[Castoff995]

Этот PR служит постоянным каналом уведомлений для автоматической проверки GitHub каждые 2 часа. Его не нужно ревьюить, мержить или закрывать, пока слежка нужна.

Высший приоритет

• FlashML-org/FreeToken PR feat(rocm): add RDNA3 and RDNA4 runtime foundation FlashML-org/FreeToken#132–feat(rocm): serve on AMD GPUs through the HIP toolchain FlashML-org/FreeToken#137: ROCm/RDNA3/RDNA4 foundation, TVM-FFI HIP JIT, backend gating, RCCL, GGUF RDNA4 и альтернативный ROCm port.
• Issues ROCm RDNA4 gfx1201 Ported. Windows only for now. Test please FlashML-org/FreeToken#82, Add scripts for laying ROCM artifacts out for Python packaging. ROCm/TheRock#120, Uplift how artifacts are handled. ROCm/TheRock#122 и roadmap FreeToken Roadmap (2026) FlashML-org/FreeToken#79.
• Новые upstream issues/PR с AMD, ROCm, HIP, RDNA4, gfx1200, gfx1201, RX 9070, Windows, Triton, TVM-FFI, pinned/mapped host memory, hipHostGetDevicePointer, launch_pdl, clang-cl, MoE/offload.

Рабочие форки

• PialGhosh2233/FreeToken-rocm-gfx1200:gfx1200-moe-gguf.
• Maxritz/FreeToken-rocm-test:main.

Зависимости

• ROCm/TheRock — Windows/RDNA4/gfx1201 wheels, PyTorch/HIP regressions and releases.
• triton-lang/triton-windows — Windows AMD/HIP support and releases.
• apache/tvm-ffi — Windows/HIP JIT and cache/toolchain changes.

Устойчивость модели

Также отслеживаются изменения, способные повлиять на первый токен и практический перевод: Qwen3.6/NVFP4, MoE offload, hybrid_radix, pinned/pageable residency, mapped device pointers, long-prefill/KV-cache hangs and engine crashes.

Уведомление публикуется только при существенном изменении: новый релевантный issue/PR, новый commit в ключевом PR/форке, изменение draft/merge/CI/state, новый релиз либо подтверждённый gfx1201/Windows результат. При отсутствии новостей комментариев не будет.

[github-actions]

@Castoff995 FreeToken ROCm Watch запущен. Проверка выполняется каждые 2 часа; при отсутствии существенных изменений комментариев не будет.

Текущий P0 baseline:

• PR Optimize CI/CD builders ROCm/TheRock#132: open, head 45417560, checks pending — feat(rocm): add RDNA3 and RDNA4 runtime foundation
• PR Do not write out a HIP/VERSION file. ROCm/TheRock#133: open / draft, head 89ad7ab0, checks pending — fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP
• PR First draft of full python packaging. ROCm/TheRock#134: open / draft, head 9a523dc8, checks pending — fix(rocm): gate CUDA-only optional backends on ROCm
• PR The codes building issue of unexpected keyword argument 'follow_symlinks' ROCm/TheRock#135: open / draft, head 9d5e3fc8, checks pending — fix(rocm): route tensor parallel communication through RCCL
• PR How to install AMD GPU driver on a clean machine when using TheRock ROCm/TheRock#136: open / draft, head 49591978, checks pending — feat(rocm): enable native GGUF kernels on RDNA4
• PR Fix Python 3.10 issue ROCm/TheRock#137: open, head 7e1d7421, checks pending — feat(rocm): serve on AMD GPUs through the HIP toolchain

Ключевые issues: FreeToken Roadmap (2026), ROCm RDNA4 gfx1201 Ported. Windows only for now. Test please, Native Windows is classified as pin-uncapped, so ROCm/TheRock#112's per-layer residency never auto-engages, Windows/ROCm on RX 9060 XT (gfx1200): packed GGUF working end-to-end (dense + MoE), root cause of the RDNA4 offload-decode TDR, benchmarks vs llama.cpp.

Дополнительно отслеживаются fork commits, ROCm/TheRock, Triton Windows, TVM-FFI, PyTorch gfx1201, Qwen3.6/NVFP4, mapped host pointers, pin-budget, long-prefill и offload stability.

[github-actions]

