ROCm / ROCm/AMDMIGraphX

Update attention fusion to support SDXL Clip2 pattern

Open
#2,967 0 comments 0 reactions 1 assignee View on GitHub

@kahmed10 is already working on this.

Since Apr 23, 2025.

Perf Improve Tier1
Dominant language
C++
Stars
333
Forks
150
Avg merge
4d 19h
Merged PRs (30d)
54

Description

@25 = gpu::code_object[code_object=5232,symbol_name=mlir_reshape_transpose_reshape_reshape_transpose_slice_reshape_transpose_dot,global=46080,local=256,](@23,@19,@24) -> half_type, {20, 77, 77}, {5929, 77, 1}, target_id=0
@26 = load[offset=591360,end=828520](@1) -> half_type, {1, 20, 77, 77}, {118580, 5929, 77, 1}, target_id=0
@27 = reshape_lazy[dims={1, 20, 77, 77}](@25) -> half_type, {1, 20, 77, 77}, {118580, 5929, 77, 1}, target_id=0
@28 = multibroadcast[out_lens={1, 20, 77, 77},out_dyn_dims={}](@20) -> float_type, {1, 20, 77, 77}, {5929, 0, 77, 1}, target_id=0
@29 = gpu::code_object[code_object=6056,symbol_name=convert_add_reduce_max_sub_exp_reduce_sum_div_convert_kernel,global=197120,local=128,](@27,@28,@26) -> half_type, {1, 20, 77, 77}, {118580, 5929, 77, 1}, target_id=0
@30 = load[offset=828528,end=1025648](@1) -> half_type, {1, 20, 77, 64}, {98560, 4928, 64, 1}, target_id=0
@31 = gpu::code_object[code_object=5656,symbol_name=mlir_reshape_transpose_slice_dot,global=25600,local=256,](@29,@19,@30) -> half_type, {1, 20, 77, 64}, {98560, 4928, 64, 1}, target_id=0

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.