model.layers.5.input_layernorm.weight": [1, "BF16", [5120], "model-00100-of-00131.safetensors"],
+  "model.layers.5.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.5.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.5.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.5.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00103-of-00131.safetensors"],
+  "model.layers.5.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00098-of-00131.safetensors"],
+  "model.layers.5.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00091-of-00131.safetensors"],
+  "model.layers.5.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00091-of-00131.safetensors"],
+  "model.layers.5.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00091-of-00131.safetensors"],
+  "model.layers.5.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00100-of-00131.safetensors"],
+  "model.layers.5.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00098-of-00131.safetensors"],
+  "model.layers.5.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00104-of-00131.safetensors"],
+  "model.layers.5.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00104-of-00131.safetensors"],
+  "model.layers.5.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00099-of-00131.safetensors"],
+  "model.layers.5.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00098-of-00131.safetensors"],
+  "model.layers.5.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00094-of-00131.safetensors"],
+  "model.layers.5.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00104-of-00131.safetensors"],
+  "model.layers.5.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00103-of-00131.safetensors"],
+  "model.layers.5.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00097-of-00131.safetensors"],
+  "model.layers.5.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00115-of-00131.safetensors"],
+  "model.layers.5.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00107-of-00131.safetensors"],
+  "model.layers.5.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00102-of-00131.safetensors"],
+  "model.layers.5.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00096-of-00131.safetensors"],
+  "model.layers.5.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00110-of-00131.safetensors"],
+  "model.layers.6.input_layernorm.weight": [1, "BF16", [5120], "model-00101-of-00131.safetensors"],
+  "model.layers.6.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.6.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.6.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.6.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00103-of-00131.safetensors"],
+  "model.layers.6.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00098-of-00131.safetensors"],
+  "model.layers.6.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00093-of-00131.safetensors"],
+  "model.layers.6.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00092-of-00131.safetensors"],
+  "model.layers.6.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00092-of-00131.safetensors"],
+  "model.layers.6.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00101-of-00131.safetensors"],
+  "model.layers.6.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00099-of-00131.safetensors"],
+  "model.layers.6.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00104-of-00131.safetensors"],
+  "model.layers.6.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00102-of-00131.safetensors"],
+  "model.layers.6.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00092-of-00131.safetensors"],
+  "model.layers.6.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00112-of-00131.safetensors"],
+  "model.layers.6.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00109-of-00131.safetensors"],
+  "model.layers.6.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00102-of-00131.safetensors"],
+  "model.layers.6.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00097-of-00131.safetensors"],
+  "model.layers.6.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00092-of-00131.safetensors"],
+  "model.layers.7.input_layernorm.weight": [1, "BF16", [5120], "model-00101-of-00131.safetensors"],
+  "model.layers.7.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.7.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.7.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.7.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00103-of-00131.safetensors"],
+  "model.layers.7.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00098-of-00131.safetensors"],
+  "model.layers.7.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00096-of-00131.safetensors"],
+  "model.layers.7.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00095-of-00131.safetensors"],
+  "model.layers.7.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00095-of-00131.safetensors"],
+  "model.layers.7.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00101-of-00131.safetensors"],
+  "model.layers.7.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00099-of-00131.safetensors"],
+  "model.layers.7.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00104-of-00131.safetensors"],
+  "model.layers.7.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00102-of-00131.safetensors"],
+  "model.layers.7.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00093-of-00131.safetensors"],
+  "model.layers.7.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00112-of-00131.safetensors"],
+  "model.layers.7.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00110-of-00131.safetensors"],
+  "model.layers.7.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00102-of-00131.safetensors"],
+  "model.layers.7.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00097-of-00131.safetensors"],
+  "model.layers.7.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00092-of-00131.safetensors"],
+  "model.layers.8.input_layernorm.weight": [1, "BF16", [5120], "model-00127-of-00131.safetensors"],
+  "model.layers.8.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.8.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.8.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.8.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00130-of-00131.safetensors"],
+  "model.layers.8.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00124-of-00131.safetensors"],
+  "model.layers.8.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00119-of-00131.safetensors"],
+  "model.layers.8.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00118-of-00131.safetensors"],
+  "model.layers.8.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00118-of-00131.safetensors"],
+  "model.layers.8.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00127-of-00131.safetensors"],
+  "model.layers.8.self_attn.g_proj.weight": [1, "BF16", [64, 5120], "model-00125-of-00131.safetensors"],
+  "model.layers.8.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00001-of-00131.safetensors"],
+  "model.layers.8.self_attn.kv_a_layernorm.weight": [1, "BF16", [1024], "model-00128-of-00131.safetensors"],
+  "model.layers.8.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [1088, 5120], "model-00118-of-00131.safetensors"],
+  "model.layers.8.self_attn.kv_b_proj.weight": [1, "BF16", [20480, 1024], "model-00007-of-00131.safetensors"],
+  "model.layers.8.self_attn.o_proj.weight": [1, "BF16", [5120, 8192], "model-00005-of-00131.safetensors"],
+  "model.layers.8.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00129-of-00131.safetensors"],
+  "model.layers.8.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00123-of-00131.safetensors"],
+  "model.layers.8.self_attn.q_b_proj.weight": [1, "BF16", [16384, 1024], "model-00117-of-00131.safetensors"],
+  "model.layers.9.input_layernorm.weight": [1, "BF16", [5120], "model-00127-of-00131.safetensors"],
+  "model.layers.9.mlp.experts.{E}.down_proj.weight": [256, "BF16", [5120, 1536], "*"],
+  "model.layers.9.mlp.experts.{E}.gate_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.9.mlp.experts.{E}.up_proj.weight": [256, "BF16", [1536, 5120], "*"],
+  "model.layers.9.mlp.gate.e_score_correction_bias": [1, "F32", [256], "model-00130-of-00131.safetensors"],
+  "model.layers.9.mlp.gate.weight": [1, "BF16", [256, 5120], "model-00124-of-00131.safetensors"],
+  "model.layers.9.mlp.shared_experts.down_proj.weight": [1, "BF16", [5120, 1536], "model-00122-of-00131.safetensors"],
+  "model.layers.9.mlp.shared_experts.gate_proj.weight": [1, "BF16", [1536, 5120], "model-00121-of-00131.safetensors"],
+  "model.layers.9.mlp.shared_experts.up_proj.weight": [1, "BF16", [1536, 5120], "model-00121-of-00131.safetensors"],
+  "model.layers.9.post_attention_layernorm.weight": [1, "BF16", [5120], "model-00127-of-00131.safetensors"],
+  "model.layers.9.self_attn.g_proj.weight": [1, "BF16", [128, 5120], "model-00124-of-00131.safetensors"],
+  "model.layers.9.self_attn.indexer.k_norm.bias": [1, "BF16", [128], "model-00130-of-00131.safetensors"],
+  "model.layers.9.self_attn.indexer.k_norm.weight": [1, "BF16", [128], "model-00131-of-00131.safetensors"],
+  "model.layers.9.self_attn.indexer.weights_proj.weight": [1, "BF16", [64, 5120], "model-00125-of-00131.safetensors"],
+  "model.layers.9.self_attn.indexer.wk.weight": [1, "BF16", [128, 5120], "model-00124-of-00131.safetensors"],
+  "model.layers.9.self_attn.indexer.wq_b.weight": [1, "BF16", [8192, 1024], "model-00120-of-00131.safetensors"],
+  "model.layers.9.self_attn.k_rope_only_layernorm.weight": [1, "BF16", [64], "model-00001-of-00131.safetensors"],
+  "model.layers.9.self_attn.kv_a_layernorm.weight": [1, "BF16", [512], "model-00130-of-00131.safetensors"],
+  "model.layers.9.self_attn.kv_a_proj_with_mqa.weight": [1, "BF16", [576, 5120], "model-00123-of-00131.safetensors"],
+  "model.layers.9.self_attn.kv_b_proj.weight": [1, "BF16", [32768, 512], "model-00118-of-00131.safetensors"],
+  "model.layers.9.self_attn.o_proj.weight": [1, "BF16", [5120, 16384], "model-00003-of-00131.safetensors"],
+  "model.layers.9.self_attn.q_a_layernorm.weight": [1, "BF16", [1024], "model-00129-of-00131.safetensors"],
+  "model.layers.9.self_attn.q_a_proj.weight": [1, "BF16", [1024, 5120], "model-00123-of-00131.safetensors"],
+  "model.layers.9.self_attn.q_b_proj.weight": [1, "BF16", [24576, 1024], "model-00005-of-00131.safetensors"],
+  "model.mtp.embed_tokens.weight": [1, "BF16", [152064, 5120], "model-00067-of-00131.safetensors"],
+  "model.norm.weight": [1, "BF16", [5120], "model-00062-of-00131.safetensors"],
+  "vision_encoder.adapter.ln_q.bias": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.adapter.ln_q.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.adapter.mlp.0.bias": [1, "BF16", [6144], "model-vision.safetensors"],
+  "vision_encoder.adapter.mlp.0.weight": [1, "BF16", [6144, 6144], "model-vision.safetensors"],
+  "vision_encoder.adapter.mlp.2.bias": [1, "BF16", [5120], "model-vision.safetensors"],
+  "vision_encoder.adapter.mlp.2.weight": [1, "BF16", [5120, 6144], "model-vision.safetensors"],
