[Quant] Remove all the dequant nodes when the ref module has multi input args #90157
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
…input [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/90157
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 0ba4c89: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
leslie-fang-intel
changed the title
leslie-fang-intel
changed the title
leslie-fang-intel
added
intel
Xia-Weiwen
approved these changes
Dec 5, 2022
…as multi input args" **Summary**: When converting a ref module into a quant module, `_lower_static_weighted_ref_module` pass assumes the `ref_node` only has 1 input node, and only remove the first `dequant` node. However, when we enable the `conv add` fusion, there will be a extra input node from `add` node besides the original input node from `conv`. Similar as did in the `_lower_quantized_binary_op` pass https://github.com/pytorch/pytorch/blob/41c3b41b92f5019f8d5e2f2846a06b87db01ca4e/torch/ao/quantization/fx/_lower_to_native_backend.py#L766-L775, We should remove all the `dequant` nodes in the `_lower_static_weighted_ref_module` pass. **Test Plan**: It's a bug fix instead of a new feature. When we enable the `conv add` fusion PR later, we will add test cases accordingly. cc jerryzh168 jianyuh raghuramank100 jamesr66a vkuzo jgong5 Xia-Weiwen mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
jgong5
approved these changes
Dec 9, 2022
…as multi input args" **Summary**: When converting a ref module into a quant module, `_lower_static_weighted_ref_module` pass assumes the `ref_node` only has 1 input node, and only remove the first `dequant` node. However, when we enable the `conv add` fusion, there will be a extra input node from `add` node besides the original input node from `conv`. Similar as did in the `_lower_quantized_binary_op` pass https://github.com/pytorch/pytorch/blob/41c3b41b92f5019f8d5e2f2846a06b87db01ca4e/torch/ao/quantization/fx/_lower_to_native_backend.py#L766-L775, We should remove all the `dequant` nodes in the `_lower_static_weighted_ref_module` pass. **Test Plan**: It's a bug fix instead of a new feature. When we enable the `conv add` fusion PR later, we will add test cases accordingly. cc jerryzh168 jianyuh raghuramank100 jamesr66a vkuzo jgong5 Xia-Weiwen mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
Collaborator
Author
|
@jerryzh168 I think this PR is ready for review, could you help to take a look? |
…as multi input args" **Summary**: When converting a ref module into a quant module, `_lower_static_weighted_ref_module` pass assumes the `ref_node` only has 1 input node, and only remove the first `dequant` node. However, when we enable the `conv add` fusion, there will be a extra input node from `add` node besides the original input node from `conv`. Similar as did in the `_lower_quantized_binary_op` pass https://github.com/pytorch/pytorch/blob/41c3b41b92f5019f8d5e2f2846a06b87db01ca4e/torch/ao/quantization/fx/_lower_to_native_backend.py#L766-L775, We should remove all the `dequant` nodes in the `_lower_static_weighted_ref_module` pass. **Test Plan**: It's a bug fix instead of a new feature. When we enable the `conv add` fusion PR later, we will add test cases accordingly. cc jerryzh168 jianyuh raghuramank100 jamesr66a vkuzo jgong5 Xia-Weiwen mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
…as multi input args" **Summary**: When converting a ref module into a quant module, `_lower_static_weighted_ref_module` pass assumes the `ref_node` only has 1 input node, and only remove the first `dequant` node. However, when we enable the `conv add` fusion, there will be a extra input node from `add` node besides the original input node from `conv`. Similar as did in the `_lower_quantized_binary_op` pass https://github.com/pytorch/pytorch/blob/41c3b41b92f5019f8d5e2f2846a06b87db01ca4e/torch/ao/quantization/fx/_lower_to_native_backend.py#L766-L775, We should remove all the `dequant` nodes in the `_lower_static_weighted_ref_module` pass. **Test Plan**: It's a bug fix instead of a new feature. When we enable the `conv add` fusion PR later, we will add test cases accordingly. cc jerryzh168 jianyuh raghuramank100 jamesr66a vkuzo jgong5 Xia-Weiwen mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
…as multi input args" **Summary**: When converting a ref module into a quant module, `_lower_static_weighted_ref_module` pass assumes the `ref_node` only has 1 input node, and only remove the first `dequant` node. However, when we enable the `conv add` fusion, there will be a extra input node from `add` node besides the original input node from `conv`. Similar as did in the `_lower_quantized_binary_op` pass https://github.com/pytorch/pytorch/blob/41c3b41b92f5019f8d5e2f2846a06b87db01ca4e/torch/ao/quantization/fx/_lower_to_native_backend.py#L766-L775, We should remove all the `dequant` nodes in the `_lower_static_weighted_ref_module` pass. **Test Plan**: It's a bug fix instead of a new feature. When we enable the `conv add` fusion PR later, we will add test cases accordingly. cc jerryzh168 jianyuh raghuramank100 jamesr66a vkuzo jgong5 Xia-Weiwen mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
leslie-fang-intel
added a commit
to leslie-fang-intel/pytorch
that referenced
this pull request
Dec 19, 2022
…input ghstack-source-id: 63fcac0 Pull Request resolved: pytorch#90157
Collaborator
Author
|
Hi @jerryzh168, Could you help to take a look of this fix? |
jerryzh168
reviewed
Dec 20, 2022
| dq_node = ref_node.args[0] | ||
| assert(isinstance(dq_node, Node)) | ||
| ref_node.replace_input_with(dq_node, dq_node.args[0]) | ||
| for arg in ref_node.args: |
Contributor
There was a problem hiding this comment.
