Files
lerobot/.github/workflows/model_profiling.yml
T
Pepijn a6dd28e8b4 fix(profiling): tolerate groot dep-install failure
groot's only policy-specific dependency is flash-attn, which has no
prebuilt wheel for torch 2.10 and requires nvcc to build from source.
The CI image is based on nvidia/cuda:12.4.1-base, which ships the
CUDA runtime but not the compiler toolkit, so the source build fails
with `/usr/local/cuda/bin/nvcc: No such file or directory`. The
repo's own pyproject.toml already carries a TODO acknowledging this:
gr00t needs bespoke flash-attn install steps.

Treat this as an environmental limitation rather than a regression:
dep-install failures for groot are logged via `::warning::` and skip
the policy without failing the job. Dep-install failures for any
other policy remain fatal, so real regressions still surface.

Made-with: Cursor
2026-04-16 21:15:14 +02:00

241 lines
9.4 KiB
YAML

# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# 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.
name: Model Profiling
on:
schedule:
- cron: "0 0 * * 0"
pull_request:
branches:
- main
- feat/libero-benchmark
paths:
- .github/workflows/model_profiling.yml
- profiling/model_profiling_specs.json
- scripts/ci/run_model_profiling.py
- src/lerobot/configs/train.py
- src/lerobot/scripts/lerobot_train.py
- src/lerobot/utils/profiling_utils.py
- tests/scripts/test_model_profiling.py
workflow_dispatch:
inputs:
git_ref:
description: Git ref to profile when no commit SHA is provided
required: false
type: string
default: main
git_commit:
description: Optional exact commit SHA to profile
required: false
type: string
default: ""
policies:
description: Optional comma-separated policy filter
required: false
type: string
default: ""
profile_mode:
description: Torch profiler mode
required: false
type: choice
options:
- trace
- summary
default: trace
publish_results:
description: Publish results to the profiling dataset when a Hub token is available
required: false
type: boolean
default: true
results_repo:
description: Dataset repo name or fully qualified repo id
required: false
type: string
default: model-profiling-history
permissions:
contents: read
concurrency:
group: ${{ github.workflow }}-${{ github.event_name }}-${{ github.event.inputs.git_commit || github.event.inputs.git_ref || github.ref_name || github.run_id }}
cancel-in-progress: true
jobs:
profile-models:
name: Weekly Model Profiling
runs-on:
group: aws-g6-4xlarge-plus
env:
HF_USER_TOKEN: ${{ secrets.LEROBOT_HF_USER }}
PROFILE_MODE: ${{ github.event_name == 'pull_request' && 'summary' || github.event.inputs.profile_mode || 'trace' }}
POLICY_FILTER: ${{ github.event_name == 'pull_request' && 'act,diffusion,pi0,pi05,smolvla,groot,xvla,wall_x' || github.event.inputs.policies || '' }}
RESULTS_REPO: ${{ github.event.inputs.results_repo || 'model-profiling-history' }}
SHOULD_PUBLISH: ${{ github.event_name == 'schedule' || (github.event_name == 'workflow_dispatch' && github.event.inputs.publish_results == 'true') }}
steps:
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
with:
persist-credentials: false
lfs: true
ref: ${{ github.event.pull_request.head.sha || github.event.inputs.git_commit || github.event.inputs.git_ref || 'main' }}
- name: Pull GPU image
run: docker pull huggingface/lerobot-gpu:latest
- name: Run model profiling
env:
HOST_GIT_COMMIT: ${{ github.event.pull_request.head.sha || github.event.inputs.git_commit || github.sha }}
PROFILE_GIT_REF: ${{ github.head_ref || github.ref_name || github.event.inputs.git_ref || 'main' }}
PROFILE_PR_NUMBER: ${{ github.event.pull_request.number || '' }}
run: |
set -eux
mkdir -p profiling-results
docker run --rm --gpus all \
--user "$(id -u):$(id -g)" \
--shm-size=16g \
-e HOME=/tmp/lerobot-home \
-e HF_HOME=/tmp/hf \
-e HF_LEROBOT_HOME=/tmp/hf-lerobot \
-e TORCH_HOME=/tmp/torch-home \
