Qwen3.8-27B-NVFP4-AWQ-AutoRound

Mixed-precision NVFP4 quantization of Qwen/Qwen3.8-27B, built with llm-compressor using AWQ activation-aware scaling followed by AutoRound.

24.7 GB — 6 GB smaller than FP8 and ~22% faster, while recovering 84% of the gap between a plain NVFP4+GPTQ build and FP8 on high-confidence token agreement.

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. The scales merge fully into weights, so unlike rotation methods (QuIP/SpinQuant) this still runs under tensor parallelism.
2. AutoRound — SignSGD-optimized rounding and clipping (200 iters) against a block-wise reconstruction loss, replacing GPTQ. Mean block loss fell ~18%.

Calibration: 1358 × 1024-token packed sequences (1.39M tokens) from 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

Measured 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 FP4 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 comparison it is anti-correlated with quality — the checkpoint with the best perplexity (RadixArk, −1.75%) has the worst certain-bucket damage measured (0.70%, 4× this model's). Do not rank FP4 checkpoints of this model by perplexity.

What AutoRound contributes

Holding the recipe fixed and swapping only the rounding algorithm, confident damage falls from 2.69% (GPTQ) → 1.85% (AutoRound) — a paired McNemar z of 9.26 over 33,812 high-confidence positions. Against a plain NVFP4+GPTQ build without AWQ (3.97%), the two passes together close 84% of the distance to FP8.

The likely reason: GPTQ minimizes a per-layer weight-MSE proxy, while AutoRound optimizes the block's actual output. Several methods we tested that sharpened the local proxy (GPTQv2, error-guided layer allocation, a 3× larger AWQ grid) all made end-to-end quality worse on this model.

Usage

from vllm import LLM
llm = LLM("TelperionAI/Qwen3.8-27B-NVFP4-AWQ-AutoRound", tensor_parallel_size=2)

Requires a Blackwell-class GPU for native NVFP4, and vLLM with compressed-tensors.

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-NVFP4-AWQ-AutoRound", 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

• AutoRound ran at effective batch size 1. The default (8) raised this model has not been supported on this architecture. Gradients are noisier than intended, so these numbers likely understate what the method can do here.
• Calibration sequences are packed to a uniform 1024 tokens (AutoRound concatenates batch elements), so they cross document boundaries.
• Single evaluation corpus. All numbers come from one self-distilled corpus.
• 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.