biondizzle d635dcbbb6 fix: keep token_indices on CPU, index with CPU sort_idx ...

CuTeDSL's cute.compile corrupts GPU memory during JIT compilation.
Keeping token_indices on CPU and using sort_idx.cpu() for indexing
avoids the corruption. The .to(device) call after indexing moves the
result back to GPU for the hidden_states indexing.

2026-05-17 08:29:18 +00:00

..

patches

fix: set _l1_activation_global_scale (with underscore) — attribute name mismatch

2026-05-17 03:35:20 +00:00

nvfp4_cutedsl.py

fix: keep token_indices on CPU, index with CPU sort_idx

2026-05-17 08:29:18 +00:00