Bump version for float8 dynamic quant and weight only quant configs #2650
+196
−102
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/2650
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
e19cb46 to
5ae457c
Compare
jerryzh168
added
the
topic: bc-breaking
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
5ae457c to
79e89ff
Compare
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
79e89ff to
d375bbb
Compare
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
d375bbb to
c464d5b
Compare
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
c464d5b to
456a77f
Compare
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
456a77f to
f6e4522
Compare
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
f6e4522 to
7cdfe0a
Compare
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
b06dafd to
5016603
Compare
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
5016603 to
b2c8536
Compare
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
b2c8536 to
912f6e5
Compare
|
Ci failures look real |
jerryzh168
added a commit
that referenced
this pull request
Aug 6, 2025
Summary: This PR changes the default VERSION for Float8DynamicActivationFloat8WeightConfig and Float8WeightOnlyConfig from 1 to 2 and makes the VERSION 1 config and VERSION 1 quantized models deprecated, more details in: #2649 Also extended current config serialization to work with multiple config versions Deprecation Note: ``` from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "torchao-testing/opt-125m-float8dq-row-v1-0.13-dev" quantized_model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype="bfloat16", device_map="cuda", ) /data/users/jerryzh/ao/torchao/core/config.py:249: UserWarning: Stored version is not the same as current default version of the config: stored_version=1, current_version=2, please check the deprecation warning warnings.warn( /data/users/jerryzh/ao/torchao/dtypes/floatx/float8_layout.py:113: UserWarning: Models quantized with VERSION 1 of Float8DynamicActivationFloat8WeightConfig is deprecated and will no longer be supported in a future release, please upgrade torchao and quantize again, or download a newer torchao checkpoint, see #2649 for more details warnings.warn( ``` Suggestion: upgrade torchao to 0.13 and later and generate the checkpoint again: ``` quantize_(model, Float8DynamicActivationFloat8WeightConfig(granularity=PerRow())) ``` Or download the checkpoint again (please let us know if the checkpoint is not updated) Test Plan: tested with serializing a model with VERSION 1 config and load it, and checks warnings are properly printed ``` python test/integration/test_loading_deprecated_checkpoint.py ``` Reviewers: Subscribers: Tasks: Tags: stack-info: PR: #2650, branch: jerryzh168/stack/14
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
912f6e5 to
16b2c4b
Compare
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
3 times, most recently
from
d3ee6c2 to
356e477
Compare
vkuzo
reviewed
Aug 6, 2025
| @@ -1506,6 +1506,9 @@ class Float8WeightOnlyConfig(AOBaseConfig): | |||
|
|
|||
| def _float8_weight_only_quant_tensor(weight, config): | |||
| if config.VERSION == 1: | |||
| warnings.warn( | |||
There was a problem hiding this comment.
I missed it in previous PRs, but why VERSION instead of version? If there is no good reason, can we change it before it's too late?
There was a problem hiding this comment.
There was a problem hiding this comment.
OK updated it to instance variable and renamed to version
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
2 times, most recently
from
b844dbd to
5f1eb8a
Compare
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
2 times, most recently
from
a76e263 to
b855b91
Compare
drisspg
reviewed
Aug 7, 2025
torchao/core/config.py
Outdated
| data_dict[f.name] = self.encode_value(getattr(o, f.name)) | ||
|
|
||
| return { | ||
| # Only store the class name for dataclasses too | ||
| "_type": o.__class__.__name__, | ||
| "_version": getattr(o.__class__, "VERSION", 1), | ||
| "_version": getattr(o, "version", 1), |
There was a problem hiding this comment.
should the 3rd arg here be the _DEFAULT_VERSION?
drisspg
reviewed
Aug 7, 2025
| @@ -253,10 +244,11 @@ def config_from_dict(data: Dict[str, Any]) -> AOBaseConfig: | |||
| f"Failed to find class {type_path} in any of the allowed modules: {allowed_modules_str}" | |||
| ) | |||
|
|
|||
| # Check version - require exact match | |||
There was a problem hiding this comment.
I feel like in general versions will error right? like shouldnt we have this path but just let instances figure out if they should error or not at deserialization
There was a problem hiding this comment.
they can still error, see quant_api.py, they can error when config.version has a deprecated version number
but erroring out here would mean older version are not supported at all
Summary: This PR changes the default VERSION for Float8DynamicActivationFloat8WeightConfig and Float8WeightOnlyConfig from 1 to 2 and makes the VERSION 1 config and VERSION 1 quantized models deprecated, more details in: #2649 Also extended current config serialization to work with multiple config versions Deprecation Note: ``` from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "torchao-testing/opt-125m-float8dq-row-v1-0.13-dev" quantized_model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype="bfloat16", device_map="cuda", ) /data/users/jerryzh/ao/torchao/core/config.py:249: UserWarning: Stored version is not the same as current default version of the config: stored_version=1, current_version=2, please check the deprecation warning warnings.warn( /data/users/jerryzh/ao/torchao/dtypes/floatx/float8_layout.py:113: UserWarning: Models quantized with VERSION 1 of Float8DynamicActivationFloat8WeightConfig is deprecated and will no longer be supported in a future release, please upgrade torchao and quantize again, or download a newer torchao checkpoint, see #2649 for more details warnings.warn( ``` Suggestion: upgrade torchao to 0.13 and later and generate the checkpoint again: ``` quantize_(model, Float8DynamicActivationFloat8WeightConfig(granularity=PerRow())) ``` Or download the checkpoint again (please let us know if the checkpoint is not updated) Test Plan: tested with serializing a model with VERSION 1 config and load it, and checks warnings are properly printed ``` python test/integration/test_loading_deprecated_checkpoint.py ``` Reviewers: Subscribers: Tasks: Tags: stack-info: PR: #2650, branch: jerryzh168/stack/14
drisspg
approved these changes
Aug 7, 2025
jerryzh168
force-pushed
the
jerryzh168/stack/14
branch
from
b855b91 to
e66af4f
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:
optional_tensor_namesin TorchAOBaseTensor #2710Bump version for float8 dynamic quant and weight only quant configs
Summary:
This PR changes the default VERSION for Float8DynamicActivationFloat8WeightConfig and Float8WeightOnlyConfig from 1 to 2
and makes the VERSION 1 config and VERSION 1 quantized models deprecated, more details in: #2649
Also extended current config serialization to work with multiple config versions
Deprecation Note:
Suggestion: upgrade torchao to 0.13 and later and generate the checkpoint again:
Or download the checkpoint again (please let us know if the checkpoint is not updated)
Test Plan:
tested with serializing a model with VERSION 1 config and load it, and checks warnings are properly printed
Reviewers:
Subscribers:
Tasks:
Tags: