Cortex-M: Fuse relu activation into quantized_add#18462
Merged
rascani merged 5 commits intopytorch:mainfrom Mar 27, 2026
Merged
Cortex-M: Fuse relu activation into quantized_add#18462rascani merged 5 commits intopytorch:mainfrom
rascani merged 5 commits intopytorch:mainfrom
Conversation
ResNet8 has skip connections with relu(add(conv(x), skip(x))). The ActivationFusionPass only fused relu into conv/linear, leaving 3 unfused relu ops that fell through to portable aten::relu.out which incorrectly clamps int8 tensors to literal 0 instead of the quantized zero_point, causing numerical mismatches on the FVP. Add fused activation patterns (relu, hardtanh, clamp) for add/add_ to quantizer_support.py BINARY_OP_PATTERNS so the quantizer produces activation-aware quantization bounds. Add aten.add.Tensor to ActivationFusionPass FUSE_OPS. Update QuantizedOpFusionPass to read activation bounds from output_qparams and pass them to quantized_add. Update the quantized_add operator (schema, meta, impl, C++) to accept activation_min/activation_max parameters. Co-authored-by: Claude <noreply@anthropic.com>
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18462
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New Failures, 1 Cancelled Job, 3 Unrelated FailuresAs of commit e527a04 with merge base 7c79395 ( NEW FAILURES - The following jobs have failed:
CANCELLED JOB - The following job was cancelled. Please retry:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
Author
|
@AdrianLundell - I still need to add tests for this, but I wanted to make sure this is the right approach for the Quantizer. |
This PR needs a
|
Add add_relu, add_relu_channels_last, add_hardtanh, and add_hardtanh_channels_last test cases to test_add.py verifying that relu/hardtanh activations are fused into quantized_add. Remove the conv_add_relu xfail from test_nn_modules.py since the fusion now works. Co-authored-by: Claude <noreply@anthropic.com>
Remove add_.Tensor + activation fused patterns from BINARY_OP_PATTERNS. Functionalization converts inplace ops to out-of-place before the quantizer runs, so these patterns are never matched. Co-authored-by: Claude <noreply@anthropic.com>
AdrianLundell
approved these changes
Mar 27, 2026
Contributor
Author
|
Failured unrelated. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
ResNet8 has skip connections with relu(add(conv(x), skip(x))). The ActivationFusionPass only fused relu into conv/linear, leaving 3 unfused relu ops that fell through to portable aten::relu.out which incorrectly clamps int8 tensors to literal 0 instead of the quantized zero_point, causing numerical mismatches on the FVP.
Add fused activation patterns (relu, hardtanh, clamp) for add/add_ to quantizer_support.py BINARY_OP_PATTERNS so the quantizer produces activation-aware quantization bounds. Add aten.add.Tensor to ActivationFusionPass FUSE_OPS. Update QuantizedOpFusionPass to read activation bounds from output_qparams and pass them to quantized_add. Update the quantized_add operator (schema, meta, impl, C++) to accept activation_min/activation_max parameters.