qwen36-27b-ft-grm-w8a8

GRM-2.6-Plus W8A8 (INT8 weights + INT8 dynamic activations) — MTP-preserved

W8A8 INT8 quantization of OrionLLM/GRM-2.6-Plus, produced with llm-compressor (SmoothQuant + GPTQ). The native MTP (Multi-Token Prediction) head is preserved as BF16 and works at 91-94 % acceptance under vLLM ≥ 0.17 — running ~26 % faster than the BF16 source while taking 35 % less memory.

Quick start

pip install "vllm>=0.17"
huggingface-cli download vadery/qwen36-27b-ft-grm-w8a8 --local-dir ./GRM-2.6-Plus-W8A8

vllm serve ./GRM-2.6-Plus-W8A8 \
  --max-model-len 262144 \
  --dtype bfloat16 \
  --gpu-memory-utilization 0.92 \
  --trust-remote-code \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3 \
  --speculative-config '{"method":"qwen3_5_mtp","num_speculative_tokens":1}'

No additional patches required — both config.json.quantization_config.ignore (covers MTP Linear modules) and actorder field are already fixed.

Performance (single H200 SXM, vLLM 0.17.1, temperature=0)

For reference, the BF16 source with the same MTP recipe reaches 102 tok/s single-stream and 749 tok/s at concurrency 8 on W2 — i.e. this W8A8 model is 26-29 % faster while using ~18 GB less weight memory.

Architecture preserved

Quantization recipe

SmoothQuantModifier(smoothing_strength=0.8, mappings=SQ_MAPPINGS,
                    ignore=[...vision, mtp, linear_attn, embed, lm_head...])
GPTQModifier(targets="Linear", scheme="W8A8",
             ignore=[same as above],
             dampening_frac=0.01)

SmoothQuant mappings explicitly cover only the 16 full-attention layers (indices 3, 7, 11, …, 63 out of 64) plus MLP on every layer — to avoid SmoothQuant trying to fuse into the linear_attn projections which have non-standard shapes.

Calibration: 512 samples × 2048 tokens from HuggingFaceH4/ultrachat_200k.

Post-process steps (already applied; documented for reproducers)

llm-compressor 0.10 drops the MTP tensors from the saved state and writes a quantization_config.ignore that doesn't cover MTP Linear modules. We post-process:

1. Restore MTP tensors — copy 15 mtp.* tensors from the BF16 source model-*.safetensors shards into the W8A8 single-shard safetensors.
2. Patch config.json — add 8 MTP Linear module names to quantization_config.ignore so vLLM treats them as un-quantized BF16 on load.

Without these two steps, vLLM either drops the MTP head (0 % acceptance) or loads garbage values (also 0 % acceptance after weights are corrupted on load).

File size

Reasoning + tool calling

Same parser flags as the BF16 source:

• --reasoning-parser qwen3 — separates <think> segments into reasoning field
• --tool-call-parser qwen3_coder + --enable-auto-tool-choice — OpenAI tool-call API

Notes

• vLLM ≥ 0.17 required (qwen3_5_mtp speculative method only landed there).
• transformers ≥ 5.x is required for qwen3_5 model_type.
• Tested on H200 (compute capability 9.0). H100 should also work. SM80 (A100) cutlass INT8 path exists but is slower per dollar than H100/H200.
• For multi-stream serving raise --max-num-seqs to taste; observed peak throughput stays linear in concurrency up to at least 8.

License

Inherits Apache 2.0 from the base model.