{"branches":{"FlashML-org/FreeToken:main":{"branch":"main","date":"2026-10-21T21:28:09Z","message":"docs: add a guide to speeding up prefill with offloaded experts (FlashML-org#785)","repo":"FlashML-org/FreeToken","sha":"b10c0c479bae49dd87c41fd99b864c58abcaa225","url":"FlashML-org@b10c0c479bae49dd87c41fd99b864c58abcaa225"},"Maxritz/FreeToken-rocm-test:main":{"branch":"main","date":"2026-09-20T16:29:33Z","message":"Merge pull request Maxritz#5 from uploadyours/fix/windows-rocm-install-and-qwen35moe-gguf","repo":"Maxritz/FreeToken-rocm-test","sha":"e8545daca33f54069eedd71c8afd8a2ff35380bc","url":"Maxritz@e8545daca33f54069eedd71c8afd8a2ff35380bc"},"PialGhosh2233/FreeToken-rocm-gfx1200:gfx1200-moe-gguf":{"branch":"gfx1200-moe-gguf","date":"2026-08-26T07:51:22Z","message":"docs: Gemma-4-26B-A4B QAT q4_0 GGUF results on gfx1200","repo":"PialGhosh2233/FreeToken-rocm-gfx1200","sha":"a338f49b93e2f098eeb3aa83423c9c5386c53ac4","url":"PialGhosh2233@a338f49b93e2f098eeb3aa83423c9c5386c53ac4"}},"discovered":{"FlashML-org/FreeToken#273%22:%7B%22body_hash%22:%2295a6bb552bd85981%22,%22comments%22:2,%22kind%22:%22issue%22,%22number%22:273,%22repo%22:%22FlashML-org/FreeToken%22,%22state%22:%22open%22,%22title%22:%22%5BFeature Request] Please add SSD offloading for KV cache to save GPU memory (like LM-Cache)","updated_at":"2026-10-20T10:51:50Z","url":"https://github.com/FlashML-org/FreeToken/issues/273"},"FlashML-org/FreeToken#311":{"body_hash":"05fb3389b262f3c4","comments":2,"kind":"pr","number":311,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(qwen4_exp): stream the PLE n-gram table from disk","updated_at":"2026-10-18T08:55:52Z","url":"https://github.com/FlashML-org/FreeToken/pull/311"},"FlashML-org/FreeToken#534":{"body_hash":"265c914cd1ae1994","comments":6,"kind":"issue","number":534,"repo":"FlashML-org/FreeToken","state":"open","title":"FP8 block-quantized MoE models have no usable kernel under moe_strategy=offload (0.1.3)","updated_at":"2026-10-18T03:57:03Z","url":"https://github.com/FlashML-org/FreeToken/issues/534"},"FlashML-org/FreeToken#535":{"body_hash":"d58b10217ef3c43c","comments":0,"kind":"pr","number":535,"repo":"FlashML-org/FreeToken","state":"open","title":"ci(rocm): add ROCm 10.0.0 Dockerfile and validation workflow","updated_at":"2026-10-19T07:26:37Z","url":"https://github.com/FlashML-org/FreeToken/pull/535"},"FlashML-org/FreeToken#544":{"body_hash":"a01bd957aa12585d","comments":3,"kind":"issue","number":544,"repo":"FlashML-org/FreeToken","state":"closed","title":"[engine bug] Decode throughput degrades from 60-120 to 11-25 tok/s over runtime; engine-global, cumulative, only reset by restart","updated_at":"2026-10-20T10:44:34Z","url":"https://github.com/FlashML-org/FreeToken/issues/544"},"FlashML-org/FreeToken#567":{"body_hash":"6af3e9c4ead76444","comments":1,"kind":"issue","number":567,"repo":"FlashML-org/FreeToken","state":"open","title":"A stream that is silent through a long prefill is dropped by Node clients and prefilled for nobody","updated_at":"2026-10-19T20:19:07Z","url":"https://github.com/FlashML-org/FreeToken/issues/567"},"FlashML-org/FreeToken#579":{"body_hash":"8b996dedc8b58ce6","comments":1,"kind":"issue","number":579,"repo":"FlashML-org/FreeToken","state":"open","title":"An offloaded prefill streams every expert of each layer, however short the chunk","updated_at":"2026-10-19T20:26:30Z","url":"https://github.com/FlashML-org/FreeToken/issues/579"},"FlashML-org/FreeToken#650":{"body_hash":"4717116335d9d5af","comments":0,"kind":"pr","number":650,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(engine): apply shrinking cache resizes before growing ones in live rebuild (Closes FlashML-org#643)","updated_at":"2026-10-20T15:51:53Z","url":"https://github.com/FlashML-org/FreeToken/pull/650"},"FlashML-org/FreeToken#661":{"body_hash":"3673dc784be5d9b3","comments":3,"kind":"issue","number":661,"repo":"FlashML-org/FreeToken","state":"open","title":"Requests stall and time out above --max-running-requests 8; GPU stays at ~28% of TDP and only one CPU thread is busy","updated_at":"2026-10-18T05:51:33Z","url":"https://github.com/FlashML-org/FreeToken/issues/661"},"FlashML-org/FreeToken#671":{"body_hash":"99ecd9f9147429d9","comments":0,"kind":"issue","number":671,"repo":"FlashML-org/FreeToken","state":"open","title":"🚧 [WIP] MoE cache, CPU experts and GGUF support","updated_at":"2026-10-20T22:20:53Z","url":"https://github.com/FlashML-org/FreeToken/issues/671"},"FlashML-org/FreeToken#677":{"body_hash":"5fa2424cec41ab37","comments":1,"kind":"issue","number":677,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(moe): prefetch the next layer's experts from a router lookahead during offloaded decode","updated_at":"2026-10-19T20:27:57Z","url":"https://github.com/FlashML-org/FreeToken/issues/677"},"FlashML-org/FreeToken#678":{"body_hash":"966245138f27565f","comments":3,"kind":"issue","number":678,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(moe): run the shared expert beside the routed-expert miss copy in offloaded decode","updated_at":"2026-10-20T16:13:31Z","url":"https://github.com/FlashML-org/FreeToken/issues/678"},"FlashML-org/FreeToken#693":{"body_hash":"2cebe71320aa6859","comments":1,"kind":"issue","number":693,"repo":"FlashML-org/FreeToken","state":"open","title":"seed and chat logprobs are silently ignored (200 with no effect); proposal for per-request seeded sampling","updated_at":"2026-10-19T20:19:11Z","url":"https://github.com/FlashML-org/FreeToken/issues/693"},"FlashML-org/FreeToken#730":{"body_hash":"805f5e2859ba911b","comments":0,"kind":"issue","number":730,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(moe): fused triton router for sigmoid + selection-bias routing","updated_at":"2026-10-17T09:45:48Z","url":"https://github.com/FlashML-org/FreeToken/issues/730"},"FlashML-org/FreeToken#731":{"body_hash":"0b907b4a25ffa4a0","comments":0,"kind":"pr","number":731,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(gguf): serve GLM-4.5, GLM-4.6 and GLM-4.7 GGUF checkpoints","updated_at":"2026-10-17T19:56:20Z","url":"https://github.com/FlashML-org/FreeToken/pull/731"},"FlashML-org/FreeToken#732":{"body_hash":"dc151e25101533d5","comments":0,"kind":"pr","number":732,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(gguf): apply the weight scales of nvfp4 gguf tensors","updated_at":"2026-10-17T20:46:50Z","url":"https://github.com/FlashML-org/FreeToken/pull/732"},"FlashML-org/FreeToken#733":{"body_hash":"c6dfee39b15803f0","comments":0,"kind":"pr","number":733,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(moe): speed up the nvfp4 moe prefill gemm on rtx gpus","updated_at":"2026-10-18T00:51:00Z","url":"https://github.com/FlashML-org/FreeToken/pull/733"},"FlashML-org/FreeToken#734":{"body_hash":"6e92214b2f20f6c7","comments":1,"kind":"pr","number":734,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(moe): bench hybrid fetch fractions per moe layer kind","updated_at":"2026-10-18T06:25:12Z","url":"https://github.com/FlashML-org/FreeToken/pull/734"},"FlashML-org/FreeToken#735":{"body_hash":"eb96c0dadb7705bd","comments":0,"kind":"pr","number":735,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(moe): pipeline the nvfp4 prefill scale loads on geforce sm_86 / sm_89","updated_at":"2026-10-18T21:39:57Z","url":"https://github.com/FlashML-org/FreeToken/pull/735"},"FlashML-org/FreeToken#736":{"body_hash":"47f88387f9fb2e25","comments":1,"kind":"pr","number":736,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(moe): prefill nvfp4 W4A4 moe kernel on sm_120","updated_at":"2026-10-19T03:10:52Z","url":"https://github.com/FlashML-org/FreeToken/pull/736"},"FlashML-org/FreeToken#737":{"body_hash":"4d2b0d73a9ec59e9","comments":0,"kind":"pr","number":737,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(linear): run nvfp4 linears on nvfp4 activations on sm_120","updated_at":"2026-10-19T06:55:08Z","url":"https://github.com/FlashML-org/FreeToken/pull/737"},"FlashML-org/FreeToken#738":{"body_hash":"cde0e6dee644a0d8","comments":0,"kind":"issue","number":738,"repo":"FlashML-org/FreeToken","state":"closed","title":"Support for orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4","updated_at":"2026-10-20T03:18:35Z","url":"https://github.com/FlashML-org/FreeToken/issues/738"},"FlashML-org/FreeToken#739":{"body_hash":"8d19c14d4567242f","comments":1,"kind":"pr","number":739,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(quant)!: keep nvfp4 global scales in fp32","updated_at":"2026-10-19T19:56:10Z","url":"https://github.com/FlashML-org/FreeToken/pull/739"},"FlashML-org/FreeToken#741":{"body_hash":"e94b6f6d71a8923d","comments":0,"kind":"pr","number":741,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(muse_glimmer)!: keep the attention gate out of the qkv fusion","updated_at":"2026-10-19T06:16:38Z","url":"https://github.com/FlashML-org/FreeToken/pull/741"},"FlashML-org/FreeToken#742":{"body_hash":"87364e1a47a713fb","comments":0,"kind":"pr","number":742,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(moe): fetch locked-layer experts through pinned staging","updated_at":"2026-10-20T21:59:06Z","url":"https://github.com/FlashML-org/FreeToken/pull/742"},"FlashML-org/FreeToken#743":{"body_hash":"0b2c88bc0c5fa714","comments":5,"kind":"issue","number":743,"repo":"FlashML-org/FreeToken","state":"closed","title":"Regression: first prefill OOMs on RTX 5090 with Qwen3.8-Flash-Next NVFP4 (TMA prefill kernel's first launch allocates ~3.7 GiB)","updated_at":"2026-10-20T16:25:36Z","url":"https://github.com/FlashML-org/FreeToken/issues/743"},"FlashML-org/FreeToken#746":{"body_hash":"8de1ebf2ecdb1042","comments":2,"kind":"issue","number":746,"repo":"FlashML-org/FreeToken","state":"closed","title":"[Bug] sm_120: the first TMA kernel launch reserves ~14 KB of stack per resident thread (1.4 GiB on 16 GB cards), and the VRAM budget does not account for it","updated_at":"2026-10-19T20:07:13Z","url":"https://github.com/FlashML-org/FreeToken/issues/746"},"FlashML-org/FreeToken#747":{"body_hash":"5b8f8e57853a1b79","comments":1,"kind":"issue","number":747,"repo":"FlashML-org/FreeToken","state":"closed","title":"[Bug] Qwen3.5/3.6: Gemma-style (1 + w) RMSNorm weights are folded into bf16 at load time; transformers applies (1 + w) in fp32","updated_at":"2026-10-20T16:13:24Z","url":"https://github.com/FlashML-org/FreeToken/issues/747"},"FlashML-org/FreeToken#748":{"body_hash":"dd09fc2d6f54b554","comments":0,"kind":"issue","number":748,"repo":"FlashML-org/FreeToken","state":"open","title":"[Feature] Hybrid radix cache: greedy outputs depend on request history even when cached_tokens is 0 (request: cache flush endpoint / deterministic mode)","updated_at":"2026-10-19T19:56:18Z","url":"https://github.com/FlashML-org/FreeToken/issues/748"},"FlashML-org/FreeToken#749":{"body_hash":"d4a3e190d19a36e5","comments":0,"kind":"issue","number":749,"repo":"FlashML-org/FreeToken","state":"open","title":"[Bug] Context overflow is checked only after tokenization in the worker: a streamed request gets 200 plus an in-stream error, and an oversized max_tokens is silently clamped","updated_at":"2026-10-19T20:18:57Z","url":"https://github.com/FlashML-org/FreeToken/issues/749"},"FlashML-org/FreeToken#750":{"body_hash":"f3392ec80519b08b","comments":0,"kind":"issue","number":750,"repo":"FlashML-org/FreeToken","state":"open","title":"[Perf] Layer-major prefill for offloaded MoE: stream each expert layer once per prompt, not once per chunk (qwen3_5_moe)","updated_at":"2026-10-19T20:25:46Z","url":"https://github.com/FlashML-org/FreeToken/issues/750"},"FlashML-org/FreeToken#751":{"body_hash":"4e104dc2364a3d0a","comments":0,"kind":"issue","number":751,"repo":"FlashML-org/FreeToken","state":"open","title":"[Perf] Offloaded decode: gather missed experts with a few TMA bulk-copy CTAs (cp.async.bulk, .shared::cta) instead of 48 x 1024 LDG threads","updated_at":"2026-10-19T20:26:08Z","url":"https://github.com/FlashML-org/FreeToken/issues/751"},"FlashML-org/FreeToken#752":{"body_hash":"3ee02754e21bcbac","comments":0,"kind":"issue","number":752,"repo":"FlashML-org/FreeToken","state":"open","title":"[Perf] Offloaded prefill: keep buffer invalidation off the prefill copy stream and give that stream top priority (short-prompt H2D 21 -> 42 GB/s under Nsight)","updated_at":"2026-10-19T20:26:51Z","url":"https://github.com/FlashML-org/FreeToken/issues/752"},"FlashML-org/FreeToken#753":{"body_hash":"8021a150dcfefe04","comments":0,"kind":"issue","number":753,"repo":"FlashML-org/FreeToken","state":"open","title":"[Perf] Offloaded decode: overlap the expert-miss gather with the same layer's shared expert, event-ordered inside the CUDA graph (bit-exact implementation)","updated_at":"2026-10-19T20:27:13Z","url":"https://github.com/FlashML-org/FreeToken/issues/753"},"FlashML-org/FreeToken#754":{"body_hash":"b4ac05e3c29ecdb4","comments":0,"kind":"issue","number":754,"repo":"FlashML-org/FreeToken","state":"open","title":"[Feature] Optional host-resident token-embedding table (pinned, mapped, UVA lookups): ~1 GB of VRAM back for offloaded expert slots","updated_at":"2026-10-19T20:28:19Z","url":"https://github.com/FlashML-org/FreeToken/issues/754"},"FlashML-org/FreeToken#755":{"body_hash":"74d56ef1a8a9cc88","comments":0,"kind":"pr","number":755,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(qwen3_5_moe)!: apply the zero-centred norm weights' +1 in fp32 at runtime","updated_at":"2026-10-20T03:24:05Z","url":"https://github.com/FlashML-org/FreeToken/pull/755"},"FlashML-org/FreeToken#756":{"body_hash":"739603743786aee8","comments":0,"kind":"pr","number":756,"repo":"FlashML-org/FreeToken","state":"closed","title":"feat(gguf): add scripts that check which gguf files load","updated_at":"2026-10-19T21:35:22Z","url":"https://github.com/FlashML-org/FreeToken/pull/756"},"FlashML-org/FreeToken#759":{"body_hash":"e6e04cc5353bc699","comments":0,"kind":"pr","number":759,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(qwen4_exp): read weights under the checkpoint's quant config","updated_at":"2026-10-20T03:18:34Z","url":"https://github.com/FlashML-org/FreeToken/pull/759"},"FlashML-org/FreeToken#760":{"body_hash":"167dc99b17b02592","comments":0,"kind":"pr","number":760,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(qwen4_exp): load gguf files with a quantized indexer","updated_at":"2026-10-20T04:14:03Z","url":"https://github.com/FlashML-org/FreeToken/pull/760"},"FlashML-org/FreeToken#761":{"body_hash":"6f36a86eb0296984","comments":0,"kind":"pr","number":761,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(mimo): add text-only MiMo v2.6 Flash support","updated_at":"2026-10-20T16:28:32Z","url":"https://github.com/FlashML-org/FreeToken/pull/761"},"FlashML-org/FreeToken#763":{"body_hash":"aa4374e043b356fb","comments":0,"kind":"pr","number":763,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(tests): support ROCm in AIME GPU memory checks","updated_at":"2026-10-21T03:48:56Z","url":"https://github.com/FlashML-org/FreeToken/pull/763"},"FlashML-org/FreeToken#764":{"body_hash":"016c59053e383956","comments":1,"kind":"pr","number":764,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(moe): prevent stale cpu expert inputs in rocm graphs","updated_at":"2026-10-21T09:44:53Z","url":"https://github.com/FlashML-org/FreeToken/pull/764"},"FlashML-org/FreeToken#765":{"body_hash":"4298148731027ae6","comments":0,"kind":"issue","number":765,"repo":"FlashML-org/FreeToken","state":"open","title":"Two tests fail on a CUDA host since FlashML-org#736 / FlashML-org#697: test_nvfp4_experts_restricted_to_triton_for_swigluoai (sm_120) and test_graph_capture_reuses_warm_offload_cache_before_capture (any Linux host with libcuda)","updated_at":"2026-10-20T15:28:36Z","url":"https://github.com/FlashML-org/FreeToken/issues/765"},"FlashML-org/FreeToken#766":{"body_hash":"9f679ed60bbf9854","comments":0,"kind":"issue","number":766,"repo":"FlashML-org/FreeToken","state":"open","title":"An FTW checkpoint can never serve the W4A4 NVFP4 MoE prefill, and the refusal tells you to