biondizzle 62abf41b03 Revert deepseek_v4_attention.py to ffc2264 — don't nuke existing patches ...

The file at ffc2264 already had our BF16 wo_a path (_apply_inv_rope_bf16 +
BMM + all-gather) with FP8 fallback. I was replacing it from the wrong
vllm source, losing all prior work. Restored to the known-good version.

2026-05-19 06:52:40 +00:00

..

kernels/linear/nvfp4

Fix activation global scale: don't double-invert input_global_scale_inv

2026-05-19 06:03:08 +00:00

patches

Revert deepseek_v4_attention.py to ffc2264 — don't nuke existing patches

2026-05-19 06:52:40 +00:00

cutedsl_quant_method.py

Replace autograd.Function with torch.library.custom_op for Dynamo compat

2026-05-19 01:54:48 +00:00

nvfp4_cutedsl.py

Replace autograd.Function with torch.library.custom_op for Dynamo compat

2026-05-19 01:54:48 +00:00