+  "vision_encoder.blocks.0.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.0.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.0.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.0.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.0.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.0.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.0.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.0.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.0.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.1.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.1.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.1.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.1.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.1.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.1.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.1.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.1.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.1.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.10.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.10.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.10.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.10.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.10.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.10.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.10.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.10.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.10.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.11.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.11.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.11.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.11.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.11.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.11.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.11.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.11.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.11.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.12.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.12.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.12.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.12.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.12.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.12.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.12.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.12.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.12.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.13.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.13.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.13.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.13.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.13.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.13.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.13.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.13.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.13.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.14.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.14.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.14.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.14.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.14.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.14.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.14.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.14.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.14.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.15.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.15.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.15.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.15.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.15.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.15.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.15.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.15.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.15.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.16.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.16.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.16.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.16.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.16.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.16.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.16.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.16.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.16.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.17.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.17.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.17.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.17.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.17.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.17.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.17.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.17.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.17.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.18.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.18.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.18.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.18.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.18.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.18.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.18.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.18.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.18.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.19.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.19.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.19.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.19.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.19.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.19.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.19.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.19.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.19.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.2.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.2.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.2.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.2.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.2.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.2.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.2.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.2.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.2.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.20.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.20.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.20.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.20.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.20.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.20.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.20.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.20.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.20.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.21.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.21.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.21.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.21.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.21.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.21.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.21.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.21.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.21.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.22.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.22.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.22.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.22.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.22.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.22.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.22.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.22.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.22.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.23.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.23.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.23.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.23.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.23.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.23.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.23.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.23.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.23.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.24.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.24.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.24.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.24.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.24.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.24.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.24.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.24.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.24.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.mlp.experts.{E}.fc1.weight": [4, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.mlp.experts.{E}.fc2.weight": [4, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.mlp.experts.{E}.fc3.weight": [4, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.mlp.gate_weight": [1, "BF16", [4, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.mlp.router_bias": [1, "F32", [4], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.25.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.mlp.experts.{E}.fc1.weight": [8, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.mlp.experts.{E}.fc2.weight": [8, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.mlp.experts.{E}.fc3.weight": [8, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.mlp.gate_weight": [1, "BF16", [8, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.mlp.router_bias": [1, "F32", [8], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.26.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.mlp.experts.{E}.fc1.weight": [12, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.mlp.experts.{E}.fc2.weight": [12, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.mlp.experts.{E}.fc3.weight": [12, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.mlp.gate_weight": [1, "BF16", [12, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.mlp.router_bias": [1, "F32", [12], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.27.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.mlp.experts.{E}.fc1.weight": [16, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.mlp.experts.{E}.fc2.weight": [16, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.mlp.experts.{E}.fc3.weight": [16, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.mlp.gate_weight": [1, "BF16", [16, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.mlp.router_bias": [1, "F32", [16], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.28.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.mlp.experts.{E}.fc1.weight": [20, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.mlp.experts.{E}.fc2.weight": [20, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.mlp.experts.{E}.fc3.weight": [20, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.mlp.gate_weight": [1, "BF16", [20, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.mlp.router_bias": [1, "F32", [20], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.29.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.3.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.3.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.3.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.3.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.3.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.3.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.3.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.3.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.3.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.mlp.experts.{E}.fc1.weight": [24, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.mlp.experts.{E}.fc2.weight": [24, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.mlp.experts.{E}.fc3.weight": [24, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.mlp.gate_weight": [1, "BF16", [24, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.mlp.router_bias": [1, "F32", [24], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.30.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.mlp.experts.{E}.fc1.weight": [28, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.mlp.experts.{E}.fc2.weight": [28, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.mlp.experts.