Qwen3.5-27B-NVFP4-QAD-W4A16 (final, step 2640) — W4A4 serving schema
NVFP4 a4schema export (NVFP4 weights + activation input scales for real
W4A4 serving) of the NV-W4A16 QAD arm final checkpoint. This is the
cross-serving variant used to test whether W4A16 training suffices for W4A4
serving on NVFP4; the weight-only variant lives at
weili-0234/Qwen3.5-27B-NVFP4-QAD-W4A16-LR1e-5-s2640.
Training provenance and the fwd-KL 0.005815 headline are documented on the
weight-only card; both variants come from the same checkpoint
(export_nvfp4_vllm.py, with and without --weight-only). Load-smoked with
vLLM 0.25.1 (first engine start pays a FlashInfer fp4_gemm autotune,
~15-25 min cold). Standardized 7-benchmark w4a4 results will be added when
the 2026-07-25 campaign completes.
