[minor] add a general FP8ScaleSweepCalibrator and its registry#1171
Open
[minor] add a general FP8ScaleSweepCalibrator and its registry#1171
Conversation
Signed-off-by: Fridah-nv <201670829+Fridah-nv@users.noreply.github.com>
Contributor
📝 WalkthroughWalkthroughThe changes introduce a new Changes
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes 🚥 Pre-merge checks | ✅ 4✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
Contributor
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
🧹 Nitpick comments (3)
modelopt/torch/quantization/model_calib.py (2)
63-80: Consider adding overwrite warning and type validation for consistency.Two observations:
- Unlike
register_quant_backend(which warns when overwriting an existing backend), this silently overwrites. For consistency and to help users debug accidental overwrites, consider adding a warning.- The type hint is
typerather thantype[FP8ScaleSweepCalibrator]. Adding a runtimeissubclasscheck would catch incorrect registrations early.♻️ Suggested improvement
+import warnings +from .calib import FP8ScaleSweepCalibrator + def register_fp8_sweep_calibrator(backend: str, calibrator_cls: type) -> None: """Register a :class:`FP8ScaleSweepCalibrator` subclass for a quantization backend. ... """ + if not issubclass(calibrator_cls, FP8ScaleSweepCalibrator): + raise TypeError( + f"calibrator_cls must be a subclass of FP8ScaleSweepCalibrator, got {calibrator_cls}" + ) + if backend in _FP8_SWEEP_CALIBRATOR_REGISTRY: + warnings.warn(f"Overwriting existing FP8 sweep calibrator for backend: {backend}") _FP8_SWEEP_CALIBRATOR_REGISTRY[backend] = calibrator_clsThis aligns with the existing
register_quant_backendpattern shown intensor_quantizer.py:82-99.🤖 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 63 - 80, The register_fp8_sweep_calibrator function currently overwrites entries silently and accepts a generic type; change it to mirror register_quant_backend by (1) validating that calibrator_cls is a subclass of FP8ScaleSweepCalibrator via issubclass(calibrator_cls, FP8ScaleSweepCalibrator) and raising a TypeError if not, and (2) emitting a warning (e.g., using warnings.warn) when _FP8_SWEEP_CALIBRATOR_REGISTRY already contains the backend key before overwriting; update the function signature/type hint to accept type[FP8ScaleSweepCalibrator] to reflect the enforced type.
651-682: Code is correct; consider extracting shared calibrator selection logic.This block mirrors the calibrator selection in
mse_calibrate()(lines 351-378). Both paths:
- Check
is_nvfp4_staticvs registry-backed backends- Instantiate the appropriate calibrator class
A small helper (e.g.,
_create_fp8_sweep_calibrator(...)) could reduce duplication, but this is optional given the clarity of the current implementation.🤖 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 651 - 682, The calibrator selection/instantiation logic is duplicated between this block and mse_calibrate(); extract a helper (e.g., _create_fp8_sweep_calibrator) that accepts (weight_quantizer, initial_amax, quant_func, error_func, is_nvfp4_static) and returns the configured calibrator instance; inside the helper, if is_nvfp4_static return NVFP4MSECalibrator configured with amax=initial_amax, axis=weight_quantizer._calibrator._axis if present, global_amax=weight_quantizer.global_amax and quant_func/error_func; otherwise lookup calibrator_cls = _FP8_SWEEP_CALIBRATOR_REGISTRY[weight_quantizer.backend] and instantiate with amax, axis, quant_func and error_func; replace the duplicated instantiation code in both this block and mse_calibrate() to call the new helper and assign to weight_quantizer._calibrator.modelopt/torch/quantization/calib/mse.py (1)
174-185: Add a named constant for the FP8 E4M3 maximum value and document inherited unused parameters.The magic number
448.0represents the maximum representable value in FP8 E4M3 format and should be defined as a module-level constant for clarity and maintainability, following the pattern used elsewhere in the codebase (e.g.,mxfp8_tensor.py).Additionally,
FP8ScaleSweepCalibratorinherits__init__fromMseCalibrator, which acceptsstep_size,start_multiplier, andstop_multiplierparameters. Since_generate_candidatesis overridden in this class, these parameters are silently ignored. Consider either overriding__init__to explicitly exclude these parameters or adding a note to the class docstring that these inherited parameters are not used.♻️ Suggested improvement
