biondizzle 53c25bee0b rewrite: cudagraph-safe runner - no dynamic slicing, no GPU scalar indices ...

- Removed all [:total_slots] dynamic slicing with GPU scalars
- slot_hidden gathers from hidden_states directly using sorted_token_ids
- scatter_add uses full sorted_token_ids (padding slots have zero weight)
- _assemble_scales_cudagraph_safe returns 2D via padded_scales.shape[0]
- Fixed padded_scales_buf allocation via float16->float8 cast
- GEMM output size: n_dim * 2 for float4_e2m1fn_x2 packed format

2026-05-16 18:44:25 +00:00

..

patches

fix: remove broken hc_head warmup (wrong tensor shape)

2026-05-16 10:11:34 +00:00

nvfp4_cutedsl.py

rewrite: cudagraph-safe runner - no dynamic slicing, no GPU scalar indices

2026-05-16 18:44:25 +00:00