Release: sync from internal （clean run) #308

Workflow file for this run

.github/workflows/gpu-tests.yml at 92c842f

	# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
	# SPDX-License-Identifier: OpenMDW-1.1

	# Cosmos3-Nano GPU test suite on a self-hosted 8×H200 runner.
	#
	# A single ``pre-commit`` (lint) job runs first; the five GPU jobs all
	# ``needs:`` it, so they wait on ONE pre-commit run and are skipped if lint
	# fails — the single self-hosted runner is never spent on a lint-failing commit.
	# The five GPU jobs then run (one at a time on the single runner):
	# * training-smoke — Nano SFT pipeline (convert -> train 5 -> export -> t2i)
	# * generator-training-regression — vision_sft_nano loss vs goldens (4-GPU subset)
	# * generator-inference-smoke — Nano multi-modality inference (t2vs + policy + forward_dynamics)
	# * reasoner-inference-smoke — Nano reasoner inference first-token logits vs golden (image-conditioned, 4-GPU)
	# * reasoner-training-regression — llava_ov loss vs goldens (4-GPU subset)
	#
	# Requires:
	# * a self-hosted runner labelled [self-hosted, gpu, h200] with 8 GPUs,
	# NVIDIA drivers, and `uv` on PATH;
	# * an `HF_TOKEN` repository secret (gated dataset/model downloads, incl. the
	# streamed LLaVA-OneVision-Data dataset).
	#
	# Inputs/checkpoints download to examples/ + the HF cache and are reused across
	# runs (the h100 goldens are reused on H200 — see _detect_arch).
	name: GPU Tests

	on:
	push:
	branches: [main]
	pull_request:
	branches: [main]

	concurrency:
	group: gpu-tests-${{ github.ref }}
	cancel-in-progress: true

	jobs:
	# Single lint gate: runs once on ubuntu-latest; every GPU job below waits on it
	# and is skipped if it fails.
	pre-commit:
	uses: ./.github/workflows/pre-commit.yml

	training-smoke:
	needs: pre-commit
	runs-on: [self-hosted, gpu, h200]
	timeout-minutes: 90
	env:
	HF_TOKEN: ${{ secrets.HF_TOKEN }}
	HF_HUB_DISABLE_XET: "1"
	steps:
	- uses: actions/checkout@v6

	- uses: astral-sh/setup-uv@v7

	- name: Sync environment (cu128-train)
	run: uv sync --all-extras --group=cu128-train

	# Full SFT pipeline: download + convert Nano->DCP, train 5 steps (loss
	# trend), export to HF safetensors, then a t2i generation from the export.
	# MAX_GPUS defaults to 8. -s streams the live process log.
	- name: Nano SFT pipeline smoke (convert -> train 5 -> export -> t2i, 8 GPU)
	run: \|
	export LD_LIBRARY_PATH=
	uv run --all-extras --group=cu128-train python -m pytest -v -s \
	tests/nano_training_smoke_test.py --num-gpus=8 --levels=2 -o addopts=

	# Clear the heavy artifacts (even on failure): examples/checkpoints (the
	# Cosmos3-Nano DCP + Wan VAE, ~30 GB) and the pytest tmp dirs (the SFT
	# checkpoint + logs). The small examples/data dataset and the HF cache are
	# intentionally kept so subsequent runs reuse them.
	- name: Clean up run outputs
	if: always()
	run: \|
	rm -rf examples/checkpoints \|\| true
	rm -rf "${TMPDIR:-/tmp}"/pytest-of-* /tmp/pytest-of-* \|\| true

	generator-training-regression:
	needs: pre-commit
	runs-on: [self-hosted, gpu, h200]
	timeout-minutes: 60
	env:
	HF_TOKEN: ${{ secrets.HF_TOKEN }}
	HF_HUB_DISABLE_XET: "1"
	# Select the 4-GPU regression test variant (uses 4 of the 8 GPUs).
	TEST_MAX_GPUS: "4"
	steps:
	- uses: actions/checkout@v6

	- uses: astral-sh/setup-uv@v7

	- name: Sync environment (cu128-train)
	run: uv sync --all-extras --group=cu128-train

	# Generator (vision_sft_nano) loss vs the h100 goldens. -s streams the live log.
	- name: Generator regression (vision_sft_nano, 4-GPU subset)
	run: \|
	export LD_LIBRARY_PATH=
	uv run --all-extras --group=cu128-train python -m pytest -v -s \
	tests/launch_regression_test.py -k vision_sft_nano \
	--num-gpus=4 --levels=2 -o addopts=

	# The h100_inputs fixture removes its DCP stage on teardown; clear the
	# pytest tmp dirs too (logs + any run output). The HF cache is kept.
	- name: Clean up run outputs
	if: always()
	run: \|
	rm -rf "${TMPDIR:-/tmp}"/pytest-of-* /tmp/pytest-of-* \|\| true

	generator-inference-smoke:
	needs: pre-commit
	runs-on: [self-hosted, gpu, h200]
	timeout-minutes: 60
	env:
	HF_TOKEN: ${{ secrets.HF_TOKEN }}
	HF_HUB_DISABLE_XET: "1"
	steps:
	- uses: actions/checkout@v6