reconvert with the option that produced it","updated_at":"2026-10-20T15:28:38Z","url":"https://github.com/FlashML-org/FreeToken/issues/766"},"FlashML-org/FreeToken#768":{"body_hash":"4a0106ea7abc7b65","comments":0,"kind":"issue","number":768,"repo":"FlashML-org/FreeToken","state":"open","title":"Proposal: vLLM scoring fields on /v1/completions (allowed_token_ids, return_tokens_as_token_ids, token-id prompts) for classifier-style clients","updated_at":"2026-10-20T15:28:43Z","url":"https://github.com/FlashML-org/FreeToken/issues/768"},"FlashML-org/FreeToken#769":{"body_hash":"e6ce20a431fb4454","comments":0,"kind":"issue","number":769,"repo":"FlashML-org/FreeToken","state":"open","title":"Proposal: load an FTW with fp16 NVFP4 global scales (upcast with a warning) instead of refusing it after FlashML-org#739","updated_at":"2026-10-20T15:28:46Z","url":"https://github.com/FlashML-org/FreeToken/issues/769"},"FlashML-org/FreeToken#771":{"body_hash":"856edbf255190eb0","comments":1,"kind":"pr","number":771,"repo":"FlashML-org/FreeToken","state":"closed","title":"fix(moe): enable prefill hit-d2d on rocm","updated_at":"2026-10-21T06:15:18Z","url":"https://github.com/FlashML-org/FreeToken/pull/771"},"FlashML-org/FreeToken#772":{"body_hash":"d97c16ca72989368","comments":0,"kind":"issue","number":772,"repo":"FlashML-org/FreeToken","state":"open","title":"Proposal: isolate logprob reads (max_tokens 1) from batch composition and the prefix cache so scoring is reproducible","updated_at":"2026-10-20T22:45:16Z","url":"https://github.com/FlashML-org/FreeToken/issues/772"},"FlashML-org/FreeToken#773":{"body_hash":"958e2e8b814b44cd","comments":0,"kind":"pr","number":773,"repo":"FlashML-org/FreeToken","state":"closed","title":"chore(release): 0.2.0","updated_at":"2026-10-21T00:02:27Z","url":"https://github.com/FlashML-org/FreeToken/pull/773"},"FlashML-org/FreeToken#774":{"body_hash":"270b9b431f941a95","comments":3,"kind":"pr","number":774,"repo":"FlashML-org/FreeToken","state":"open","title":"feat(gguf): serve DeepSeek-V4 Flash and Pro GGUF checkpoints on deepseek4","updated_at":"2026-10-21T17:22:12Z","url":"https://github.com/FlashML-org/FreeToken/pull/774"},"FlashML-org/FreeToken#775":{"body_hash":"bb57eb0132d4b99a","comments":0,"kind":"issue","number":775,"repo":"FlashML-org/FreeToken","state":"open","title":"Capture-time cuGraphUpload (FlashML-org#697) widens inter-kernel gaps in replayed decode graphs: ~1-2% slower decode","updated_at":"2026-10-21T02:05:40Z","url":"https://github.com/FlashML-org/FreeToken/issues/775"},"FlashML-org/FreeToken#776":{"body_hash":"516d6f592bf684c8","comments":0,"kind":"pr","number":776,"repo":"FlashML-org/FreeToken","state":"open","title":"fix(ple): add rocm disk wait-sync support","updated_at":"2026-10-21T05:08:19Z","url":"https://github.com/FlashML-org/FreeToken/pull/776"},"FlashML-org/FreeToken#777":{"body_hash":"29f36d82972d7c98","comments":0,"kind":"pr","number":777,"repo":"FlashML-org/FreeToken","state":"open","title":"test(models): cover the Ornith-1.5-122B NVFP4 export","updated_at":"2026-10-21T18:37:10Z","url":"https://github.com/FlashML-org/FreeToken/pull/777"},"FlashML-org/FreeToken#778":{"body_hash":"9907a184fcff9599","comments":0,"kind":"pr","number":778,"repo":"FlashML-org/FreeToken","state":"closed","title":"perf(moe): scan the hybrid eviction pool in chunks and add fetch rounding modes","updated_at":"2026-10-21T18:38:14Z","url":"https://github.com/FlashML-org/FreeToken/pull/778"},"FlashML-org/FreeToken#779":{"body_hash":"a59bfa513c94bac0","comments":0,"kind":"pr","number":779,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(kernel): add opt-in hip fp8 decode kernels for r9700","updated_at":"2026-10-21T08:20:08Z","url":"https://github.com/FlashML-org/FreeToken/pull/779"},"FlashML-org/FreeToken#780":{"body_hash":"afdbc68310c135ca","comments":0,"kind":"issue","number":780,"repo":"FlashML-org/FreeToken","state":"open","title":"[Perf] Fuse the BF16 attention output gate in short gfx1201 decode graphs","updated_at":"2026-10-21T07:56:16Z","url":"https://github.com/FlashML-org/FreeToken/issues/780"},"FlashML-org/FreeToken#781":{"body_hash":"24423cb92efaf411","comments":0,"kind":"pr","number":781,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(attention): fuse short bf16 decode output gates on gfx1201","updated_at":"2026-10-21T08:34:09Z","url":"https://github.com/FlashML-org/FreeToken/pull/781"},"FlashML-org/FreeToken#782":{"body_hash":"9d0c77a8ebae4856","comments":2,"kind":"issue","number":782,"repo":"FlashML-org/FreeToken","state":"open","title":"[Desktop] weights failed to load: Qwen3.6-35B-A3B NVFP4","updated_at":"2026-10-21T19:24:53Z","url":"https://github.com/FlashML-org/FreeToken/issues/782"},"FlashML-org/FreeToken#783":{"body_hash":"f2d37277681db800","comments":0,"kind":"issue","number":783,"repo":"FlashML-org/FreeToken","state":"open","title":"perf(qsa): prefill indexer scoring pads 4 heads to 16 per row; 16-row tiles are ~5x faster and bitwise identical","updated_at":"2026-10-21T17:28:44Z","url":"https://github.com/FlashML-org/FreeToken/issues/783"},"FlashML-org/FreeToken#784":{"body_hash":"aa6db6e7f82e34ce","comments":1,"kind":"issue","number":784,"repo":"FlashML-org/FreeToken","state":"open","title":"[Desktop] weights failed to load: Qwen3.6-35B-A3B NVFP4","updated_at":"2026-10-21T19:20:27Z","url":"https://github.com/FlashML-org/FreeToken/issues/784"}},"errors":["dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 9417:2A8AA4:C5B4:154F1:6AD94955 and timestamp 2026-10-21 23:23:01 UTC. For more on scraping GitHub and how it may affect your rights, please review our Terms of Service (https://docs.github.com/en/site-policy/github-terms/github-terms-of-service)"],"external_tracked_issues":{"ROCm/legacy-rocm-build#6461":{"body_hash":"0ac6e80ee9b489cd","comments":6,"kind":"issue","number":6461,"relevance":"high","relevance_reason":"Windows + gfx1201 + hipBLASLt/rocBLAS sequence/state-dependent GEMM failure","repo":"ROCm/legacy-rocm-build","state":"closed","title":"[Issue]: rocBLAS bf16/f16 GEMM fails with hipErrorInvalidValue on 2nd call, gfx1201 / Windows / ROCm 7.14","updated_at":"2026-09-23T17:39:44Z","url":"https://github.com/ROCm/legacy-rocm-build/issues/6461"}},"last_checked":"2026-10-21T23:22:25+00:00","releases":{"FlashML-org/FreeToken":{"draft":false,"name":"v0.2.0","prerelease":false,"published_at":"2026-10-21T00:37:56Z","tag":"v0.2.0","url":"https://github.com/FlashML-org/FreeToken/releases/tag/v0.2.0"},"ROCm/ROCm":{"draft":false,"name":"ROCm 7.14.0 Release","prerelease":false,"published_at":"2026-07-16T03:28:58Z","tag":"rocm-7.14.0","url":"https://github.com/ROCm/legacy-rocm-build/releases/tag/rocm-7.14.0"},"ROCm/TheRock":{"draft":false,"name":"ROCm Core SDK 10.1.0 Release","prerelease":false,"published_at":"2026-10-05T10:42:56Z","tag":"therock-10.1","url":"https://github.com/ROCm/TheRock/releases/tag/therock-10.1"},"apache/tvm-ffi":{"draft":false,"name":"v0.1.15","prerelease":false,"published_at":"2026-10-18T17:22:03Z","tag":"v0.1.15","url":"https://github.com/apache/tvm-ffi/releases/tag/v0.1.15"},"ggml-org/llama.cpp":{"draft":false,"name":"v0.7.0","prerelease":false,"published_at":"2026-10-15T17:13:19Z","tag":"v0.7.0","url":"https://github.com/ggml-org/llama.cpp/releases/tag/v0.7.0"},"triton-lang/triton-windows":{"draft":false,"name":"v3.8.0-windows.post29","prerelease":false,"published_at":"2026-09-28T22:00:36Z","tag":"v3.8.0-windows.post29","url":"https://github.com/triton-lang/triton-windows/releases/tag/v3.8.0-windows.post29"}},"tracked_issues":{"110":{"body_hash":"d6672f16a90f2854","comments":3,"kind":"issue","number":110,"repo":"FlashML-org/FreeToken","state":"open","title":"Unhandled CUDA OOM in prefill expert workspace kills the engine; server keeps accepting requests that never return (root cause behind FlashML-org#20-style headless state; likely also FlashML-org#72)","updated_at":"2026-10-06T10:03:58Z","url":"https://github.com/FlashML-org/FreeToken/issues/110"},"111":{"body_hash":"b01b977d143435c7","comments":3,"kind":"issue","number":111,"repo":"FlashML-org/FreeToken","state":"open","title":"Requests longer than the KV pool are queued forever with no error — and --moe-cache-auto leaves only ~8k tokens of KV on a 96GB GPU","updated_at":"2026-08-31T02:09:35Z","url":"https://github.com/FlashML-org/FreeToken/issues/111"},"120":{"body_hash":"a7caa8c0bca18470","comments":1,"kind":"issue","number":120,"repo":"FlashML-org/FreeToken","state":"open","title":"Native Windows is classified as pin-uncapped, so FlashML-org#112's per-layer residency never auto-engages","updated_at":"2026-08-24T03:04:21Z","url":"https://github.com/FlashML-org/FreeToken/issues/120"},"122":{"body_hash":"4d42515b782747ed","comments":6,"kind":"issue","number":122,"repo":"FlashML-org/FreeToken","state":"open","title":"Windows/ROCm on RX 9060 XT (gfx1200): packed GGUF working end-to-end (dense + MoE), root cause of the RDNA4 offload-decode TDR, benchmarks vs llama.cpp","updated_at":"2026-08-27T07:18:05Z","url":"https://github.com/FlashML-org/FreeToken/issues/122"},"123":{"body_hash":"09dd1b9b8cf6560f","comments":1,"kind":"issue","number":123,"repo":"FlashML-org/FreeToken","state":"closed","title":"Request hang (silent, zero-log) with hybrid_radix cache + moe-backend offload + nvfp4 triton backend — reproduced on 2 independent models, not explained by FlashML-org#110","updated_at":"2026-08-29T05:13:10Z","url":"https://github.com/FlashML-org/FreeToken/issues/123"},"124":{"body_hash":"64a1de8500f135a3","comments":0,"kind":"issue","number":124,"repo":"FlashML-org/FreeToken","state":"closed","title":"NVFP4 checkpoints fail to load: model.safetensors.index.json is not downloaded, but the NVFP4 bank loader requires it","updated_at":"2026-09-02T02:43:20Z","url":"https://github.com/FlashML-org/FreeToken/issues/124"},"79":{"body_hash":"ba47c9bf723f23e8","comments":17,"kind":"issue","number":79,"repo":"FlashML-org/FreeToken","state":"open","title":"FreeToken Roadmap (2026)","updated_at":"2026-10-18T01:11:22Z","url":"https://github.com/FlashML-org/FreeToken/issues/79"},"82":{"body_hash":"ff59a7ff00722b01","comments":19,"kind":"issue","number":82,"repo":"FlashML-org/FreeToken","state":"open","title":"ROCm RDNA4 gfx1201 Ported. Windows only for now. Test please","updated_at":"2026-09-15T15:50:45Z","url":"https://github.com/FlashML-org/FreeToken/issues/82"}},"tracked_prs":{"112":{"base_sha":"f0abe587a11cca53bb3c37a9596fad24973ace62","body_hash":"aa539235301fea3d","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":1,"commits":3,"draft":false,"head_sha":"831d38a66bc45543411078c6805de8c42e7603fa","merged":true,"number":112,"review_comments":0,"state":"closed","title":"feat(moe): per-layer host-bank residency","updated_at":"2026-08-24T00:39:32Z","url":"https://github.com/FlashML-org/FreeToken/pull/112"},"118":{"base_sha":"e0a3bbc04652ec0622d932adcbf2772848e3c2e2","body_hash":"d73a76c00cdb2f94","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":0,"commits":1,"draft":false,"head_sha":"6241178a7550e5a0ffadd85f0eb04af0160f9988","merged":false,"number":118,"review_comments":0,"state":"open","title":"fix(scheduler): clamp admission to the actual KV pool so oversized prompts fail loudly","updated_at":"2026-08-24T00:52:33Z","url":"https://github.com/FlashML-org/FreeToken/pull/118"},"125":{"base_sha":"bd372b630a028e3faa51f4ab0ef6a98c2f2de501","body_hash":"793491958f083651","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":0,"commits":1,"draft":false,"head_sha":"cb44bdeedb7ea6ed2b37c5c49864ad40065e1985","merged":false,"number":125,"review_comments":0,"state":"open","title":"test(pinned): use mapped memory for UVA pointer check","updated_at":"2026-08-24T05:48:44Z","url":"https://github.com/FlashML-org/FreeToken/pull/125"},"129":{"base_sha":"bd372b630a028e3faa51f4ab0ef6a98c2f2de501","body_hash":"d5c4ac29d87557fc","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":3,"commits":1,"draft":false,"head_sha":"d30726e006b2708d4fbfead120985bc23c935541","merged":false,"number":129,"review_comments":0,"state":"closed","title":"fix(download): include json files when fetching weights, so NVFP4's index.json is no longer silently skipped","updated_at":"2026-09-02T02:45:39Z","url":"https://github.com/FlashML-org/FreeToken/pull/129"},"131":{"base_sha":"bd372b630a028e3faa51f4ab0ef6a98c2f2de501","body_hash":"33bc52d171e8ca00","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":8,"commits":36,"draft":false,"head_sha":"bb432e8c473c557cdef6517abccf0c129bd8c544","merged":false,"number":131,"review_comments":0,"state":"open","title":"GGUF: support all quant types, add qwen35moe","updated_at":"2026-08-29T22:27:37Z","url":"https://github.com/FlashML-org/FreeToken/pull/131"},"132":{"base_sha":"cc1f5c2c91855f2cc7787ad6b909f7e46a5d5825","body_hash":"f5b92547538a8571","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":8,"commits":11,"draft":false,"head_sha":"c6be4afb541b34f36fb811cd2f6efc01784e9e5d","merged":true,"number":132,"review_comments":7,"state":"closed","title":"feat(rocm): add RDNA3 and RDNA4 runtime foundation","updated_at":"2026-09-26T00:49:46Z","url":"https://github.com/FlashML-org/FreeToken/pull/132"},"133":{"base_sha":"377a3bc4284f032e9d30777ce099828a158e7e4c","body_hash":"31c9b77d36e57cab","checks":{"failed":[],"pending":[],"state":"pending","total":1},"comments":3,"commits":3,"draft":false,"head_sha":"5309ca655c2c7b21e09bbb91a88b94d51b941b3d","merged":false,"number":133,"review_comments":0,"state":"closed","title":"fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP","updated_at":"2026-10-08T20:32:48Z","url":"https://github.com/FlashML-org/FreeToken/pull/133"},"134":{"base_sha":"377a3bc4284f032e9d30777ce099828a158e7e4c","body_hash":"cf8d7ae82d59f62c","checks":{"failed":[],"pending":[],"state":"pending","total":1},"comments":1,"commits":2,"draft":false,"head_sha":"decc3d4e1337891fc31a45aa8a02fa383f24ae3f","merged":true,"number":134,"review_comments":0,"state":"closed","title":"fix(rocm): gate CUDA-only optional backends on ROCm","updated_at":"2026-10-09T02:55:57Z","url":"https://github.com/FlashML-org/FreeToken/pull/134"},"135":{"base_sha":"cc1f5c2c91855f2cc7787ad6b909f7e46a5d5825","body_hash":"3a0be6c359e61828","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":5,"commits":12,"draft":true,"head_sha":"3675d03e17110adfcaa532cc965e5c79895cc8f3","merged":false,"number":135,"review_comments":0,"state":"open","title":"fix(rocm): route tensor parallel communication through RCCL","updated_at":"2026-10-06T10:49:08Z","url":"https://github.com/FlashML-org/FreeToken/pull/135"},"136":{"base_sha":"cc1f5c2c91855f2cc7787ad6b909f7e46a5d5825","body_hash":"c875bbbd1cb60d21","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":5,"commits":13,"draft":true,"head_sha":"76bf2ac398464f8762148356b4a4019e3327b2a9","merged":false,"number":136,"review_comments":0,"state":"closed","title":"feat(rocm): enable native GGUF kernels on ROCm (validated on RDNA4)","updated_at":"2026-09-24T06:02:29Z","url":"https://github.com/FlashML-org/FreeToken/pull/136"},"137":{"base_sha":"f7c31e92dbf296de3a5bb1b9c5fcb58f988e3a07","body_hash":"a0597a23e4607628","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":0,"commits":1,"draft":false,"head_sha":"5eec2582ce89376e3f5016d47ff73ca1b4e8d445","merged":false,"number":137,"review_comments":0,"state":"open","title":"feat(rocm): serve on AMD GPUs through the HIP toolchain","updated_at":"2026-08-26T09:03:28Z","url":"https://github.com/FlashML-org/FreeToken/pull/137"},"27":{"base_sha":"0ab982f10905fa775962a4eddcb44caa50065251","body_hash":"e85637efb5176c06","checks":{"failed":[],"pending":[],"state":"pending","total":0},"comments":4,"commits":1,"draft":false,"head_sha":"d76d5e0105235e9f13edd534fec3dfa86f25cb15","merged":false,"number":27,"review_comments":0,"state":"closed","title":"feat(moe): partial-pin serving to beat the Windows/WSL2 pinned-memory quota (--pin-exempt-layers)","updated_at":"2026-08-24T01:28:09Z","url":"https://github.com/FlashML-org/FreeToken/pull/27"}},"version":1}

