[Bugfix][MoE] Clamp BLOCK_SIZE_K to the FP8 quantization block in the batched Triton MoE (#60733)
This commit is contained in:
@@ -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),
|
||||
|
||||
Reference in New Issue
Block a user