[Cadence] add reference quantized fully connected out #9018
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
Differential Revision: D70503708 Pull Request resolved: #8888
Differential Revision: D70528853 Pull Request resolved: #8908
Differential Revision: D70538475 Pull Request resolved: #8918
…8616) * Qualcomm AI Engine Direct - Meta CI for Mobilebert and W2L * variable update
The TOSA compiler previously had a bug that caused a segmentation fault when loading the Wav2Letter model on Ethos-U85. This issue has now been fixed. Enable the test that previously failed due to this bug. Co-authored-by: Martin Lindström <[email protected]>
* Update using-executorch-building-from-source.md * Update using-executorch-building-from-source.md
…8411) * [ARM backend] Update fuse_batchnorm_pass to create new placeholders - This allows to fuse bn+convs with multiple users of the same weights - Adds new util functions create/delete_const_placeholders to take care of updating the GraphSignature and state_dict/constants dict when handling constant placholders. - Adds and updates related tests Change-Id: I8e550614d9741de840786d9dca9f30af9eb95a64 * Move create/delete_constant_node utils to shared folder Change-Id: I3a82f58f9796e421bd205f030f7d79d72a2f7ed9 * Add buck dependency * Fix bazel build --------- Co-authored-by: Digant Desai <[email protected]>
Currently the result has large variance from outliers, so only use 80% samples in the middle (trimmean 0.2)
* up * up * up * up * up * up * up * up * up * up * up * up * up
) This is not supported, so we shouldn't partition it. Add an expectedFailure test to indicate that this is not supported Differential Revision: [D70343584](https://our.internmc.facebook.com/intern/diff/D70343584/) ghstack-source-id: 269356867 Pull Request resolved: #8891 Co-authored-by: Digant Desai <[email protected]>
Doesn't seem to be any reason not to allow optimized ops for this one.
[Tensor.cpp] add algorithm include to get stable_sort and iter_swap on windows
Differential Revision: D69947096 Pull Request resolved: #8896
Pull Request resolved: #8892 Differential Revision: [D70372220](https://our.internmc.facebook.com/intern/diff/D70372220/) ghstack-source-id: 269599293 Co-authored-by: Digant Desai <[email protected]>
We don't need a second isnan; see code comment. (This is a small optimization.)
Differential Revision: D70528015 Pull Request resolved: #8906
Differential Revision: D70540908 Pull Request resolved: #8943
Update [ghstack-poisoned]
See class comment. In brief, this adds an iterable range to make broadcasting ops convenient and efficient to implement.
#8941) Fix Timing Adapter settings depending on the memory mode & placement in the linker script
Pull Request resolved: #8887 PteDataMap is the NamedDataMap that will live in the runtime. It is used to give delegates access to opaque named data stored in the PTE file. Open to alternative naming suggestions, maybe 'PTEDataMap' or 'ProgramDataMap'? **Usage** The PteDataMap is owned by the program, and instantiated at program load time if named_data exists in the PTE file. We introduce usage of 'std::optional' here. I think we can also use executorch::aten::optional to avoid adding standard lib ? When initializing delegates, the PteDataMap is given to delegate_init. Delegates can retrieve opaque delegate data by key using 'get_data'. This gives them a FreeableBuffer that they can free later. **Testing** This test uses the C++ flatbuffer API to build a fake program containing named data. We also creates a temp file with sample data that the data loader can wrap around. TODO: e2e test once delegate aot is ready and we can generate a file with named data. **Note** As the PteDataMap wraps around flatbuffer constructs, the Program must outlive the PteDataMap. PteDataMap does not implement - get_metadata; currently, all data stored is opaque. Later, we can implement get_metadata if a backend stores plain tensor data. - load_into; this is mostly used for the training case, and isn't used by delegates, at least not at the moment ghstack-source-id: 269779453 Differential Revision: [D70213646](https://our.internmc.facebook.com/intern/diff/D70213646/) Co-authored-by: lucylq <[email protected]>
Differential Revision: D70541550 Pull Request resolved: #8921
Differential Revision: D70597013 Pull Request resolved: #8952
Differential Revision: D70577129 Pull Request resolved: #8940
* fix windows build issue * revert xnnpack cmakelist and fix file_data_loader * fix lint warning * undo non-relevant changes * linter error - extra newline --------- Co-authored-by: Chao Zhang <[email protected]>
partner engineers are calling ET via LlamaModule: https://github.com/pytorch/executorch/blob/main/extension/android/src/main/java/org/pytorch/executorch/LlamaModule.java This is a wrapper around the runner: https://www.internalfb.com/code/fbsource/[90d251fc01a84871b679406d6dc855eb5ded82fd]/fbcode/executorch/examples/models/llama/runner/runner.cpp?lines=47 Differential Revision: [D70596210](https://our.internmc.facebook.com/intern/diff/D70596210/) ghstack-source-id: 269741205 Pull Request resolved: #8953 Co-authored-by: lucylq <[email protected]>
don't use invalid flags on windows
Differential Revision: D70334196 Pull Request resolved: #8478
Previously it was copied in several places per executorch_srcs.cmake. Needed for #8932 Test Plan: Compare cmake-out/executorch_srcs.cmake before/after for my usual testing cmake config with "all the CPU stuff" on; found that thread_parallel.cpp is now duplicated only in one place instead of multiple (it's in llama_runner, which needs a general fixup because it's duplicating several extensions).
