[Minor] Improve local hessian and mse calibration#976
Open
[Minor] Improve local hessian and mse calibration#976
Conversation
Signed-off-by: Fridah-nv <201670829+Fridah-nv@users.noreply.github.com>
Signed-off-by: Fridah-nv <201670829+Fridah-nv@users.noreply.github.com>
Contributor
📝 WalkthroughWalkthroughThe changes update quantization calibration behavior to apply MSE calibration exclusively to weight quantizers while activation quantizers retain max-calibration amax values. A new utility function replaces MSE calibrators with max calibrators after calibration steps, and docstrings clarify this weight-only calibration scope across MSE and Hessian calibration configurations. Changes
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes Important Pre-merge checks failedPlease resolve all errors before merging. Addressing warnings is optional. ❌ Failed checks (1 error)
✅ Passed checks (3 passed)
✨ Finishing Touches
🧪 Generate unit tests (beta)
Tip Try Coding Plans. Let us write the prompt for your AI agent so you can ship faster (with fewer bugs). Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@modelopt/torch/quantization/model_calib.py`:
- Around line 580-615: The captured_quantized[0] may still be None if
capture_forward is not invoked during _forward_no_local_hessian; add a defensive
guard before calling self.hessian_helper.accumulate_hessian: check
captured_quantized[0] is not None and only then call
accumulate_hessian(captured_quantized[0]); if it is None fall back to the raw
input path (use input.to_local() if available) and then restore
self.input_quantizer.forward to original_forward if set; update the block that
follows out = self._forward_no_local_hessian(...) to perform this None-check and
fallback while preserving restoration of the original_forward.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
Run ID: c0dbb264-f19a-4f5d-a597-2d4860ebdde3
📒 Files selected for processing (2)
modelopt/torch/quantization/config.pymodelopt/torch/quantization/model_calib.py
Comment on lines
+580
to
+615
| # Forward without weight quantization during caching | ||
| if LocalHessianHelper.cache_mode: | ||
| self.weight_quantizer.disable() | ||
|
|
||
| # Capture quantized input from the forward pass for Hessian collection | ||
| captured_quantized = [None] | ||
| original_forward = None | ||
| if ( | ||
| self.hessian_helper.is_enabled | ||
| and hasattr(self, "input_quantizer") | ||
| and self.input_quantizer.is_enabled | ||
| ): | ||
| original_forward = self.input_quantizer.forward | ||
|
|
||
| def capture_forward(input_tensor): | ||
| quantized = original_forward(input_tensor) | ||
| captured_quantized[0] = ( | ||
| quantized.to_local() if hasattr(quantized, "to_local") else quantized | ||
| ) | ||
| return quantized | ||
|
|
||
| self.input_quantizer.forward = capture_forward | ||
|
|
||
| out = self._forward_no_local_hessian(input, *args, **kwargs) | ||
|
|
||
| # Collect Hessian from the quantized input that was used in forward pass | ||
| if self.hessian_helper.is_enabled: | ||
| if hasattr(self, "input_quantizer") and self.input_quantizer.is_enabled: | ||
| self.hessian_helper.accumulate_hessian(captured_quantized[0]) | ||
| if original_forward is not None: | ||
| self.input_quantizer.forward = original_forward | ||
| else: | ||
| # No input_quantizer, use raw input | ||
| input_local = input.to_local() if hasattr(input, "to_local") else input | ||
| self.hessian_helper.accumulate_hessian(input_local) | ||
|
|
Contributor
There was a problem hiding this comment.
Consider adding a guard for captured_quantized[0] being None.
The monkey-patching approach to capture quantized inputs is sound and correctly restores the original forward method. However, if capture_forward is never invoked during _forward_no_local_hessian (an unlikely edge case), captured_quantized[0] would remain None, causing accumulate_hessian to fail on input_tensor.reshape().
Consider adding a defensive check:
🛡️ Proposed defensive guard
# Collect Hessian from the quantized input that was used in forward pass
if self.hessian_helper.is_enabled:
if hasattr(self, "input_quantizer") and self.input_quantizer.is_enabled:
- self.hessian_helper.accumulate_hessian(captured_quantized[0])
+ if captured_quantized[0] is not None:
+ self.hessian_helper.accumulate_hessian(captured_quantized[0])
if original_forward is not None:
self.input_quantizer.forward = original_forward
else:🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@modelopt/torch/quantization/model_calib.py` around lines 580 - 615, The
captured_quantized[0] may still be None if capture_forward is not invoked during
_forward_no_local_hessian; add a defensive guard before calling
self.hessian_helper.accumulate_hessian: check captured_quantized[0] is not None
and only then call accumulate_hessian(captured_quantized[0]); if it is None fall
back to the raw input path (use input.to_local() if available) and then restore
self.input_quantizer.forward to original_forward if set; update the block that
follows out = self._forward_no_local_hessian(...) to perform this None-check and
fallback while preserving restoration of the original_forward.
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #976 +/- ##
==========================================
- Coverage 72.12% 72.09% -0.04%
==========================================
Files 209 209
Lines 23628 23652 +24
==========================================
+ Hits 17042 17052 +10
- Misses 6586 6600 +14 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
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
1 participant
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.
What does this PR do?
Type of change: ? Bug fix
Testing
Before your PR is "Ready for review"
Make sure you read and follow Contributor guidelines and your commits are signed (
git commit -s -S).Make sure you read and follow the Security Best Practices (e.g. avoiding hardcoded
trust_remote_code=True, usingtorch.load(..., weights_only=True), avoidingpickle, etc.).Additional Information
Summary by CodeRabbit
Refactor
Documentation