{E}.fc3.weight": [28, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.mlp.gate_weight": [1, "BF16", [28, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.mlp.router_bias": [1, "F32", [28], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.31.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.mlp.experts.{E}.fc1.weight": [32, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.mlp.experts.{E}.fc2.weight": [32, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.mlp.experts.{E}.fc3.weight": [32, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.mlp.gate_weight": [1, "BF16", [32, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.mlp.router_bias": [1, "F32", [32], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.32.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.mlp.experts.{E}.fc1.weight": [36, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.mlp.experts.{E}.fc2.weight": [36, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.mlp.experts.{E}.fc3.weight": [36, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.mlp.gate_weight": [1, "BF16", [36, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.mlp.router_bias": [1, "F32", [36], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.33.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.mlp.experts.{E}.fc1.weight": [40, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.mlp.experts.{E}.fc2.weight": [40, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.mlp.experts.{E}.fc3.weight": [40, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.mlp.gate_weight": [1, "BF16", [40, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.mlp.router_bias": [1, "F32", [40], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.34.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.mlp.experts.{E}.fc1.weight": [44, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.mlp.experts.{E}.fc2.weight": [44, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.mlp.experts.{E}.fc3.weight": [44, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.mlp.gate_weight": [1, "BF16", [44, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.mlp.router_bias": [1, "F32", [44], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.35.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.mlp.experts.{E}.fc1.weight": [48, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.mlp.experts.{E}.fc2.weight": [48, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.mlp.experts.{E}.fc3.weight": [48, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.mlp.gate_weight": [1, "BF16", [48, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.mlp.router_bias": [1, "F32", [48], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.36.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.mlp.experts.{E}.fc1.weight": [52, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.mlp.experts.{E}.fc2.weight": [52, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.mlp.experts.{E}.fc3.weight": [52, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.mlp.gate_weight": [1, "BF16", [52, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.mlp.router_bias": [1, "F32", [52], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.37.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.mlp.experts.{E}.fc1.weight": [56, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.mlp.experts.{E}.fc2.weight": [56, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.mlp.experts.{E}.fc3.weight": [56, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.mlp.gate_weight": [1, "BF16", [56, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.mlp.router_bias": [1, "F32", [56], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.38.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.mlp.experts.{E}.fc1.weight": [60, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.mlp.experts.{E}.fc2.weight": [60, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.mlp.experts.{E}.fc3.weight": [60, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.mlp.gate_weight": [1, "BF16", [60, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.mlp.router_bias": [1, "F32", [60], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.39.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.4.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.4.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.4.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.4.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.4.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.4.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.4.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.4.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.4.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.mlp.experts.{E}.fc1.weight": [64, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.mlp.experts.{E}.fc2.weight": [64, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.mlp.experts.{E}.fc3.weight": [64, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.mlp.gate_weight": [1, "BF16", [64, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.mlp.router_bias": [1, "F32", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.40.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.mlp.experts.{E}.fc1.weight": [64, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.mlp.experts.{E}.fc2.weight": [64, "BF16", [1536, 2112], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.mlp.experts.{E}.fc3.weight": [64, "BF16", [2112, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.mlp.gate_weight": [1, "BF16", [64, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.mlp.router_bias": [1, "F32", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.41.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.5.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.5.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.5.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.5.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.5.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.5.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.5.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.5.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.5.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.6.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.6.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.6.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.6.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.6.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.6.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.6.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.6.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.6.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.7.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.7.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.7.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.7.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.7.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.7.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.7.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.7.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.7.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.8.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.8.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.8.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.8.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.8.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.8.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.8.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.8.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.8.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.9.attn.k_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.9.attn.proj.weight": [1, "BF16", [1536, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.9.attn.q_norm.weight": [1, "BF16", [64], "model-vision.safetensors"],
+  "vision_encoder.blocks.9.attn.qkv.weight": [1, "BF16", [4608, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.9.mlp.fc1.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.9.mlp.fc2.weight": [1, "BF16", [1536, 4224], "model-vision.safetensors"],
+  "vision_encoder.blocks.9.mlp.fc3.weight": [1, "BF16", [4224, 1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.9.norm_1.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.blocks.9.norm_2.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.patch_embed.norm.weight": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.patch_embed.proj.bias": [1, "BF16", [1536], "model-vision.safetensors"],
+  "vision_encoder.patch_embed.proj.weight": [1, "BF16", [1536, 3, 14, 14], "model-vision.safetensors"],
+  "vision_encoder.post_trunk_norm.weight": [1, "BF16", [1536], "model-vision.safetensors"]
+ }
+}
diff --git a/tests/vllm/models/test_dots3_note_scaffold.cpp b/tests/vllm/models/test_dots3_note_scaffold.cpp
index 97f5a51f0..df9bbff78 100644
--- a/tests/vllm/models/test_dots3_note_scaffold.cpp
+++ b/tests/vllm/models/test_dots3_note_scaffold.cpp
@@ -1,7 +1,8 @@
-// dots3-note (`Dots3NoteForCausalLM`) W1 STRUCTURAL gate — issue #699, spec
-// `.agents/specs/dots3-note.md` §4 / §7 W1.
+// dots3-note (`Dots3NoteForCausalLM`) W1+W2 STRUCTURAL gate — issue #699, spec
+// `.agents/specs/dots3-note.md` §4 (traps, evidence §4.1) and §4.4 (the whole
+// weight map, evidence and mutation table).
 //
-// WHY THIS FILE IS THE WHOLE GATE FOR W1. Spec §6.4 records the decision that
+// WHY THIS FILE IS THE WHOLE GATE FOR THE ROW SO FAR. Spec §6.4 records the decision that
 // this row has NO oracle: the checkpoint is ~576 GB bf16 / ~290 GB fp8 and the
 // biggest host this project owns is 122 GiB, so vLLM cannot run this model
 // anywhere we can reach. There is therefore no token gate downstream of these
@@ -17,24 +18,37 @@
 //   (2) the REAL released config.json parses (committed byte-for-byte as a
 //       fixture) — the 46-entry 13-full/33-sliding schedule, BOTH MLA
 //       geometries, the MoE dims, and ALL SIX §4 TRAPS;
-//   (3) the on-disk NAME MAP is faithful: every one of the 1614 tensors in the
-//       committed slice of the released `model.safetensors.index.json` is
-//       CLAIMED by a named consumer, and nothing is enumerated that the
-//       checkpoint does not ship;
+//   (3) the on-disk NAME MAP is faithful over the WHOLE released
+//       `model.safetensors.index.json` — all 38006 tensors, bucketed
+//       35381 language / 2195 vision / 430 audio and asserted BY NUMBER, with
+//       nothing unaccounted and nothing enumerated that the checkpoint does not
+//       ship. W1's 1614-tensor slice is kept beside it as a cross-check;
+//   (3a) W2 only: the two tower files are NAMED W6/W7 DEFERRAL RECORDS — prefix,
+//       the one file each ships in, the owing brick — not integer counters, and
+//       the load refusal prints the table. Folding them into `language` is the
+//       mutation the #1805 review found passing, and it is what 38006-scale
+//       accounting would otherwise hide;
+//   (3b) W2 only: the backbone has exactly FOUR distinct layer shapes, the
+//       full/sliding split read off the shipped `indexer.wk` tensors equals
+//       `config.layer_types`, and the checkpoint's memory format is 37944 BF16
+//       plus 62 F32 in TWO families;
 //   (4) the geometry the params imply is the geometry the released safetensors
-//       HEADERS carry (shapes, and the one F32 tensor in a BF16 tower);
+//       HEADERS carry (shapes, and the F32 tensors in a BF16 tower);
 //   (5) the padded MLA KV row is the SLIDING row, not the full one;
 //   (6) load succeeds only with 100% accounting, the forward REFUSES BY NAME
 //       through the REAL model the factory returns, and GGUF refuses by name.
 //
-// The fixtures are the released `config.json` verbatim and a HEADERS-ONLY
-// projection of the shard index at revision
-// `1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b`; see `index.json`'s `_provenance`.
