Qwen3.8-27B-INT4-AWQ-GPTQ

Mixed-precision INT4 (W4A16) quantization of Qwen/Qwen3.8-27B, built with llm-compressor using AWQ activation-aware scaling followed by GPTQ.

25.1 GB, and measurably more faithful to the BF16 model than the official FP8 release on high-confidence token agreement — at 5.8 GB less. Runs on anything Turing or newer (no FP8/FP4 hardware required).

Recipe

Two passes:

1. AWQ — per-input-channel scaling on post_attention_layernorm → {gate_proj, up_proj} and up_proj → down_proj. Gate and up share one input, so the reciprocal scale folds into the norm weights: zero size and zero throughput cost.
2. GPTQ — Hessian-based error compensation, actorder="static", dampening_frac=0.01.

Calibration: 924 sequences × 1024 tokens of a balanced Nemotron-v2 blend (25% code, 25% math, 20% STEM, 20% chat, 10% multilingual).

lm_head and embed_tokens stay BF16, matching Qwen's own official FP8 release.

Benchmarks

Against the BF16 base on 142,727 tokens of self-distilled thinking-mode output plus 200 free greedy generations. vLLM 0.27.1, TP=2, 2×B300.

Bold marks the best value in each column among the 4-bit checkpoints; the FP8 row is a reference at a different precision and size class, so it is excluded from the comparison. All sizes are on-disk tensor bytes and include the ~0.85 GB BF16 MTP head, which every checkpoint in this table ships. Subtract ~0.85 GB for a no-MTP comparison.

Columns. top-1 is raw argmax agreement with BF16. The four bucket columns are disagreement rates, split by how confident the base model was at that position (top1−top2 logprob margin): near-tie <0.5, moderate 0.5–2, confident 2–5, certain >5. Only confident and certain are real damage — a flip where the base model was itself nearly tied is numerical noise. divmed is the median token index at which free greedy generation first diverges from BF16 (higher is better).

Perplexity is deliberately excluded. On this model it is anti-correlated with quality — the FP4 checkpoint with the best perplexity we measured also had the worst certain-bucket damage. Do not rank quantizations of this model by perplexity.

Why INT4 rather than NVFP4

At ~4.6 effective bits, INT4 group-32 asymmetric gives 16 uniformly-spaced levels plus a per-group zero point. NVFP4 gives 8 non-uniform FP4 levels at 4.5 bits with no zero point. On this model the zero point matters — MLP weight distributions are skewed — and the same recipe in NVFP4 lands at 1.85% confident versus 0.93% here.

NVFP4's advantage is hardware, not numerics: on Blackwell it decodes at 10590 tok/s against 4617 here, because native NVFP4 tensor-core paths beat Marlin INT4. Pick NVFP4 if you are throughput-bound on Blackwell; pick this if you want fidelity, or you are on Ampere/Ada where FP4 and FP8 do not exist.

Usage

from vllm import LLM
llm = LLM("TelperionAI/Qwen3.8-27B-INT4-AWQ-GPTQ", tensor_parallel_size=2)

W4A16 needs only compute capability ≥ 7.5 (Turing), so this runs on 3090/4090/A100/H100 as well as Blackwell. Throughput above was measured on B300 and is not representative of Ampere, where Marlin INT4 is comparatively better tuned.

Speculative decoding (MTP)

The model's MTP (multi-token prediction) head is included, in BF16, and works with vLLM's mtp speculative decoding:

from vllm import LLM
llm = LLM("TelperionAI/Qwen3.8-27B-INT4-AWQ-GPTQ", tensor_parallel_size=2,
          speculative_config={"method": "mtp", "num_speculative_tokens": 2})

Qwen3_5ForConditionalGeneration does not carry mtp.* in its state dict, so llm-compressor never sees it and it is silently dropped, even though config.json still declares mtp_num_hidden_layers: 1. It is grafted back in here from the base checkpoint and excluded from quantization (re:.*mtp.* in quantization_config.ignore; without that exclusion the quantization target regexes also match mtp.layers.0.mlp.* and vLLM fails to load). Draft quality drives acceptance rate, so it is kept at full precision rather than quantized.

Acceptance rate has not been measured; the head is verified to load and generate.

Limitations

• 25.1 GB does not fit a single 24 GB card. The GDN-4bit variant is 22.6 GB, which is closer but still tight once KV cache is counted.
• Throughput on Blackwell is ~2.3× below NVFP4. This checkpoint trades speed for fidelity.
• Single evaluation corpus. All numbers come from one self-distilled corpus. The margins over FP8 and cyankiwi are statistically solid but have not been replicated on a second distribution, nor on downstream task benchmarks.
• Vision tower untouched (BF16); evaluated as a text model.

KV-cache quantization (calibrated scales included)

This checkpoint carries per-layer k_scale / v_scale in model-kv-scales.safetensors, with kv_cache_scheme recorded under quantization_config — the on-disk contract llm-compressor emits and vLLM reads in BaseKVCacheMethod. With no scales present vLLM falls back to 1.0.

--kv-cache-dtype fp8      # 32 KiB/token, from 64 KiB at fp16

Qwen3.8-27B is a hybrid stack: only 16 of its 64 layers are full_attention and carry a KV cache at all, so this is 32 scalars in a 4 KB file.

Scales were observed on this quantized model, at the point vLLM caches K/V (post-RoPE, at the attention dispatch), over a 2353-document / 3,713,778-token corpus, taking a running absmax per layer (memoryless_minmax, llm-compressor's default KV observer); k_scale = amax/448.

Corpus size matters more than it looks, and only for V. absmax is a tail statistic: it sits flat, then jumps when a rare document appears, so a small calibration set looks converged when it is not. A 48-document pass under-estimated V by up to 1.9x and a 231-document pass was still up to 1.5x low; at this size no tensor set a new maximum over the final 50 documents. Under-estimating clips exactly the outliers a cache most needs to keep. K is converged by ~100 documents and needs none of this; V's range also grows ~13x from the first attention layer to the last, which is why the scales are per-layer rather than global.