MXFP4 vLLM-schema export of the MX-W4A16 QAD arm final checkpoint (weight-only fake-quant training) from the stage-3 W4A4-vs-W4A16 training study on dense Qwen/Qwen3.5-27B. Activation quantization is applied dynamically by vLLM at serving time, so this single artifact serves both W4A4 and W4A16 modes.

Training provenance

QAD temp-1.0 pure-KL distillation (teacher = BF16 Qwen3.5-27B), QATFactory @ 8e7cf1a + save fix + the 2026-07-24 context patches.
configs/llm/qwen3_5_27b_mxfp4_qad.yaml (W4A16 training), lr 1.0e-5 cosine (1% warmup), 2640 steps (~1 epoch), global batch 32, seq 8192, seed 42, BF16 FSDP2, save/250, eval/100.
Data: openperfectblend_100k Qwen3.5-9B-think JSONLs (train md5 0406bb3a7a482352360716a1bc5e9e04; disjoint eval split).
8x B200 (research-common-b200-jbom-5c-01), crash-free 13.2h run, completed 2026-07-25 16:23 PT, exit 0.

Result headline

Final held-out eval fwd-KL 0.008904 vs the BF16 teacher (own W4A16 training quant config; plateau ~0.0089 from step 1300). The matched MX-W4A4-trained twin reached 0.0316 in its own (harder) W4A4 config — the decisive cross-comparison (both finals served at real W4A4) is being produced by the standardized 7-benchmark campaign and will be added here.

Reproduction (export step)

bash

python scripts/export_mxfp4_vllm.py \
  --source  /scratch/wxu/mxfp4/outputs/27b-mxfp4-w4a16-lr1.0e-5-2640 \
  --model-assets /scratch/wxu/mxfp4/models/Qwen3.5-27B \
  --output  /scratch/wxu/mxfp4/exports/27b-mxw4a16-s2640 --device cuda:0

Load-smoked with vLLM 0.25.1: loads clean, completes text correctly.

weili-0234/Qwen3.5-27B-MXFP4-QAD-W4A16-LR1e-5-s2640

Qwen3.5-27B-MXFP4-QAD-W4A16 (final, step 2640)

Training provenance

Result headline

Reproduction (export step)

weili-0234/
Qwen3.5-27B-MXFP4-QAD-W4A16-LR1e-5-s2640