+// The fixtures are the released `config.json` verbatim, a HEADERS-ONLY
+// projection of the WHOLE shard index (`index_full.json`, all 38006 tensors
+// joined to all 133 safetensors headers) and W1's four-layer slice
+// (`index.json`), both at revision
+// `1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b`; see each one's `_provenance`.
 // No tensor byte of the checkpoint was ever read.
 //
-// Upstream anchors are at vLLM `origin/main` =
-// `c205726108df54bb6fbf15b19e725a4a3add2b18`. `dots3_note` does NOT exist at
-// our parity pin `555967922`.
+// Upstream anchors: W1's are at vLLM `origin/main` =
+// `c205726108df54bb6fbf15b19e725a4a3add2b18`, W2's at `185cada36b`, and each
+// citation says which. `dots3_note` does NOT exist at our parity pin
+// `555967922`.
 #include "vllm/model_executor/models/dots3_note.h"
 
 #include <doctest/doctest.h>
@@ -42,13 +56,16 @@
 #include <nlohmann/json.hpp>
 
 #include <algorithm>
+#include <chrono>
 #include <cmath>
 #include <cstdint>
 #include <cstdio>
 #include <filesystem>
 #include <fstream>
 #include <map>
+#include <random>
 #include <set>
+#include <sstream>
 #include <string>
 #include <system_error>
 #include <utility>
@@ -98,12 +115,44 @@ nlohmann::json ReadJson(const std::string& path) {
 
 // A throwaway config.json on disk, so every case drives the SAME LoadHfConfig
 // the engine uses rather than hand-building an HfConfig.
+// A temp directory unique to THIS PROCESS, not merely to this object.
+//
+// WHY A BARE COUNTER IS A DEFECT AND NOT A STYLE POINT (#1860, review F1 on
+// #1847). `static int counter` is per-process state, so two concurrent runs of
+// this same binary land on the SAME path: each constructor rewrites a file the
+// other has already mmapped through `SafetensorsFile::Open`, and each
+// destructor `remove_all()`s the other's directory. The reader takes SIGBUS and
+// exits 135, and doctest's stdout is block-buffered, so the summary dies with
+// the process and the run reads as NO RESULT rather than as a failure. A fresh
+// reviewer reproduced it 3/3 with two concurrent runs at 47 GB free and 61 GB
+// RAM available, both processes sharing one `/tmp/dots3_note_cfg_8` — so it is
+// not the disk pressure this row's spec first blamed. It matters more here than
+// elsewhere: spec §6.4 records that this row has NO oracle, so this file is the
+// only instrument it has, and a second agent building on the same box is
+// routine.
+//
+// No `getpid()`: this file is compiled by MSVC too, where that spelling is
+// `_getpid()` behind `<process.h>`. A random word plus the steady clock needs no
+// platform header and is unique for the same reason a pid is.
+// The house-wide sweep over the other files with this shape is #1860, not this
+// row.
+std::filesystem::path UniqueTempDir(const std::string& stem) {
+  static const std::string kToken = [] {
+    std::random_device rd;
+    std::ostringstream os;
+    os << std::hex << rd() << "_"
+       << std::chrono::steady_clock::now().time_since_epoch().count();
+    return os.str();
+  }();
+  static int counter = 0;
+  return std::filesystem::temp_directory_path() /
+         (stem + kToken + "_" + std::to_string(counter++));
+}
+
 class TempConfig {
  public:
   explicit TempConfig(const nlohmann::json& doc) {
-    static int counter = 0;
-    dir_ = std::filesystem::temp_directory_path() /
-           ("dots3_note_cfg_" + std::to_string(counter++));
+    dir_ = UniqueTempDir("dots3_note_cfg_");
     std::filesystem::create_directories(dir_);
     std::ofstream(dir_ / "config.json") << doc.dump();
   }
@@ -893,6 +942,501 @@ TEST_CASE("dots3-note enumeration: the shapes it implies are the ones on disk")
         std::vector<int64_t>{h, 2 * h});
 }
 
+
+// ─────────────────────────────────────────────────────────────────────────────
+// W2 — the WHOLE released `model.safetensors.index.json`, not W1's slice.
+//
+// W1 gated 1614 tensors over four layers. Everything below reads the complete
+// index of `dots-studio/dots3-note-prev` @ `1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b`
+// joined to the safetensors HEADER of all 133 files it names: 38006 tensors,
+// the 42 backbone layers the slice never saw, and the 2625 tensors of the two
+// tower files. HEADERS ONLY — 4770592 bytes over 266 HTTP Range requests, and
+// not one tensor byte; the checkpoint is ~576 GB and was never downloaded.
+//
+// WHY THE BUCKET COUNTS ARE ASSERTED BY NUMBER. "Nothing was left over" is also
+// true of a classifier that counts all 2625 tower tensors as LANGUAGE, and that
+// is the mutation the #1805 review found passing (M15). At W1's scale it was a
+// latent hole; at W2's it is the whole gate, because 100% accounting would then
+// cover 2625 weights nobody loads. So every case below states three numbers.
+// ─────────────────────────────────────────────────────────────────────────────
+
+namespace {
+
+// One entry of the full-index fixture: `[count, dtype, shape, file]`, with the
+// ROUTED-EXPERT index (and only that index) collapsed to `{E}`.
+struct FullEntry {
+  std::string dtype;
+  std::vector<int64_t> shape;
+  std::string file;  // "*" when an {E} family spans more than one shard
+};
+
+nlohmann::json FullIndexDoc() { return ReadJson(FixtureDir() + "/index_full.json"); }
+
+// Expand the fixture back into one entry per REAL tensor name.
+std::map<std::string, FullEntry> ExpandFullIndex(const nlohmann::json& index) {
+  std::map<std::string, FullEntry> out;
+  for (const auto& [pattern, meta] : index.at("tensors").items()) {
+    REQUIRE(meta.is_array());
+    REQUIRE(meta.size() == 4);
+    const auto count = meta.at(0).get<int64_t>();
+    FullEntry e{meta.at(1).get<std::string>(),
+                meta.at(2).get<std::vector<int64_t>>(),
+                meta.at(3).get<std::string>()};
+    const std::string marker = "{E}";
+    const size_t at = pattern.find(marker);
+    if (at == std::string::npos) {
+      REQUIRE_MESSAGE(count == 1, "non-expert family with count != 1: " << pattern);
+      out.emplace(pattern, e);
+      continue;
+    }
+    for (int64_t i = 0; i < count; ++i) {
+      std::string name = pattern;
+      name.replace(at, marker.size(), std::to_string(i));
+      out.emplace(std::move(name), e);
+    }
+  }
+  return out;
+}
+
+bool HasPrefix(const std::string& s, const std::string& prefix) {
+  return s.rfind(prefix, 0) == 0;
+}
+
+bool IsTowerName(const std::string& s) {
+  return HasPrefix(s, "vision_encoder.") || HasPrefix(s, "audio_encoder.");
+}
+
+}  // namespace
+
+TEST_CASE("dots3-note W2: all 38006 tensors of the WHOLE released index are claimed") {
+  const Dots3NoteParams p = FixtureParams();
+  const nlohmann::json index = FullIndexDoc();
+
+  // The fixture states what it is, and the numbers are the released
+  // checkpoint's own rather than this port's arithmetic.
