[Bugfix][MoE] Clamp BLOCK_SIZE_K to the FP8 quantization block in the batched Triton MoE (#60733)

This commit is contained in:
Anirudh Ravichandran
2026-10-13 17:02:17 -04:00
committed by GitHub
parent 292da21f9d
commit e90158d912
2 changed files with 9 additions and 1 deletions
@@ -450,6 +450,8 @@ def invoke_moe_batched_triton_kernel(
BLOCK_M = config["BLOCK_SIZE_M"]
BLOCK_N = config["BLOCK_SIZE_N"]
BLOCK_K = config["BLOCK_SIZE_K"]
if block_shape is not None:
BLOCK_K = min(BLOCK_K, min(block_shape[0], block_shape[1]))
grid = (
expert_num_tokens.size(0),