Qwen3.8-27B-INT4-W4A16

INT4 weight-only (W4A16) quantization of Qwen/Qwen3.8-27B — a dense hybrid model (GatedDeltaNet linear-attention + periodic full-attention) with a native vision tower and an MTP head.

At BF16 the base model needs ~56 GB and does not fit two 24 GB cards. This checkpoint is ~20 GB, so the weights fit on a single 24 GB GPU and leave room for a large KV cache at tensor-parallel 2.

Quantization

• AWQ activation-aware scale search + symmetric INT4 weights, group_size 32, MSE observer (llm-compressor), calibrated on 128 × 512-token Magicoder samples.
• Quantized: all MLP, the full-attention self_attn, and the GatedDeltaNet out_proj.
• Kept bf16 (quality-sensitive): GatedDeltaNet in_proj_a/in_proj_b, lm_head, embeddings, vision tower, MTP head.
• Format: compressed-tensors (pack-quantized). Full recipe: recipe.yaml.

The MTP head is included (mtp.safetensors, bf16) and is listed in quantization_config.ignore, so vLLM loads it unquantized and speculative decoding works out of the box. Quantizers silently drop this head — if you re-quantize the base yourself, graft it back or spec-decode will accept 0 %.

Validation

Served with vLLM tensor-parallel 2 on 2 × RTX 3090 and exercised over the OpenAI API:

• Chinese / English generation and translation: coherent.
• Vision path (image input through the native VL tower): working.
• reasoning_effort (low / xhigh) and enable_thinking pass through correctly.
• MTP speculative decoding: mean accept length ≈ 2.4 / 3.0 at K=2.

GSM8K: 96.4 % (241/250). Served with vLLM tp2 and scored over the OpenAI API — thinking enabled, the base model's recommended sampling (temperature=1.0, top_p=0.95, top_k=20), default reasoning_effort, max_tokens=24576, and the question suffixed with "Put your final numeric answer inside \boxed{}". The same protocol scores the INT8-W8A8 build of this model at 96.8 % (242/250), i.e. the two are indistinguishable at this sample size.

Note the protocol matters more than it looks: without the \boxed{} instruction this model often states the answer in bold and then keeps talking, so a "last number in the answer" extractor scores many correct answers wrong (the same 250 questions land at ~90 % that way). MMLU-Pro has not been run on this checkpoint.

Throughput

2 × RTX 3090, tensor-parallel 2, single stream, prefix caching disabled, no speculative decoding:

Decode is int4-weight bandwidth-bound and is the same for W4A16 and W4A8; the int8 activation path buys +61 % … +70 % prefill.

Usage (vLLM)

# W4A16 — int4 weights, fp16 activations
vllm serve Avesed/Qwen3.8-27B-INT4-W4A16 \
  --tensor-parallel-size 2 --trust-remote-code --reasoning-parser qwen3

On Ampere the same checkpoint can be served as W4A8 (int4 weights + int8 dynamic activations — faster prefill and batched serving) with the vllm-ampere-optimized fork:

vllm serve Avesed/Qwen3.8-27B-INT4-W4A16 \
  --tensor-parallel-size 2 --marlin-input-dtype int8 --trust-remote-code --reasoning-parser qwen3

Add --speculative-config '{"method":"mtp","num_speculative_tokens":2}' to use the MTP head.

Sampling follows the base model: thinking temperature=1.0, top_p=0.95, top_k=20; non-thinking temperature=0.7, top_p=0.80, top_k=20, presence_penalty=1.5.

Quantized with vllm-ampere-optimized/quantize.