+_FP8_E4M3_MAX = 448.0 # Maximum representable value in FP8 E4M3 format + class FP8ScaleSweepCalibrator(MseCalibrator): """MSE calibrator that sweeps 126 valid FP8 E4M3 candidates of ``initial_amax``. Candidate amax values are ``initial_amax * candidate`` + + Note: + The ``step_size``, ``start_multiplier``, and ``stop_multiplier`` parameters + inherited from :class:`MseCalibrator` are not used by this calibrator. """ def _generate_candidates(self, device: torch.device) -> torch.Tensor: - """Generate 126 valid FP8 E4M3 scale candidates.""" + """Generate 126 valid FP8 E4M3 scale candidates (all nonzero finite values / max).""" uint8_values = torch.arange(0, 128, dtype=torch.uint8, device=device) fp8_values = uint8_values.view(torch.float8_e4m3fn).float() valid_mask = torch.isfinite(fp8_values) & (fp8_values > 0) - return fp8_values[valid_mask] / 448.0 + return fp8_values[valid_mask] / _FP8_E4M3_MAX🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@modelopt/torch/quantization/calib/mse.py` around lines 174 - 185, Replace the magic literal 448.0 with a module-level constant named FP8_E4M3_MAX (or similar) and use that constant in FP8ScaleSweepCalibrator._generate_candidates so the maximum representable FP8 E4M3 value is documented and reusable (matching the pattern in mxfp8_tensor.py); also update the FP8ScaleSweepCalibrator class docstring to note that the inherited MseCalibrator __init__ parameters (step_size, start_multiplier, stop_multiplier) are intentionally unused by this subclass, or alternatively override __init__ in FP8ScaleSweepCalibrator to explicitly accept and ignore those params so their omission is explicit.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Nitpick comments:
In `@modelopt/torch/quantization/calib/mse.py`:
- Around line 174-185: Replace the magic literal 448.0 with a module-level
constant named FP8_E4M3_MAX (or similar) and use that constant in
FP8ScaleSweepCalibrator._generate_candidates so the maximum representable FP8
E4M3 value is documented and reusable (matching the pattern in mxfp8_tensor.py);
also update the FP8ScaleSweepCalibrator class docstring to note that the
inherited MseCalibrator __init__ parameters (step_size, start_multiplier,
stop_multiplier) are intentionally unused by this subclass, or alternatively
override __init__ in FP8ScaleSweepCalibrator to explicitly accept and ignore
those params so their omission is explicit.
In `@modelopt/torch/quantization/model_calib.py`:
- Around line 63-80: The register_fp8_sweep_calibrator function currently
overwrites entries silently and accepts a generic type; change it to mirror
register_quant_backend by (1) validating that calibrator_cls is a subclass of
FP8ScaleSweepCalibrator via issubclass(calibrator_cls, FP8ScaleSweepCalibrator)
and raising a TypeError if not, and (2) emitting a warning (e.g., using
warnings.warn) when _FP8_SWEEP_CALIBRATOR_REGISTRY already contains the backend
key before overwriting; update the function signature/type hint to accept
type[FP8ScaleSweepCalibrator] to reflect the enforced type.
- Around line 651-682: The calibrator selection/instantiation logic is
duplicated between this block and mse_calibrate(); extract a helper (e.g.,
_create_fp8_sweep_calibrator) that accepts (weight_quantizer, initial_amax,
quant_func, error_func, is_nvfp4_static) and returns the configured calibrator
instance; inside the helper, if is_nvfp4_static return NVFP4MSECalibrator
configured with amax=initial_amax, axis=weight_quantizer._calibrator._axis if
present, global_amax=weight_quantizer.global_amax and quant_func/error_func;
otherwise lookup calibrator_cls =
_FP8_SWEEP_CALIBRATOR_REGISTRY[weight_quantizer.backend] and instantiate with
amax, axis, quant_func and error_func; replace the duplicated instantiation code
in both this block and mse_calibrate() to call the new helper and assign to
weight_quantizer._calibrator.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
Run ID: 6a31f4b3-e4ad-4edc-b8f6-48c7553b2e41
📒 Files selected for processing (2)
modelopt/torch/quantization/calib/mse.pymodelopt/torch/quantization/model_calib.py
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #1171 +/- ##
===========================================
- Coverage 70.20% 54.51% -15.69%
===========================================
Files 230 348 +118
Lines 26098 39789 +13691
===========================================
+ Hits 18322 21692 +3370
- Misses 7776 18097 +10321
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ 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: ?
Usage
# Add a code snippet demonstrating how to use thisTesting
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,torch.load(..., weights_only=False),pickle, etc.).CONTRIBUTING.md: ✅ / ❌ / N/AAdditional Information
Summary by CodeRabbit
New Features
Refactor