AMDMIGraphX
AMDMIGraphX copied to clipboard
find_split_concat fix
Fix a bug that caused an exception in torch-migraphx benchmarks. Closes Issue https://github.com/ROCm/AMDMIGraphX/issues/3009
Codecov Report
All modified and coverable lines are covered by tests :white_check_mark:
Project coverage is 91.78%. Comparing base (
b7ab6e7) to head (4d497ca). Report is 158 commits behind head on develop.
Additional details and impacted files
@@ Coverage Diff @@
## develop #3067 +/- ##
========================================
Coverage 91.78% 91.78%
========================================
Files 486 486
Lines 18936 18938 +2
========================================
+ Hits 17381 17383 +2
Misses 1555 1555
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
| Test | Batch | Rate new 4d497c |
Rate old 06eef0 |
Diff | Compare |
|---|---|---|---|---|---|
| torchvision-resnet50 | 64 | 2,789.94 | 2,790.88 | -0.03% | :white_check_mark: |
| torchvision-resnet50_fp16 | 64 | 6,203.11 | 6,211.36 | -0.13% | :white_check_mark: |
| torchvision-densenet121 | 32 | 2,087.12 | 2,090.96 | -0.18% | :white_check_mark: |
| torchvision-densenet121_fp16 | 32 | 3,621.99 | 3,612.27 | 0.27% | :white_check_mark: |
| torchvision-inceptionv3 | 32 | 1,596.72 | 1,596.79 | -0.00% | :white_check_mark: |
| torchvision-inceptionv3_fp16 | 32 | 2,559.00 | 2,560.84 | -0.07% | :white_check_mark: |
| cadene-inceptionv4 | 16 | 716.39 | 715.97 | 0.06% | :white_check_mark: |
| cadene-resnext64x4 | 16 | 677.62 | 678.10 | -0.07% | :white_check_mark: |
| slim-mobilenet | 64 | 5,814.32 | 5,816.86 | -0.04% | :white_check_mark: |
| slim-nasnetalarge | 64 | 154.27 | 154.28 | -0.01% | :white_check_mark: |
| slim-resnet50v2 | 64 | 2,577.05 | 2,577.84 | -0.03% | :white_check_mark: |
| bert-mrpc-onnx | 8 | 968.28 | 969.10 | -0.08% | :white_check_mark: |
| bert-mrpc-tf | 1 | 406.57 | 412.54 | -1.45% | :white_check_mark: |
| pytorch-examples-wlang-gru | 1 | 394.80 | 392.40 | 0.61% | :white_check_mark: |
| pytorch-examples-wlang-lstm | 1 | 370.17 | 372.63 | -0.66% | :white_check_mark: |
| torchvision-resnet50_1 | 1 | 600.86 | 599.43 | 0.24% | :white_check_mark: |
| cadene-dpn92_1 | 1 | 387.71 | 385.42 | 0.59% | :white_check_mark: |
| cadene-resnext101_1 | 1 | 323.23 | 325.87 | -0.81% | :white_check_mark: |
| onnx-taau-downsample | 1 | 306.65 | 306.83 | -0.06% | :white_check_mark: |
| dlrm-criteoterabyte | 1 | 28.49 | 28.53 | -0.16% | :white_check_mark: |
| dlrm-criteoterabyte_fp16 | 1 | 47.21 | 47.19 | 0.04% | :white_check_mark: |
| agentmodel | 1 | 7,877.61 | 7,446.79 | 5.79% | :high_brightness: |
| unet_fp16 | 2 | 57.59 | 57.58 | 0.01% | :white_check_mark: |
| resnet50v1_fp16 | 1 | 894.07 | 878.48 | 1.77% | :white_check_mark: |
| resnet50v1_int8 | 1 | 788.57 | 789.53 | -0.12% | :white_check_mark: |
| bert_base_cased_fp16 | 64 | 1,021.93 | 1,022.15 | -0.02% | :white_check_mark: |
| bert_large_uncased_fp16 | 32 | 298.98 | 299.07 | -0.03% | :white_check_mark: |
| bert_large_fp16 | 1 | 156.27 | 156.36 | -0.06% | :white_check_mark: |
| distilgpt2_fp16 | 16 | 1,830.36 | 1,830.80 | -0.02% | :white_check_mark: |
| yolov5s | 1 | 472.92 | 472.82 | 0.02% | :white_check_mark: |
| tinyllama | 1 | 32.99 | 33.02 | -0.08% | :white_check_mark: |
| vicuna-fastchat | 1 | 157.26 | 158.10 | -0.53% | :white_check_mark: |
| whisper-tiny-encoder | 1 | 349.91 | 352.27 | -0.67% | :white_check_mark: |
| whisper-tiny-decoder | 1 | 401.47 | 402.57 | -0.27% | :white_check_mark: |
Check results before merge :high_brightness:
:red_circle:bert_large_uncased_fp16: FAILED: MIGraphX is not within tolerance - check verbose output
I would assume a test for concat(relu, slice, relu, slice) should reproduce the issue. If that doesnt trigger the issue then we can investigate why. We can use MIGRAPHX_TRACE_MATCHES=1 to see if another matcher is being triggered over this one, and from there we can tweak the test to avoid that.