[ET-VK][build] Fix Windows Vulkan/Android build and llama ETDump profiling#21148
Merged
Conversation
…iling
Summary:
Three additive, backward-compatible fixes needed to build the ExecuTorch Vulkan
backend and profile the llama runner from a Windows host. Each is guarded so it
is a no-op on non-Windows / non-profiling builds.
- backends/vulkan/runtime/gen_vulkan_spv.py: normalize CRLF -> LF at the top of
preprocess() before the trailing-backslash macro-continuation regex. GLSL
templates checked out with CRLF otherwise leave a lone backslash that escape()
turns into an unterminated Python string literal
("SyntaxError: unterminated string literal") during shader codegen.
- third-party/CMakeLists.txt: allow importing a prebuilt host flatc/flatcc via
FLATC_EXECUTABLE / FLATCC_EXECUTABLE instead of building them from source. The
existing WIN32 host-tool build only supports the Visual Studio generator
(VS-only CMAKE_GENERATOR_PLATFORM, .exe-less BUILD_BYPRODUCTS), which fails
under the Ninja generator required by the Android NDK cross-compile.
- examples/models/llama/CMakeLists.txt: link etdump and define
ET_EVENT_TRACER_ENABLED when the etdump target exists (i.e. core libs built
with EXECUTORCH_BUILD_DEVTOOLS=ON + EXECUTORCH_ENABLE_EVENT_TRACER=ON). This
enables ETDump collection and makes the Vulkan delegate emit per-dispatch GPU
timestamps for per-shader profiling.
Test Plan:
Cross-compiled ExecuTorch core libs + llama_main for Android arm64 with the
Vulkan backend enabled from a Windows host (NDK r29, Ninja generator, imported
host flatc/flatcc). Confirmed libvulkan_backend.a is linked and llama_main runs
on a Galaxy S24 (Adreno 750): Qwen3 0.6B ~80 tok/s decode, Llama 3.2 1B 8da4w
~60 tok/s decode. With DEVTOOLS/EVENT_TRACER on, collected an ETDump containing
per-shader Vulkan events and analyzed prefill/decode breakdown via
custom_inspector.py --mode llm.
Reviewers:
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
pytorch-bot
Bot
added
the
module: vulkan
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21148
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ No FailuresAs of commit f4a0e68 with merge base 86c3470 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
digantdesai
approved these changes
Jul 23, 2026
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
2 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.
Three additive, backward-compatible fixes needed to build the ExecuTorch Vulkan backend and profile the llama runner from a Windows host. Each change is guarded so it is a no-op on non-Windows / non-profiling builds.
Fixes
backends/vulkan/runtime/gen_vulkan_spv.py— normalize CRLF to LF at the top ofpreprocess(), before the trailing-backslash macro-continuation regex runs. GLSL templates checked out with CRLF line endings otherwise leave a lone backslash thatescape()turns into an unterminated Python string literal (SyntaxError: unterminated string literal) during shader codegen.third-party/CMakeLists.txt— allow importing a prebuilt hostflatc/flatccviaFLATC_EXECUTABLE/FLATCC_EXECUTABLEinstead of building them from source. The existingWIN32host-tool build only supports the Visual Studio generator (VS-onlyCMAKE_GENERATOR_PLATFORM,.exe-lessBUILD_BYPRODUCTS), which fails under the Ninja generator required by the Android NDK cross-compile.examples/models/llama/CMakeLists.txt— linketdumpand defineET_EVENT_TRACER_ENABLEDwhen theetdumptarget exists (i.e. core libs built withEXECUTORCH_BUILD_DEVTOOLS=ON+EXECUTORCH_ENABLE_EVENT_TRACER=ON). This enables ETDump collection and makes the Vulkan delegate emit per-dispatch GPU timestamps for per-shader profiling.All three changes are additive and guarded, so they are no-ops on non-Windows hosts and on non-profiling builds.
Test plan
Cross-compiled ExecuTorch core libs +
llama_mainfor Android arm64 with the Vulkan backend enabled from a Windows host (NDK r29, Ninja generator, imported hostflatc/flatcc). Confirmedlibvulkan_backend.ais linked andllama_mainruns on a Galaxy S24 (Adreno 750): Qwen3 0.6B ~80 tok/s decode, Llama 3.2 1B 8da4w ~60 tok/s decode. With DEVTOOLS/EVENT_TRACER on, collected an ETDump containing per-shader Vulkan events and analyzed the prefill/decode breakdown viacustom_inspector.py --mode llm.This PR was authored with Claude Code.
🤖 Generated with Claude Code
cc @manuelcandales @digantdesai @cbilgin