Technical baseline updated automatically: 2026-10-21T23:22:25+00:00. This comment is edited in place and does not represent a notification.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T15:02:31+00:00).

• [Issue] Windows gfx120X-all Test Artifacts: hipInfo.exe hangs / GPU not detected → cascade failure across 27 jobs: обновлён (state open→open, comments 5→6).
• Новый релевантный issue: bench_offload_cache_copy.py default sweep aborts on minimax-m2.5-marlin and skips the last four profiles.
• Новый релевантный issue: [XPU] inductor workloads can fail intermittently with UR_RESULT_ERROR_PROFILING_INFO_NOT_AVAILABLE (error 17).
• [BUG] preparing weights when converting: обновлён (state open→open, comments 3→4).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T16:56:53+00:00).

• fix(download): include json files when fetching weights, so NVFP4's index.json is no longer silently skipped: обновлены описание/review/discussion (comments 0→1, review 0→0).
• Новый релевантный issue: AssertionError: cache budget too small.
• torch.nn.LSTM.forward produces incorrect outputs (or crashes) on ROCm: обновлён (state open→open, comments 13→14).
• bench_offload_cache_copy.py default sweep aborts on minimax-m2.5-marlin and skips the last four profiles: обновлён (state open→open, comments 0→1).
• feat(server): Add OpenAI /v1/models/{model_id} endpoint: обновлён (state open→open, comments 1→2).
• [Issue]: Windows: rocm-sdk silently drops device files from the devel tree when the link target exceeds MAX_PATH: обновлён (state open→closed, comments 0→0).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T18:31:06+00:00).

• fix(download): include json files when fetching weights, so NVFP4's index.json is no longer silently skipped: обновлены описание/review/discussion (comments 1→2, review 0→0).
• Новый релевантный pr: fix(bench): skip an unsupported slot count instead of aborting the sweep.
• Новый релевантный issue: Eval bug: Windows HIP gfx1201: Qwen3.6-35B-A3B NVFP4 loads, then first MUL_MAT fails with ROCm invalid argument.
• bench_offload_cache_copy.py default sweep aborts on minimax-m2.5-marlin and skips the last four profiles: обновлён (state open→open, comments 1→5).
• Новый релевантный pr: fix(kernel): clear the latched CUDA error before raising in pinned_tensor.
• Новый релевантный pr: CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3.
• Новый релевантный issue: Eval bug: inline </think> tag contained in thinking block gets mis-parsed as the end of reasoning content.
• Новый релевантный issue: Misc. bug: RPC + Metal: main node retains the full model with default mmap load mode, capping split models to a single node's memory.
• Новый релевантный pr: AVX2: Speed up large batch size prompt processing of IQ models.
• Новый релевантный pr: CUDA: extend MOE fusion to specdec, earlier MOE glu fusion and topk-router fusion were restricted to 1 token.
• Новый релевантный issue: Feature Request: server in Router Mode: Multiple presets for the same model (loaded once).
• Новый релевантный issue: SYCL: MoE expert tensors never get the Q8_0 reorder path (opt_for_reorder_id).
• Новый релевантный issue: HIP/ROCm backend produces corrupted output on gfx1151: two dense architectures fail, Vulkan is correct with byte-identical flags.
• Новый релевантный pr: CUDA: XOR swizzle flash attn K,V smem fp16 tiles.
• Новый релевантный pr: CUDA: Update Q4_K and Q5_K to use branchless computation, which stops the scale unpack being re-executed for every column in mmvq, improving perf at batch sizes > 1.
• Новый релевантный pr: vulkan: add f16 B-type matmul pipelines and warp tile size tuning for Intel coopmat1.
• Новый релевантный issue: Vulkan: batched decode throughput cliff at n_tokens=9 on many-expert MoE (fixed 8-token thresholds in MMV dispatch).
• Новый релевантный issue: Eval bug: Unexpected massive performace drop with Vulkan Flash Attention.
• Новый релевантный pr: mamba2 : Flatten in/out projections to dispatch GEMM instead of GEMV.
• Новый релевантный issue: Perf: CUDA: Mamba2 batched decode dispatch GEMV instead of GEMM.
• Новый релевантный pr: hexagon: add alternative Hexagon NPU backend implementation.
• Новый релевантный pr: tests: add mul_mat, flash-attention and copy shapes that real models emit.
• Новый релевантный pr: CUDA: Add experimental SM120 CUTLASS MoE prefill for MXFP4 and NVFP4.
• Новый релевантный issue: Misc. bug: Q6_K MUL_MAT produces wrong results on Adreno 740 OpenCL - test-backend-ops fails.
• Новый релевантный issue: Documented Linux HIP Compile Command Fails with pip-installed ROCm.
• Новый релевантный pr: opencl: use right-sized matmul kernels for speculative decoding's verification step.
• Новый релевантный issue: [CUDA][Blackwell] Qwen3.8-27B-NVFP4 decode hangs, CPU spin, no GPU work.
• Новый релевантный pr: metal : skinny-batch mul_mm tiles + deterministic split-K for small-batch decode.
• Новый релевантный issue: Vulkan dense prefill on gfx1151 (RDNA3 iGPU): ~1.8x from a larger K-quant int-mmq tile.
• Новый релевантный issue: Eval bug: MMVQ host/device nwarps mismatch causes NaN on gfx11-generic / RDNA3.5.
• Новый релевантный pr: opencl: tune the quant paths for Intel Xe-LP GPUs to improve its TG and PP performance.
• Новый релевантный pr: opencl: build one program per kernel where the compiler cannot hold several (for Adreno A6x).
• Новый релевантный pr: convert: fix get block count error for Nemotron 3 Ultra.
• Новый релевантный pr: Quant: OCP FP8 E4M3 support.
• Новый релевантный pr: ggml: add support for MXFP8 CPU.
• …ещё 11 изменений скрыто, чтобы комментарий не превращался в свалку.

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T19:01:35+00:00).

• CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3: обновлён (state open→open, comments 12→14).
• CUDA: XOR swizzle flash attn K,V smem fp16 tiles: обновлён (state open→open, comments 21→22).
• [BUG] preparing weights when converting: обновлён (state open→open, comments 4→5).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T20:49:53+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit 0f26f7ec → 6f6c8640.
• FreeToken Roadmap (2026): comments 2→3.
• Windows/ROCm on RX 9060 XT (gfx1200): packed GGUF working end-to-end (dense + MoE), root cause of the RDNA4 offload-decode TDR, benchmarks vs llama.cpp: comments 0→1.
• Новый релевантный pr: [ROCm][Inductor] Add gfx950 FlexAttention backward default configs.
• Новый релевантный pr: quantization: complete MXFP4 and NVFP4.
• feat(ci): support runtime-only host-ASAN presubmit builds and tests: обновлён (state open→open, comments 4→5).
• Новый релевантный issue: [RFC] Transition ROCm nightly wheels to wheel-based ROCm install (7.14+).
• CUDA: extend MOE fusion to specdec, earlier MOE glu fusion and topk-router fusion were restricted to 1 token: обновлён (state open→open, comments 2→3).
• Новый релевантный pr: Fix hipify skipping sources under Windows subst drives/symlinks in CUDAExtension.

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-24T22:45:58+00:00).

• Новый релевантный pr: [ROCm][c10d] Enable NCCL2 windows and enforce allocator provenance.
• Новый релевантный pr: [ROCm][Inductor] Apply tt.pointer_range guards to template kernels.
• [ROCm][Inductor] Add gfx950 FlexAttention backward default configs: обновлён (state open→open, comments 6→8).
• Новый релевантный issue: Request: RX 6000 (RDNA2) Windows support - native runtime proven (gfx1030/1031/1032).
• Новый релевантный issue: DISABLED test_workspace_runtime_integration (main.TestNVUniversalGemmEpilogueFusion).
• Новый релевантный issue: Please help support older hardware (1080/2080 series).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T01:53:38+00:00).

