Qwen3.8-27B-W4A16-AWQ-GPTQ

Quantized version of Qwen/Qwen3.8-27B, a Qwen3.5-family hybrid-attention VLM, produced with llm-compressor and served with vLLM.

Recipe

• Scheme: INT4 weight-only (W4A16), pack-quantized, group size 128, AWQ then GPTQ.
• Calibration: HuggingFaceH4/ultrachat_200k, 512 samples at 2048 tokens.
• What is quantized: the text-decoder Linear layers only. The vision tower (re:.*visual.*), the hybrid linear-attention mixers (re:.*linear_attn.*), and lm_head stay in bf16. The full VLM (with vision_config) is saved in the compressed-tensors format, and the base MTP predictor is preserved for speculative decoding.
• Hardware: Runs on any CUDA GPU with a Marlin or compressed-tensors int4 kernel.

from transformers import AutoModelForImageTextToText, AutoTokenizer
from llmcompressor import oneshot
from llmcompressor.modifiers.gptq import GPTQModifier
from llmcompressor.modifiers.transform.awq import AWQModifier

MODEL_ID = "Qwen/Qwen3.8-27B"
model = AutoModelForImageTextToText.from_pretrained(MODEL_ID, dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
# ... load 512 ultrachat_200k samples, chat-templated, at 2048 tokens ...

recipe = [
    AWQModifier(duo_scaling="both", n_grid=20),
    GPTQModifier(targets=["Linear"], scheme="W4A16", block_size=128,
                 dampening_frac=0.01, actorder="static",
                 ignore=["lm_head", "re:.*visual.*", "re:.*linear_attn.*"]),
]
oneshot(model=model, dataset=ds, recipe=recipe,
        max_seq_length=2048, num_calibration_samples=512)
model.save_pretrained("Qwen3.8-27B-W4A16-AWQ-GPTQ", save_compressed=True, save_original_format=False)
tokenizer.save_pretrained("Qwen3.8-27B-W4A16-AWQ-GPTQ")

Serving (vLLM)

vllm serve soyrsoyr/Qwen3.8-27B-W4A16-AWQ-GPTQ

from vllm import LLM, SamplingParams

llm = LLM(model="soyrsoyr/Qwen3.8-27B-W4A16-AWQ-GPTQ")
out = llm.generate(["The capital of France is"], SamplingParams(max_tokens=32))
print(out[0].outputs[0].text)

Performance

Recovery vs. the bf16 base, evaluated through the vLLM backend with lm-evaluation-harness (OpenLLM v1) and lighteval (generative reasoning at temperature 0.6, top_p 0.95, up to 32k tokens).

OpenLLM Leaderboard v1

Benchmark	Qwen3.8-27B	W4A16-AWQ-GPTQ	Recovery
ARC-Challenge (25-shot), acc_norm	50.68	50.09	98.8%
HellaSwag (10-shot), acc_norm	71.99	71.94	99.9%
TruthfulQA-mc2 (0-shot), acc	61.25	60.33	98.5%
Winogrande (5-shot), acc	76.87	76.64	99.7%
Average	65.20	64.75	99.3%

MMLU and GSM8K are omitted. Qwen3.8-27B is a reasoning model, so under the OpenLLM v1 protocol GSM8K has its chain of thought truncated and MMLU's loglikelihood is measured where the model wants to emit its think block, both of which collapse to a harness artifact rather than a real score. Math and knowledge are captured by the generative reasoning suite instead.

Reasoning suite (generative)

Benchmark	Qwen3.8-27B	W4A16-AWQ-GPTQ	Recovery
AIME-24, avg@4	95.00	90.83	95.6%
AIME-25, avg@4	93.33	85.00	91.1%
MATH-500, pass@1	82.00	80.80	98.5%
Average	90.11	85.54	94.9%

AIME-24 and AIME-25 are reported as avg@4 (mean accuracy over 4 samples per problem, temperature 0.6, up to 32k tokens, stderr about 4 points) to average out the large single-sample variance of a 30-problem test. The base and the quantized model are scored in the same run under identical settings, so the small remaining gap reflects the int4 quantization rather than sampling noise. MATH-500 (500 problems) is pass@1.

Community results

User-reported serving results, not verified by me.

4x RTX 3090 24GB, vLLM, TP4 (from discussion #1, thanks to @mwyborski):

• vLLM selected the Marlin int4 kernel on all four ranks, with FlashAttention and bf16 KV. Serving window 256K tokens, no YaRN.
• Decode about 66.3 tok/s. Cold prefill 1,439 tok/s at 33.5K and 1,340 tok/s at 100.4K. Warm startup 191 s.
• GPU KV capacity about 846K tokens, roughly 3.31x a full 256K context.
• TP2xPP2 alternative: decode about 46.0 tok/s, cold prefill 2,376 tok/s at 33.5K. TP4 was kept because decode is about 44% faster, which wins long coding turns.
• Prefix caching works (a repeated 1,092-token probe fell from 1.34s to 0.22s, 71.8% cache hits). OpenAI/Anthropic tool calling, tool-result round trips, MCP schemas, and reasoning extraction all passed.