pytorch

OSSForks/pytorch

Fork 0

mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00

Commit Graph

Author	SHA1	Message	Date
Justin Chu	fbbab794ef	[ONNX] Implement Attention-23 (#156431 ) Implement Attention-23 using sdpa and flexattention. - I used copilot for this. - Also updated the conversion logic to remove trailing None inputs. @gramalingam @kunal-vaishnavi @titaiwangms Pull Request resolved: https://github.com/pytorch/pytorch/pull/156431 Approved by: https://github.com/titaiwangms Co-authored-by: kunal-vaishnavi <115581922+kunal-vaishnavi@users.noreply.github.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>	2025-06-20 23:54:57 +00:00
Justin Chu	0e805aad7f	[ONNX] Support float4 (#151069 ) - Support exporting float4 models (note: currently we use IR version 10 universally in the exporter, which does not include float 4 support. Eventually when onnx runtime and the ecosystem moves to support the new IR version 11 we should bump our version to 11 in the exporter as well) - The shape of the type is set according to https://github.com/pytorch/pytorch/pull/148791#discussion_r2038704986 (added last dim with size 2) - Use ml_dtypes types when converting to numpy for consistency with ONNX IR Fix https://github.com/pytorch/pytorch/issues/150202 Pull Request resolved: https://github.com/pytorch/pytorch/pull/151069 Approved by: https://github.com/titaiwangms	2025-05-18 03:19:35 +00:00
Justin Chu	010963032c	[ONNX] Create onnx_symbolic (#148905 ) In the old exporter we allow users to define a symbolic() method to bypass JIT tracing for a block of logic. We can allow users to do similar things by creating symbolic ops at export. This PR implements `torch.onnx.ops.symbolic` and `torch.onnx.ops.symbolic_multi_out` to allow users to create onnx nodes symbolically with pt2 & fx. The custom pytorch ops were designed such that the attributes are encoded to be part of a valid fx op. Users provide shape and dtype for the meta function to produce the currect fake tensor during export. An example is ![image](https://github.com/user-attachments/assets/c62f5f21-e038-456e-a71d-b9a5d0a7cd9d) Pull Request resolved: https://github.com/pytorch/pytorch/pull/148905 Approved by: https://github.com/titaiwangms	2025-03-18 21:32:06 +00:00

Author

SHA1

Message

Date

Justin Chu

fbbab794ef

[ONNX] Implement Attention-23 (#156431 )

Implement Attention-23 using sdpa and flexattention.

- I used copilot for this.
- Also updated the conversion logic to remove trailing None inputs.

@gramalingam @kunal-vaishnavi @titaiwangms
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156431
Approved by: https://github.com/titaiwangms

Co-authored-by: kunal-vaishnavi <115581922+kunal-vaishnavi@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

2025-06-20 23:54:57 +00:00

Justin Chu

0e805aad7f

[ONNX] Support float4 (#151069 )

- Support exporting float4 models (note: currently we use IR version 10 universally in the exporter, which does not include float 4 support. Eventually when onnx runtime and the ecosystem moves to support the new IR version 11 we should bump our version to 11 in the exporter as well)
- The shape of the type is set according to https://github.com/pytorch/pytorch/pull/148791#discussion_r2038704986 (added last dim with size 2)
- Use ml_dtypes types when converting to numpy for consistency with ONNX IR

Fix https://github.com/pytorch/pytorch/issues/150202

Pull Request resolved: https://github.com/pytorch/pytorch/pull/151069
Approved by: https://github.com/titaiwangms

2025-05-18 03:19:35 +00:00

Justin Chu

010963032c

[ONNX] Create onnx_symbolic (#148905 )

In the old exporter we allow users to define a symbolic() method to bypass JIT tracing for a block of logic. We can allow users to do similar things by creating symbolic ops at export.

This PR implements `torch.onnx.ops.symbolic` and `torch.onnx.ops.symbolic_multi_out` to allow users to create onnx nodes symbolically with pt2 & fx. The custom pytorch ops were designed such that the attributes are encoded to be part of a valid fx op. Users provide shape and dtype for the meta function to produce the currect fake tensor during export.

An example is

![image](https://github.com/user-attachments/assets/c62f5f21-e038-456e-a71d-b9a5d0a7cd9d)

Pull Request resolved: https://github.com/pytorch/pytorch/pull/148905
Approved by: https://github.com/titaiwangms

2025-03-18 21:32:06 +00:00

3 Commits