Add float int32 vulkan shader

[mohammed-saalim]

Summary
Adds missing float to int32 conversion support in the Vulkan backend's view_convert_buffer shader.

Problem
When exporting models that require float-to-int32 dtype conversion to the Vulkan backend, execution fails at runtime with:

Exception raised from get_shader_info at /pytorch/executorch/backends/vulkan/runtime/api/ShaderRegistry.cpp:54:
(it != listings_.end()) is false! Could not find ShaderInfo with name view_convert_buffer_float_int32
This occurs because the view_convert_buffer shader only supports conversions FROM int32/uint8 TO float/half, but not the reverse direction.

Solution
Added float → int32 conversion variant to the shader generation configuration in
view_convert_buffer.yaml
.

Changed file:

backends/vulkan/runtime/graph/ops/glsl/view_convert_buffer.yaml
Change:

yaml
combos:

• parameter_values: [int32, float]
• parameter_values: [int32, half]
• parameter_values: [uint8, float]
• parameter_values: [uint8, half]
• parameter_values: [uint8, int32]
• parameter_values: [float, int32] # <-- Added
  Testing
  Unable to test on Windows due to build environment constraints (requires CMake + MSVC). The change follows the existing pattern for other dtype conversion combinations in the same file.

Test case that would trigger this:

Models with operations requiring float→int32 conversion (e.g., certain quantization or indexing operations)
Export to Vulkan backend using VulkanPartitioner
Additional Context
This issue was encountered while exporting a depth estimation model (PanDA) to the Vulkan backend for Android deployment. The model worked with XNNPACK but failed with Vulkan due to this missing shader variant.

cc @SS-JIA @manuelcandales @digantdesai @cbilgin

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17382

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 2 New Failures, 1 Unrelated Failure

As of commit 0534658 with merge base eb888f8 ():

NEW FAILURES - The following jobs have failed:

• pull / test-samsung-models-linux / linux-job (gh)
  RuntimeError: Command docker exec -t 59fcf94b91a5319e1ade81f962fc611c7fa80c944fcdf9688a7153c6ee7e129d /exec failed with exit code 1
• pull / test-samsung-quantmodels-linux / linux-job (gh)
  RuntimeError: Command docker exec -t 6aa4ec9afa6f37e8abea0c5020f7f05db13d812ee6429c256e6ed02fc26bd99b /exec failed with exit code 1

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / test-llama-runner-qnn-linux (fp32, qnn_16a16w, qnn) / linux-job (gh) (trunk failure)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[Copilot]

Pull request overview

Adds the missing float -> int32 shader variant for the Vulkan backend’s view_convert_buffer op so models requiring that dtype conversion no longer fail at runtime due to an unregistered shader name.

Changes:

• Add float,int32 to the generated shader combination list for view_convert_buffer.
• Remove an editable-install warning block from the “building from source” docs.

Reviewed changes

Copilot reviewed 1 out of 1 changed files in this pull request and generated 2 comments.

File	Description
docs/source/using-executorch-building-from-source.md	Removes an editable-install warning/workaround block from the build-from-source guide.
backends/vulkan/runtime/graph/ops/glsl/view_convert_buffer.yaml	Adds the float -> int32 combo so view_convert_buffer_float_int32 gets generated/registered.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Adding the float→int32 combo makes view_convert_buffer perform a direct GLSL cast from float to int (OUT_T(t_inp[...]) where OUT_T becomes int). In GLSL, float-to-int conversion is undefined/unspecified for NaNs and for values outside the representable int32 range, which can lead to non-deterministic results. Consider updating the shader logic for this combo to explicitly define behavior (e.g., clamp to int32 min/max before casting and decide how to handle NaNs), so the new variant is robust and matches the expected runtime semantics.

Suggested change

- parameter_values: [float, int32]

[Copilot]

This change fixes a runtime crash path (missing shader variant), but there doesn’t appear to be a Vulkan test that exercises float→int32 view/copy conversion. Adding a small unit/integration test that runs a minimal graph/model requiring float→int32 conversion and validates output (and/or that the shader is present in the registry) would help prevent regressions.

[mohammed-saalim]

@pytorchbot label "release notes: vulkan"