• FreeToken Roadmap (2026): comments 3→4.
• [ROCm] Fix hipify skipping sources under Windows subst drives/symlinks in CUDAExtension: обновлён (state open→open, comments 3→3).
• [ROCm][c10d] Enable NCCL2 windows and enforce allocator provenance: обновлён (state open→open, comments 3→3).
• Новый релевантный issue: Eval bug: Vulkan regression on AMD Strix Halo UMA between b9660 and b9672: model loads on GPU but inference shows no GPU compute.
• Новый релевантный issue: Eval bug: CUDA error: unsupported value or parameter in cublasSgemm_v2 during large context processing.
• Новый релевантный issue: Eval bug: Starting from B9992, long contexts trigger a CUDA error: the function failed to launch on the GPU.
• feat(server)!: add --gpu to choose the GPU on multi-GPU machines: обновлён (state open→closed, comments 0→1).
• Difficult to load model despite more than enough vram/ram available: обновлён (state open→open, comments 5→6).
• FlashML-org/FreeToken:main: bd372b63 → 2757bb5f — feat(server)!: add --gpu to choose the GPU on multi-GPU machines (feat(server)!: add --gpu to choose the GPU on multi-GPU machines FlashML-org/FreeToken#117).
• Upstream main: релевантные изменения bd372b63 → 2757bb5f — files: python/freetoken/engine/engine.py, python/freetoken/kernel/triton/e4m3_compat.py, python/freetoken/kernel/triton/sampling.py, python/freetoken/moe/bench_profile.py, python/freetoken/moe/benchbw.py, python/freetoken/moe/host_banks.py, python/freetoken/scheduler/scheduler.py.

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T03:13:46+00:00).

• FreeToken Roadmap (2026): comments 4→5.
• Новый релевантный issue: GGUF loader only supports gemma4 — Qwen2 / Qwen2.5 GGUF fails with "architecture 'qwen2' is not supported".
• CUDA: XOR swizzle flash attn K,V smem fp16 tiles: обновлён (state open→open, comments 22→23).
• Новый релевантный issue: Eval bug: Excessive compute buffer reservation in MTP draft context on ROCm HIP unnecessarily reduces fitted context size.
• Prefill hangs on RTX 4060 Ti (SM89) for prompts beyond warmup length: обновлён (state open→open, comments 2→3).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T04:55:02+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit 6f6c8640 → b2f84751.
• Новый релевантный issue: Add explicit detection/support for compressed-tensors NVFP4A16 DeepSeek-V4 REAM checkpoints.
• Новый релевантный issue: Qwen3.5 MoE modelopt_mixed loader assumes shared experts use routed-expert NVFP4 quantization.
• [Bug]: Detect hardware fails with cudaErrorNotSupported on NVIDIA RTX 4060 Lapt: обновлён (state open→closed, comments 2→2).
• GGUF loader only supports gemma4 — Qwen2 / Qwen2.5 GGUF fails with "architecture 'qwen2' is not supported": обновлён (state open→closed, comments 0→1).
• CUDA: extend MOE fusion to specdec, earlier MOE glu fusion and topk-router fusion were restricted to 1 token: обновлён (state open→open, comments 3→4).
• Новый релевантный issue: Eval bug: MTP retains inter-request state causing non-deterministic output and model degradation (Qwen3.6-35B-A3B-MTP).
• Новый релевантный issue: Arbitrary Code Execution in PyTorch via torch.package.PackageImporter.load_pickle.
• Новый релевантный issue: torch.distributions.kl_divergence Fails with MultivariateNormal in Dynamo Due to _infer_size Type Error.
• Новый релевантный issue: Model catalog shows models the local machine cannot run (no capability pre-check before display).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T07:09:50+00:00).

• Новый релевантный pr: ggml-cuda: enable MMA FlashAttention for head dim 256 on AMD RDNA.
• Новый релевантный pr: CI: fix workflow_dispatch runner fallback for PyTorch wheel _ci builds + tolerate newer ninja.
• [ROCm][inductor] Fix Inductor compiler selection for ROCm on Windows: обновлён (state open→open, comments 5→7).
• Новый релевантный pr: HIP: RDNA 4 MUL_MAT optimizations.
• Новый релевантный pr: CUDA: enable the CUB path on HIP via hipCUB.
• Новый релевантный issue: Feature request: AMD GPU support.
• Новый релевантный issue: Vulkan: model load hangs indefinitely on RDNA4 (gfx1201) with plain Unsloth Q6_K quant, or OOMs host with --direct-io.
• Новый релевантный pr: cuda : dynamic MMVQ nwarps for narrow matrices.

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T09:01:32+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit b2f84751 → dd48aacb.
• Новый релевантный pr: Make the drafter probabilistic and the target verify by rejection sampling for simple draft and MTP.
• Новый релевантный issue: Eval bug: qwen35moe (hybrid recurrent/attention) models produce empty generation when prompt exceeds ~16K tokens via llama-server (Metal).
• Новый релевантный pr: fix(engine): load prefill triton kernels at startup, not mid-request.
• Новый релевантный issue: CUDA OoM with larger --kv-reserve-tokens (workaround: empiric lower --memory-ratio).
• Новый релевантный pr: Vulkan: add hoisting support for row IDs and expert count in shaders.
• Новый релевантный pr: server: add recurrent state shrink/expand for prompt cache (#22746).
• Новый релевантный pr: vulkan : transpose A-matrix data layout for K-quant mul_mat performance.
• Новый релевантный issue: Silent GTT fallback when context + MTP exceeds VRAM — no error at load, massive throughput collapse at runtime.
• Новый релевантный issue: Eval bug: --batch-size and --ubatch-size ignored when set to 1024/1024.
• Новый релевантный issue: draft-mtp speculative decoding not activating on Turing (sm_75) with hybrid SSM+attention model (Qwen3.6-35B-A3B).
• Новый релевантный pr: WIP: kv-cache apply_ubtach change from O(n) to O(1).
• Новый релевантный issue: Eval bug: vulkan performance drop pp512 560 -> 180.
• Новый релевантный pr: llama + spec: MTP Support.
• CUDA: enable the CUB path on HIP via hipCUB: обновлён (state open→open, comments 46→47).

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T10:51:05+00:00).

• HIP/ROCm backend produces corrupted output on gfx1151: two dense architectures fail, Vulkan is correct with byte-identical flags: обновлён (state open→open, comments 7→8).
• Eval bug: Windows HIP gfx1201: Qwen3.6-35B-A3B NVFP4 loads, then first MUL_MAT fails with ROCm invalid argument: обновлён (state open→open, comments 0→4).
• Новый релевантный issue: Long prefill (>48k tokens) crashes the whole server — CUDA driver error: device not ready in fla/l2norm kernel.
• Новый релевантный issue: [Issue] Linux gfx103X: GPU not accessible to runner — rocminfo exits 8 / "No GPU found matching.
• Новый релевантный issue: torch.cholesky_inverse fails with HIPBLAS_STATUS_ALLOC_FAILED at batch 20 on ROCm due to fixed hipBLAS workspace.
• Новый релевантный pr: [ROCm] Enable CK SDPA on gfx1200/gfx1201 (RDNA4).
• Новый релевантный issue: [Issue]: Access violation (0xC0000005) in bmm/baddbmm on gfx1201 (RX 9070 XT) under Windows, crashes in both hipBLASLt and rocBLAS.
• Новый релиз: ggml-org/llama.cpp v0.3.0.

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T11:11:49+00:00).

• Новый явно отслеживаемый external issue: [Issue]: rocBLAS bf16/f16 GEMM fails with hipErrorInvalidValue on 2nd call, gfx1201 / Windows / ROCm 7.14. High relevance.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 8024:3880E4:B12B:25494:6A8D788F and timestamp 2026-08-25 11:12:15 UTC. For…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T13:13:08+00:00).

• feat(rocm): add RDNA3 and RDNA4 runtime foundation: обновлены описание/review/discussion (comments 1→2, review 0→0).
• Новый релевантный issue: Feature: Predictive Expert Offloading.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID F440:3A3DE2:D8F1C6:2D7E43E:6A8D950E and timestamp 2026-08-25 13:13:50 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T15:07:11+00:00).

• FreeToken Roadmap (2026): comments 5→6.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 0407:272BC6:1AC4A9F:5964A5E:6A8DAFBB and timestamp 2026-08-25 15:07:39 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T16:57:34+00:00).

• Model catalog shows models the local machine cannot run (no capability pre-check before display): обновлён (state open→open, comments 0→1).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 283F:2B3BAA:2009A4A:6A360EB:6A8DC9A1 and timestamp 2026-08-25 16:58:09 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T18:58:44+00:00).

• Новый релевантный pr: fix(deps): use PyTorch 2.13 to fix SM89 hangs.
• Новый релевантный issue: Apodex-1.1-mini-NVFP4 fails: dense_quant hardcoded to NVFP4 when experts are NVFP4, but shared_expert is FP8 (modelopt MIXED_PRECISION).
• Difficult to load model despite more than enough vram/ram available: обновлён (state open→open, comments 6→7).
• Новый релевантный issue: FP8 row-wise _scaled_mm hangs on SM89 with PyTorch 2.11.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID C819:1EE7B3:2EC3E0B:9D340E1:6A8DE60D and timestamp 2026-08-25 18:59:25 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T20:47:37+00:00).

• fix(deps): use PyTorch 2.13 to fix SM89 hangs: обновлён (state open→open, comments 0→0).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 7830:16ACDE:3C41178:CA29304:6A8DFF87 and timestamp 2026-08-25 20:48:07 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-25T22:46:42+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit dd48aacb → 9952a39d.
• Maxritz/FreeToken-rocm-test:main: 45675e47 → 3a5616ec — Merge pull request Tweak globals so that LLVM builds with the proper shared library settings. ROCm/TheRock#1 from Castoff995/codex/gfx1201-offload-decode.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID CC18:2EB04F:4D5504F:100358CF:6A8E1B72 and timestamp 2026-08-25 22:47:14 UT…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T01:59:36+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit 9952a39d → 7da6b77e.
• feat(rocm): enable native GGUF kernels on RDNA4: обновлены описание/review/discussion (comments 0→1, review 0→0).
• Requests longer than the KV pool are queued forever with no error — and --moe-cache-auto leaves only ~8k tokens of KV on a 96GB GPU: comments 1→2.
• Новый релевантный pr: feat(kvcache): 8-bit KV cache (q8_0 / fp8_e4m3) behind --kv-cache-dtype.
• FlashML-org/FreeToken:main: 2757bb5f → f7c31e92 — fix(fla): stop l2norm recompiling per token count.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 7431:947DB:5671D36:1210F83B:6A8E48B1 and timestamp 2026-08-26 02:00:17 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T03:19:34+00:00).

• feat(rocm): add RDNA3 and RDNA4 runtime foundation: обновлены описание/review/discussion (comments 2→3, review 0→0).
• feat(rocm): serve on AMD GPUs through the HIP toolchain: новый commit 7e1d7421 → 5eec2582.
• Новый релевантный issue: Qwen3.6-35B-A3B hits KV-cache limit during prolonged reasoning on LeetCode problem.
• Model catalog shows models the local machine cannot run (no capability pre-check before display): обновлён (state open→open, comments 1→2).
• Новый релевантный issue: [AppImage] Global LD_LIBRARY_PATH export breaks system-curl child processes (engine installer fails on non-Ubuntu distros).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID B00B:2DC8AF:65DCB0D:15086008:6A8E5B63 and timestamp 2026-08-26 03:20:03 UT…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T04:57:21+00:00).

• Новый релевантный issue: [Linux] AppImage blank window / WebKitWebProcess abort: "Could not create default EGL display: EGL_BAD_PARAMETER" (bundled libwayland-client on Mesa 25+ hosts).
• Новый релевантный issue: Loading nvidia/Qwen3.6-35B-A3B-NVFP4 fails: "Incomplete fp8 fusions: ['model.layers.37.linear_attn.in_proj_qkvz']".
• Qwen3.6-35B-A3B hits KV-cache limit during prolonged reasoning on LeetCode problem: обновлён (state open→open, comments 0→0).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID D828:1EDC:4C5ACCD:FD5C08A:6A8E724C and timestamp 2026-08-26 04:57:48 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T07:10:22+00:00).

• GGUF: support all quant types, add qwen35moe: новый commit 7da6b77e → bb432e8c.
• FreeToken Roadmap (2026): body edited.
• Новый релевантный pr: qwen3_5_moe: load mixed-precision compressed-tensors checkpoints (unsloth NVFP4+FP8).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID CAC4:33DF9:7ABF9D8:19628807:6A8E9187 and timestamp 2026-08-26 07:11:03 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T09:04:06+00:00).

• Bug in PyTorch 2.11 causes SM89 hangs: обновлён (state open→open, comments 0→0).
• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 2→3).
• FlashML-org/FreeToken:main: f7c31e92 → 9ef36513 — chore(assets): update wechat group QR code.
• Maxritz/FreeToken-rocm-test:main: 3a5616ec → 89f14790 — Merge pull request Enable the GH build. ROCm/TheRock#2 from PialGhosh2233/main.
• PialGhosh2233/FreeToken-rocm-gfx1200:gfx1200-moe-gguf: 544b83df → a338f49b — docs: Gemma-4-26B-A4B QAT q4_0 GGUF results on gfx1200.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 0846:30A2EB:30974E:A35D01:6A8EAC23 and timestamp 2026-08-26 09:04:35 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T10:52:44+00:00).

