Decompose after export in export_llama #15951
Merged
+35
−17
Conversation
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/15951
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 1 Unrelated FailureAs of commit afc924f with merge base 04f1e4d ( NEW FAILURES - The following jobs have failed:
UNSTABLE - The following job is marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
This PR needs a
|
lucylq
force-pushed
the
lfq.decomp-after-export
branch
2 times, most recently
from
3d09dc1 to
10783cb
Compare
lucylq
requested review from
digantdesai,
jackzhxng,
larryliu0820 and
mergennachin
as code owners
Contributor
|
LGTM! Address any failing CI tests before landing. |
metascroy
approved these changes
Nov 24, 2025
facebook-github-bot
pushed a commit
that referenced
this pull request
Dec 2, 2025
Summary: `unwrap_tensor_subclass` was not unwrapping nested lora linears. This meant qdata/scale/zero were bundled together in the subclass, and separated at run decompositions inside to_edge_transform_and_lower. This is after nodes are tagged, meaning that the scales were not tagged, and remained in the PTE file after the rest of the weights were moved to a PTD file. It's recommended to move away from `unwrap_tensor_subclass` and rely on export + decomps. This PR adds a decomp after exporting in export_llama, and removes cases of `unwrap_tensor_subclass`. TODO: remove all cases of `unwrap_tensor_subclass` in ET. Test Plan: Add check that quantized weights are in PTD file (not PTE file) after quantization. This is a simple check, nested linears seem to be the real issue that decomposing resolves. TODO to add a test for that (probably e2e test with stories in subsequent PR) ``` python -m unittest executorch.backends.xnnpack.test.passes.test_propagate_custom_meta_pass ``` Reviewed By: metascroy Differential Revision: D87826410 Pulled By: lucylq
facebook-github-bot
force-pushed
the
lfq.decomp-after-export
branch
from
10783cb to
22a9e78
Compare
facebook-github-bot
pushed a commit
that referenced
this pull request
Dec 8, 2025
Summary: `unwrap_tensor_subclass` was not unwrapping nested lora linears. This meant qdata/scale/zero were bundled together in the subclass, and separated at run decompositions inside to_edge_transform_and_lower. This is after nodes are tagged, meaning that the scales were not tagged, and remained in the PTE file after the rest of the weights were moved to a PTD file. It's recommended to move away from `unwrap_tensor_subclass` and rely on export + decomps. This PR adds a decomp after exporting in export_llama, and removes cases of `unwrap_tensor_subclass`. TODO: remove all cases of `unwrap_tensor_subclass` in ET. Test Plan: Add check that quantized weights are in PTD file (not PTE file) after quantization. This is a simple check, nested linears seem to be the real issue that decomposing resolves. TODO to add a test for that (probably e2e test with stories in subsequent PR) ``` python -m unittest executorch.backends.xnnpack.test.passes.test_propagate_custom_meta_pass ``` Reviewed By: metascroy Differential Revision: D87826410 Pulled By: lucylq
facebook-github-bot
force-pushed
the
lfq.decomp-after-export
branch
from
22a9e78 to
d9e7c6c
Compare
Summary: `unwrap_tensor_subclass` was not unwrapping nested lora linears. This meant qdata/scale/zero were bundled together in the subclass, and separated at run decompositions inside to_edge_transform_and_lower. This is after nodes are tagged, meaning that the scales were not tagged, and remained in the PTE file after the rest of the weights were moved to a PTD file. It's recommended to move away from `unwrap_tensor_subclass` and rely on export + decomps. This PR adds a decomp after exporting in export_llama, and removes cases of `unwrap_tensor_subclass`. TODO: remove all cases of `unwrap_tensor_subclass` in ET. Test Plan: Add check that quantized weights are in PTD file (not PTE file) after quantization. This is a simple check, nested linears seem to be the real issue that decomposing resolves. TODO to add a test for that (probably e2e test with stories in subsequent PR) ``` python -m unittest executorch.backends.xnnpack.test.passes.test_propagate_custom_meta_pass ``` Reviewed By: metascroy Differential Revision: D87826410 Pulled By: lucylq
facebook-github-bot
force-pushed
the
lfq.decomp-after-export
branch
from
d9e7c6c to
afc924f
Compare
lucylq
added a commit
that referenced
this pull request
Dec 10, 2025
### Summary Use qwen0.6B with unsloth (instead of llama1B with torchtune) for lora test. 1. Smaller model / quicker test. 2. Eventually remove dependency on torchtune. 3. Qwen is not gated on HF. TODO: add quantized test after #15951 ``` Expected result prefix: <|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant To calculate 15% of 80, we can multiply 80 by 0.15. 80 * 0.15 = 12 So, 15% of 80 is 12. #### 12 The answer is: 12<|im_end|> + echo 'Actual result: <|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant To calculate 15% of 80, we can multiply 80 by 0.15. 80 * 0.15 = 12 So, 15% of 80 is 12. #### 12 The answer is: 12<|im_end|> PyTorchObserver {"prompt_tokens":15,"generated_tokens":65,"model_load_start_ms":1765320124550,"model_load_end_ms":1765320127516,"inference_start_ms":1765320152867,"inference_end_ms":1765320178119,"prompt_eval_end_ms":1765320153334,"first_token_ms":1765320153334,"aggregate_sampling_time_ms":19,"SCALING_FACTOR_UNITS_PER_SECOND":1000}' Actual result: <|im_start|>user Calculate 15% of 80?<|im_end|><|im_start|>assistant To calculate 15% of 80, we can multiply 80 by 0.15. 80 * 0.15 = 12 So, 15% of 80 is 12. #### 12 The answer is: 12<|im_end|> PyTorchObserver {"prompt_tokens":15,"generated_tokens":65,"model_load_start_ms":1765320124550,"model_load_end_ms":1765320127516,"inference_start_ms":1765320152867,"inference_end_ms":1765320178119,"prompt_eval_end_ms":1765320153334,"first_token_ms":1765320153334,"aggregate_sampling_time_ms":19,"SCALING_FACTOR_UNITS_PER_SECOND":1000} + echo Success Success ```
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
unwrap_tensor_subclasswas not unwrapping nested lora linears. This meant qdata/scale/zero were bundled together in the subclass, and separated at run decompositions inside to_edge_transform_and_lower. This is after nodes are tagged, meaning that the scales were not tagged, and remained in the PTE file after the rest of the weights were moved to a PTD file.It's recommended to move away from
unwrap_tensor_subclassand rely on export + decomps. This PR adds a decomp after exporting in export_llama, and removes cases ofunwrap_tensor_subclass.TODO: remove all cases of
unwrap_tensor_subclassin ET.Test plan
Add check that quantized weights are in PTD file (not PTE file) after quantization. This is a simple check, nested linears seem to be the real issue that decomposing resolves. TODO to add a test for that (probably e2e test with stories in subsequent PR)