[quant][graphmode] Fp16 quant support - match numerics with eager mode #41049
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
Summary: In eager mode there is no cast operator for the activation tensor in the fbgemm fp16 operator Remove it from graph mode. This makes numerics match between debug model and final quantized model. Test Plan: python test/test_quantization.py test_linear_dynamic_fp16 Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
facebook-github-bot
added
the
oncall: jit
supriyar
added a commit
that referenced
this pull request
Jul 7, 2020
Summary: In eager mode there is no cast operator for the activation tensor in the fbgemm fp16 operator Remove it from graph mode. This makes numerics match between debug model and final quantized model. Test Plan: python test/test_quantization.py test_linear_dynamic_fp16 Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 1a7c6c6 Pull Request resolved: #41049
…cast op" Summary: In eager mode there is no cast operator for the activation tensor in the fbgemm fp16 operator Remove it from graph mode. This makes numerics match between debug model and final quantized model. Test Plan: python test/test_quantization.py test_linear_dynamic_fp16 Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
supriyar
added a commit
that referenced
this pull request
Jul 7, 2020
Summary: In eager mode there is no cast operator for the activation tensor in the fbgemm fp16 operator Remove it from graph mode. This makes numerics match between debug model and final quantized model. Test Plan: python test/test_quantization.py test_linear_dynamic_fp16 Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: c973a37 Pull Request resolved: #41049
💊 CI failures summary and remediationsAs of commit 31c0a60 (more details on the Dr. CI page): ✅ None of the CI failures appear to be your fault 💚
❄️ 7 failures tentatively classified as flakybut reruns have not yet been triggered to confirm:
|
…cast op" Summary: In eager mode there is no cast operator for the activation tensor in the fbgemm fp16 operator Remove it from graph mode. This makes numerics match between debug model and final quantized model. Test Plan: python test/test_quantization.py test_linear_dynamic_fp16 Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
supriyar
changed the title
…h eager mode" Summary: In eager mode there is no cast operator for the activation tensor in the fbgemm fp16 operator Remove it from graph mode. For the weight tensor we handle saturation by clipping the values to fp16 range This makes numerics match between debug model and final quantized model. Test Plan: python test/test_quantization.py test_linear_dynamic_fp16 Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
…h eager mode" Summary: In eager mode there is no cast operator for the activation tensor in the fbgemm fp16 operator Remove it from graph mode. For the weight tensor we handle saturation by clipping the values to fp16 range This makes numerics match between debug model and final quantized model. Test Plan: python test/test_quantization.py test_linear_dynamic_fp16 Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
supriyar
added a commit
that referenced
this pull request
Jul 7, 2020
Summary: In eager mode there is no cast operator for the activation tensor in the fbgemm fp16 operator Remove it from graph mode. For the weight tensor we handle saturation by clipping the values to fp16 range This makes numerics match between debug model and final quantized model. Test Plan: python test/test_quantization.py test_linear_dynamic_fp16 Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 510d046 Pull Request resolved: #41049
jerryzh168
reviewed
Jul 7, 2020
Comment on lines
+237
to
+238
| // We don't need to insert cast operators for activation tensors for fp16 | ||
| // quant. |
Contributor
There was a problem hiding this comment.
can we filter this in a different place? e.g. we don't insert observer for activation tensors?
Contributor
Author
There was a problem hiding this comment.
Is it possible for user to not specify activation observer in the qconfig? Does the prepare_jit pass ensure observers aren't inserted in that case for activation tensors?
Contributor
There was a problem hiding this comment.
I mean doing checks like this: https://codebrowser.bddppq.com/pytorch/pytorch/torch/csrc/jit/passes/quantization/insert_observers.cpp.html#1187
Contributor
Author
There was a problem hiding this comment.
That may have issues since for FP16 quant we don't specify dtype anywhere in the qconfig. We set quant type to dynamic so there is no way to distinguish int8 dynamic quant from fp16 dynamic quant. Hence was wondering if not specifying any activation observer (since we don't want it observed) would work here.
Contributor
There was a problem hiding this comment.
Looks like right now we are checking noop observer to do fp16 quantization, this sounds like a hack, can we expose fp16 as an argument to the API?
Contributor
There was a problem hiding this comment.
i feel it makes more sense to expose this in the top level API, why don't we do that?
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.
Stack from ghstack:
Summary:
In eager mode there is no cast operator for the activation tensor in the fbgemm fp16 operator
Remove it from graph mode.
For the weight tensor we handle saturation by clipping the values to fp16 range
This makes numerics match between debug model and final quantized model.
Test Plan:
python test/test_quantization.py test_linear_dynamic_fp16
Reviewers:
Subscribers:
Tasks:
Tags: