Qwen3.6-35B-A3B · DDTree (Diffusion Draft Tree) · BLOCKED (hybrid GDN cache)

# DDTree harness inside the trusted SGLang cu130 torch env (has flash_attn + datasets, built for GB10 sm_121a).
# Target + drafter load fine; generation fails in the target forward.
scripts/bench-ddtree.sh Qwen/Qwen3.6-35B-A3B z-lab/Qwen3.6-35B-A3B-DFlash mt-bench 20 512 64,256 0.0 q35b
# -> ValueError: `has_previous_state` can only be called on LinearAttention layers,
#    and the current Cache seem to only contain Attention layers.  (transformers qwen3_5_moe)

BLOCKED — the open DDTree research harness cannot run Qwen3.6’s hybrid GatedDeltaNet target. Qwen3.6-35B-A3B is model_type: qwen3_5_moe (Qwen3_5MoeForConditionalGeneration): a hybrid stack of GatedDeltaNet linear-attention layers interleaved with full-attention layers. The DDTree/DFlash harness (dflash.py / ddtree.py) verifies speculative blocks by rolling the target KV cache back to the accepted prefix with past_key_values.crop(start) after every block. That rollback is the wall:

• Vanilla cache mismatch. The harness allocates a plain transformers.DynamicCache. On the first decode step the model’s _update_linear_attn_mask calls cache.has_previous_state(), which raises because a DynamicCache holds only attention layers, not the LinearAttention state the hybrid model expects.
• Deeper than a cache-class swap. Even with the correct hybrid cache, crop(start) cannot rewind a GatedDeltaNet layer to an arbitrary token: linear-attention keeps a recurrent state, not per-token KV, so there is nothing to slice back to position start. Speculative decoding with rejection requires rewindable state; recurrent layers don’t provide it without explicit per-step checkpointing.
• This is the same GatedDeltaNet wall seen elsewhere. The AEON vLLM fork only runs DFlash on this model by purpose-building a unified attention page (block size 1136) that reconciles the draft’s small page with the mamba/GDN page (see …ultimate-dflash). The research harness has no such integration, and no serving engine implements DDTree at all.

Consequence for the money chart: the DDTree line cannot be measured for Qwen3.6-35B-A3B on the Spark today. DDTree here would require either (a) a hybrid-aware re-implementation of the tree verifier with GatedDeltaNet state checkpointing, or (b) DDTree landing inside a purpose-built engine (e.g. AEON’s DFlash fork extended to trees). Neither exists.

What we measured instead: DDTree on the harness’s own non-hybrid target Qwen/Qwen3-Coder-30B-A3B-Instruct (qwen3_moe, standard attention) + z-lab/Qwen3-Coder-30B-A3B-DFlash — a same-size MoE and a coding workload — giving the real tree-vs-single-line-DFlash accept-len + single-stream speedup on this box. See qwen3-coder-30b-a3b-ddtree. The dense Qwen3.6-27B (qwen3_5, also hybrid) is blocked for the identical reason — …27b-ddtree-blocked.