Add rms_norm op + named-data upload + x86 CI (#19893)#19893
Open
JulianCloudNTH wants to merge 1 commit into
Open
Add rms_norm op + named-data upload + x86 CI (#19893)#19893JulianCloudNTH wants to merge 1 commit into
JulianCloudNTH wants to merge 1 commit into
Conversation
JulianCloudNTH
requested review from
kirklandsign and
larryliu0820
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19893
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. |
|
|
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@JulianCloudNTH has exported this pull request. If you are a Meta employee, you can view the originating Diff in D106887028. |
This PR needs a
|
mergennachin
changed the title
meta-codesync
Bot
changed the title
JulianCloudNTH
force-pushed
the
export-D106887028
branch
2 times, most recently
from
1abcd20 to
38e33e5
Compare
Summary: Adds the `et_vk.rms_norm.default` operator to the WebGPU backend (a WGSL compute shader using a cooperative tree reduction, one workgroup per row), fixes constant upload so the op's weight loads correctly, and wires the backend into CI. The Vulkan serializer that the WebGPU backend reuses stores every non-empty constant (e.g. the rms_norm weight) in the PTE's named-data map with `offset == UINT64_MAX` and a `named_key`, rather than inline in the VK00 blob. `WebGPUGraph::build` previously handled only inline constants, so the weight was never uploaded and the op returned all zeros. `build` now also fetches named-data constants via `NamedDataMap::get_data`, mirroring the path `VulkanBackend` already uses. `aten.add` was unaffected since it has no constant tensors. The shader mirrors the Vulkan implementation (`backends/vulkan/runtime/graph/ops/impl/RmsNorm.cpp`, `backends/vulkan/runtime/graph/ops/glsl/rms_norm_buffer.glsl`); indexing assumes contiguous fp32 inputs. The handler fails loud (throws, mirroring Vulkan's `VK_CHECK_COND`) on invalid shape/dtype/dispatch-limit conditions, and defaults `eps` to the float32 machine epsilon. Also adds a simple x86 Linux CI job, mirroring the Vulkan delegate: `backends/test/suite/flows/webgpu.py` plus a `WebGPUTester`, run by `oss/.github/workflows/test-backend-webgpu.yml` on SwiftShader (a software Vulkan adapter, via `wgpu-native`, minimal dependencies, no GPU). Two fixes were needed for SwiftShader's downlevel limits: request the adapter's full `requiredLimits` at device creation (software adapters default storage-buffer limits to 0), and lower the `add` op `workgroup_size` from 256 to 64 (256 exceeded SwiftShader's 128-invocation cap; the Vulkan delegate uses 64). Differential Revision: D106887028
meta-codesync
Bot
changed the title
JulianCloudNTH
force-pushed
the
export-D106887028
branch
from
38e33e5 to
d6f278e
Compare
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
1 participant
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.
Summary:
Adds the
et_vk.rms_norm.defaultoperator to the WebGPU backend (a WGSL compute shader using a cooperative tree reduction, one workgroup per row), fixes constant upload so the op's weight loads correctly, and wires the backend into CI.The Vulkan serializer that the WebGPU backend reuses stores every non-empty constant (e.g. the rms_norm weight) in the PTE's named-data map with
offset == UINT64_MAXand anamed_key, rather than inline in the VK00 blob.WebGPUGraph::buildpreviously handled only inline constants, so the weight was never uploaded and the op returned all zeros.buildnow also fetches named-data constants viaNamedDataMap::get_data, mirroring the pathVulkanBackendalready uses.aten.addwas unaffected since it has no constant tensors.The shader mirrors the Vulkan implementation (
backends/vulkan/runtime/graph/ops/impl/RmsNorm.cpp,backends/vulkan/runtime/graph/ops/glsl/rms_norm_buffer.glsl); indexing assumes contiguous fp32 inputs. The handler fails loud (throws, mirroring Vulkan'sVK_CHECK_COND) on invalid shape/dtype/dispatch-limit conditions, and defaultsepsto the float32 machine epsilon.Also adds a simple x86 Linux CI job, mirroring the Vulkan delegate:
backends/test/suite/flows/webgpu.pyplus aWebGPUTester, run byoss/.github/workflows/test-backend-webgpu.ymlon SwiftShader (a software Vulkan adapter, viawgpu-native, minimal dependencies, no GPU). Two fixes were needed for SwiftShader's downlevel limits: request the adapter's fullrequiredLimitsat device creation (software adapters default storage-buffer limits to 0), and lower theaddopworkgroup_sizefrom 256 to 64 (256 exceeded SwiftShader's 128-invocation cap; the Vulkan delegate uses 64).Differential Revision: D106887028