webhie/Qwen3.6-27B-int4-AutoRound

Qwen3.6-27B INT4 AutoRound (Best Recipe)

A W4A16 (INT4 weight, FP16 activation) quantization of Qwen/Qwen3.6-27B, produced with Intel's AutoRound.

Key difference from other AutoRound quants of this model: This was quantized using the auto-round-best preset — 1000 iterations and 512 calibration samples instead of the standard 200/128. This preset runs ~4–5× slower but achieves the best possible accuracy at INT4, as it performs a more thorough weight rounding optimization. MTP (speculative decoding) and image/vision inputs work out of the box with no post-processing required.

TL;DR

• Base: Qwen3.6-27B (27B dense VLM)
• Quant: INT4 W4A16, group_size 128, symmetric
• Tool: auto-round-best (1000 iters, 512 samples, torch.compile)
• Size: ~18 GB (down from ~54 GB BF16) — 3× reduction
• MTP: Native Multi-Token Prediction head preserved in BF16 — enables native speculative decoding in vLLM (~85–90% draft acceptance, ~2× throughput)
• Vision: Image inputs work via the MoonViT encoder (weights kept at original BF16/FP16 precision)

Quick inference with vLLM (with MTP speculative decoding)

Requires vLLM v0.19.1+ with Qwen3_5 MTP support. Set the following environment variables before starting:

export VLLM_USE_FLASHINFER_SAMPLER=1
export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
export VLLM_FLOAT32_MATMUL_PRECISION=high
export PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True,max_split_size_mb:512"
export VLLM_NO_USAGE_STATS=1
export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=1
export VLLM_MARLIN_USE_ATOMIC_ADD=1
export OMP_NUM_THREADS=1
export CUDA_DEVICE_MAX_CONNECTIONS=8
export NCCL_CUMEM_ENABLE=0
export NCCL_P2P_DISABLE=1

vllm serve webhie/Qwen3.6-27B-int4-AutoRound \
  --served-model-name qwen3.6-27b \
  --host 0.0.0.0 --port 11434 \
  --trust-remote-code \
  --dtype auto \
  --quantization auto_round \
  --max-model-len 200704 \
  --gpu-memory-utilization 0.92 \
  --max-num-seqs 4 \
  --kv-cache-dtype fp8_e4m3 \
  --attention-backend flashinfer \
  --performance-mode throughput \
  --max-num-batched-tokens 2048 \
  --enable-chunked-prefill \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3 \
  --default-chat-template-kwargs '{"preserve_thinking":true}' \
  --override-generation-config '{"temperature":0.6,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":0.0,"repetition_penalty":1.0}' \
  --enable-prompt-tokens-details \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

Remove --speculative-config to disable MTP speculative decoding. See the vllm-blackwell-guide repo for a full Docker Compose setup with all env vars pre-configured.

OpenAI-compatible request

from openai import OpenAI
client = OpenAI(base_url="http://localhost:11434/v1", api_key="EMPTY")
r = client.chat.completions.create(
    model="qwen3.6-27b",
    messages=[{"role": "user", "content": "Write a quicksort in Python."}],
    max_tokens=512,
)
print(r.choices[0].message.content)

Transformers (no spec decoding)

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained(
    "webhie/Qwen3.6-27B-int4-AutoRound",
    trust_remote_code=True,
    device_map="auto",
)
tok = AutoTokenizer.from_pretrained("webhie/Qwen3.6-27B-int4-AutoRound")
msg = [{"role": "user", "content": "Explain quantum computing briefly."}]
ids = tok.apply_chat_template(msg, add_generation_prompt=True, return_tensors="pt").to(m.device)
print(tok.decode(m.generate(ids, max_new_tokens=256)[0]))

Quantization details

Field	Value
Base	Qwen/Qwen3.6-27B
Method	AutoRound (intel/auto-round), best recipe
Scheme	W4A16 (4-bit weights, FP16 activations)
Bits	4
Group size	128
Symmetric	yes
Packing format	auto_round:auto_gptq
Unquantized layers	linear_attn.in_proj_a/b, all LayerNorms, RMSNorms, router gates
Calibration samples	512
Iterations	1000
torch.compile	enabled
GPU used for quant	1× RTX 5090 (32 GB, SM120), low_gpu_mem_usage=True

Unquantized layers — why

• linear_attn.in_proj_a/b: low-rank projections in Qwen3.6's Gated DeltaNet whose shapes aren't divisible by 32 (group_size), so AutoRound skips them automatically. Tiny fraction of total parameters.
• Norms, routers: precision-sensitive and very small — kept at full precision.

Performance

Benchmarked on 1× RTX 5090 (32 GB) with vLLM + FP8 KV cache + MTP n=3:

Config	Throughput
vLLM + MTP n=3	~150 tok/s
vLLM (MTP disabled)	~70 tok/s

The ~2× speedup comes from ~85–90% draft acceptance via MTP speculative decoding with num_speculative_tokens: 3.

Reproduction

pip install auto-round
auto-round-best \
  --model Qwen/Qwen3.6-27B \
  --scheme W4A16 \
  --format auto_round \
  --output_dir Qwen3.6-27B-int4-AutoRound \
  --enable_torch_compile \
  --low_gpu_mem_usage \
  --device_map 0

No post-processing needed — MTP and image inputs work out of the box.

Acknowledgements

• Alibaba / Qwen team for the base Qwen3.6-27B model
• Intel AutoRound team for the quantization framework
• Lorbus for the original AutoRound quant of this model that inspired this release
• @eugr for the spark-vllm-docker fork and TurboQuant KV cache work
• vLLM project for the inference engine and Qwen3_5 MTP support

License

Apache 2.0 — same as Qwen3.6-27B base.

Citation

If you use this quant, please cite the original Qwen3.6 release (see base model card) and the AutoRound paper:

@article{cheng2023autoround,
  title   = {Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs},
  author  = {Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi},
  journal = {arXiv preprint arXiv:2309.05516},
  year    = {2023}
}