I feel this might be a bit hacky, generally we want to identify a specific pattern and just lower that pattern. this function (_lower_static_weighted_ref_module) is assuming that we have "dq -> ref_fp32_module -> q" pattern I think, could you
1). add some checks to this function to make sure this is the case, e.g. check len(ref_node.args) == 1 or something
2). add another lowering function for conv -> add that uses this code?
Collaborator
Author
There was a problem hiding this comment.
Thanks for the suggestions, @jerryzh168. Followed up these 2 steps:
- In this PR, I only add the the check of
len(ref_node.args) == 1for this pass (_lower_static_weighted_ref_module) - I have added another lowering pass for
conv -> addnamed as_lower_static_weighted_ref_module_with_two_dq_inputsin this [Quant][FX] Lower QConvAdd2d for onednn backend #91153. Could you take a look of this lowering pass as we talked about here? Does it look good to you?
Collaborator
Author
There was a problem hiding this comment.
Hi @jerryzh168, Could you help to take a look of this fix again?
Contributor
There was a problem hiding this comment.
sounds great, could you update the summary for this PR as well
This was referenced Dec 20, 2022
…as multi input args" **Summary**: When converting a ref module into a quant module, `_lower_static_weighted_ref_module` pass assumes the `ref_node` only has 1 input node, and only remove the first `dequant` node. However, when we enable the `conv add` fusion, there will be a extra input node from `add` node besides the original input node from `conv`. Similar as did in the `_lower_quantized_binary_op` pass https://github.com/pytorch/pytorch/blob/41c3b41b92f5019f8d5e2f2846a06b87db01ca4e/torch/ao/quantization/fx/_lower_to_native_backend.py#L766-L775, We should remove all the `dequant` nodes in the `_lower_static_weighted_ref_module` pass. **Test Plan**: It's a bug fix instead of a new feature. When we enable the `conv add` fusion PR later, we will add test cases accordingly. cc jerryzh168 jianyuh raghuramank100 jamesr66a vkuzo jgong5 Xia-Weiwen mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
…as multi input args" **Summary**: When converting a ref module into a quant module, `_lower_static_weighted_ref_module` pass assumes the `ref_node` only has 1 input node, and only remove the first `dequant` node. However, when we enable the `conv add` fusion, there will be a extra input node from `add` node besides the original input node from `conv`. Similar as did in the `_lower_quantized_binary_op` pass https://github.com/pytorch/pytorch/blob/41c3b41b92f5019f8d5e2f2846a06b87db01ca4e/torch/ao/quantization/fx/_lower_to_native_backend.py#L766-L775, We should remove all the `dequant` nodes in the `_lower_static_weighted_ref_module` pass. **Test Plan**: It's a bug fix instead of a new feature. When we enable the `conv add` fusion PR later, we will add test cases accordingly. cc jerryzh168 jianyuh raghuramank100 jamesr66a vkuzo jgong5 Xia-Weiwen mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
leslie-fang-intel
added a commit
to leslie-fang-intel/pytorch
that referenced
this pull request
Dec 20, 2022
…input ghstack-source-id: 08a8267 Pull Request resolved: pytorch#90157
…as multi input args" **Summary**: When converting a ref module into a quant module, `_lower_static_weighted_ref_module` pass assumes the `ref_node` only has 1 input node, and only remove the first `dequant` node. However, when we enable the `conv add` fusion, there will be a extra input node from `add` node besides the original input node from `conv`. Similar as did in the `_lower_quantized_binary_op` pass https://github.com/pytorch/pytorch/blob/41c3b41b92f5019f8d5e2f2846a06b87db01ca4e/torch/ao/quantization/fx/_lower_to_native_backend.py#L766-L775, We should remove all the `dequant` nodes in the `_lower_static_weighted_ref_module` pass. **Test Plan**: It's a bug fix instead of a new feature. When we enable the `conv add` fusion PR later, we will add test cases accordingly. cc jerryzh168 jianyuh raghuramank100 jamesr66a vkuzo jgong5 Xia-Weiwen mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
jerryzh168
approved these changes
Jan 4, 2023
Contributor
jerryzh168
left a comment
jerryzh168
left a comment
There was a problem hiding this comment.
thanks, please update the summary as well
Collaborator
Author
Updated the summary for this PR. |
Collaborator
Author
|
@pytorchbot rebase |
Collaborator
|
@pytorchbot successfully started a rebase job. Check the current status here |
…as multi input args" **Summary**: When converting a ref module into a quant module, `_lower_static_weighted_ref_module` pass assumes the `ref_node` only has 1 input node, and only remove the first `dequant` node. We add a check in this PR to ensure this is the case for `_lower_static_weighted_ref_module` pass. **Test Plan**: We only add a check in this PR, there is no new added test case. cc jerryzh168 jianyuh raghuramank100 jamesr66a vkuzo jgong5 Xia-Weiwen mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
Collaborator
|
Successfully rebased |
Collaborator
Author
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Collaborator
|
The merge job was canceled. If you believe this is a mistake,then you can re trigger it through pytorch-bot. |
Collaborator
Author
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
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 (oldest at bottom):
Summary:
When converting a ref module into a quant module,
_lower_static_weighted_ref_modulepass assumes theref_nodeonly has 1 input node, and only remove the firstdequantnode. We add a check in this PR to ensure this is the case for_lower_static_weighted_ref_modulepass.Test Plan:
We only add a check in this PR, there is no new added test case.
cc @jerryzh168 @jianyuh @raghuramank100 @jamesr66a @vkuzo @jgong5 @Xia-Weiwen @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10