• Новый релевантный pr: feat(models): add Qwen3-Next-80B-A3B support.
• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 3→5).
• Новый релиз: ROCm/TheRock therock-10.0.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID B008:2BB36A:73A408:77B54F:6A8EC59E and timestamp 2026-08-26 10:53:18 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T13:18:13+00:00).

• FreeToken Roadmap (2026): comments 6→7.
• Feature request: AMD GPU support: обновлён (state open→open, comments 11→12).
• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 5→7).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 0400:15153F:A5D123:230714B:6A8EE7C0 and timestamp 2026-08-26 13:18:56 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-08-26T15:55:00+00:00).

• [Linux] AppImage blank window / WebKitWebProcess abort: "Could not create default EGL display: EGL_BAD_PARAMETER" (bundled libwayland-client on Mesa 25+ hosts): обновлён (state open→open, comments 0→1).
• feat(models): support TP for qwen3_5_moe: обновлён (state open→open, comments 7→8).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID D82D:361ECC:232D6C:2C6E23:6A8F0C76 and timestamp 2026-08-26 15:55:34 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-15T21:25:10+00:00).

• feat(gguf): serve GGUF checkpoints on vendored ggml-cuda kernels: обновлён (state open→open, comments 3→4).
• Новый релиз: ggml-org/llama.cpp v0.7.0.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID ABCB:298B0E:2E366AA:961F5DC:6AD144D6 and timestamp 2026-10-15 21:25:43 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-16T00:25:40+00:00).

• feat(moe): support gsq q2_0 routed experts: обновлён (state open→closed, comments 1→1).
• 🚧 [WIP] MoE cache and CPU expert refactor: обновлён (state open→open, comments 0→0).
• feat(gguf): serve GGUF checkpoints on vendored ggml-cuda kernels: обновлён (state open→closed, comments 4→5).
• FlashML-org/FreeToken:main: 0bca88f7 → e5d3bd3f — feat(gguf): serve GGUF checkpoints on vendored ggml-cuda kernels (feat(gguf): serve GGUF checkpoints on vendored ggml-cuda kernels FlashML-org/FreeToken#690).
• Upstream main: релевантные изменения 0bca88f7 → e5d3bd3f — commits: feat(gguf): serve GGUF checkpoints on vendored ggml-cuda kernels (feat(gguf): serve GGUF checkpoints on vendored ggml-cuda kernels FlashML-org/FreeToken#690) | files: pyproject.toml, python/freetoken/engine/config.py, python/freetoken/engine/engine.py, python/freetoken/moe/expert_banks.py, python/freetoken/moe/fused_q4_0.py, python/freetoken/moe/offload_cache.py, python/freetoken/moe/offload_kernels.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 283E:2B2DB2:1C84F8:5D92FB:6AD16F1F and timestamp 2026-10-16 00:26:07 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-16T07:12:42+00:00).

• 🚧 [WIP] MoE cache, CPU experts and GGUF support: обновлён (state open→open, comments 0→0).
• Новый релевантный pr: fix(engine): upload cuda graphs at capture so a stream wait cannot block the first launch.
• Новый релевантный pr: feat(gguf): run experts whose gate_up and down mix float and quantized types.
• Новый релевантный pr: feat(moe): serve experts of several formats through per-format slot pools.
• Новый релевантный pr: fix(engine): load prefill triton kernels at startup, not mid-request.
• Новый релевантный pr: feat(gguf): serve Qwen3.5 dense and MoE GGUF checkpoints.
• Новый релевантный pr: Fix GGUF user-defined token atomicity.
• Новый релевантный pr: build(deps): upgrade transformers to 5.18.
• fix(rocm): clamp QSA KV tile to the device LDS budget on RDNA (gfx11/gfx12): обновлён (state open→closed, comments 1→3).
• FlashML-org/FreeToken:main: e5d3bd3f → 9b68e007 — fix(engine): upload cuda graphs at capture so a stream wait cannot block the first launch (fix(engine): upload cuda graphs at capture so a stream wait cannot block the first launch FlashML-org/FreeToken#697).
• Upstream main: релевантные изменения e5d3bd3f → 9b68e007 — commits: fix(rocm): clamp QSA KV tile to the device LDS budget on RDNA (fix(rocm): clamp QSA KV tile to the device LDS budget on RDNA (gfx11/gfx12) FlashML-org/FreeToken#587); feat(gguf): serve Qwen3.5 dense and MoE GGUF checkpoints (feat(gguf): serve Qwen3.5 dense and MoE GGUF checkpoints FlashML-org/FreeToken#701) | files: pyproject.toml, python/freetoken/engine/graph.py, python/freetoken/kernel/triton/qsa/attend.py, python/freetoken/models/qwen3_5_moe/init.py, python/freetoken/models/qwen3_5_moe/gdn.py, python/freetoken/models/qwen3_5_moe/gguf.py, python/freetoken/models/qwen3_5_moe/model.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 10AC:36200D:271B76:29E4F5:6AD1CE8A and timestamp 2026-10-16 07:13:14 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-16T13:16:54+00:00).

• Новый релевантный pr: feat(models): add nvidia/Qwen3.8-Flash-Next-FP4-Local checkpoint support.
• 🚧 [WIP] MoE cache, CPU experts and GGUF support: обновлён (state open→open, comments 0→0).
• Новый релевантный pr: feat(gguf): serve Qwen2, Qwen3 and Qwen3 MoE GGUF checkpoints.
• Новый релевантный pr: feat(cpu-moe): serve GGUF experts on the cpu and hybrid backends through ggml-cpu.
• feat(moe): serve experts of several formats through per-format slot pools: обновлён (state open→open, comments 1→2).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 2C30:3A280D:38F6CE4:B9C00ED:6AD223F2 and timestamp 2026-10-16 13:17:38 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-16T18:04:38+00:00).

• Новый релевантный issue: qwen3_coder: non-streaming cuts a tool-call argument at a literal </tool_call>, or <parameter= inside the value.
• Новый релевантный issue: qwen3_coder: a bare <function=NAME> block without <tool_call> is a call plus raw markup in non-streaming, and only text in streaming.
• Новый релевантный issue: qwen3_coder: non-streaming drops the text between two tool calls (streaming keeps it).
• Новый релевантный issue: qwen3 reasoning parser: streamed content starts with "\n\n" after and streamed reasoning keeps its trailing newline; non-streaming strips both.
• Новый релевантный issue: Reasoning parsers re-enter reasoning on a <think> after </think> when streaming, and delete it when not (qwen3, glm, minimax, deepseekv32, gemma4, minimax_m3).
• Новый релевантный issue: Design: how chat logprobs should behave when a reasoning or tool parser is active.
• Новый релевантный issue: Streaming /v1/messages: message_start reports input_tokens 0 and no cache fields.
• Новый релевантный issue: qwen3_coder: max_tokens truncation leaks raw tool-call markup into content and arguments.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 7011:1FEAE9:99C520:A50164:6AD2675A and timestamp 2026-10-16 18:05:14 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-16T21:12:12+00:00).

• Новый релевантный pr: fix(kernel): correct fp4 cache scale rounding.
• Новый релевантный pr: feat(gguf): serve GLM-5.3-Flash GGUF checkpoints on glm5-next.
• 🚧 [WIP] MoE cache, CPU experts and GGUF support: обновлён (state open→open, comments 0→0).
• feat(gguf): serve Qwen2, Qwen3 and Qwen3 MoE GGUF checkpoints: обновлён (state open→closed, comments 0→0).
• FlashML-org/FreeToken:main: 9b68e007 → d5eb18fe — fix(kernel): correct fp4 cache scale rounding (fix(kernel): correct fp4 cache scale rounding FlashML-org/FreeToken#720).
• Upstream main: релевантные изменения 9b68e007 → d5eb18fe — commits: feat(gguf): serve Qwen2, Qwen3 and Qwen3 MoE GGUF checkpoints (feat(gguf): serve Qwen2, Qwen3 and Qwen3 MoE GGUF checkpoints FlashML-org/FreeToken#707); fix(kernel): correct fp4 cache scale rounding (fix(kernel): correct fp4 cache scale rounding FlashML-org/FreeToken#720) | files: python/freetoken/kernel/triton/dsv41/pack.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 8049:D6E0:927475:1DB17E5:6AD29356 and timestamp 2026-10-16 21:12:54 UTC. F…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-17T00:30:53+00:00).

• Новый релевантный pr: perf(attention): copy triton metadata without blocking the host.
• 🚧 [WIP] MoE cache, CPU experts and GGUF support: обновлён (state open→open, comments 0→0).
• feat(cpu-moe): serve GGUF experts on the cpu and hybrid backends through ggml-cpu: обновлён (state open→open, comments 0→0).
• feat(moe): serve experts of several formats through per-format slot pools: обновлён (state open→closed, comments 2→2).
• Новый релевантный pr: feat(gguf): serve Muse-Glimmer GGUF checkpoints.
• Новый релевантный pr: feat(gguf): serve MiniMax-M2 and MiniMax-M3 GGUF checkpoints.
• Новый релевантный pr: perf(kernel): rewrite the triton nvfp4 moe prefill gemm.
• FlashML-org/FreeToken:main: d5eb18fe → bc0e2feb — perf(attention): copy triton metadata without blocking the host (perf(attention): copy triton metadata without blocking the host FlashML-org/FreeToken#726).
• Upstream main: релевантные изменения d5eb18fe → bc0e2feb — commits: feat(moe): serve experts of several formats through per-format slot pools (feat(moe): serve experts of several formats through per-format slot pools FlashML-org/FreeToken#700); perf(attention): copy triton metadata without blocking the host (perf(attention): copy triton metadata without blocking the host FlashML-org/FreeToken#726) | files: python/freetoken/engine/engine.py, python/freetoken/models/qwen3_5_moe/weight.py, python/freetoken/moe/benchbw.py, python/freetoken/moe/cpu_executor.py, python/freetoken/moe/expert_banks.py, python/freetoken/moe/expert_pieces.py, python/freetoken/moe/host_banks.py, python/freetoken/moe/offload_cache.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 0403:2A9060:678CA0:15436B2:6AD2C1DF and timestamp 2026-10-17 00:31:27 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-17T06:58:52+00:00).

• Новый релевантный pr: perf(kernel): split triton decode attention by context length and sm count.
• feat(cpu-moe): serve GGUF experts on the cpu and hybrid backends through ggml-cpu: обновлён (state open→closed, comments 0→1).
• perf(kernel): rewrite the triton nvfp4 moe prefill gemm: обновлён (state open→closed, comments 0→1).
• 🚧 [WIP] MoE cache, CPU experts and GGUF support: обновлён (state open→open, comments 0→0).
• feat(gguf): serve MiniMax-M2 and MiniMax-M3 GGUF checkpoints: обновлён (state open→closed, comments 0→0).
• feat(gguf): serve Muse-Glimmer GGUF checkpoints: обновлён (state open→closed, comments 0→0).
• FlashML-org/FreeToken:main: bc0e2feb → 11c1adb8 — perf(kernel): split triton decode attention by context length and sm count (perf(kernel): split triton decode attention by context length and sm count FlashML-org/FreeToken#729).
• Upstream main: релевантные изменения bc0e2feb → 11c1adb8 — commits: perf(kernel): rewrite the triton nvfp4 moe prefill gemm (perf(kernel): rewrite the triton nvfp4 moe prefill gemm FlashML-org/FreeToken#723); feat(cpu-moe): serve GGUF experts on the cpu and hybrid backends through ggml-cpu (feat(cpu-moe): serve GGUF experts on the cpu and hybrid backends through ggml-cpu FlashML-org/FreeToken#704); perf(kernel): split triton decode attention by context length and sm count (perf(kernel): split triton decode attention by context length and sm count FlashML-org/FreeToken#729) | files: python/freetoken/engine/engine.py, python/freetoken/kernel/triton/attention.py, python/freetoken/kernel/triton/fp8_blockscale_moe.py, python/freetoken/kernel/triton/nvfp4_fused_moe.py, python/freetoken/moe/benchbw.py, python/freetoken/moe/cpu_executor.py, python/freetoken/moe/expert_banks.py, python/freetoken/moe/fused_nvfp4.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 6802:0C5F:19C1E:2D9EA:6AD31CC6 and timestamp 2026-10-17 06:59:18 UTC. For …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-17T12:13:00+00:00).

