[bong-water-water-bong]

Lets a Qwen3.8-Flash-Next (qwen4exp) MTP draft head draft from part of the vocabulary.

• A draft head GGUF may carry blk.<L>.nextn.shared_head_head.weight with K rows and an I64 d2t [K] (draft row -> token id). Without d2t nothing changes.
• The draft pass runs the K-row head, then scatters the logits into a full-vocabulary row (-inf elsewhere) with the same ggml_fill + ggml_set_rows pattern dflash.cpp uses.
• Our code: src/models/qwen4exp-draft-vocab.{h,cpp} (Apache notice); qwen4exp.cpp gets two hooks.
• The engine side (tool that writes these files, id list, NOTICE) is a separate engine PR.

Measured

Strix Halo, Vulkan, Qwen3.8-Flash-Next UD-Q4_K_XL, Unsloth's shared Q8_0 MTP head, greedy, draft length 3, -lm mmap -lzm on, coder-llm stopped, box.lock held. Median of rounds 2-3 of 3.

Prompt	Full head (644 MiB)	65,536-id head (170 MiB)
code, 256 tok	47.8 tok/s, 186/207 accepted	52.1, 186/207
prose, 256 tok	38.6, 169/258	42.2, 169/258
code, 768 tok	45.8, 552/642	48.3, 544/669
French, 26 tok	46.4, 18/21	39.6, 18/27
ms per decode step (code)	76.2	69.9

The id list is English + code (dime-online/qwen3.8-Flash-DGX-UltraFast, Apache-2.0), so French drafts get rejected more often. Without a draft: 23.7 tok/s.

Output

Greedy text compared with decoding without a draft:

• The full head is not token-identical there either. Prose and long code each flip once at a near-tie, from batched verification.
• The 65K head gives exactly the full head's text in rounds 2-3. In round 1 it flipped at two other near-ties.

Unsloth's shared Q4_K_M head was also tried: no faster (46.7 code) and more flips, so not used.

Tests

• Built on strixhalo (llama-server, llama-cli, test-llama-archs).
• Real-model runs above.
• The engine PR adds a unit test for the file writer.

🤖 Generated with Claude Code

[bong-water-water-bong]

Review (PR-Agent duty). The approach is good: the K-row head plus a scatter to a full row of -inf reuses DFlash's set_rows path. All the code is in our headed files (qwen4exp-draft-vocab.{h,cpp}), with two hooks in qwen4exp.cpp. And the target still verifies every token.

One fix before merge:

• Validate the d2t ids. qwen4exp_draft_vocab_rows checks d2t's type, rank and length, but not its values. A d2t id that's negative or >= n_vocab makes ggml_set_rows write outside the full-vocabulary row, and duplicate ids silently drop a row. d2t is tiny (K int64s), so check it once after tensor data is loaded (0 <= id < n_vocab, and no duplicates) and fail the load with a clear message. A malformed or hand-made draft GGUF shouldn't be able to corrupt memory.

Notes:

1. Non-English. French drops 46.4 -> 39.6 tok/s (-15%) because its tokens fall outside the 65,536-id list. Since the reduced head only exists when the draft GGUF carries it, the default is unaffected. Please keep it that way: 1bit serve shouldn't auto-pick a reduced-head draft, and the docs that mention it should state the non-English cost next to the speedup.
2. Provenance of the id list. The 65,536-id list comes from dime-online/qwen3.8-Flash-DGX-UltraFast. Before we publish a reduced-head GGUF (HF org) or commit the list, check that repo's license, and credit it in the model card and NOTICE (per-model license checks, as with every HF upload).
3. Measurements. "Rounds 2-3 of 3" means the warm rounds. Fine, but say so in the docs table too.

Mergeable once the d2t value check is in.

[bong-water-water-bong]

Post-merge: the d2t value validation from the review is still missing (see engine ggml-org#269). Needed before the engine pins this.