Copy weights and preserve device for 8da4w QAT linear #211
Conversation
This file contains 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
facebook-github-bot
added
the
CLA Signed
jerryzh168
reviewed
May 2, 2024
| @@ -47,6 +47,7 @@ def prepare( | |||
| *args: Any, | |||
| **kwargs: Any | |||
| ) -> torch.nn.Module: | |||
| state_dict = model.state_dict() | |||
There was a problem hiding this comment.
oh OK, right, this function have to do create_quantized_state_dict() and load_state_dict() in the gpt-fast API, but I feel we could also change _replace_linear_8da4w to instantiate from the existing floating point module, that might be clearer
There was a problem hiding this comment.
Yeah I agree, updated
jerryzh168
reviewed
May 2, 2024
| ) | ||
| break | ||
| if should_load_state_dict: | ||
| model.load_state_dict(state_dict) |
There was a problem hiding this comment.
will this work if we use tensor subclasses?
jerryzh168
reviewed
May 2, 2024
| # on the meta device, in which case there is no need to | ||
| # load the state dict, and doing so will lead to an error | ||
| should_load_state_dict = True | ||
| for k, v in state_dict.items(): |
There was a problem hiding this comment.
this can probably be simplified a bit:
should_load_state_dict = all(not v.is_meta for v in state_dict.values())
jerryzh168
approved these changes
May 2, 2024
Summary: This fixes two correctness bugs. First, we never copied over the weights from the existing linear, so we would start from random weights even when loading from checkpoints. Second, we never preserved the device of the original linear. This is important for settings like FSDP, where we expect non-zero ranks to have their parameters on the meta device in order to initialize these parameters correctly. Test Plan: python test/quantization/test_qat.py -k test_qat_8da4w_quantizer python test/quantization/test_qat.py -k test_qat_8da4w_quantizer_meta_weights Reviewers: jerryzh168, cpuhrsch Subscribers: jerryzh168, cpuhrsch, supriyar
andrewor14
force-pushed
the
8da4w_qat_state_dict
branch
from
bac2301 to
cd50d7a
Compare
|
@andrewor14 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
jerryzh168
approved these changes
May 2, 2024
|
@andrewor14 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
dbyoung18
pushed a commit
to dbyoung18/ao
that referenced
this pull request
Jul 31, 2024
* Copy weights and preserve device for 8da4w QAT linear Summary: This fixes two correctness bugs. First, we never copied over the weights from the existing linear, so we would start from random weights even when loading from checkpoints. Second, we never preserved the device of the original linear. This is important for settings like FSDP, where we expect non-zero ranks to have their parameters on the meta device in order to initialize these parameters correctly. Test Plan: python test/quantization/test_qat.py -k test_qat_8da4w_quantizer python test/quantization/test_qat.py -k test_qat_8da4w_quantizer_meta_weights Reviewers: jerryzh168, cpuhrsch Subscribers: jerryzh168, cpuhrsch, supriyar * Update test_qat.py
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.
Summary: This fixes two correctness bugs. First, we never copied over the weights from the existing linear, so we would start from random weights even when loading from checkpoints. Second, we never preserved the device of the original linear. This is important for settings like FSDP, where we expect non-zero ranks to have their parameters on the meta device in order to initialize these parameters correctly.
Test Plan:
python test/quantization/test_qat.py -k test_qat_8da4w_quantizer python test/quantization/test_qat.py -k test_qat_8da4w_quantizer_load_state_dict_meta
Reviewers: jerryzh168, cpuhrsch
Subscribers: jerryzh168, cpuhrsch, supriyar