Skip to content

[AIMIGRAPHX-885] Add slice squeeze matcher#5004

Merged
causten merged 10 commits into
developfrom
add_slice_squeeze_matcher_clean
Jun 29, 2026
Merged

[AIMIGRAPHX-885] Add slice squeeze matcher#5004
causten merged 10 commits into
developfrom
add_slice_squeeze_matcher_clean

Conversation

@TedThemistokleous

@TedThemistokleous TedThemistokleous commented Jun 22, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator

Motivation

Further split of the changes regarding the MLP tower components

#4723

Handle the slice squeeze matching such that we can

rewrite slice->squeeze->pointwise/reduce into slice->pointwise/reduce->squeeze

This is used in the MLP tower before we perform a horizontal fusion down the pipeline for the pointwise activation op. In that case a SILU, but this can be used for other activations

Technical Details

Changelog Category

Add a CHANGELOG.md entry for any option other than Not Applicable

    • Added: New functionality.
    • Changed: Changes to existing functionality.
    • Removed: Functionality or support that has been removed. (Compared to a previous release)
    • Optimized: Component performance that has been optimized or improved.
    • Resolved Issues: Known issues from a previous version that have been resolved.
    • Not Applicable: This PR is not to be included in the changelog.

@TedThemistokleous
Add find_slice_squeeze matcher to simplify_reshapes
daa1388 
Port the find_slice_squeeze matcher from the MLP_prediction_towers branch.
This matcher rewrites slice->squeeze->pointwise/reduce into
slice->pointwise/reduce->squeeze (unsqueezing the other inputs), which lets
the squeeze propagate downstream and parallel slice branches merge back
together. Includes the associated unit tests.
@TedThemistokleous
Reuse find_op_shape_transform_op::is_reduce in find_slice_squeeze
ea45bce 
Replace the brittle "not pointwise" check with the shared is_reduce helper
from find_op_shape_transform_op so reduce/argmin/argmax detection is precise
and consistent with the rest of the pass.
@TedThemistokleous
Reuse find_op_shape_transform_op::insert for axis remapping
53e3c11 
Delegate the reduce/argmin axis remapping in find_slice_squeeze to the shared
insert() helper by building a source->common axes map for the unsqueeze.
This removes the hand-rolled axis-shifting logic, keeps behavior consistent
with find_op_shape_transform_op, and additionally handles layout permutations
for free.
@codecov

codecov Bot commented Jun 22, 2026 •
edited
Loading

Copy link
Copy Markdown

Codecov Report

✅ All modified and coverable lines are covered by tests.

Additional details and impacted files 
@@             Coverage Diff             @@
##           develop    #5004      +/-   ##
===========================================
+ Coverage    92.69%   92.69%   +0.01%     
===========================================
  Files          596      596              
  Lines        31603    31631      +28     
===========================================
+ Hits         29292    29320      +28     
  Misses        2311     2311
Files with missing lines Coverage Δ
src/simplify_reshapes.cpp 98.06% <100.00%> (+0.06%) ⬆️
🚀 New features to boost your workflow:
  • ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

@TedThemistokleous TedThemistokleous added the bugfix Fixes a bug found in the code. label Jun 22, 2026
@gh-app-migraphx-bot-pr-write

gh-app-migraphx-bot-pr-write Bot commented Jun 23, 2026 •
edited
Loading

Copy link
Copy Markdown
Test Batch New Rate (5ee371) Old Rate (90b175)* Diff Status
torchvision-resnet50 64 3,154.84 886.14 256.02% 🔆
torchvision-resnet50_fp16 64 6,639.57 1,735.02 282.68% 🔆
torchvision-densenet121 32 2,695.44 700.76 284.65% 🔆
torchvision-densenet121_fp16 32 4,536.64 1,439.55 215.14% 🔆
torchvision-inceptionv3 32 1,799.18 538.77 233.94% 🔆
torchvision-inceptionv3_fp16 32 2,819.61 182.21 1447.41% 🔆
cadene-inceptionv4 16 822.82 205.30 300.78% 🔆
cadene-resnext64x4 16 783.45 179.33 336.88% 🔆
slim-mobilenet 64 8,389.35 1,806.11 364.50% 🔆
slim-nasnetalarge 64 228.93 115.20 98.73% 🔆
slim-resnet50v2 64 3,165.04 1,029.11 207.55% 🔆
bert-mrpc-onnx 8 1,171.24 137.82 749.83% 🔆
bert-mrpc-tf 1 485.62 17.79 2629.72% 🔆
pytorch-examples-wlang-gru 1 329.27 35.57 825.58% 🔆
pytorch-examples-wlang-lstm 1 450.98 145.48 209.99% 🔆
torchvision-resnet50_1 1 748.50 45.24 1554.46% 🔆
cadene-dpn92_1 1 441.14 27.50 1504.13% 🔆
cadene-resnext101_1 1 366.42 28.35 1192.51% 🔆
onnx-taau-downsample 1 400.33 103.47 286.89% 🔆
dlrm-criteoterabyte 1 32.40 23.55 37.59% 🔆
dlrm-criteoterabyte_fp16 1 51.90 28.67 81.03% 🔆
agentmodel 1 10,218.53 640.03 1496.56% 🔆
unet_fp16 2 57.01 12.84 344.07% 🔆
resnet50v1_fp16 1 936.73 186.82 401.42% 🔆
resnet50v1_int8 1 941.90 150.02 527.86% 🔆
bert_base_cased_fp16 64 1,097.77 424.43 158.65% 🔆
bert_large_uncased_fp16 32 346.48 191.64 80.79% 🔆
bert_large_fp16 1 203.74 203.46 0.14% ✅
distilgpt2_fp16 16 2,094.56 791.60 164.60% 🔆
yolov5s 1 594.95 157.53 277.67% 🔆
tinyllama 1 16.94 28.67 -40.91% 🔴
vicuna-fastchat 1 43.87 33.14 32.39% 🔆
whisper-tiny-encoder 1 417.64 418.36 -0.17% ✅
whisper-tiny-decoder 1 415.17 128.92 222.04% 🔆
llama2_7b 1 20.34 11.70 73.90% 🔆
qwen1.5-7b 1 13.50 15.16 -10.97% 🔴
phi3-3.8b 1 26.70 13.69 95.02% 🔆
llama3-8b 1 8.55 15.19 -43.73% 🔴
whisper-large-encoder 1 4.15 5.78 -28.24% 🔴
whisper-large-decoder 1 24.42 31.17 -21.65% 🔴
mistral-7b 1 23.75 21.52 10.38% 🔆
FLUX.1-schnell 1 771.91 586.06 31.71% 🔆

