Qwen3.6-35B-A3B — DFlash vs MTP

Qwen3.6-35B-A3B — DFlash vs MTP vs baseline (same model)

Date: 2026-07-06 Setup: target unsloth/Qwen3.6-35B-A3B-GGUF UD-Q4_K_M + draft williamliao/Qwen3.6-35B-A3B-DFlash-GGUF Q8_0. Stack: production llama-hipgraphs build 0eca4d4, gfx906/MI50. All three spec strategies on the identical model → clean head-to-head. Models deleted after; Ornith restored.

Compatibility (checked before download — this pairing is the runnable one)

draft	verdict	why
williamliao DFlash-GGUF	✅ runs	arch dflash (loadable format), vocab 248320 = target, our build has a working DFlash runtime (common_speculative_impl_draft_dflash engaged, non-zero accept)
z-lab / Anbeeld DFlash-GGUF	❌	arch dflash-draft — rejected by our stock build
modal-labs DFlash	❌	safetensors only (no GGUF), DFlashDraftModel for vLLM
DSpark-AEON draft	❌	safetensors + vLLM patches, wrong (AEON) base

The williamliao draft is a real block-diffusion DFlash (block_size 16, n_extract 8), Q8_0 421 MB, 6-layer.

n-max tune (temp-0 code prompt), DFlash

n2 = 81.3 t/s (74% accept) · n3 = 82.7 (63%) · n4 = 78.7 (58%). Best n-max 3. For reference, MTP on the same model hit 89–91 t/s (92% accept) → MTP already ~10% faster on code.

benchy decode vs context — DFlash (n3) vs MTP (n2) vs baseline

depth	baseline	DFlash	MTP	DFlash vs base	vs MTP
0	63.5	76.7	82.3	+21%	−7%
512	63.4	72.1	82.5	+14%	−13%
1k	63.2	78.1	85.0	+24%	−8%
2k	62.2	70.6	84.4	+13%	−16%
3k	60.9	78.6	84.8	+29%	−7%
4096	60.1	💥 server crashed	74.6	—	—
… 65k	38.9	(dead)	61.6	—	—

DFlash avg prefill (0–3k): 752 t/s (comparable to MTP’s 706 / baseline’s 820).

Verdict — MTP beats DFlash on this model, on BOTH axes

1. Speed: DFlash is faster than baseline (+13–29%), but slower than MTP at every measured depth (and on the code probe, 83 vs 90). Root cause: DFlash’s draft acceptance is 63% vs the official embedded MTP head’s 92% — MTP proposes better tokens, so more of them stick.
2. Stability: the DFlash server crashed at depth 4096 (completed 0–3072, then connection-refused for every deeper depth). MTP ran cleanly to 65k. The crash is at the -b/-ub 4096 boundary — likely a DFlash block-diffusion / batch-size interaction in our build; not investigated further since MTP already won.

This REVERSES the earlier gemma-4-26B-A4B result (where DFlash beat MTP). Difference: Qwen3.6 ships an official, well-tuned embedded MTP head (92% accept), whereas the gemma MTP head was weaker relative to its DFlash. So “DFlash vs MTP” is model-specific — it comes down to which head has higher acceptance. On Qwen3.6-35B-A3B, the embedded MTP wins decisively.

Files

• Charts: qwen36_spec_comparison.png (the 3-way, same model) + DFlash line added to mi50_model_sweep_decode.png.
• Raw JSON: qwen36full_DFlash_benchy.json (0–3k), qwen36full_MTP_benchy.json, qwen36full_benchy.json.