masahi commented on code in PR #14274:
URL: https://github.com/apache/tvm/pull/14274#discussion_r1145421269
##########
python/tvm/contrib/cutlass/gemm_operation.py:
##########
@@ -369,7 +369,8 @@ def instantiate_gemm_template(attrs):
{
"bias_decl": "void* ptr_bias = (void*)(${bias_arg}->data);\n",
"ptr_c": "ptr_bias",
- "c_stride": "${bias_arg}->ndim == 1 ? 0 : " + attrs["ldc"],
+ # "c_stride": "${bias_arg}->ndim == 1 ? 0 : " + attrs["ldc"],
Review Comment:
yes, otherwise there is an ambiguity in which of bias add or residual add
should be used for `matmul + add` pattern. This causes a subtle issue for
`FusesOpsByPattern`, which applies rewriting over a list of patterns which
should be ordered by the "size" of a pattern.
We are using a residual add pattern only for matmul + bias_add + residual
add (two add).
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
