Qwen3.8-27B-AWQ-INT4

Qwen3.8-27B — AWQ INT4 (compressed-tensors)

4-bit weight quantisation of Qwen/Qwen3.8-27B, built so the model fits a single 24 GB consumer GPU.

Note on the qwen3_5 tag: Qwen3.8 is implemented with the Qwen3.5 architecture (model_type: qwen3_5, Qwen3_5ForConditionalGeneration), so the Hub derives that family tag automatically — the upstream Qwen/Qwen3.8-27B repo carries it too. This is Qwen 3.8.

55.6 GB → 18 GB. Verified serving on an RTX 3090 (sm_86) under vLLM.

Usage

vLLM (recommended)

vllm serve avyukth/Qwen3.8-27B-AWQ-INT4 \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.93 \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml

Text-only serving reclaims the vision encoder cache (~5k extra KV tokens on a 24 GB card):

  --limit-mm-per-prompt '{"image": 0}'

transformers

from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "avyukth/Qwen3.8-27B-AWQ-INT4", device_map="auto")
processor = AutoProcessor.from_pretrained("avyukth/Qwen3.8-27B-AWQ-INT4")

Measured results

Measured on an RTX 3090 (24 GB, sm_86), vLLM, max_model_len=8192.

The GSM8K figure is on 40 problems, not the full 1319-item test set. It is enough to show reasoning survived quantisation; it is not a leaderboard number, and no bf16 side-by-side was run, so the exact delta from the base model is unmeasured.

Throughput (RTX 3090, 256 tokens/request, ignore_eos)

Throughput saturates at concurrency 8; beyond that, aggregate is flat and only latency grows.

Limitations

• 48 of 64 layers received INT4 without activation-aware scaling. Qwen3.5/3.8 interleaves 16 full-attention layers with 48 linear_attn (Qwen3_5GatedDeltaNet) layers. AWQ's scaling search must replay a parent module to collect reference outputs, and Qwen3_5GatedDeltaNet.forward has the runtime signature (self, *args, **kwargs) — a decorator drops the real one — so llm-compressor's captured arguments collapse into a single nested kwargs key and replay fails. Those layers are therefore quantised by round-to-nearest rather than AWQ. Their MLPs and all full-attention layers do get proper AWQ scaling. No measurable reasoning cost was found (GSM8K above), but this is not the same as a full AWQ model.
• Vision tower is bf16. Its intermediate_size is 4304, which is not divisible by group_size=128 (4304/128 = 33.625; the only divisor under 256 is 16), so those layers cannot be group-quantised at all. This costs 0.92 GB. Per-channel quantisation would be the workaround if that matters.
• Marlin thread-tile padding. vLLM warns that some GDN projection shapes need padding, so those layers pad/slice on every forward. Correctness is unaffected; some throughput is lost.
• Symmetric W4A16, chosen for Marlin support on Ampere. Asymmetric (W4A16_ASYM) tracks AWQ's usual formulation slightly more closely.
• Calibration: 128 samples × 512 tokens from HuggingFaceH4/ultrachat_200k. Calibration data is general chat, not domain-specific.

Build recipe

Produced with llm-compressor 0.13.0. Two details that are easy to get wrong on this architecture:

from llmcompressor.modifiers.transform.awq import AWQModifier   # NOT modifiers.awq
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.modifiers.transform.awq.dynamic_mappings import (
    build_hybrid_attention_mappings)

# The dynamic builder emits layer-index-scoped rules per attention kind.
# Drop only the linear-attention rule; the rest keep proper AWQ scaling.
maps = [m for m in build_hybrid_attention_mappings(model)
        if not any("linear_attn" in b for b in m.balance_layers)]

recipe = [
    AWQModifier(mappings=maps, duo_scaling="both"),
    QuantizationModifier(targets=["Linear"], scheme="W4A16",
                         ignore=["lm_head", "re:.*visual.*", "re:.*mtp.*"]),
]

Load the model with the class its config declares (Qwen3_5ForConditionalGeneration), not AutoModelForCausalLM — the latter resolves to the text-only class, which silently drops the vision tower and writes a config that no longer matches the weights.

Save the tokenizer from a fresh instance. HF fast tokenizers persist truncation state, so saving the tokenizer used for calibration bakes "truncation": {"max_length": 512} into tokenizer.json, which silently clips every prompt beyond the calibration length.

License

Apache 2.0, inherited from the base model. All credit for the model itself goes to the Qwen team; this repository contains only a quantisation of their weights.