Summary

Polish pass on the Vulkan-compute PyTorch backend. Focus: a real reproducible-build bug in the shader toolchain, removing hardcoded/leaked developer paths, and adding CI for what can be checked without a GPU.

Changes

Shader build (real bug, fixed + verified)

• csrc/shaders/compile.sh used glslangValidator -V, which emits SPIR-V 1.0. The 12 subgroup-using quantized kernels (matmul_q4k/q5k/q6k, their _batch variants, and matmul_gpuq4/5/6, plus argmax) use GL_KHR_shader_subgroup reductions that require SPIR-V >= 1.3, so the script could not rebuild them. Switched to --target-env vulkan1.1 (SPIR-V 1.3). Added set -euo pipefail and a non-zero exit if any shader fails.
• Verified locally with the system glslangValidator: all 38 shaders compile (12 of which failed before).

Hardcoded / leaked paths removed (portability)

• csrc/vulkan_engine.cpp hardcoded /home/raz/projects/torch-vulkan/csrc/shaders/. Removed it; the engine now falls back to the TORCH_VULKAN_SHADER_DIR env var.
• torch_vulkan/__init__.py now exports the resolved bundled-shader directory into TORCH_VULKAN_SHADER_DIR so the lazily-constructed VulkanEngine finds the same shaders the Kompute context uses.
• tests/test_algo_cache.py and tests/bench_layer.py hardcoded /home/raz/builds/pytorch-gfx1150 on sys.path. Replaced with an optional TORCH_VULKAN_PYTORCH_PATH env var.

Tests

• tests/test_mm.py: the old test_cpu_fallback claimed "relu isn't implemented" — but relu is wired to the Vulkan backend. Split into test_relu (real correctness check vs CPU) and test_unimplemented_op_falls_back_to_cpu (uses sign, which has no Vulkan impl, to actually exercise the boxed CPU fallback).

Lint / cleanup

• Removed dead imports (numpy, time in persistent_pipeline.py; sys in setup.py) and unused benchmark locals.
• Aligned the __init__.py module docstring with the README: .to("vulkan") is the supported tensor-creation path; torch.randn(..., device="vulkan") and .vulkan() are only partially wired.
• README quickstart now compiles shaders and uses cmake -S . -B build (the old cd build referenced a gitignored, non-existent dir).

CI (new)

• .github/workflows/ci.yml: compiles every shader with --target-env vulkan1.1 (guards the regression above) and runs ruff + py_compile.
• pyproject.toml: ruff config (E,F,I; E402 ignored because the package intentionally imports torch, then loads _C, then aliases into torch.vulkan).

Verified

• Shader compilation: all 38 .comp compile with the new flag (ran locally).
• Ruff: clean (ruff 0.15.15, ran locally on the full tree).
• py_compile: passes for all Python files.
• Committed .spv binaries were intentionally left untouched — recompiling them locally produced different bytes (different glslang build) that I can't validate against the target GPU, and the committed ones are the verified artifacts per the README.

UNVERIFIED (no toolchain on this machine)

• C++ extension build NOT run. cmake is not installed, the Kompute dependency is absent, and there is no Vulkan dev headers / GPU here. The C++ edits are minimal and mechanical (one #include <cstdlib>, a std::getenv fallback block, a comment-only change in torch_vulkan.cpp) but were not compiled.
• Runtime test suite (pytest tests/) NOT run — requires a Vulkan GPU + Kompute + the custom PyTorch build. No correctness/perf numbers were re-measured; none are claimed here.

TODOs left (not addressed — intent unclear / out of scope)

• The many "written but not wired" shaders (Q5_K/Q6_K dispatch, rope, rmsnorm, silu_gate, kv-cache attention, etc.) remain roadmap items with no host dispatch.
• mm_raw (raw VulkanEngine path) is implemented but still not registered to any aten op.

🤖 Generated with Claude Code