Add FLOAT4E2M1 data type #6318
Merged
Add FLOAT4E2M1 data type #6318
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
Signed-off-by: Yuan Yao <yuanyao@nvidia.com>
justinchuby
approved these changes
Aug 24, 2024
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #6318 +/- ##
==========================================
+ Coverage 56.95% 57.26% +0.30%
==========================================
Files 506 507 +1
Lines 30467 31354 +887
Branches 4592 4679 +87
==========================================
+ Hits 17353 17954 +601
- Misses 12285 12550 +265
- Partials 829 850 +21 ☔ View full report in Codecov by Sentry. |
justinchuby
reviewed
Aug 24, 2024
There was a problem hiding this comment.
@yuanyao-nv I just noticed - the test data shouldn’t change in this PR?
There was a problem hiding this comment.
This was auto generated by update_doc.sh. I've seen cases where some seemingly irrelevant pb files get updated by the script. Do you understand why?
There was a problem hiding this comment.
The pb files can be different in bytes when generated on different OS. I think that’s what’s happening here.
There was a problem hiding this comment.
I see. I wonder if it's possible to let the pipeline auto-generate these test files instead and reduce the number of files changed in each PR.
lutzroeder
added a commit
to lutzroeder/netron
that referenced
this pull request
Aug 24, 2024
lutzroeder
added a commit
to lutzroeder/netron
that referenced
this pull request
Aug 24, 2024
andife
pushed a commit
to andife/onnx
that referenced
this pull request
Aug 26, 2024
### Description - Add FLOAT4E2M1 as a new data type to proto as well as relevant helper functions and tests. - This PR splits out the portion of onnx#6283 relevant to data type updates to reduce the PR's size. ### Motivation and Context Narrow precision data types with sub-byte bit widths are becoming solutions to the rising cost, performance, and deployment challenges of LLMs. ONNX already has INT4/UINT4. FP4 is another commonly used narrow-precision data type for compressing both the weights and activations of LLMs. For example [OCP](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf) MXFP4 uses E2M1 as element type. Similar to INT4/UNIT4, FP4 weights/inputs are expected to be packed. Signed-off-by: Yuan Yao <yuanyao@nvidia.com> Signed-off-by: Andreas Fehlner <fehlner@arcor.de>
github-merge-queue bot
pushed a commit
that referenced
this pull request
Aug 27, 2024
### Description - FLOAT4E2M1 has been added to proto in #6318 - This PR adds FLOAT4E2M1 support for QuantizeLinear, DequantizeLinear, Cast, CastLike (opset 23). - Also add support to non-compute ops: Constant, ConstantOfShape, Identity, Reshape, Shape, Size, If, Loop, Scan, Flatten, Pad, Squeeze, Unsqueeze, Transpose (opset 23). Similar to INT4/UNIT4, FP4 weights/inputs are expected to be packed. --------- Signed-off-by: Yuan Yao (yuanyao) <yuanyao@nvidia.com> Signed-off-by: Yuan Yao <yuanyao@nvidia.com>
linshokaku
pushed a commit
to linshokaku/onnx
that referenced
this pull request
Oct 2, 2024
### Description - Add FLOAT4E2M1 as a new data type to proto as well as relevant helper functions and tests. - This PR splits out the portion of onnx#6283 relevant to data type updates to reduce the PR's size. ### Motivation and Context Narrow precision data types with sub-byte bit widths are becoming solutions to the rising cost, performance, and deployment challenges of LLMs. ONNX already has INT4/UINT4. FP4 is another commonly used narrow-precision data type for compressing both the weights and activations of LLMs. For example [OCP](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf) MXFP4 uses E2M1 as element type. Similar to INT4/UNIT4, FP4 weights/inputs are expected to be packed. Signed-off-by: Yuan Yao <yuanyao@nvidia.com> Signed-off-by: Linsho Kaku <linsho@preferred.jp>
linshokaku
pushed a commit
to linshokaku/onnx
that referenced
this pull request
Oct 2, 2024
### Description - FLOAT4E2M1 has been added to proto in onnx#6318 - This PR adds FLOAT4E2M1 support for QuantizeLinear, DequantizeLinear, Cast, CastLike (opset 23). - Also add support to non-compute ops: Constant, ConstantOfShape, Identity, Reshape, Shape, Size, If, Loop, Scan, Flatten, Pad, Squeeze, Unsqueeze, Transpose (opset 23). Similar to INT4/UNIT4, FP4 weights/inputs are expected to be packed. --------- Signed-off-by: Yuan Yao (yuanyao) <yuanyao@nvidia.com> Signed-off-by: Yuan Yao <yuanyao@nvidia.com> Signed-off-by: Linsho Kaku <linsho@preferred.jp>
hariharans29
added a commit
to microsoft/onnxruntime
that referenced
this pull request
Sep 3, 2025
### Description onnx/onnx#6318 and onnx/onnx#6283 added FP4 support to ONNX. This change introduces the FP4 type in ORT and adds type support to one relevant operator (`Cast`) as a proof-of-concept for the type integration into ORT. More op support will be added on a need-basis. This change took inspiration from the following PRs: #14731 #22228 #20362 Some notes: 1) Only `tensor` type gets support for FP4 initially. Secondary types like `seq(tensor)`, `sparse_tensor`, `optional` do not get support (so as to not introduce unnecessary bloat to the framework without a solid use-case) 2) Flatbuffer related files receive no updates in this PR ### Motivation and Context Be able to run FP4 models with ORT
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
Motivation and Context
Narrow precision data types with sub-byte bit widths are becoming solutions to the rising cost, performance, and deployment challenges of LLMs. ONNX already has INT4/UINT4. FP4 is another commonly used narrow-precision data type for compressing both the weights and activations of LLMs. For example OCP MXFP4 uses E2M1 as element type.
Similar to INT4/UNIT4, FP4 weights/inputs are expected to be packed.