Qwen3.6-27B — Dense Quick Test

Qwen3.6-27B (dense) + own MTP — MI50 quick test + DSpark compatibility

Date: 2026-07-06 Requested: test unsloth/Qwen3.6-27B-GGUF with the Hikari07jp/DSpark-Qwen3.6-27B-AEON-draft draft. Stack: production llama-hipgraphs:upstream-rocm-7.2.4 (build 0eca4d4), gfx906/MI50. Model deleted after; Ornith restored.

DSpark draft — INCOMPATIBLE (checked first, not downloaded)

The DSpark-AEON draft cannot be paired with the 27B on our stack, on three independent grounds (all verified from the repo file list + card):

1. No GGUF / no llama.cpp. Repo is model.safetensors + dflash.py + markov_head.py + vllm_patches/ — a patched-vLLM-only artifact (“stock vLLM does not implement the Markov semi-autoregressive drafting path”). We run llama.cpp; gfx906 vLLM is a dead-end anyway.
2. Wrong base. Distilled to AEON logits (AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored), not stock Qwen3.6-27B → hidden-state mismatch → degraded/zero acceptance even if it loaded (same base-mismatch rule as the Ornith-DFlash note).
3. NVIDIA/DGX-Spark target (“DSpark”), like the ROCmFP4 Deckard model.

Substitute tested instead: the model’s OWN embedded MTP (unsloth/Qwen3.6-27B-MTP-GGUF, Q4_K_M, arch qwen35 dense, block_count 65 with blk.64.nextn.* head) — the working self-speculative path, no external draft needed.

Findings (basic bench + n-max probe; full benchy skipped by choice — slow non-contender)

• Dense 27B (qwen35, 27.3B all-active, hidden 5120, hybrid Gated-DeltaNet/Attention). llama-bench Q4_K_M: pp512 = 244 t/s, tg128 = 23.2 t/s (no spec).
• With MTP (--spec-type draft-mtp, embedded head): n-max 2 = 23.9 t/s (+3%) at 78% draft acceptance.

Verdict — MTP is ~net-neutral on a dense 27B

Despite 78% acceptance, MTP delivers only +3% because on a dense model every verify token costs the full 27B forward pass — the verify cost eats almost all the speculation benefit. This is the same dense penalty seen on gemma-4-12b (MTP only +5%) and the mirror of the A3B MoE models where MTP gives +30–58% (verify only activates ~3B there). Spec-decode’s payoff scales with how cheap verify is → big on MoE, marginal on dense.

Separately, this 27B is the slowest model in the sweep (23 t/s decode, 244 t/s prefill) — the dense 27B + hybrid-DeltaNet arch is heavy and appears less optimized in our build than standard attention. Not a serving contender on the MI50 vs the ~60-80 t/s A3B MoE models.

Bottom line: the DSpark pairing was impossible; the 27B’s own MTP works but barely helps because it’s dense. The MI50’s sweet spot remains A3B MoE + MTP.