[diwakergupta]

What this changes

The native CUDA qwen3_5 engine now serves dense EXL3 checkpoints (turboderp's Qwen3.8-27B-exl3) and attaches the DFlash2 drafter to EXL3 targets. This is part (a) of the split of #1407 that @ashhart suggested; (b) prompt tiles, (c) memory and (d) copy policy follow as separate PRs.

• cuda_geometry.zig: admits non-MoE qwen3_5 EXL3 checkpoints (pocketed ones are still refused).
• cuda_exl3_weights.zig: loads the dense SwiGLU MLP (mlp.{gate,up,down}_proj) as EXL3 linears, and allocates the grouped-expert scratch only for MoE geometry.
• cuda_native.zig: lists exl3 in the family's formats, and openDrafter no longer refuses EXL3 targets.
• cuda_draft_frame.zig: the drafter reads the target's EXL3 head over its two vocabulary spans as column views of the stored matrix. It uses the same packing path as the affine head.
• exl3.zig: adds Weight.columns(from, n), a view of whole 128-column blocks that keeps the parent's strides, with a host test.
• Prompt arithmetic is unchanged: prompt rows still take 128-row blocks of the decode linear. So the Qwen3.5-MoE EXL3 checkpoints keep their bits.
• Docs: the 27B recipe gains a "Native EXL3 checkpoints" section. exl3.md and cuda.md no longer say the native engine serves no EXL3 pack.

Licences: turboderp/Qwen3.8-27B-exl3 is Apache-2.0 (base Qwen/Qwen3.8-27B), and the drafter z-lab/Qwen3.8-27B-DFlash2 is Apache-2.0. No code or tables were taken from other engines.

Receipt

• Environment. RTX 3090 Ti (24 GB, sm_86, CUDA 13.1 nvcc), zig build native -Dsm=86. Checkpoint turboderp/Qwen3.8-27B-exl3 branch 3.50bpw (rev 8351c54e), drafter z-lab/Qwen3.8-27B-DFlash2 (rev 50307d4c) at --drafter-bits 4. bf16 KV cache (default), --context 32768, one rank. The binary was built from this branch before its rebase onto ee896951; the rebase changed no file in this PR.
• Exactness, drafted vs "draft": false and concurrent vs solo. tools/bench_concurrent.py --levels 2 --tokens 256 --alone --serial, greedy and sampled:
  • The default code and chat cells: all alone and serial checks equal, none unequal or failed.
  • --prompts with two long prompts (5,080 and 2,278 tokens, so one spans a 4,096-row cut): all equal.
• Exactness, resumed vs fresh. The engine cuts prompt passes every 4,096 rows and keeps a state at each cut.
  • Turn 1 is a ~5,100-token document. Turn 2 shares its first 4,096 tokens and ends 16, 17, 64 or 300 rows past the cut (sized with /tokenize). Each tail length runs greedy and seeded-sampled, 128 tokens.
  • Every resumed turn 2 reported cached_tokens 4096. The fresh runs went to a restarted server and reported cached_tokens 0.
  • token_sha256 was equal in all 8 cases.
• Decode speed (2 streams, bench_concurrent, per-stream tok/s): code 155 / 122 (sampled / greedy), chat 86 / 90. There is no baseline: main refuses this checkpoint.
• Prompt speed: n/a. The prompt arithmetic is unchanged from main's EXL3 path; cold prefill of an 8k prompt is 14.5 s here, and (b) changes it.
• Tests. zig build test test-kernel-contract test-golden test-server-security test-pull-security fuzz-cortheon test-cuda-host: exit 0, 447/447 build steps succeeded with the same single skip as main at ee896951. python3 tools/zig/lean_new.py against the base: 0 new problems.
• Not run. Metal (no Mac). GB10 / sm_121. The Qwen3.5-MoE EXL3 checkpoints: no copy here, but their prompt and decode code paths are unchanged.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo.
• No precision traded for speed: the head's column views read the same stored words.
• No GPU kernel changed.
• Prompt processing is no slower than the last release (same arithmetic).
• Platforms: CUDA sm_86 only; Metal and GB10 not run.
• New tests fail before the change: the Weight.columns host test, and the EXL3 dense admission in qwen35_formats_host.zig.
• Comments and docstrings are one line.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

[ashhart]

Thanks @diwakergupta, and thanks for splitting #1407 up so quickly. Unfortunately this part now overlaps #1465, which makes the same change: the dense admission in cuda_geometry.zig, the dense MLP as three EXL3 linears, the grouped-expert scratch only for MoE, and DFlash2 on dense EXL3 targets. #1465 is already queued for our exactness receipts on a GB10, so we'll take the dense path from there and close this one. Sorry for the duplication, since #1407 proposed it first.

Two things here aren't in #1465, and we'd welcome them as one small PR on top of it:

1. The drafter reading the EXL3 head over its two vocabulary spans (Weight.columns at zig/src/cuda/exl3.zig:100 with its host test, and the .x3 arm at cuda_draft_frame.zig:162). On qwen3_5 cuda: dense EXL3 checkpoints (Qwen3.8-27B), with DFlash2 drafts #1465 that head still projects all 248,320 columns. Please add drafted == "draft": false and a draft-speed comparison against qwen3_5 cuda: dense EXL3 checkpoints (Qwen3.8-27B), with DFlash2 drafts #1465. The 384-column tail plans 8 K splits where the full head has 1, so its proposals' bits change, which is fine.
2. exl3 in the family's formats (cuda_native.zig:27) and the exl3.md/cuda.md lines.

Please keep #1465's refusal for MoE EXL3 targets. This branch removes it from both openDrafter and openMtp (cuda_native.zig:161-180), and no MoE EXL3 drafter has been checked. We've asked for #1475 and #1476 to be rebased onto #1465.