feat: YAML-driven torch op codegen with canonical naming and exposed semantic params#595

Open

voltjia wants to merge 31 commits into

masterfrom

feat/torch-codegen

Commits on May 9, 2026

build: add PyYAML build dependency
voltjia
committed
refactor(swiglu): move `Sigmoid` helper to `detail::`
voltjia
committed
feat(operator): graceful handling of unknown device types
voltjia
committed
feat: add YAML-driven torch op codegen
voltjia
committed
build: integrate torch op codegen into CMake
voltjia
committed
feat(scripts): generate pybind bindings for generated torch ops
voltjia
committed
fix(scripts): order pybind overloads from specific to permissive
voltjia
committed
test: add data-driven coverage for generated torch ops
voltjia
committed
feat(scripts): drop overload-name suffix from generated class names
voltjia
committed
feat(scripts): store visible non-tensor params as base members
voltjia
committed
feat(scripts): expose default-valued non-optional params
voltjia
committed
fix(scripts): preserve `std::vector<int64_t>` params in pybind generation
voltjia
committed
fix(scripts): gate generated-base scan on `--with-torch`
voltjia
committed
fix(scripts): rename ATen `self` parameter to `input`
voltjia
committed
perf(scripts): gate torch op instantiations by active device
voltjia
committed
chore(scripts): drop `AUTO-GENERATED` header from emitted files
voltjia
committed
fix(scripts): pipe generated headers through clang-format
voltjia
committed
style(scripts): apply `ruff format` from latest
voltjia
committed
build: throttle torch source compilation via Ninja job pool
voltjia
committed
build: add `clang-format` to build-system requires
voltjia
committed
fix(scripts): pip-install `clang-format` if missing at codegen time
voltjia
committed
fix(scripts): make `clang-format` optional at codegen time
voltjia
committed
fix(build): break cyclic dep between `infiniops` and torch object lib
voltjia
committed
build: skip PyTorch auto-detection on Cambricon
voltjia
committed
test: skip overloads that the harness cannot drive cleanly
voltjia
committed
test: resolve dtype attributes lazily for older torch forks
voltjia
committed
style(tests): apply ruff format
voltjia
committed
test: skip `mode` op (`torch_musa` kernel hangs on MUSA)
voltjia
committed

Commits on May 10, 2026

build: probe system Python for `torch` when build-isolated env lacks it
voltjia
committed
feat: pin codegen schema to the locally installed torch version
voltjia
committed
fix(scripts): fall back to latest stable when torch version tag is unreachable
voltjia
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: YAML-driven torch op codegen with canonical naming and exposed semantic params#595

feat: YAML-driven torch op codegen with canonical naming and exposed semantic params#595
voltjia wants to merge 31 commits into
masterfrom
feat/torch-codegen

Commits on May 9, 2026

build: add PyYAML build dependency

refactor(swiglu): move `Sigmoid` helper to `detail::`

feat(operator): graceful handling of unknown device types

feat: add YAML-driven torch op codegen

build: integrate torch op codegen into CMake

feat(scripts): generate pybind bindings for generated torch ops

fix(scripts): order pybind overloads from specific to permissive

test: add data-driven coverage for generated torch ops

feat(scripts): drop overload-name suffix from generated class names

feat(scripts): store visible non-tensor params as base members

feat(scripts): expose default-valued non-optional params

fix(scripts): preserve `std::vector<int64_t>` params in pybind generation

fix(scripts): gate generated-base scan on `--with-torch`

fix(scripts): rename ATen `self` parameter to `input`

perf(scripts): gate torch op instantiations by active device

chore(scripts): drop `AUTO-GENERATED` header from emitted files

fix(scripts): pipe generated headers through clang-format

style(scripts): apply `ruff format` from latest

build: throttle torch source compilation via Ninja job pool

build: add `clang-format` to build-system requires

fix(scripts): pip-install `clang-format` if missing at codegen time

fix(scripts): make `clang-format` optional at codegen time

fix(build): break cyclic dep between `infiniops` and torch object lib

build: skip PyTorch auto-detection on Cambricon

test: skip overloads that the harness cannot drive cleanly

test: resolve dtype attributes lazily for older torch forks

style(tests): apply ruff format

test: skip `mode` op (`torch_musa` kernel hangs on MUSA)

Commits on May 10, 2026

build: probe system Python for `torch` when build-isolated env lacks it

feat: pin codegen schema to the locally installed torch version

fix(scripts): fall back to latest stable when torch version tag is unreachable