Summary - Remove redundant directory Co-authored-by: DannyYuyang-quic <[email protected]>
* [executorch][runtime] Introduce PteDataMap for weight sharing Pull Request resolved: #8887 PteDataMap is the NamedDataMap that will live in the runtime. It is used to give delegates access to opaque named data stored in the PTE file. Open to alternative naming suggestions, maybe 'PTEDataMap' or 'ProgramDataMap'? **Usage** The PteDataMap is owned by the program, and instantiated at program load time if named_data exists in the PTE file. We introduce usage of 'std::optional' here. I think we can also use executorch::aten::optional to avoid adding standard lib ? When initializing delegates, the PteDataMap is given to delegate_init. Delegates can retrieve opaque delegate data by key using 'get_data'. This gives them a FreeableBuffer that they can free later. **Testing** This test uses the C++ flatbuffer API to build a fake program containing named data. We also creates a temp file with sample data that the data loader can wrap around. TODO: e2e test once delegate aot is ready and we can generate a file with named data. **Note** As the PteDataMap wraps around flatbuffer constructs, the Program must outlive the PteDataMap. PteDataMap does not implement - get_metadata; currently, all data stored is opaque. Later, we can implement get_metadata if a backend stores plain tensor data. - load_into; this is mostly used for the training case, and isn't used by delegates, at least not at the moment Differential Revision: [D70213646](https://our.internmc.facebook.com/intern/diff/D70213646/) ghstack-source-id: 269691307 * [executorch][runtime] Add get_named_data_map to Program Pull Request resolved: #8853 Add to the program interface, to allow users to retrieve the NDM. Differential Revision: [D70276106](https://our.internmc.facebook.com/intern/diff/D70276106/) ghstack-source-id: 269693108 --------- Co-authored-by: lucylq <[email protected]>
IMO, Buck visibility is just inverse deps. We should trust that people have a good reason to add deps rather than attempt to police them and require double entry in both deps and visibility, especially since we seem to be committed to APIs by default in OSS anyway. Specific motivation is that #8712 would otherwise have to ad-hoc slap ExecuTorch-wide visibility on a lot of targets, but I've held this view for a long time. Differential Revision: D70647462
Fix java build
* [Windows Build] Implement MMAP for mmap_data_loader.cpp There is no sys/mman.h or posix-compatible mmap() implementation on Windows. The extension data loaders use it to map in data files, so adding an implementation. Test-run, & .\cmake-out\extension\data_loader\test\Debug\extension_data_loader_test.exe --gtest_brief=1 --gtest_filter=MmapDataLoader* Running main() from ...\executorch\third-party\googletest\googletest\src\gtest_main.cc [==========] 8 tests from 1 test suite ran. (50 ms total) [ PASSED ] 8 tests. * apply code suggestions * fix src -> srcs typo * try to fix build * fix src/headers brackets --------- Co-authored-by: Jeff Whiteside <[email protected]>
* Fix phi4mini test model * Remove pull trigger since done testing
Program schema should stay a private dep.
…ass_manager (#8997) Add FuseViewCopyTransform and FuseConstantsPass in arm_pass_manager These passes both removes redundant ops from the graph: - FuseViewCopyTransform pass is added from backends/transforms to merge sequential view ops. - FuseConstantOpsPass is created to compute ops with constant inputs AOT - This is not done in cases where the result is a larger tensor, to avoid increasing the constant memory size. - For BI, ops are quantized with the q/dq-ops as to not change the behaviour of the graph. - Pass order is important: the pass must be placed after all passes which may add constant ops, but before the InsertTableOpsPass, since it doesn't handle TOSA _table-ops. Signed-off-by: Adrian Lundell <[email protected]>
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/9018
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
zonglinpeng
had a problem deploying
to
upload-benchmark-results
GitHub Actions
Failure
zonglinpeng
had a problem deploying
to
upload-benchmark-results
GitHub Actions
Failure
This was referenced Mar 17, 2025
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.
add reference quantized fully connected out
test plan
python3 -m examples.cadence.models.rnnt_joiner