+  CHECK(index.at("revision").get<std::string>() ==
+        "1e1e7b0cd37a3a48a6c8d7fa55d5f9d14377006b");
+  CHECK(index.at("checkpoint_total_tensors").get<int64_t>() == 38006);
+  CHECK(index.at("checkpoint_total_size_bytes").get<int64_t>() == 576886825984LL);
+  CHECK(index.at("shard_file_count").get<int64_t>() == 133);
+  CHECK(index.at("bucket_totals").at("language").get<int64_t>() == 35381);
+  CHECK(index.at("bucket_totals").at("vision").get<int64_t>() == 2195);
+  CHECK(index.at("bucket_totals").at("audio").get<int64_t>() == 430);
+
+  const auto released = ExpandFullIndex(index);
+  REQUIRE(released.size() == 38006);
+
+  // The language name map, over EVERY backbone layer, the root and the nextn
+  // tail — the production enumeration, not a slice.
+  const std::vector<Dots3NoteTensor> enumerated = EnumerateDots3NoteTensors(p);
+  std::set<std::string> seen;
+  std::vector<std::string> invented;
+  std::vector<std::string> nameless;
+  std::vector<std::string> twice;
+  for (const Dots3NoteTensor& t : enumerated) {
+    if (t.consumer.empty()) nameless.push_back(t.name);
+    if (!seen.insert(t.name).second) twice.push_back(t.name);
+    if (released.count(t.name) == 0) invented.push_back(t.name);
+  }
+  CHECK_MESSAGE(nameless.empty(), "enumerated with no named consumer, first: "
+                                      << (nameless.empty() ? std::string("-")
+                                                           : nameless.front()));
+  CHECK_MESSAGE(twice.empty(), "enumerated twice, first: "
+                                   << (twice.empty() ? std::string("-")
+                                                     : twice.front()));
+  CHECK_MESSAGE(invented.empty(),
+                "enumerated tensors the checkpoint does not ship, first: "
+                    << (invented.empty() ? std::string("-") : invented.front())
+                    << " (" << invented.size() << " total)");
+  CHECK(enumerated.size() == 35381);
+  CHECK(seen.size() == 35381);
+
+  // Nothing on disk is left without an owner: either the language map claims
+  // it, or a REGISTERED deferral does.
+  std::vector<std::string> orphan;
+  for (const auto& [name, meta] : released) {
+    (void)meta;
+    if (seen.count(name) == 0 && vllm::Dots3NoteDeferralFor(name) == nullptr) {
+      orphan.push_back(name);
+    }
+  }
+  CHECK_MESSAGE(orphan.empty(),
+                "checkpoint tensors no consumer and no deferral claims, first: "
+                    << (orphan.empty() ? std::string("-") : orphan.front())
+                    << " (" << orphan.size() << " total)");
+
+  // The classifier the LOADER runs, over the whole index, with the three
+  // buckets stated by number. If the towers folded into `language` this reads
+  // 38006 / 0 / 0 and every "nothing left over" assertion below still passes.
+  std::vector<std::string> names;
+  names.reserve(released.size());
+  for (const auto& [name, meta] : released) {
+    (void)meta;
+    names.push_back(name);
+  }
+  std::vector<int64_t> backbone;
+  for (int64_t l = 0; l < p.num_hidden_layers; ++l) backbone.push_back(l);
+  const Dots3NoteAccounting acc = AccountDots3NoteTensors(p, names, backbone);
+  CHECK(acc.language == 35381);
+  CHECK(acc.vision == 2195);
+  CHECK(acc.audio == 430);
+  CHECK(acc.total() == 38006);
+  CHECK(acc.unaccounted.empty());
+  CHECK(acc.missing.empty());
+  CHECK(acc.duplicated.empty());
+}
+
+TEST_CASE("dots3-note W2: the two tower files are NAMED W6/W7 deferrals, all 2625") {
+  // A deferral is a record that names the brick, the file and what the tower
+  // is. A bare counter is not one: it cannot tell a reader whether 2625 weights
+  // are deferred on purpose or lost.
+  const auto& towers = vllm::Dots3NoteDeferredTowers();
+  REQUIRE(towers.size() == 2);
+  CHECK(std::string(towers[0].prefix) == "vision_encoder.");
+  CHECK(std::string(towers[0].file) == "model-vision.safetensors");
+  CHECK(std::string(towers[0].brick) == "W6");
+  CHECK(std::string(towers[1].prefix) == "audio_encoder.");
+  CHECK(std::string(towers[1].file) == "model-audio.safetensors");
+  CHECK(std::string(towers[1].brick) == "W7");
+  for (const auto& t : towers) {
+    CHECK_MESSAGE(std::string(t.what).size() > 10,
+                  "the deferral for " << t.prefix << " says nothing about what it is");
+  }
+
+  const nlohmann::json index = FullIndexDoc();
+  const auto released = ExpandFullIndex(index);
+  // The fixture's own tower_files map and the port's table must agree, so a
+  // renamed file in one place cannot pass unnoticed in the other.
+  CHECK(index.at("tower_files").at("vision_encoder.").get<std::string>() ==
+        std::string(towers[0].file));
+  CHECK(index.at("tower_files").at("audio_encoder.").get<std::string>() ==
+        std::string(towers[1].file));
+
+  int64_t vision = 0;
+  int64_t audio = 0;
+  int64_t language = 0;
+  // One assertion per DEFECT CLASS with the first offender named, not one per
+  // tensor: 38006 doctest assertions cost minutes at -O0 and print thousands of
+  // lines when they fail. Each list below still covers all 38006 names.
+  std::vector<std::string> lang_with_deferral;
+  std::vector<std::string> lang_in_tower_file;
+  std::vector<std::string> tower_without_deferral;
+  std::vector<std::string> wrong_brick;
+  std::vector<std::string> wrong_file;
+  for (const auto& [name, meta] : released) {
+    const vllm::Dots3NoteDeferredTower* d = vllm::Dots3NoteDeferralFor(name);
+    if (!IsTowerName(name)) {
+      if (d != nullptr) lang_with_deferral.push_back(name);
+      // A language weight never ships in a tower file.
+      if (meta.file == std::string(towers[0].file) ||
+          meta.file == std::string(towers[1].file)) {
+        lang_in_tower_file.push_back(name);
+      }
+      ++language;
+      continue;
+    }
+    if (d == nullptr) {
+      tower_without_deferral.push_back(name);
+      continue;
+    }
+    const int which = HasPrefix(name, "vision_encoder.") ? 0 : 1;
+    if (std::string(d->brick) != std::string(towers[which].brick)) {
+      wrong_brick.push_back(name);
+    }
+    if (std::string(d->file) != std::string(towers[which].file)) {
+      wrong_file.push_back(name);
+    }
+    // Measured, not assumed: each tower ships whole in ONE standalone file,
+    // never spread across the 131 numbered language shards.