• FreeToken Roadmap (2026): body edited.
• Новый релевантный pr: feat(gguf): serve GLM-4.5, GLM-4.6 and GLM-4.7 GGUF checkpoints.
• Новый релевантный issue: perf(moe): fused triton router for sigmoid + selection-bias routing.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID C00E:2EE512:1CC3090:5D8461B:6AD36674 and timestamp 2026-10-17 12:13:40 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-17T15:39:10+00:00).

• Новый релевантный issue: Requests stall and time out above --max-running-requests 8; GPU stays at ~28% of TDP and only one CPU thread is busy.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 8401:2E87DF:2CF35B5:91FC444:6AD396B6 and timestamp 2026-10-17 15:39:34 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-17T18:31:35+00:00).

• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 2C04:68D00:D98E7:ED502:6AD3BF26 and timestamp 2026-10-17 18:32:06 UTC. For…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-17T22:21:27+00:00).

• 🚧 [WIP] MoE cache, CPU experts and GGUF support: обновлён (state open→open, comments 0→0).
• Новый релевантный pr: feat(gguf): apply the weight scales of nvfp4 gguf tensors.
• feat(gguf): serve GLM-4.5, GLM-4.6 and GLM-4.7 GGUF checkpoints: обновлён (state open→closed, comments 0→0).
• FlashML-org/FreeToken:main: 11c1adb8 → 31c6ef4f — feat(gguf): apply the weight scales of nvfp4 gguf tensors (feat(gguf): apply the weight scales of nvfp4 gguf tensors FlashML-org/FreeToken#732).
• Upstream main: релевантные изменения 11c1adb8 → 31c6ef4f — commits: feat(gguf): apply the weight scales of nvfp4 gguf tensors (feat(gguf): apply the weight scales of nvfp4 gguf tensors FlashML-org/FreeToken#732) | files: python/freetoken/engine/engine.py, python/freetoken/models/qwen3_5_moe/gguf.py, python/freetoken/moe/cpu_executor.py, python/freetoken/moe/offload_cache.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 77C1:1D369:133F5E:3F1026:6AD3F50E and timestamp 2026-10-17 22:22:06 UTC. F…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-18T04:37:12+00:00).

• FreeToken Roadmap (2026): body edited.
• 🚧 [WIP] MoE cache, CPU experts and GGUF support: обновлён (state open→open, comments 0→0).
• Новый релевантный issue: FP8 block-quantized MoE models have no usable kernel under moe_strategy=offload (0.1.3).
• Новый релевантный pr: perf(moe): speed up the nvfp4 moe prefill gemm on rtx gpus.
• FlashML-org/FreeToken:main: 31c6ef4f → ac6a15de — perf(moe): speed up the nvfp4 moe prefill gemm on rtx gpus (perf(moe): speed up the nvfp4 moe prefill gemm on rtx gpus FlashML-org/FreeToken#733).
• Upstream main: релевантные изменения 31c6ef4f → ac6a15de — commits: perf(moe): speed up the nvfp4 moe prefill gemm on rtx gpus (perf(moe): speed up the nvfp4 moe prefill gemm on rtx gpus FlashML-org/FreeToken#733) | files: python/freetoken/kernel/triton/nvfp4_fused_moe.py, python/freetoken/moe/fused_nvfp4.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID B425:17C709:D7822E:2BB28B1:6AD44D12 and timestamp 2026-10-18 04:37:38 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-18T11:10:04+00:00).

• Новый релевантный pr: feat(qwen4_exp): stream the PLE n-gram table from disk.
• Новый релевантный pr: perf(moe): prefill nvfp4 moe experts on nvfp4 activations on sm_120.
• Новый релевантный pr: perf(moe): pipeline the nvfp4 prefill scale loads on geforce sm_86 / sm_89.
• 🚧 [WIP] MoE cache, CPU experts and GGUF support: обновлён (state open→open, comments 0→0).
• Новый релевантный pr: feat(moe): bench hybrid fetch fractions per moe layer kind.
• Requests stall and time out above --max-running-requests 8; GPU stays at ~28% of TDP and only one CPU thread is busy: обновлён (state open→open, comments 2→3).
• FlashML-org/FreeToken:main: ac6a15de → 8f49e0af — feat(moe): bench hybrid fetch fractions per moe layer kind (feat(moe): bench hybrid fetch fractions per moe layer kind FlashML-org/FreeToken#734).
• Upstream main: релевантные изменения ac6a15de → 8f49e0af — commits: feat(moe): bench hybrid fetch fractions per moe layer kind (feat(moe): bench hybrid fetch fractions per moe layer kind FlashML-org/FreeToken#734) | files: python/freetoken/engine/engine.py, python/freetoken/moe/bench_profile.py, python/freetoken/moe/benchbw.py, python/freetoken/moe/offload_cache.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 7C45:3CFF03:F5E763:3219735:6AD4A927 and timestamp 2026-10-18 11:10:31 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-18T15:53:56+00:00).

• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID F820:3A1AC7:BBDCF9:CB1F37:6AD4EBB3 and timestamp 2026-10-18 15:54:27 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-18T18:38:59+00:00).

• Новый релиз: apache/tvm-ffi v0.1.15.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 8044:1C8625:21DB03F:6DA3A3C:6AD5126B and timestamp 2026-10-18 18:39:39 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-18T22:26:50+00:00).

• perf(moe): pipeline the nvfp4 prefill scale loads on geforce sm_86 / sm_89: обновлён (state open→closed, comments 0→0).
• FlashML-org/FreeToken:main: 8f49e0af → fd410ee6 — perf(moe): pipeline the nvfp4 prefill scale loads on geforce sm_86 / sm_89 (perf(moe): pipeline the nvfp4 prefill scale loads on geforce sm_86 / sm_89 FlashML-org/FreeToken#735).
• Upstream main: релевантные изменения 8f49e0af → fd410ee6 — commits: perf(moe): pipeline the nvfp4 prefill scale loads on geforce sm_86 / sm_89 (perf(moe): pipeline the nvfp4 prefill scale loads on geforce sm_86 / sm_89 FlashML-org/FreeToken#735) | files: python/freetoken/kernel/triton/nvfp4_fused_moe.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 5415:3CC12C:32848AD:A44630A:6AD547C9 and timestamp 2026-10-18 22:27:21 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-19T04:44:03+00:00).

• Новый релевантный pr: perf(linear): run nvfp4 linears on nvfp4 activations on sm_120.
• perf(moe): prefill nvfp4 W4A4 moe kernel on sm_120: обновлён (state open→closed, comments 0→1).
• Новый релевантный pr: fix(quant)!: keep nvfp4 global scales in fp32.
• Новый релевантный issue: Support for orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4.
• FlashML-org/FreeToken:main: fd410ee6 → e4f69d9e — perf(moe): prefill nvfp4 moe experts on nvfp4 activations on sm_120 (perf(moe): prefill nvfp4 W4A4 moe kernel on sm_120 FlashML-org/FreeToken#736).
• Upstream main: релевантные изменения fd410ee6 → e4f69d9e — commits: perf(moe): prefill nvfp4 moe experts on nvfp4 activations on sm_120 (perf(moe): prefill nvfp4 W4A4 moe kernel on sm_120 FlashML-org/FreeToken#736) | files: python/freetoken/kernel/triton/nvfp4_act_quant.py, python/freetoken/moe/expert_pieces.py, python/freetoken/moe/fused_nvfp4.py, python/freetoken/moe/legacy_format.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 748A:16D6DB:64632D0:14590CC9:6AD5A03C and timestamp 2026-10-19 04:44:44 UT…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-19T13:21:39+00:00).

• perf(linear): run nvfp4 linears on nvfp4 activations on sm_120: обновлён (state open→closed, comments 0→0).
• Новый релевантный pr: fix(muse_glimmer)!: keep the attention gate out of the qkv fusion.
• FlashML-org/FreeToken:main: e4f69d9e → 4d2e2e21 — perf(linear): run nvfp4 linears on nvfp4 activations on sm_120 (perf(linear): run nvfp4 linears on nvfp4 activations on sm_120 FlashML-org/FreeToken#737).
• Upstream main: релевантные изменения e4f69d9e → 4d2e2e21 — commits: perf(linear): run nvfp4 linears on nvfp4 activations on sm_120 (perf(linear): run nvfp4 linears on nvfp4 activations on sm_120 FlashML-org/FreeToken#737) | files: python/freetoken/kernel/triton/nvfp4_act_quant.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 680C:2D4920:D6C998:2CAAE03:6AD6198A and timestamp 2026-10-19 13:22:18 UTC.…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-19T19:57:37+00:00).

• Новый релевантный issue: [Feature] Hybrid radix cache: greedy outputs depend on request history even when cached_tokens is 0 (request: cache flush endpoint / deterministic mode).
• fix(quant)!: keep nvfp4 global scales in fp32: обновлён (state open→closed, comments 0→1).
• Новый релевантный issue: [Bug] Qwen3.5/3.6: Gemma-style (1 + w) RMSNorm weights are folded into bf16 at load time; transformers applies (1 + w) in fp32.
• Новый релевантный issue: [Bug] sm_120: the first TMA kernel launch reserves ~14 KB of stack per resident thread (1.4 GiB on 16 GB cards), and the VRAM budget does not account for it.
• Новый релевантный issue: Regression: first prefill OOMs on RTX 5090 with Qwen3.8-Flash-Next NVFP4 (TMA prefill kernel's first launch allocates ~3.7 GiB).
• FlashML-org/FreeToken:main: 4d2e2e21 → 0db5023b — fix(quant)!: keep nvfp4 global scales in fp32 (fix(quant)!: keep nvfp4 global scales in fp32 FlashML-org/FreeToken#739).
• Upstream main: релевантные изменения 4d2e2e21 → 0db5023b — commits: fix(quant)!: keep nvfp4 global scales in fp32 (fix(quant)!: keep nvfp4 global scales in fp32 FlashML-org/FreeToken#739) | files: python/freetoken/kernel/triton/nvfp4_dequant.py, python/freetoken/kernel/triton/nvfp4_fused_moe.py, python/freetoken/kernel/triton/nvfp4_linear.py, python/freetoken/models/qwen3_5_moe/weight.py, python/freetoken/moe/benchbw.py, python/freetoken/moe/cpu_executor.py, python/freetoken/moe/offload_cache.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 7BC3:26F93F:3B2C685:BF31FA7:6AD67655 and timestamp 2026-10-19 19:58:13 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-19T23:44:01+00:00).

• Новый релевантный pr: feat(gguf): add scripts that check which gguf files load.
• Новый релевантный pr: fix(qwen3_5_moe)!: apply the zero-centred norm weights' +1 in fp32 at runtime.
• feat(gguf): serve GLM-5.3-Flash GGUF checkpoints on glm5-next: обновлён (state open→open, comments 1→2).
• Новый релевантный issue: [Feature] Optional host-resident token-embedding table (pinned, mapped, UVA lookups): ~1 GB of VRAM back for offloaded expert slots.
• Новый релевантный issue: perf(moe): prefetch the next layer's experts from a router lookahead during offloaded decode.
• Новый релевантный issue: perf(moe): run the shared expert beside the routed-expert miss copy in offloaded decode.
• Новый релевантный issue: [Perf] Offloaded decode: overlap the expert-miss gather with the same layer's shared expert, event-ordered inside the CUDA graph (bit-exact implementation).
• Новый релевантный issue: [Perf] Offloaded prefill: keep buffer invalidation off the prefill copy stream and give that stream top priority (short-prompt H2D 21 -> 42 GB/s under Nsight).
• Новый релевантный issue: An offloaded prefill streams every expert of each layer, however short the chunk.
• Новый релевантный issue: [Perf] Offloaded decode: gather missed experts with a few TMA bulk-copy CTAs (cp.async.bulk, .shared::cta) instead of 48 x 1024 LDG threads.
• Новый релевантный issue: [Perf] Layer-major prefill for offloaded MoE: stream each expert layer once per prompt, not once per chunk (qwen3_5_moe).
• seed and chat logprobs are silently ignored (200 with no effect); proposal for per-request seeded sampling: обновлён (state open→open, comments 0→1).
• Новый релевантный issue: A stream that is silent through a long prefill is dropped by Node clients and prefilled for nobody.
• Новый релевантный issue: [Bug] Context overflow is checked only after tokenization in the worker: a streamed request gets 200 plus an in-stream error, and an oversized max_tokens is silently clamped.
• [Bug] sm_120: the first TMA kernel launch reserves ~14 KB of stack per resident thread (1.4 GiB on 16 GB cards), and the VRAM budget does not account for it: обновлён (state open→closed, comments 0→2).
• Regression: first prefill OOMs on RTX 5090 with Qwen3.8-Flash-Next NVFP4 (TMA prefill kernel's first launch allocates ~3.7 GiB): обновлён (state open→open, comments 1→4).
• FlashML-org/FreeToken:main: 0db5023b → 5596e6b3 — feat(gguf): add scripts that check which gguf files load (feat(gguf): add scripts that check which gguf files load FlashML-org/FreeToken#756).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 9C49:3DF140:31A1A29:9E06A45:6AD6AB6B and timestamp 2026-10-19 23:44:43 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-20T04:54:09+00:00).

