XPU use stock pytorch instead of Intel Extension for PyTorch - #7877
Merged
sfc-gh-truwase merged 19 commits intoMar 2, 2026
Merged
XPU use stock pytorch instead of Intel Extension for PyTorch#7877sfc-gh-truwase merged 19 commits into
sfc-gh-truwase merged 19 commits into
Conversation
Copy link
Copy Markdown
Collaborator
Sorry, something went wrong.
Replace all intel_extension_for_pytorch (IPEX) imports and APIs with stock PyTorch equivalents for XPU accelerator detection, device operations, and JIT compilation of SYCL kernels. Key changes: - Use torch.xpu.is_available() for XPU detection instead of ipex._C._has_xpu() - Replace DPCPPExtension/DpcppBuildExtension with SyclExtension/BuildExtension - Replace IPEX's cpp_extension.load() with torch.utils.cpp_extension.load() - Set CXX=icpx for JIT compilation of .cpp/.dp.cpp SYCL source files - Add _sycl_env_paths() to resolve SYCL ABI mismatch between system oneAPI compiler and Python environment's libsycl.so - Replace ipex.h includes with c10/xpu/XPUStream.h in C++ sources - Use torch.nn.functional.scaled_dot_product_attention for flash attention - Keep IPEX dependency only in inference.py for pre-compiled kernels Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
…b mismatch Replace AOT compilation (spir64_gen + device-specific targets like pvc/bmg) with pure JIT compilation (spir64) for portability across Intel GPU architectures. Add a wrapper header csrc/xpu/includes/sycl/feature_test.hpp that undefs SYCL_EXT_ONEAPI_BFLOAT16_MATH_FUNCTIONS after including the real header. This prevents c10::BFloat16 from using sycl::ext::oneapi::bfloat16 intrinsics, which would emit __devicelib_ConvertBF16ToFINTEL calls that fail at JIT time when the icpx compiler version (2025.3) is newer than the SYCL runtime shipped with PyTorch (2025.1). The bitwise fallback path is functionally identical. Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
The feature_test.hpp wrapper that #undef'd SYCL_EXT_ONEAPI_BFLOAT16_MATH_FUNCTIONS is no longer needed: PyTorch 2.10+xpu ships intel-sycl-rt 2025.3.1 which matches the system icpx 2025.3.1, eliminating the compiler/runtime version mismatch that caused unresolved __devicelib_ConvertBF16ToFINTEL symbols. Update the Intel XPU section of the accelerator setup guide to reflect the switch from IPEX to stock PyTorch, document the icpx version matching requirement, and add Arc Pro B60 to the verified GPU list. Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
…d dependency Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
…Torch Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Copy link
Copy Markdown
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2d30d045f3
ℹ️ About Codex in GitHubCodex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
Sorry, something went wrong.
Comment thread
accelerator/xpu_accelerator.py
delock
force-pushed
the
gma/xpu_use_stock_pytorch
branch
from
2d30d04 to
e0954c6
Compare
February 27, 2026 06:32
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
Copy link
Copy Markdown
Collaborator
Author
Sorry, something went wrong.
…vice_type=...) Signed-off-by: Ma, Guokai <guokai.ma@intel.com>
sfc-gh-truwase
approved these changes
Feb 28, 2026
Copy link
Copy Markdown
Collaborator
Author
|
Currently there is no XPU runner in DeepSpeed. For the record, here is the test we ran on an Intel ARC B60x4 system:
|
Sorry, something went wrong.