+    if (meta.file != std::string(d->file)) {
+      wrong_file.push_back(name + " ships in " + meta.file + ", not " + d->file);
+    }
+    if (which == 0) {
+      ++vision;
+    } else {
+      ++audio;
+    }
+  }
+  const auto first_of = [](const std::vector<std::string>& v) {
+    return v.empty() ? std::string("-") : v.front();
+  };
+  CHECK_MESSAGE(lang_with_deferral.empty(),
+                "a language tensor resolved to a deferral, first: "
+                    << first_of(lang_with_deferral) << " ("
+                    << lang_with_deferral.size() << " total)");
+  CHECK_MESSAGE(lang_in_tower_file.empty(),
+                "a language tensor ships in a TOWER file, first: "
+                    << first_of(lang_in_tower_file) << " ("
+                    << lang_in_tower_file.size() << " total)");
+  CHECK_MESSAGE(tower_without_deferral.empty(),
+                "a tower tensor resolved to NO deferral, first: "
+                    << first_of(tower_without_deferral) << " ("
+                    << tower_without_deferral.size() << " total)");
+  CHECK_MESSAGE(wrong_brick.empty(),
+                "a tower tensor resolved to the wrong BRICK, first: "
+                    << first_of(wrong_brick) << " (" << wrong_brick.size()
+                    << " total)");
+  CHECK_MESSAGE(wrong_file.empty(),
+                "a tower tensor resolved to the wrong FILE, first: "
+                    << first_of(wrong_file) << " (" << wrong_file.size()
+                    << " total)");
+  CHECK(language == 35381);
+  CHECK(vision == 2195);
+  CHECK(audio == 430);
+
+  // ...and the language name map never claims a name inside a deferred tower.
+  // THIS LOOP IS THE WHOLE GUARD for that invariant, and it is here rather than
+  // in the classifier on purpose. A name that is both loaded and deferred would
+  // take the classifier's language branch, so the tower count would drop by one
+  // while the total still read 38006 — silently. W2 first wrote the invariant as
+  // a runtime `VT_CHECK` inside `AccountDots3NoteTensors` and measured it DEAD:
+  // no config makes the enumeration emit a tower-prefixed name, and deleting the
+  // check left the gate green (spec §4.4, M12). Adding such a name to
+  // `EnumerateDots3NoteTensors` is what has to fail, and it fails here.
+  const Dots3NoteParams p = FixtureParams();
+  std::vector<std::string> claimed_tower;
+  for (const Dots3NoteTensor& t : EnumerateDots3NoteTensors(p)) {
+    if (vllm::Dots3NoteDeferralFor(t.name) != nullptr) claimed_tower.push_back(t.name);
+  }
+  CHECK_MESSAGE(claimed_tower.empty(),
+                "the language map claims a deferred tower tensor, first: "
+                    << first_of(claimed_tower) << " (" << claimed_tower.size()
+                    << " total)");
+}
+
+TEST_CASE("dots3-note W2: the 42 backbone layers W1's slice never saw") {
+  // W1 gated layers 0, 1, 2 and 46 and RECORDED that "the remaining 42 backbone
+  // layers repeat layers 1/2 exactly". That was a claim about a checkpoint
+  // nobody had read. This measures it, from the shipped tensor names and the
+  // shipped shapes, and cross-checks the answer against `config.layer_types`.
+  const Dots3NoteParams p = FixtureParams();
+  const auto released = ExpandFullIndex(FullIndexDoc());
+
+  // Group the language tensors by backbone layer, collapsing the expert index.
+  std::map<int64_t, std::set<std::string>> suffixes;
+  std::map<int64_t, std::set<int64_t>> expert_ids;
+  std::map<int64_t, std::map<std::string, std::vector<int64_t>>> shapes;
+  const std::string kPre = "model.layers.";
+  for (const auto& [name, meta] : released) {
+    if (!HasPrefix(name, kPre)) continue;
+    const size_t dot = name.find('.', kPre.size());
+    REQUIRE(dot != std::string::npos);
+    const int64_t layer = std::stoll(name.substr(kPre.size(), dot - kPre.size()));
+    std::string suffix = name.substr(dot + 1);
+    const std::string kExp = "mlp.experts.";
+    if (HasPrefix(suffix, kExp)) {
+      const size_t end = suffix.find('.', kExp.size());
+      REQUIRE(end != std::string::npos);
+      expert_ids[layer].insert(
+          std::stoll(suffix.substr(kExp.size(), end - kExp.size())));
+      suffix = kExp + "{E}" + suffix.substr(end);
+    }
+    suffixes[layer].insert(suffix);
+    shapes[layer][suffix] = meta.shape;
+  }
+  // 46 backbone layers plus the one nextn layer, and no other index.
+  REQUIRE(suffixes.size() == 47);
+  CHECK(suffixes.begin()->first == 0);
+  CHECK(suffixes.rbegin()->first == 46);
+
+  // The checkpoint's OWN answer to "which layers are full attention": the DSA
+  // indexer ships only on the full class (model.py:430-432).
+  std::vector<int64_t> full_by_checkpoint;
+  std::vector<int64_t> moe_by_checkpoint;
+  for (const auto& [layer, sufs] : suffixes) {
+    if (layer >= p.num_hidden_layers) continue;  // the nextn tail is not scheduled
+    if (sufs.count("self_attn.indexer.wk.weight") != 0) {
+      full_by_checkpoint.push_back(layer);
+    }
+    if (!expert_ids[layer].empty()) moe_by_checkpoint.push_back(layer);
+  }
+  CHECK(full_by_checkpoint ==
+        std::vector<int64_t>{0, 1, 5, 9, 13, 17, 21, 25, 29, 33, 37, 41, 45});
+
+  // ...and it agrees with `config.layer_types`, which is an INDEPENDENT source:
+  // one is the released weight list, the other the released config.
+  std::vector<int64_t> full_by_config;
+  for (int64_t l = 0; l < p.num_hidden_layers; ++l) {
+    if (p.kind_of(l) == Dots3NoteLayerKind::kFullAttention) full_by_config.push_back(l);
+  }
+  CHECK(full_by_config == full_by_checkpoint);
+  CHECK(full_by_config.size() == 13);
+
+  // The dense/MoE split the same way: `first_k_dense_replace` is 1, so layer 0
+  // is the only dense backbone layer, and every other layer carries 256 experts.
+  CHECK(moe_by_checkpoint.size() == 45);
+  CHECK(moe_by_checkpoint.front() == 1);
+  for (int64_t l = 0; l < p.num_hidden_layers; ++l) {
+    const bool moe = !expert_ids[l].empty();
+    CHECK_MESSAGE(moe == p.is_moe_layer(l),
+                  "layer " << l << ": checkpoint says moe=" << moe
+                           << ", config says " << p.is_moe_layer(l));
+    if (!moe) continue;
+    // The expert index is a contiguous 0..255 block on every MoE layer, so the
+    // `{E}` collapse in the fixture is lossless and the 256-expert map is real.