-e TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor-cache \
-e UV_PROJECT_ENVIRONMENT=/tmp/lerobot-venv \
-e UV_CACHE_DIR=/tmp/uv-cache \
-e UV_PYTHON_PREFERENCE=only-system \
-e XDG_DATA_HOME=/tmp/xdg-data \
-e XDG_CACHE_HOME=/tmp/xdg-cache \
-e HOST_GIT_COMMIT="${HOST_GIT_COMMIT}" \
-e PROFILE_GIT_REF="${PROFILE_GIT_REF}" \
-e PROFILE_PR_NUMBER="${PROFILE_PR_NUMBER}" \
-e HF_USER_TOKEN="${HF_USER_TOKEN}" \
-e HF_TOKEN="${HF_USER_TOKEN}" \
-e PROFILE_MODE="${PROFILE_MODE}" \
-e POLICY_FILTER="${POLICY_FILTER}" \
-e RESULTS_REPO="${RESULTS_REPO}" \
-e SHOULD_PUBLISH="${SHOULD_PUBLISH}" \
-v "${GITHUB_WORKSPACE}:/workspace" \
-w /workspace \
huggingface/lerobot-gpu:latest \
bash -c '
set -euxo pipefail
mkdir -p "${HOME}" "${HF_HOME}" "${HF_LEROBOT_HOME}" "${TORCH_HOME}" "${UV_CACHE_DIR}" "${XDG_CACHE_HOME}" "${XDG_DATA_HOME}" "${TORCHINDUCTOR_CACHE_DIR}"
rm -rf /tmp/lerobot-src
cp -a /workspace/. /tmp/lerobot-src
cd /tmp/lerobot-src
if [[ -n "${HF_USER_TOKEN:-}" ]]; then
hf auth login --token "${HF_USER_TOKEN}" --add-to-git-credential 2>/dev/null || true
fi
policies_to_run=()
if [[ -n "${POLICY_FILTER}" ]]; then
IFS="," read -ra policies_to_run <<< "${POLICY_FILTER}"
else
policies_to_run=(act diffusion groot multi_task_dit pi0 pi0_fast pi05 smolvla wall_x xvla)
fi
policy_extras() {
case "$1" in
act) ;;
diffusion) echo "diffusion" ;;
groot) echo "groot" ;;
multi_task_dit) echo "multi_task_dit" ;;
pi0|pi0_fast|pi05) echo "pi" ;;
smolvla) echo "smolvla" ;;
wall_x) echo "wallx" ;;
xvla) echo "xvla" ;;
*)
echo "Unknown profiling policy $1" >&2
return 1
;;
esac
}
# Policies whose dep-install may fail due to environment constraints
# (e.g. groot requires compiling flash-attn, which needs nvcc; the CI
# image only ships the CUDA runtime). Install failures for these are
# logged as warnings and do not fail the job. See the TODO next to
# `lerobot[groot]` in pyproject.toml.
is_install_failure_tolerated() {
case "$1" in
groot) return 0 ;;
*) return 1 ;;
esac
}
overall_status=0
for raw_policy in "${policies_to_run[@]}"; do
policy="$(echo "${raw_policy}" | xargs)"
[[ -z "${policy}" ]] && continue
echo "::group::Profile ${policy}"
extra="$(policy_extras "${policy}")" || { overall_status=1; echo "::endgroup::"; continue; }
# Fresh, isolated dependency resolution per policy so that
# incompatible extras (e.g. flash-attn for groot) never block
# the rest of the matrix.
sync_cmd=(uv sync --locked --extra training --extra test)
if [[ -n "${extra}" ]]; then
sync_cmd+=(--extra "${extra}")
fi
# flash-attn does not declare torch as a build-time dep, so its
# isolated build env fails with ModuleNotFoundError. Torch is a
# core lerobot dep and is already resolved here, so we disable
# build isolation for flash-attn specifically.
sync_cmd+=(--no-build-isolation-package flash-attn)
if ! "${sync_cmd[@]}"; then
if is_install_failure_tolerated "${policy}"; then
echo "::warning::Dependency install failed for ${policy} (known-fragile); skipping."
else
echo "Dependency install failed for ${policy}; skipping." >&2
overall_status=1
fi
echo "::endgroup::"
continue
fi
cmd=(
uv run python scripts/ci/run_model_profiling.py
--output_dir=/workspace/profiling-results
--hub_org=lerobot
--results_repo="${RESULTS_REPO}"
--profile_mode="${PROFILE_MODE}"
--git_commit="${HOST_GIT_COMMIT}"
--git_ref="${PROFILE_GIT_REF}"
--pr_number="${PROFILE_PR_NUMBER}"
--policies "${policy}"
)
if [[ "${SHOULD_PUBLISH}" == "true" && -n "${HF_USER_TOKEN:-}" ]]; then
cmd+=(--publish)
fi
if ! "${cmd[@]}"; then
echo "Profiling failed for ${policy}." >&2
overall_status=1
fi
echo "::endgroup::"
done
exit "${overall_status}"
'
- name: Upload profiling artifacts
if: always()
uses: actions/upload-artifact@v4 # zizmor: ignore[unpinned-uses]
with:
name: model-profiling-results
path: profiling-results
if-no-files-found: warn