sfc-gh-truwase
pushed a commit
that referenced
this pull request
Mar 2, 2026
Pytorch now provide torch.amp https://docs.pytorch.org/docs/stable/amp.html as recommended AMP API instead of torch.<device_type>.amp which is used in DeepSpeed abstract accelerator amp(). Some PyTorch backend such as XPU does not provide the legacy `torch.xpu.amp` module. This PR replace `get_accelerator().amp()` by `torch.amp` which is the recommended way of using AMP. Related issues and PRs #7876 #7877 --------- Signed-off-by: Ma, Guokai <guokai.ma@intel.com> Co-authored-by: Olatunji Ruwase <tjruwase@gmail.com>
nathon-lee
pushed a commit
to nathon-lee/DeepSpeed_woo
that referenced
this pull request
Mar 7, 2026
…edai#7877) With Intel Extension for PyTorch retiring, XPU device would be supported by PyTorch 2.8+ and dependency to Intel Extension for PyTorch would not be needed. This PR removed IPEX dependency, adapt to builder protocol in PyTorch for XPU, and updated documents and tests accordingly. Note after this update, DeepSpeed will not work with previous PyTorch+IPEX on XPU devices. Suggest user to upgrade to latest PyTorch to get latest XPU features on XPU devices. Come with this PR is removal of InferenceBuilder, the kernel needed by InferenceBuilder is supported through Intel Extension for PyTorch. --------- Signed-off-by: Ma, Guokai <guokai.ma@intel.com> Co-authored-by: Olatunji Ruwase <tjruwase@gmail.com> Co-authored-by: Olatunji Ruwase <tunji.ruwase@snowflake.com> Signed-off-by: nathon-lee <leejianwoo@gmail.com>
nathon-lee
pushed a commit
to nathon-lee/DeepSpeed_woo
that referenced
this pull request
Mar 7, 2026
Pytorch now provide torch.amp https://docs.pytorch.org/docs/stable/amp.html as recommended AMP API instead of torch.<device_type>.amp which is used in DeepSpeed abstract accelerator amp(). Some PyTorch backend such as XPU does not provide the legacy `torch.xpu.amp` module. This PR replace `get_accelerator().amp()` by `torch.amp` which is the recommended way of using AMP. Related issues and PRs deepspeedai#7876 deepspeedai#7877 --------- Signed-off-by: Ma, Guokai <guokai.ma@intel.com> Co-authored-by: Olatunji Ruwase <tjruwase@gmail.com> Signed-off-by: nathon-lee <leejianwoo@gmail.com>
nathon-lee
pushed a commit
to nathon-lee/DeepSpeed_woo
that referenced
this pull request
Mar 28, 2026
…edai#7877) With Intel Extension for PyTorch retiring, XPU device would be supported by PyTorch 2.8+ and dependency to Intel Extension for PyTorch would not be needed. This PR removed IPEX dependency, adapt to builder protocol in PyTorch for XPU, and updated documents and tests accordingly. Note after this update, DeepSpeed will not work with previous PyTorch+IPEX on XPU devices. Suggest user to upgrade to latest PyTorch to get latest XPU features on XPU devices. Come with this PR is removal of InferenceBuilder, the kernel needed by InferenceBuilder is supported through Intel Extension for PyTorch. --------- Signed-off-by: Ma, Guokai <guokai.ma@intel.com> Co-authored-by: Olatunji Ruwase <tjruwase@gmail.com> Co-authored-by: Olatunji Ruwase <tunji.ruwase@snowflake.com> Signed-off-by: nathon-lee <leejianwoo@gmail.com>
nathon-lee
pushed a commit
to nathon-lee/DeepSpeed_woo
that referenced
this pull request
Mar 28, 2026
Pytorch now provide torch.amp https://docs.pytorch.org/docs/stable/amp.html as recommended AMP API instead of torch.<device_type>.amp which is used in DeepSpeed abstract accelerator amp(). Some PyTorch backend such as XPU does not provide the legacy `torch.xpu.amp` module. This PR replace `get_accelerator().amp()` by `torch.amp` which is the recommended way of using AMP. Related issues and PRs deepspeedai#7876 deepspeedai#7877 --------- Signed-off-by: Ma, Guokai <guokai.ma@intel.com> Co-authored-by: Olatunji Ruwase <tjruwase@gmail.com> Signed-off-by: nathon-lee <leejianwoo@gmail.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Reviewers
3 participants
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.
With Intel Extension for PyTorch retiring, XPU device would be supported by PyTorch 2.8+ and dependency to Intel Extension for PyTorch would not be needed.
This PR removed IPEX dependency, adapt to builder protocol in PyTorch for XPU, and updated documents and tests accordingly.
Note after this update, DeepSpeed will not work with previous PyTorch+IPEX on XPU devices. Suggest user to upgrade to latest PyTorch to get latest XPU features on XPU devices.
Come with this PR is removal of InferenceBuilder, the kernel needed by InferenceBuilder is supported through Intel Extension for PyTorch.