+    CHECK_MESSAGE(static_cast<int64_t>(expert_ids[l].size()) == p.n_routed_experts,
+                  "layer " << l << " ships " << expert_ids[l].size() << " experts");
+    CHECK(*expert_ids[l].begin() == 0);
+    CHECK(*expert_ids[l].rbegin() == p.n_routed_experts - 1);
+  }
+
+  // Every layer's geometry, not just the four W1 read: the q_b_proj row count
+  // is the head count times the qk head dim, and it separates the geometries.
+  for (const auto& [layer, sh] : shapes) {
+    const bool is_full = layer < p.num_hidden_layers &&
+                         p.kind_of(layer) == Dots3NoteLayerKind::kFullAttention;
+    const vllm::Dots3NoteAttnParams& g = is_full ? p.full : p.swa;
+    const std::vector<int64_t> want_qb{g.num_attention_heads * g.qk_head_dim(),
+                                       g.q_lora_rank};
+    const std::vector<int64_t> want_kv{g.latent_row(), p.hidden_size};
+    const auto it = sh.find("self_attn.q_b_proj.weight");
+    REQUIRE(it != sh.end());
+    CHECK_MESSAGE(it->second == want_qb,
+                  "layer " << layer << " q_b_proj is not the expected geometry");
+    const auto kv = sh.find("self_attn.kv_a_proj_with_mqa.weight");
+    REQUIRE(kv != sh.end());
+    CHECK_MESSAGE(kv->second == want_kv,
+                  "layer " << layer << " kv_a_proj_with_mqa is not the expected width");
+    // The two dots3-only attention tensors exist on EVERY layer, including the
+    // nextn one — DeepSeek has neither.
+    CHECK_MESSAGE(sh.count("self_attn.g_proj.weight") == 1, "layer " << layer);
+    CHECK_MESSAGE(sh.count("self_attn.k_rope_only_layernorm.weight") == 1,
+                  "layer " << layer);
+  }
+
+  // Exactly FOUR distinct layer shapes exist in the whole backbone, and this is
+  // the fact W1 could not measure. A fifth would mean a layer that breaks the
+  // repeating pattern, and the port would be reading it with the wrong map.
+  std::map<std::string, std::vector<int64_t>> classes;
+  for (const auto& [layer, sufs] : suffixes) {
+    std::string key;
+    for (const std::string& s : sufs) {
+      key += s;
+      key += '|';
+      for (const int64_t d : shapes[layer].at(s)) key += std::to_string(d) + ",";
+      key += ';';
+    }
+    classes[key].push_back(layer);
+  }
+  CHECK(classes.size() == 4);
+  std::set<std::vector<int64_t>> memberships;
+  for (const auto& [key, members] : classes) {
+    (void)key;
+    memberships.insert(members);
+  }
+  CHECK(memberships.count(std::vector<int64_t>{0}) == 1);   // dense + full
+  CHECK(memberships.count(std::vector<int64_t>{46}) == 1);  // the nextn tail
+  CHECK(memberships.count(std::vector<int64_t>{1, 5, 9, 13, 17, 21, 25, 29, 33,
+                                               37, 41, 45}) == 1);
+  CHECK(memberships.count(std::vector<int64_t>{
+            2,  3,  4,  6,  7,  8,  10, 11, 12, 14, 15, 16, 18,
+            19, 20, 22, 23, 24, 26, 27, 28, 30, 31, 32, 34, 35,
+            36, 38, 39, 40, 42, 43, 44}) == 1);
+}
+
+TEST_CASE("dots3-note W2: the memory format of the WHOLE checkpoint") {
+  // porting.md, "mirror the memory format": a loader that assumed one dtype for
+  // the checkpoint reads these wrong, and no token gate can see the difference.
+  // W1 found one F32 tensor in its slice. The whole index carries 62, in TWO
+  // families, and the second one is in the vision tower.
+  const auto released = ExpandFullIndex(FullIndexDoc());
+
+  std::vector<std::string> lang_f32;
+  std::vector<std::string> vision_f32;
+  std::vector<std::string> audio_f32;
+  std::vector<std::string> other_dtype;
+  for (const auto& [name, meta] : released) {
+    if (meta.dtype == "BF16") continue;
+    if (meta.dtype != "F32") {
+      other_dtype.push_back(name + " (" + meta.dtype + ")");
+    } else if (HasPrefix(name, "vision_encoder.")) {
+      vision_f32.push_back(name);
+    } else if (HasPrefix(name, "audio_encoder.")) {
+      audio_f32.push_back(name);
+    } else {
+      lang_f32.push_back(name);
+    }
+  }
+  CHECK_MESSAGE(other_dtype.empty(),
+                "an unexpected dtype in a BF16 checkpoint, first: "
+                    << (other_dtype.empty() ? std::string("-") : other_dtype.front()));
+
+  // Language: the noaux_tc per-expert bias, one per MoE layer, layers 1..45.
+  CHECK(lang_f32.size() == 45);
+  for (int64_t l = 1; l <= 45; ++l) {
+    const std::string n = "model.layers." + std::to_string(l) +
+                          ".mlp.gate.e_score_correction_bias";
+    CHECK_MESSAGE(std::find(lang_f32.begin(), lang_f32.end(), n) != lang_f32.end(),
+                  n << " is not F32");
+  }
+
+  // Vision: `router_bias` on the 17 pyramid-MoE blocks, and NOTHING else. This
+  // is a second F32 family W1's language-only slice could not see, and it is
+  // the spec's R5 shape — the vision MoE keeps its router statistics in fp32
+  // inside an otherwise BF16 tower.
+  CHECK(vision_f32.size() == 17);
+  for (const std::string& n : vision_f32) {
+    CHECK_MESSAGE(n.rfind(".mlp.router_bias") == n.size() - std::string(".mlp.router_bias").size(),
+                  "an unexpected F32 tensor in the vision tower: " << n);
+  }
+  // Its width is the block's routed-expert count, and the pyramid runs
+  // 4, 8, ... 64, 64 over blocks 25..41 (spec §1.2, from `vision_config`).
+  const std::vector<int64_t> pyramid{4,  8,  12, 16, 20, 24, 28, 32, 36,
+                                     40, 44, 48, 52, 56, 60, 64, 64};
+  for (size_t i = 0; i < pyramid.size(); ++i) {
+    const std::string n = "vision_encoder.blocks." + std::to_string(25 + i) +
+                          ".mlp.router_bias";
+    const auto it = released.find(n);
+    REQUIRE_MESSAGE(it != released.end(), n << " is absent");
+    CHECK(it->second.dtype == "F32");
+    const std::vector<int64_t> want{pyramid[i]};
+    CHECK_MESSAGE(it->second.shape == want, n);
+  }
+  // Audio is BF16 throughout.
