Repository navigation
Qualcomm AI Engine Direct - Fix a QNN crash on AMD Hosts - #22543
Merged
Merged
Conversation
qti-horodnic
requested review from
abhinaykukkadapu and
psiddh
as code owners
September 3, 2026 23:57
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22543
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New Failures, 1 Unclassified FailureAs of commit 2053ea8 with merge base 843f77e ( NEW FAILURES - The following jobs have failed:
UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
Author
|
@pytorchbot label "release notes: qualcomm" |
pytorch-bot
Bot
added
the
release notes: qualcomm
psiddh
approved these changes
Sep 4, 2026
psiddh
pushed a commit
that referenced
this pull request
Sep 8, 2026
…r Snapdragon 7+ Gen 3 (#22542) ### Summary Adding SoC support for SM7675 (Snapdragon 7+ Gen 3) and SM8635 (Snapdragon 8s Gen 3), both HTP V73. Scope of Support: - Quantized HTP: supported*. Quantized HTP validated across 8a8w/16a4w/16a2w, per-channel, block-wise, and QAT paths, plus the full quantized utils suite. See test cases in the Test Plan section. I did not run the whole quantized ops/model suite, which is why support is denoted with a (*). - FP16: not supported. QNN rejects these SoCs for FP16 (`"The SocModel doesn't support FP16"`), so FP16 lowering does not produce a delegated graph. Note it also does not degrade gracefully today: models containing ops on the partitioner's non-decompose list (e.g. `linear`, `layer_norm`) fail to export rather than falling back to CPU. Note that these changes need the fix added in #22543 for the tests to pass. ### Test plan ``` python backends/qualcomm/tests/test_qnn_delegate.py \ TestQNNQuantizedOperator.test_qnn_backend_16a4w_conv2d \ TestQNNQuantizedOperator.test_qnn_backend_16a4w_linear \ TestQNNQuantizedOperator.test_qnn_backend_16a4w_layer_norm \ TestQNNQuantizedOperator.test_qnn_backend_16a4w_conv2d \ -v --soc_model SM8635 --host aisw-local-vm3 --device 4603496b \ --build_folder build-android python backends/qualcomm/tests/test_qnn_delegate.py \ TestQNNQuantizedOperator.test_qnn_backend_sort \ TestQNNQuantizedOperator.test_qnn_backend_conv2d \ TestQNNQuantizedOperator.test_qnn_backend_linear \ TestQNNQuantizedOperator.test_qnn_backend_layer_norm \ TestQNNQuantizedOperator.test_qnn_backend_element_wise_add \ -v --soc_model SM8635 --host aisw-local-vm3 --device 4603496b \ --build_folder build-android python backends/qualcomm/tests/test_qnn_delegate.py \ TestQNNQuantizedOperator.test_qnn_backend_16a2w_conv2d \ TestQNNQuantizedOperator.test_qnn_backend_16a2w_linear \ TestQNNQuantizedOperator.test_qnn_backend_16a4w_per_channel_linear \ TestQNNQuantizedOperator.test_qnn_backend_16a4w_per_channel_linear_with_bias \ TestQNNQuantizedOperator.test_qnn_backend_16a4w_conv2d_qat \ TestQNNQuantizedOperator.test_qnn_backend_16a4w_block_conv2d_qat \ -v --soc_model SM8635 --host aisw-local-vm3 --device 4603496b \ --build_folder build-android python backends/qualcomm/tests/test_qnn_delegate.py TestQNNQuantizedUtils -v --soc_model SM8635 --host aisw-local-vm3 --device 4603496b --build_folder build-android ```
qti-horodnic
changed the title
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
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
Fixed a crash on AMD hosts when PyTorch's global backend flags are frozen.
QnnQuantizer.__init__callsdisable_mkldnn_on_amd(), which assignstorch.backends.mkldnn.enabled = False. PyTorch forbids that assignment oncedisable_global_flags()has been called, which happens whenevertorch.testing._internal.common_utilsis imported — as the Qualcomm test suite does. The result is that constructing a quantizer on an AMD host fails with:The assignment is now guarded, so a frozen-flags environment leaves the existing MKLDNN setting in place instead of raising. Disabling MKLDNN is a workaround for AMD-host crashes, not a correctness requirement, so skipping it when the flag is locked is safe.
Note that this is a temporary workaround for an issue introduced in a previous pr, not a root-cause fix, I haven't validated that this change has not broken anything else.
Test plan
Ran the tests mentioned in the test plan section of #22542.