Add round_scales_to_power_of_2 option for float quantization #2323
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/ao/2323
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 4d77255 with merge base 282d04f ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
drisspg
added a commit
that referenced
this pull request
Jun 6, 2025
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <[email protected]> stack-info: PR: #2323, branch: drisspg/stack/67
facebook-github-bot
added
the
CLA Signed
drisspg
added a commit
that referenced
this pull request
Jun 6, 2025
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <[email protected]> stack-info: PR: #2323, branch: drisspg/stack/67
drisspg
added a commit
that referenced
this pull request
Jun 6, 2025
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <[email protected]> stack-info: PR: #2323, branch: drisspg/stack/67
drisspg
force-pushed
the
drisspg/stack/67
branch
2 times, most recently
from
80c391c to
2902ef7
Compare
drisspg
added a commit
that referenced
this pull request
Jun 6, 2025
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <[email protected]> stack-info: PR: #2323, branch: drisspg/stack/67
drisspg
added a commit
that referenced
this pull request
Jun 7, 2025
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. stack-info: PR: #2323, branch: drisspg/stack/67
drisspg
added a commit
that referenced
this pull request
Jun 7, 2025
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. stack-info: PR: #2323, branch: drisspg/stack/67
drisspg
added a commit
that referenced
this pull request
Jun 7, 2025
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. stack-info: PR: #2323, branch: drisspg/stack/67
drisspg
added a commit
that referenced
this pull request
Jun 7, 2025
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. stack-info: PR: #2323, branch: drisspg/stack/67
danielvegamyhre
approved these changes
Jun 9, 2025
|
|
||
| config = config_factory() | ||
| if isinstance( | ||
| config, Float8DynamicActivationFloat8SemiSparseWeightConfig |
There was a problem hiding this comment.
Shouldn't the min_sm check above already handle this? Or is this because the cutlass kernel is only built on sm90 / 90a?
Test case for Float8SemiSparse is failing on h100 due to CUDA backend not supporting operator (kernel not built?) so just wondering
drisspg
added a commit
that referenced
this pull request
Jun 11, 2025
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. stack-info: PR: #2323, branch: drisspg/stack/67
drisspg
force-pushed
the
drisspg/stack/67
branch
from
195187b to
9ed84e7
Compare
drisspg
added a commit
that referenced
this pull request
Jun 16, 2025
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. stack-info: PR: #2323, branch: drisspg/stack/67
drisspg
force-pushed
the
drisspg/stack/67
branch
from
9ed84e7 to
4bb19d6
Compare
Open
drisspg
force-pushed
the
drisspg/stack/67
branch
3 times, most recently
from
511db54 to
df31f6d
Compare
drisspg
force-pushed
the
drisspg/stack/67
branch
from
df31f6d to
5f784eb
Compare
This adds support for rounding scaling factors down to the nearest power of 2 for float quantization, following the pattern established in Float8LinearConfig. Key changes: - Add round_scales_to_power_of_2 parameter to all float quantization configs - Update choose_qparams_affine_floatx and to_scaled_tc_floatx functions to apply power of 2 rounding - Thread the parameter through all relevant function calls in quant_api.py - Maintain backward compatibility with default value of False This helps reduce quantization error by avoiding rounding errors when multiplying/dividing by scaling factors and ensures consistent quantization between forward and backward passes. stack-info: PR: #2323, branch: drisspg/stack/67
drisspg
force-pushed
the
drisspg/stack/67
branch
from
5f784eb to
4d77255
Compare
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.
Stacked PRs:
Add round_scales_to_power_of_2 option for float quantization
This adds support for rounding scaling factors down to the nearest power of 2
for float quantization, following the pattern established in Float8LinearConfig.
Key changes:
Lets users who train with this setting run in inference