Fix divide-by-zero NaN in _fp8_act_quant_kernel

s = tl.max(tl.abs(x)) / 448.0 is zero when a 128-block is all-zero. Then y = x/s = NaN propagates through every downstream FP8 matmul.

-    s = tl.max(tl.abs(x)) / 448.0  # float8_e4m3fn max
-    y = (x / s).to(y_ptr.dtype.element_ty)
+    amax = tl.max(tl.abs(x))
+    s = tl.maximum(amax / 448.0, 1e-12)  # eps floor; float8_e4m3fn max = 448
+    y = (x / s).to(y_ptr.dtype.element_ty)

Same hunk in build/torch-cuda, build/torch-xpu, build/torch-rocm. Matches _per_token_group_quant_fp8 (vllm) and _per_token_group_quant_8bit (sglang).

Trigger

Attention masks zero out hidden states at padded positions before the first FP8 quant. Fires on Qwen/Qwen3-8B-FP8, Qwen/Qwen3.5-9B-FP8, Qwen/Qwen3.5-27B-FP8, RedHatAI/Qwen3-8B-FP8-block when a batch contains any padding.

Validation

H100 80GB, Qwen/Qwen3.5-27B-FP8 via transformers git main, Triton path forced.

Reproducer in this PR: tests/test_act_quant_zero_block.py (four pytest cases, CUDA required).

References

Originally diagnosed and retracted in huggingface/transformers#42831. The retraction was about a separate lm-eval padding issue; the divide-by-zero kernel bug is real and never landed a fix.