Regressions detected 🔴

* No develop baseline was found for this PR's branch point; compared against the latest available develop run instead.

@gh-app-migraphx-bot-pr-write

gh-app-migraphx-bot-pr-write Bot commented Jun 23, 2026 •
edited
Loading

Copy link
Copy Markdown
Test Status Result
bert-mrpc-onnx ✅ PASSED: MIGraphX meets tolerance
bert-mrpc-tf ✅ PASSED: MIGraphX meets tolerance
pytorch-examples-wlang-gru ✅ PASSED: MIGraphX meets tolerance
pytorch-examples-wlang-lstm ✅ PASSED: MIGraphX meets tolerance
dlrm-criteoterabyte ✅ PASSED: MIGraphX meets tolerance
agentmodel ✅ PASSED: MIGraphX meets tolerance
unet ✅ PASSED: MIGraphX meets tolerance
resnet50v1 ✅ PASSED: MIGraphX meets tolerance
bert_base_cased_fp16 ✅ PASSED: MIGraphX meets tolerance
bert_large_uncased_fp16 🔴 FAILED: MIGraphX is not within tolerance - check verbose output
bert_large ✅ PASSED: MIGraphX meets tolerance
yolov5s ✅ PASSED: MIGraphX meets tolerance
tinyllama ✅ PASSED: MIGraphX meets tolerance
vicuna-fastchat ✅ PASSED: MIGraphX meets tolerance
whisper-tiny-encoder ✅ PASSED: MIGraphX meets tolerance
whisper-tiny-decoder ✅ PASSED: MIGraphX meets tolerance
distilgpt2_fp16 🔴 FAILED: MIGraphX is not within tolerance - check verbose output
llama2_7b ✅ PASSED: MIGraphX meets tolerance
qwen1.5-7b ✅ PASSED: MIGraphX meets tolerance
phi3-3.8b ✅ PASSED: MIGraphX meets tolerance
llama3-8b ✅ PASSED: MIGraphX meets tolerance
whisper-large-decoder ✅ PASSED: MIGraphX meets tolerance
mistral-7b ✅ PASSED: MIGraphX meets tolerance
FLUX.1-schnell ✅ PASSED: MIGraphX meets tolerance

@TedThemistokleous TedThemistokleous changed the title [AIMIGRAPHX-885] Add slice squeeze matcher clean [AIMIGRAPHX-885] Add slice squeeze matcher Jun 23, 2026
@TedThemistokleous TedThemistokleous self-assigned this Jun 23, 2026
@TedThemistokleous TedThemistokleous added high priority A PR with high priority for review and merging. and removed bugfix Fixes a bug found in the code. labels Jun 23, 2026
@TedThemistokleous TedThemistokleous marked this pull request as ready for review June 23, 2026 04:13
@TedThemistokleous TedThemistokleous requested a review from a team as a code owner June 23, 2026 04:19
pfultz2
pfultz2 requested changes Jun 23, 2026
View reviewed changes
Comment thread src/simplify_reshapes.cpp Outdated
Comment thread src/simplify_reshapes.cpp Outdated
@causten causten requested a review from pfultz2 June 25, 2026 20:43
shivadbhavsar
shivadbhavsar reviewed Jun 25, 2026
View reviewed changes
Comment thread test/optimize_module_test.cpp
@TedThemistokleous TedThemistokleous mentioned this pull request Jun 25, 2026
Closed
@causten causten merged commit 2b90a79 into develop Jun 29, 2026
38 of 40 checks passed
@causten causten deleted the add_slice_squeeze_matcher_clean branch June 29, 2026 15:16
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@pfultz2 pfultz2 pfultz2 approved these changes
@shivadbhavsar shivadbhavsar shivadbhavsar approved these changes
@causten causten Awaiting requested review from causten causten is a code owner

Assignees

@TedThemistokleous TedThemistokleous

Labels

high priority A PR with high priority for review and merging.

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@TedThemistokleous @pfultz2 @shivadbhavsar @causten