[QDQ Optimizer] Update WeightBiasQuantization to skip Conv/Gemm if downstream node is not QuantizeLinear #24537
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
Contributor
Author
|
Hi @vraspar FYI, to be cherry-picked for ORT 1.22.0 |
vraspar
pushed a commit
that referenced
this pull request
Apr 28, 2025
…wnstream node is not QuantizeLinear (#24537) ### Description Updates the WeightBiasQuantization optimizer to skip processing on Conv/Gemm nodes if the downstream child node is not a QuantizeLinear. #### Before this PR Original graph: ``` input_0 -> DQ -> Conv -> graph_output (or non-Q node) ^ ^ | | weights_f32------+ | bias_f32------------+ ``` Becomes: ``` input_0 -> DQ ------> Conv -> graph_output (or non-Q node) ^ ^ | | weights_quant -> DQ --+ | bias_quant -> DQ --------+ ``` The above is **NOT** a valid QDQ node unit for Conv because the Conv's output is not consumed by a QuantizeLinear node. #### With this PR The above example graph remains unchanged after L1 optimizations: ``` input_0 -> DQ -> Conv -> graph_output (or non-Q node) ^ ^ | | weights_f32------+ | bias_f32------------+ ``` ### Motivation and Context Caused inaccuracy for a customer model. Automatically quantizing the weights and biases of a Conv/Gemm is detrimental if the output of the Conv/Gemm is not consumed by a QuantizeLinear node. In this scenario, the whole node group is not considered a valid QDQ node unit, and so the EP has to run the Conv/Gemm as float32/float16 anyway. If the Conv/Gemm is running as float32/float16, then quantizing the weights and biases introduces inaccuracy for no gain. PR that originally added this optimizer: #22969
jywu-msft
pushed a commit
that referenced
this pull request
Apr 30, 2025
### Description Cherry pick the following into [rel-1.22.0](https://github.com/microsoft/onnxruntime/tree/rel-1.22.0) - (#24487) - (#24466) - (#24493) - (#24484) - (#24494) - (#24489) - (#24504) - (#24510) - (#24456) - (#24537) - (#24501) - (#24519) - (#24513) - (#24539) - (#24514) - (#24542) - (#24585) Not added: Planning to cherry pick Cuda Matmulnbits PRs once the fix for failing cuda pipeline is ready - (#24491) - (#24509) - (#24564) --------- Co-authored-by: Adrian Lizarraga <adlizarraga@microsoft.com> Co-authored-by: minfhong-quic <quic_minfhong@quicinc.com> Co-authored-by: minfhong-quic <minfhong-quic@quicinc.com> Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com> Co-authored-by: Prathik Rao <prathik.rao@gmail.com> Co-authored-by: Edward Chen <18449977+edgchen1@users.noreply.github.com> Co-authored-by: Ankan Banerjee <ankan.ban@gmail.com> Co-authored-by: Maximilian Müller <maximilianm@nvidia.com> Co-authored-by: Gaurav Garg <gaugarg@nvidia.com> Co-authored-by: iraut <iraut@nvidia.com> Co-authored-by: Hrishikesh Manohar <hrishikeshm@nvidia.com> Co-authored-by: Maximilian Müller <44298237+gedoensmax@users.noreply.github.com> Co-authored-by: Scott McKay <skottmckay@gmail.com> Co-authored-by: Jiajia Qin <jiajiaqin@microsoft.com> Co-authored-by: kunal-vaishnavi <115581922+kunal-vaishnavi@users.noreply.github.com> Co-authored-by: xhcao <xinghua.cao@intel.com>
jatinwadhwa921
pushed a commit
to intel/onnxruntime
that referenced
this pull request
Apr 30, 2025
