Support for accumulate nodes traversal and to access op names in the compare function (#57685)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/57685
- Accumulate traversal : `minimizer.settings.traverse_method = "accumulate" `
- Feature
- net_min_tests
- Return op name to the compare function so that we can map the cosine similarity to the individual ops
- Fix the settings combinations in net_min_tests
Test Plan:
buck test glow/fb/nnpi/lowering:net_min_tests
NNPI_LOG_LEVEL=5 USE_INF_API=1 buck run mode/opt -j 12 --config fbcode//cxx.link_weight=3 --config misc.strip_binaries=debug-non-line -c glow.nnpi_project_name='fb-nnpi-nextgen' ai_codesign/video/inference:xrayvideo_2019a_eval -- --job create --model_a model_prod --device_a PTCPU --trace_a none --model_b model_v3 --device_b NNPI --trace_b fusion --replace_b true --log_level INFO --use_scrambled false --save_repro false --num_ab_runs 0 --symbolic_trace_b true --save_modified_model_b false
USE_INF_API=1 buck test glow/fb/nnpi/lowering:net_min_tests
Reviewed By: 842974287
Differential Revision: D27867010
fbshipit-source-id: 6a756468b1f1fe24ef0400669d911825a7562484