Qwen3.8-27B-NVFP4-AWQ-AutoRound-allfp4

Qwen3.8-27B-NVFP4-AWQ-AutoRound-allfp4

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

23.8 GB — all 64 MLP layers at 4 bits, with no FP8 last-band protection, and statistically indistinguishable from the 24.7 GB variant that keeps it. 7 GB smaller than FP8 and ~25% faster.

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 ~23.8 GB checkpoints; FP8 and the 24.7 GB sibling are shown for reference. 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.

The FP8 last band is not needed

Most NVFP4 recipes for this model protect MLP layers 56–63 at FP8. This build removes that, putting all 64 MLP layers at 4 bits. Against the otherwise-identical sibling that keeps it: confident 1.85% → 2.00% (paired McNemar z = 1.88, not significant) and certain 0.17% → 0.16% (z = 0.20, not significant). The protection costs 0.87 GB and 3% throughput and buys nothing measurable here.

It is load-bearing for round-to-nearest builds, which is presumably why it became convention: among RTN checkpoints of this model, those protecting the band sit at 0.25% certain and those that do not sit at 0.70%.

What the algorithm contributes

RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead has a byte-identical precision map to this model — same 23.8 GB, NVFP4 on all 64 MLP layers, FP8 attention and GDN, BF16 lm_head — differing only in using round-to-nearest:

At identical size and precision, the algorithm is worth a 4.4× reduction in the worst damage bucket.

AWQ rescales per input channel; because gate_proj and up_proj share one input, the reciprocal folds into post_attention_layernorm at zero size cost. AutoRound optimises rounding and clipping with SignSGD against a block-wise reconstruction loss rather than GPTQ's per-layer weight-MSE proxy.

Usage

from vllm import LLM
llm = LLM("TelperionAI/Qwen3.8-27B-NVFP4-AWQ-AutoRound-allfp4", 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-allfp4", 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 under quantization_config — the on-disk contract llm-compressor emits and vLLM reads in BaseKVCacheMethod. Without them vLLM falls back to a scale of 1.0.

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

Provenance: these scales were measured on Qwen3.8-27B-NVFP4-AWQ-AutoRound, a different quantization of the same base model, over a 2,353-document / 3.7M-token corpus — the size at which the post-RoPE V absmax stops moving. K/V range is a property of the model rather than of the weight format (per-layer amax agrees to a median ratio of 1.0002 across our trellis / FP8 / INT4 / NVFP4 builds), so they transfer. They were not measured on this checkpoint specifically; cross-build agreement is excellent at the median but ranges 0.88–1.36 in the tails.