### Description Cherry pick the following into [rel-1.22.0](https://github.com/microsoft/onnxruntime/tree/rel-1.22.0) - (microsoft#24487) - (microsoft#24466) - (microsoft#24493) - (microsoft#24484) - (microsoft#24494) - (microsoft#24489) - (microsoft#24504) - (microsoft#24510) - (microsoft#24456) - (microsoft#24537) - (microsoft#24501) - (microsoft#24519) - (microsoft#24513) - (microsoft#24539) - (microsoft#24514) - (microsoft#24542) - (microsoft#24585) Not added: Planning to cherry pick Cuda Matmulnbits PRs once the fix for failing cuda pipeline is ready - (microsoft#24491) - (microsoft#24509) - (microsoft#24564) --------- Co-authored-by: vraspar <vrajang@outlook.com> Co-authored-by: Adrian Lizarraga <adlizarraga@microsoft.com> Co-authored-by: minfhong-quic <quic_minfhong@quicinc.com> Co-authored-by: minfhong-quic <minfhong-quic@quicinc.com> Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com> Co-authored-by: Prathik Rao <prathik.rao@gmail.com> Co-authored-by: Edward Chen <18449977+edgchen1@users.noreply.github.com> Co-authored-by: Ankan Banerjee <ankan.ban@gmail.com> Co-authored-by: Maximilian Müller <maximilianm@nvidia.com> Co-authored-by: Gaurav Garg <gaugarg@nvidia.com> Co-authored-by: iraut <iraut@nvidia.com> Co-authored-by: Hrishikesh Manohar <hrishikeshm@nvidia.com> Co-authored-by: Maximilian Müller <44298237+gedoensmax@users.noreply.github.com> Co-authored-by: Scott McKay <skottmckay@gmail.com> Co-authored-by: Jiajia Qin <jiajiaqin@microsoft.com> Co-authored-by: kunal-vaishnavi <115581922+kunal-vaishnavi@users.noreply.github.com> Co-authored-by: xhcao <xinghua.cao@intel.com>
ankitm3k
pushed a commit
to intel/onnxruntime
that referenced
this pull request
May 12, 2025
…wnstream node is not QuantizeLinear (microsoft#24537) ### Description Updates the WeightBiasQuantization optimizer to skip processing on Conv/Gemm nodes if the downstream child node is not a QuantizeLinear. #### Before this PR Original graph: ``` input_0 -> DQ -> Conv -> graph_output (or non-Q node) ^ ^ | | weights_f32------+ | bias_f32------------+ ``` Becomes: ``` input_0 -> DQ ------> Conv -> graph_output (or non-Q node) ^ ^ | | weights_quant -> DQ --+ | bias_quant -> DQ --------+ ``` The above is **NOT** a valid QDQ node unit for Conv because the Conv's output is not consumed by a QuantizeLinear node. #### With this PR The above example graph remains unchanged after L1 optimizations: ``` input_0 -> DQ -> Conv -> graph_output (or non-Q node) ^ ^ | | weights_f32------+ | bias_f32------------+ ``` ### Motivation and Context Caused inaccuracy for a customer model. Automatically quantizing the weights and biases of a Conv/Gemm is detrimental if the output of the Conv/Gemm is not consumed by a QuantizeLinear node. In this scenario, the whole node group is not considered a valid QDQ node unit, and so the EP has to run the Conv/Gemm as float32/float16 anyway. If the Conv/Gemm is running as float32/float16, then quantizing the weights and biases introduces inaccuracy for no gain. PR that originally added this optimizer: microsoft#22969
Contributor
|
This PR has been included in the |
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.
Description
Updates the WeightBiasQuantization optimizer to skip processing on Conv/Gemm nodes if the downstream child node is not a QuantizeLinear.
Before this PR
Original graph:
Becomes:
The above is NOT a valid QDQ node unit for Conv because the Conv's output is not consumed by a QuantizeLinear node.
With this PR
The above example graph remains unchanged after L1 optimizations:
Motivation and Context
Caused inaccuracy for a customer model. Automatically quantizing the weights and biases of a Conv/Gemm is detrimental if the output of the Conv/Gemm is not consumed by a QuantizeLinear node. In this scenario, the whole node group is not considered a valid QDQ node unit, and so the EP has to run the Conv/Gemm as float32/float16 anyway. If the Conv/Gemm is running as float32/float16, then quantizing the weights and biases introduces inaccuracy for no gain.
PR that originally added this optimizer: #22969