Jeff Huang 141f77aa12 [CK Tile] Fix FMHA LSE calculation and potential division by zero (#3326) ...

This commit addresses numerical stability issues in the BlockFmhaPipelineQRKSVS pipeline when bias has -inf masking values:
1. Explicitly handle the case where the accumulated exponential sum (l) is zero. In this case, the LSE is now correctly set to negative infinity, preventing log(0) errors.
2. Extend the zero-check protection in the normalization step to cover the ELEMENTWISE_BIAS case, preventing potential division by zero.

2026-01-13 13:52:26 +08:00

..

block

[FMHA] Batch Prefill Support Improvements: Change KV Cache Layout & Large Page Size Support (#3442)

2026-01-05 18:41:47 +08:00

kernel

[FMHA] Batch Prefill Support Improvements: Change KV Cache Layout & Large Page Size Support (#3442)

2026-01-05 18:41:47 +08:00

pipeline

[CK Tile] Fix FMHA LSE calculation and potential division by zero (#3326)

2026-01-13 13:52:26 +08:00