+  CHECK(audio_f32.empty());
+}
+
+TEST_CASE("dots3-note W2: the released index states an indexer RoPE layout upstream never reads") {
+  // A FINDING, pinned here so it cannot be lost between W2 and W3. The released
+  // `model.safetensors.index.json` carries two keys in its `metadata` block
+  // beside `total_size`:
+  //
+  //     "indexer_rope_layout": "leading", "indexer_rope_converted_from": "tail"
+  //
+  // `git grep indexer_rope_layout` over vLLM `origin/main` returns nothing, so
+  // upstream never reads either key: it is the publisher telling a loader how
+  // the DSA indexer's projections are laid out. It matches what upstream's code
+  // does anyway — `DeepseekV2Indexer` splits q and k as `[..., :rope_dim]` for
+  // the rotated half and `[..., rope_dim:]` for the rest
+  // (`deepseek_v2.py:805,:814`), which is a LEADING rope slice — and it says
+  // the published weights were re-ordered from a TAIL layout to reach it.
+  //
+  // W3 owns the consumption. The assertion exists so that a W3 implementer who
+  // slices the indexer's 128-wide head the other way fails here first, and so
+  // that a re-published checkpoint that flips the layout cannot land silently.
+  // This is separate from §4 trap 2, which is about which PAIRS the rope
+  // rotates (GPT-J vs NeoX), not which HALF of the head it rotates.
+  const nlohmann::json meta = FullIndexDoc().at("index_metadata");
+  CHECK(meta.at("total_size").get<int64_t>() == 576886825984LL);
+  CHECK(meta.at("indexer_rope_layout").get<std::string>() == "leading");
+  CHECK(meta.at("indexer_rope_converted_from").get<std::string>() == "tail");
+}
+
+TEST_CASE("dots3-note W2: the full index and W1's committed slice agree") {
+  // Two fixtures projected from the same revision by different passes. If they
+  // disagree on a name, a dtype or a shape, one of them was mis-generated and
+  // every count that rests on it is unsafe.
+  const auto full = ExpandFullIndex(FullIndexDoc());
+  const auto slice = ExpandIndexFixture(ReadJson(FixtureDir() + "/index.json"));
+  REQUIRE(slice.size() == 1614);
+  for (const auto& [name, meta] : slice) {
+    const auto it = full.find(name);
+    REQUIRE_MESSAGE(it != full.end(), "the slice has a name the full index lacks: " << name);
+    CHECK_MESSAGE(it->second.dtype == meta.first, name);
+    CHECK_MESSAGE(it->second.shape == meta.second, name);
+  }
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 // The unported arms, and the refusal driven through the REAL loaded model.
 // ─────────────────────────────────────────────────────────────────────────────
@@ -939,9 +1483,10 @@ void WriteSafetensors(const std::vector<StEntry>& entries,
 class TempCheckpoint {
  public:
   explicit TempCheckpoint(const std::vector<std::string>& names) {
-    static int counter = 0;
-    dir_ = std::filesystem::temp_directory_path() /
-           ("dots3_note_ckpt_" + std::to_string(counter++));
+    // Process-unique, for the reason `UniqueTempDir` states: a concurrent run
+    // of this binary would otherwise rewrite a safetensors file this one has
+    // mmapped, and the SIGBUS reads as no result rather than as a failure.
+    dir_ = UniqueTempDir("dots3_note_ckpt_");
     std::filesystem::create_directories(dir_);
     std::vector<StEntry> entries;
     entries.reserve(names.size());
@@ -970,6 +1515,42 @@ std::vector<std::string> AllLanguageNames(const Dots3NoteParams& p) {
 
 }  // namespace
 
+TEST_CASE("dots3-note W2: the whole index loads through the PRODUCTION entry point") {
+  // The accounting above is a helper call. This drives the SAME 38006 names
+  // through `ModelRegistry::Resolve(...).factory->load_weights`, which is the
+  // path a real load takes, so the map is proved reachable and not merely
+  // correct in a unit.
+  TempConfig cfg(FixtureConfigDoc());
+  const HfConfig config = LoadHfConfig(cfg.path());
+  const vllm::ModelRegistration& reg = ModelRegistry::Resolve(config);
+
+  const auto released = ExpandFullIndex(FullIndexDoc());
+  std::vector<std::string> names;
+  names.reserve(released.size());
+  for (const auto& [name, meta] : released) {
+    (void)meta;
+    names.push_back(name);
+  }
+  REQUIRE(names.size() == 38006);
+
+  TempCheckpoint ckpt(names);
+  std::vector<vllm::SafetensorsFile> shards;
+  shards.push_back(vllm::SafetensorsFile::Open(ckpt.file()));
+  const vllm::ModelSource source = vllm::ModelSource::FromSafetensors(shards);
+  std::unique_ptr<vllm::LoadedModel> model;
+  REQUIRE_NOTHROW(model = reg.factory->load_weights(reg, config, source));
+  REQUIRE(model != nullptr);
+
+  // WHAT THIS CASE CANNOT SEE, said rather than implied. `Dots3NoteLoadedModel`
+  // lives in an anonymous namespace in the registry TU, so the bucket counts
+  // cannot be read back off the object here — the case proves the production
+  // load ACCEPTS all 38006, and the case above proves the classifier splits
+  // them 35381 / 2195 / 430. The REFUSAL side is the "unported arms" case
+  // below, which already builds a whole-tower checkpoint and reads the message
+  // the deferral table prints; a second 38006-tensor safetensors here would
+  // cost ~35 s at -O0 for a path that is already gated.
+}
+
 TEST_CASE("dots3-note: the unported arms REFUSE BY NAME") {
   TempConfig cfg(FixtureConfigDoc());
   const HfConfig config = LoadHfConfig(cfg.path());
@@ -992,8 +1573,23 @@ TEST_CASE("dots3-note: the unported arms REFUSE BY NAME") {
     std::vector<vllm::SafetensorsFile> shards;
     shards.push_back(vllm::SafetensorsFile::Open(ckpt.file()));
     const vllm::ModelSource source = vllm::ModelSource::FromSafetensors(shards);
-    CHECK_THROWS_WITH_AS(reg.factory->load_weights(reg, config, source),
-                         doctest::Contains("mystery_proj"), std::runtime_error);
+    // ONE load, several assertions over the message: each `CHECK_THROWS_WITH_AS`
+    // re-runs a whole-tower load, and this case is the slowest in the file.
+    std::string refusal;
+    try {
+      reg.factory->load_weights(reg, config, source);
+    } catch (const std::runtime_error& e) {
+      refusal = e.what();
+    }
+    REQUIRE_MESSAGE(!refusal.empty(), "the load accepted an unclaimed tensor");
+    CHECK_MESSAGE(refusal.find("mystery_proj") != std::string::npos, refusal);
+    // W2: the message distinguishes UNKNOWN from DEFERRED by printing the
+    // deferral table, so a reader is not left guessing which of the two the
+    // loader thinks it hit.
+    CHECK_MESSAGE(refusal.find("vision_encoder.* (W6)") != std::string::npos,
+                  "the refusal does not name the vision deferral: " << refusal);
+    CHECK_MESSAGE(refusal.find("audio_encoder.* (W7)") != std::string::npos,
+                  "the refusal does not name the audio deferral: " << refusal);
   }
 
   SUBCASE("a MISSING enumerated tensor refuses at LOAD, naming the tensor") {