metax.yml

# Copyright 2026 FlagOS Contributors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

display_name: MetaX

setup_script: .github/scripts/set_env.sh
setup_platform: metax

ci_image: harbor.baai.ac.cn/flagos-dev/pytorch-plugin-fl:manual-20260731-metax-dev

runner_labels:
  - mx-8g-cicd-pytorch

# Platform-specific code selects this hardware lane on PRs.
# Shared or unrecognized code selects every lane conservatively.
paths_watch:
  - torch_fl/accelerator/metax/
  - torch_fl/configs/backends_metax.conf
  - csrc/aten/backends/metax/
  - csrc/runtime/accelerator/metax/
  - csrc/runtime/allocator/backends/metax_memory.h
  - csrc/runtime/allocator/backends/cuda_memory.h
  - .github/scripts/hooks/set_env_metax.sh
  - .github/scripts/set_env_metax.sh
  - scripts/vendor/bundle_maca_libtorch.sh

container_volumes: []

# Issue #253. The MetaX image exports TORCH_ALLOW_TF32_CUBLAS_OVERRIDE=1, which
# ATen reads while it initialises its TF32 state, so
# torch.backends.cuda.matmul.allow_tf32 is already True before user code runs
# and MACA's GEMM takes its TF32 kernel for an fp32 mm/bmm. That is the
# documented meaning of the variable and MACA honors it -- the kernel is not the
# defect -- but it makes every numeric assertion in this job a TF32 assertion:
# measured on the C550, an fp32 64x128 @ 128x32 mm lands 1.5e-02 from an fp64
# reference with the variable set and 1.3e-05 without.
#
# "0" is the explicit-off spelling, and it is what overrides an inherited 1. A
# non-numeric value such as "false" or an empty string raises a UserWarning and
# is treated as off, so both are load-bearing spellings and neither is used
# here; unset and 0 are numerically identical but unset cannot cancel the
# image's value.
#
# This is the environment, not set_env_metax.sh, because the setup script runs
# as its own step: each step gets a fresh shell, so an unset inside the script
# reaches nothing, and $GITHUB_ENV would only be a second copy of this line.
# Nothing in the build reads the flag. tests/integration/test_dtype_coverage.py
# pins the resulting precision, so a dropped pin fails a test -- and fails it on
# the step below that names test_dtype_coverage.py -- rather than passing
# silently.
integration_environment:
  TORCH_ALLOW_TF32_CUBLAS_OVERRIDE: "0"

container_options: >-
  --privileged
  --ipc=host
  --shm-size=512g
  --ulimit memlock=-1
  --cap-add=SYS_PTRACE

timeout_minutes: 60
# Release/manual wheel-build artifact; CI tests the wheel in the build job.
wheel_artifact_name: wheel-metax

# What this platform's environment has to look like before any test runs, read
# by the first step of .github/integration-manifest.yml through
# .github/scripts/check_platform_environment.py. The vocabulary is documented on
# that script; the values are this platform's, and they assert the same things
# the hand-written heredoc that used to sit here did.
#
# MetaX runs in an isolated /opt/venv with stock +cpu torch, and reads its
# devices through BOTH surfaces: torch.cuda (the MACA shim) and the flagos
# device. The FlagGems group cannot run without both halves -- the generated
# kernels call flag_gems.<name>, and the Triton that compiles them has to be the
# vendor one carrying the metax backend. `require_imports` is where that second
# half is stated: the import itself is the check.
environment_check:
  torch_version_prefix: "2.10.0+cpu"
  torch_path_prefix: /opt/venv/
  torch_path_forbidden:
    - /opt/conda/
  flagos_devices:
    exactly: 8
  cuda_available: true
  cuda_devices:
    exactly: 8
  triton_backend: metax
  require_imports:
    - flag_gems