cyx-6 commented on code in PR #14150:
URL: https://github.com/apache/tvm/pull/14150#discussion_r1125160077
##########
python/tvm/contrib/cutlass/attention_operation.py:
##########
@@ -24,12 +24,12 @@ def instantiate_attention_template(attrs, func_args):
based on a template and the provided attribute map."""
bias_template = """
- CHECK(${arg3}->ndim == 4); // B, S, N, S'
+ CHECK(${arg3}->ndim == 4); // B, N, S, S'
p.attn_bias_ptr = reinterpret_cast<T *>(${arg3}->data);
- p.bias_strideH = p.num_keys; // S'
- p.bias_strideM = p.bias_strideH * p.num_heads; // S' * N
- p.bias_strideB = p.bias_strideM * p.num_queries; // S' * N * S
+ p.bias_strideM = p.num_keys; // S'
+ p.bias_strideH = p.bias_strideM * p.num_queries; // S' * S
Review Comment:
The broadcast for bias is added. Now we accept the bias with layout of
`BS'`, `BSS'` and `BNSS'`.
--
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]
