groxaxo/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-W8A8

Qwen3.6-27B-uncensored-heretic-v2 — W8A8 (AutoRound INT8 Dynamic, MTP preserved)

INT8 W8A8 dynamic (per-token activation) quantization of llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved, produced with AutoRound 0.13.1 and exported in the compressed-tensors (auto_round:llm_compressor) format for vLLM.

Same recipe as the Darwin-28B W8A8 build.

Quantization recipe

• Scheme: INT8 — channel-wise INT8 weights, INT8 dynamic per-token activations (CUTLASS INT8 path in vLLM).
• Algorithm: RTN (round-to-nearest, data-free). INT8 + dynamic activations are near-lossless.
• Quantized: language-model Linear weights only (DeltaNet in_proj_*/out_proj + full-attn q/k/v/o_proj + MLP gate/up/down).
• Preserved in BF16: native MTP module (mtp.*, via --ignore_layers mtp), the vision tower (model.visual.*, auto-skipped by AutoRound's text-module-only VLM path), lm_head, embed_tokens, all norms.
• Hardware: 2 × RTX 3090.

Architecture: Qwen3.5-generation hybrid (qwen3_5, 64 layers, 3:1 DeltaNet/full-attention, hidden 5120, vocab 248320), multimodal, with a 1-layer native MTP head for speculative decoding.

Serving (vLLM, TP2) — recommended: dense (no MTP)

vllm serve groxaxo/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-W8A8 \
  --tensor-parallel-size 2 \
  --max-model-len 262144 \
  --kv-cache-dtype fp8 \
  --reasoning-parser qwen3 \
  --trust-remote-code

MTP note (measured)

The native MTP head is preserved in BF16, but on this finetune it is non-predictive: measured ~0% draft-token acceptance (2 of 6254), so enabling speculative decoding is actually slower (~22.5 tok/s with MTP vs ~27.7 tok/s dense, single-stream on 2×RTX 3090). This was isolated to the checkpoint's MTP head, not the quantization or server:

• Same vLLM + method:mtp on the standard Qwen3.6-27B (AEON FP8) gives 78% acceptance / ~2.1× speedup — so the setup is correct.
• 0% acceptance persists with kv-cache-dtype=auto (not a KV-quant artifact).
• The W8A8 quant is near-lossless (KL≈0.02 nats/token) and preserves the MTP head structure intact.

The heretic-v2 finetune appears to have diverged its main weights from the base MTP head, so the head's drafts no longer match. Serve dense for best speed. The MTP weights remain in this repo for anyone who retrains/distills the head against this finetune. The vision tower is intact (drop --hf-overrides '{"language_model_only": true}' to use it; add it to save VRAM for text-only).