[Torch] Lora kernels compilation#4100
Open
daniil-lyakhov wants to merge 2 commits into
Open
Conversation
daniil-lyakhov
force-pushed
the
dl/lora_compile
branch
from
15914d2 to
709a45c
Compare
Contributor
There was a problem hiding this comment.
Pull request overview
This PR improves performance of Torch LoRA quantization kernels by enabling torch.compile on the LoRA quantize functions, making ReferenceQuantize more compilation-friendly, and preventing unsupported nested compilation in the shared CompilationWrapper.
Changes:
- Add a nested-compilation guard to
CompilationWrapperto avoid callingtorch.compilewhile already in a compile context. - Refactor
ReferenceQuantizesummation helper into class methods to reducetorch.compilegraph breaks. - Split LoRA quantize functions into internal implementations and wrap them with
CompilationWrapper.
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.
| File | Description |
|---|---|
src/nncf/torch/utils.py |
Prevents nested torch.compile usage inside CompilationWrapper. |
src/nncf/torch/quantization/reference.py |
Refactors summation logic to be more torch.compile friendly. |
src/nncf/torch/quantization/quantize_functions.py |
Wraps LoRA quantize kernels with CompilationWrapper for speedup. |
Comment on lines
+54
to
+55
| def _sum_like(self, tensor_to_sum: GeneralizedTensor, ref_tensor: GeneralizedTensor): | ||
| """Warning: may modify tensor_to_sum""" |
Comment on lines
+73
to
+74
| def _sum_like_fp32(self, tensor_to_sum: GeneralizedTensor, ref_tensor: GeneralizedTensor): | ||
| """Warning: may modify tensor_to_sum""" |
Comment on lines
+507
to
+508
| _asymmetric_quantize_lora = CompilationWrapper(_asymmetric_quantize_lora) | ||
| _symmetric_quantize_lora = CompilationWrapper(_symmetric_quantize_lora) |
Comment on lines
+306
to
+307
| return _asymmetric_quantize_lora( | ||
| input_, |
Comment on lines
+364
to
+365
| return _symmetric_quantize_lora(input_, input_shape, A, B, scale, level_low, level_high, levels, eps) | ||
|
|
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.
Changes
asymmetric_quantize_loraandsymmetric_quantize_loraare wrapped by theCompilationWrapperReferenceQuantizeis updated to betorch.compilefriendly: previous code was making graph brakes in the compiled graphCompilationWrapperis updated to skip compilation for nested compiled functions as it is not supported by the PyTorchReason for changes
Example https://github.com/openvinotoolkit/nncf/tree/develop/examples/llm_compression/torch/distillation_qat_with_lora
Before:
After:
Aprox ~25% speed up
With unlimited cache
(torch._dynamo.config.cache_size_limit = 100
torch._dynamo.config.accumulated_cache_size_limit = 100):
HW:
Intel(R) Core(TM) i9-10980XE CPU @ 3.00GHz
3x RTX 3090
Related tickets
Tests
Test examples: https://github.com/openvinotoolkit/nncf/actions/runs/27631571445
Test WC https://github.com/openvinotoolkit/nncf/actions/runs/27631606234
Weight compression - success
Test examples - success