• feat(gguf): serve GLM-5.3-Flash GGUF checkpoints on glm5-next: обновлён (state open→open, comments 2→3).
• Новый релевантный pr: feat(mimo): add text-only MiMo v2.6 Flash support.
• Новый релевантный pr: fix(qwen4_exp): load gguf files with a quantized indexer.
• [Bug] Qwen3.5/3.6: Gemma-style (1 + w) RMSNorm weights are folded into bf16 at load time; transformers applies (1 + w) in fp32: обновлён (state open→closed, comments 0→0).
• fix(qwen3_5_moe)!: apply the zero-centred norm weights' +1 in fp32 at runtime: обновлён (state open→closed, comments 0→0).
• Support for orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4: обновлён (state open→closed, comments 0→0).
• Новый релевантный pr: fix(qwen4_exp): read weights under the checkpoint's quant config.
• FlashML-org/FreeToken:main: 5596e6b3 → bb15d1d3 — fix(server): accept localhost authority in local file URIs (fix(server): accept localhost authority in local file URIs FlashML-org/FreeToken#758).
• Upstream main: релевантные изменения 5596e6b3 → bb15d1d3 — commits: fix(qwen3_5_moe)!: apply the zero-centred norm weights' +1 in fp32 at runtime (fix(qwen3_5_moe)!: apply the zero-centred norm weights' +1 in fp32 at runtime FlashML-org/FreeToken#755) | files: python/freetoken/models/qwen3_5_moe/attention.py, python/freetoken/models/qwen3_5_moe/gguf.py, python/freetoken/models/qwen3_5_moe/model.py, python/freetoken/models/qwen3_5_moe/weight.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 9411:3EEDE3:24D4DC:75E89A:6AD6F40E and timestamp 2026-10-20 04:54:38 UTC. …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-20T11:41:31+00:00).

• perf(moe): run the shared expert beside the routed-expert miss copy in offloaded decode: обновлён (state open→open, comments 1→2).
• Новый релевантный issue: [engine bug] Decode throughput degrades from 60-120 to 11-25 tok/s over runtime; engine-global, cumulative, only reset by restart.
• Новый релевантный pr: fix(tests): support ROCm in AIME GPU memory checks.
• 🚧 [WIP] MoE cache, CPU experts and GGUF support: обновлён (state open→open, comments 0→0).
• FlashML-org/FreeToken:main: bb15d1d3 → cc56b6bd — feat(gguf): serve GLM-5.3-Flash GGUF checkpoints on glm5-next (feat(gguf): serve GLM-5.3-Flash GGUF checkpoints on glm5-next FlashML-org/FreeToken#721).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 5C11:2FF62B:B5557:250FBF:6AD75388 and timestamp 2026-10-20 11:42:00 UTC. F…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-20T17:11:03+00:00).

• Новый релевантный pr: fix(moe): enable prefill hit-d2d on rocm.
• feat(mimo): add text-only MiMo v2.6 Flash support: обновлён (state open→open, comments 0→0).
• Regression: first prefill OOMs on RTX 5090 with Qwen3.8-Flash-Next NVFP4 (TMA prefill kernel's first launch allocates ~3.7 GiB): обновлён (state open→closed, comments 4→5).
• perf(moe): run the shared expert beside the routed-expert miss copy in offloaded decode: обновлён (state open→open, comments 2→3).
• [Bug] Qwen3.5/3.6: Gemma-style (1 + w) RMSNorm weights are folded into bf16 at load time; transformers applies (1 + w) in fp32: обновлён (state closed→closed, comments 0→1).
• Новый релевантный pr: fix(engine): apply shrinking cache resizes before growing ones in live rebuild (Closes #643).
• Новый релевантный issue: Proposal: load an FTW with fp16 NVFP4 global scales (upcast with a warning) instead of refusing it after #739.
• Новый релевантный issue: Proposal: vLLM scoring fields on /v1/completions (allowed_token_ids, return_tokens_as_token_ids, token-id prompts) for classifier-style clients.
• Новый релевантный issue: An FTW checkpoint can never serve the W4A4 NVFP4 MoE prefill, and the refusal tells you to reconvert with the option that produced it.
• Новый релевантный issue: Two tests fail on a CUDA host since #736 / #697: test_nvfp4_experts_restricted_to_triton_for_swigluoai (sm_120) and test_graph_capture_reuses_warm_offload_cache_before_capture (any Linux host with libcuda).
• Новый релевантный pr: fix(moe): prevent stale cpu expert inputs in rocm graphs.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID BC01:3E434C:B6BFF:242B47:6AD7A0C5 and timestamp 2026-10-20 17:11:33 UTC. F…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-20T21:39:08+00:00).

• Новый релевантный pr: feat(moe): fetch locked-layer experts through pinned staging.
• Новый релевантный issue: [Feature Request] Please add SSD offloading for KV cache to save GPU memory (like LM-Cache).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID AFC4:101246:E8FEA:F7630:6AD7DF9F and timestamp 2026-10-20 21:39:43 UTC. Fo…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-21T00:38:23+00:00).

• Новый релевантный pr: chore(release): 0.2.0.
• Новый релевантный issue: Proposal: isolate logprob reads (max_tokens 1) from batch composition and the prefix cache so scoring is reproducible.
• 🚧 [WIP] MoE cache, CPU experts and GGUF support: обновлён (state open→open, comments 0→0).
• feat(moe): fetch locked-layer experts through pinned staging: обновлён (state open→closed, comments 0→0).
• FlashML-org/FreeToken:main: cc56b6bd → 127aa0a9 — chore(release): 0.2.0 (chore(release): 0.2.0 FlashML-org/FreeToken#773).
• Новый релиз: FlashML-org/FreeToken v0.2.0.
• Upstream main: релевантные изменения cc56b6bd → 127aa0a9 — commits: feat(moe): fetch locked-layer experts through pinned staging (feat(moe): fetch locked-layer experts through pinned staging FlashML-org/FreeToken#742) | files: pyproject.toml, python/freetoken/engine/config.py, python/freetoken/engine/engine.py, python/freetoken/kernel/csrc/jit/fast_index_copy.cuh, python/freetoken/kernel/fast_index_copy.py, python/freetoken/kernel/pinned.py, python/freetoken/moe/bench_profile.py, python/freetoken/moe/expert_banks.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 0C02:1B0FD:6B9C1:B048F:6AD8099A and timestamp 2026-10-21 00:38:50 UTC. For…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-21T07:41:39+00:00).

• Новый релевантный pr: feat(gguf): serve DeepSeek-V4-Flash GGUF checkpoints on deepseek4.
• Новый релевантный pr: perf(kernel): add opt-in hip fp8 decode kernels for r9700.
• fix(moe): enable prefill hit-d2d on rocm: обновлён (state open→closed, comments 0→1).
• Новый релевантный pr: fix(ple): add rocm disk wait-sync support.
• Новый релевантный pr: perf(moe): scan the hybrid eviction pool in chunks and add fetch rounding modes.
• Новый релевантный pr: test(models): cover the Ornith-1.5-122B NVFP4 export.
• fix(tests): support ROCm in AIME GPU memory checks: обновлён (state open→closed, comments 0→0).
• Новый релевантный issue: Capture-time cuGraphUpload (#697) widens inter-kernel gaps in replayed decode graphs: ~1-2% slower decode.
• FlashML-org/FreeToken:main: 127aa0a9 → cc4aa085 — fix(moe): enable prefill hit-d2d on rocm (fix(moe): enable prefill hit-d2d on rocm FlashML-org/FreeToken#771).
• Upstream main: релевантные изменения 127aa0a9 → cc4aa085 — commits: fix(tests): support ROCm in AIME GPU memory checks (fix(tests): support ROCm in AIME GPU memory checks FlashML-org/FreeToken#763); fix(moe): enable prefill hit-d2d on rocm (fix(moe): enable prefill hit-d2d on rocm FlashML-org/FreeToken#771) | files: python/freetoken/engine/config.py, python/freetoken/kernel/csrc/jit/batch_memcpy.cuh, python/freetoken/moe/offload_cache.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 73C7:20D467:360B4A4:AEF4E44:6AD86CD9 and timestamp 2026-10-21 07:42:17 UTC…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-21T13:50:02+00:00).

• Новый релевантный issue: [Desktop] weights failed to load: Qwen3.6-35B-A3B NVFP4.
• fix(moe): prevent stale cpu expert inputs in rocm graphs: обновлён (state open→open, comments 0→1).
• Новый релевантный pr: perf(attention): fuse short bf16 decode output gates on gfx1201.
• perf(kernel): add opt-in hip fp8 decode kernels for r9700: обновлён (state open→open, comments 0→0).
• Новый релевантный issue: [Perf] Fuse the BF16 attention output gate in short gfx1201 decode graphs.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Atriton-lang/triton-windows%20launch_pdl&sort=updated&order=desc&per_page=20: HTTP 403: You have exceeded a secondary rate limit. Please wait a few minutes before you try again. For more on scraping GitHub and how it may affect your rights, please review our Terms of …

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-21T18:40:21+00:00).

• perf(moe): scan the hybrid eviction pool in chunks and add fetch rounding modes: обновлён (state open→closed, comments 0→0).
• [Desktop] weights failed to load: Qwen3.6-35B-A3B NVFP4: обновлён (state open→open, comments 1→2).
• Новый релевантный issue: perf(qsa): prefill indexer scoring pads 4 heads to 16 per row; 16-row tiles are ~5x faster and bitwise identical.
• feat(gguf): serve DeepSeek-V4 Flash and Pro GGUF checkpoints on deepseek4: обновлён (state open→open, comments 1→3).
• FlashML-org/FreeToken:main: cc4aa085 → 161ffa5a — perf(moe): scan the hybrid eviction pool in chunks and add fetch rounding modes (perf(moe): scan the hybrid eviction pool in chunks and add fetch rounding modes FlashML-org/FreeToken#778).
• Upstream main: релевантные изменения cc4aa085 → 161ffa5a — commits: perf(moe): scan the hybrid eviction pool in chunks and add fetch rounding modes (perf(moe): scan the hybrid eviction pool in chunks and add fetch rounding modes FlashML-org/FreeToken#778) | files: python/freetoken/moe/offload_kernels.py.
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID C83F:D8C6C:E9621:14B7BD:6AD9073B and timestamp 2026-10-21 18:40:59 UTC. Fo…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.

[github-actions]

@Castoff995 FreeToken ROCm Watch: обнаружены существенные изменения (2026-10-21T23:22:25+00:00).

• Новый релевантный issue: [Desktop] weights failed to load: Qwen3.6-35B-A3B NVFP4.
• FlashML-org/FreeToken:main: 161ffa5a → b10c0c47 — docs: add a guide to speeding up prefill with offloaded experts (docs: add a guide to speeding up prefill with offloaded experts FlashML-org/FreeToken#785).
• Upstream main: релевантные изменения 161ffa5a → b10c0c47 — commits: docs: add a guide to speeding up prefill with offloaded experts (docs: add a guide to speeding up prefill with offloaded experts FlashML-org/FreeToken#785).
• ⚠️ Новый сбой части мониторинга: dependency discovery: ApiError: GET https://api.github.com/search/issues?q=repo%3Aggml-org/llama.cpp%20MUL_MAT&sort=updated&order=desc&per_page=20: HTTP 403: API rate limit exceeded for installation. If you reach out to GitHub Support for help, please include the request ID 9417:2A8AA4:C5B4:154F1:6AD94955 and timestamp 2026-10-21 23:23:01 UTC. For…

Перед следующим Codex-таском стоит сначала сверить эти изменения с текущим blocker/патчем.