	- uses: astral-sh/setup-uv@v7

	- name: Sync environment (cu128-train)
	run: uv sync --all-extras --group=cu128-train

	# One inference call over t2vs (+sound), action policy, and forward_dynamics; checks each output.
	# MAX_GPUS defaults to 8. -s streams the live process log.
	# Reuse the same input-asset cache dir as the unittest job.
	- name: Nano inference smoke (t2vs + action policy + forward_dynamics, 8 GPU)
	run: \|
	export LD_LIBRARY_PATH=
	export COSMOS_DOWNLOAD_CACHE_DIR="$RUNNER_WORKSPACE/cosmos_input_cache"
	uv run --all-extras --group=cu128-train python -m pytest -v -s \
	tests/nano_inference_smoke_test.py --num-gpus=8 --levels=2 -o addopts=

	# Inference writes only the pytest tmp dir (the t2vs video + logs); the
	# checkpoint download stays in the HF cache (kept). No examples/ artifacts.
	- name: Clean up run outputs
	if: always()
	run: \|
	rm -rf "${TMPDIR:-/tmp}"/pytest-of-* /tmp/pytest-of-* \|\| true

	reasoner-inference-smoke:
	needs: pre-commit
	runs-on: [self-hosted, gpu, h200]
	timeout-minutes: 60
	env:
	HF_TOKEN: ${{ secrets.HF_TOKEN }}
	HF_HUB_DISABLE_XET: "1"
	# 4-GPU reasoner inference test (uses 4 of the 8 GPUs); the gpus(4) marker
	# requires MAX_GPUS == 4.
	TEST_MAX_GPUS: "4"
	steps:
	- uses: actions/checkout@v6

	- uses: astral-sh/setup-uv@v7

	- name: Sync environment (cu128-train)
	run: uv sync --all-extras --group=cu128-train

	# Image-conditioned Cosmos3-Nano reasoner inference: compare the first
	# decoded token's logits against a committed golden (exact argmax +
	# allclose rtol/atol=1e-3). -s streams the live process log. The text-only
	# variant lives in the same file but is not exercised in CI. Cache the
	# downloaded image asset (reasoner_image.json's remote vision_path) in the
	# persistent input-asset dir shared with the generator-inference-smoke job.
	- name: Nano reasoner inference (image-conditioned first-token logits golden, 4 GPU)
	run: \|
	export LD_LIBRARY_PATH=
	export COSMOS_DOWNLOAD_CACHE_DIR="$RUNNER_WORKSPACE/cosmos_input_cache"
	uv run --all-extras --group=cu128-train python -m pytest -v -s \
	tests/nano_reasoner_inference_smoke_test.py::test_nano_reasoner_image_first_token_logits \
	--num-gpus=4 --levels=2 -o addopts=

	# Reasoner inference writes only the pytest tmp dir (generated text + logs);
	# the checkpoint download stays in the HF cache (kept).
	- name: Clean up run outputs
	if: always()
	run: \|
	rm -rf "${TMPDIR:-/tmp}"/pytest-of-* /tmp/pytest-of-* \|\| true

	reasoner-training-regression:
	needs: pre-commit
	runs-on: [self-hosted, gpu, h200]
	timeout-minutes: 60
	env:
	HF_TOKEN: ${{ secrets.HF_TOKEN }}
	HF_HUB_DISABLE_XET: "1"
	# Select the 4-GPU regression test variant (uses 4 of the 8 GPUs).
	TEST_MAX_GPUS: "4"
	steps:
	- uses: actions/checkout@v6

	- uses: astral-sh/setup-uv@v7

	- name: Sync environment (cu128-train)
	run: uv sync --all-extras --group=cu128-train

	# Reasoner (llava_ov) loss vs the h100 goldens. -s streams the live log.
	- name: Reasoner regression (llava_ov, 4-GPU subset)
	run: \|
	export LD_LIBRARY_PATH=
	uv run --all-extras --group=cu128-train python -m pytest -v -s \
	tests/launch_regression_test.py -k llava_ov \
	--num-gpus=4 --levels=2 -o addopts=

	# The h100_inputs fixture removes its DCP stage on teardown; clear the
	# pytest tmp dirs too (logs + any run output). The HF cache is kept.
	- name: Clean up run outputs
	if: always()
	run: \|
	rm -rf "${TMPDIR:-/tmp}"/pytest-of-* /tmp/pytest-of-* \|\| true

	# Co-located unit tests: every *_test.py under cosmos_framework/ (CPU and GPU
	# together) in one pytest invocation, plus two torchrun steps for the
	# distributed tests that hardcode their world size. Runs parallel to the five
	# jobs above (all gated on the single pre-commit lint).
	unittest:
	needs: pre-commit
	runs-on: [self-hosted, gpu, h200]
	# 60 (not 30) so the first cold-cache run can download the Cosmos3-Nano
	# checkpoint for scripts/_test's convert->export->inference pipeline; steady
	# state (warm HF cache) is a few minutes.
	timeout-minutes: 60
	env:
	HF_TOKEN: ${{ secrets.HF_TOKEN }}
	HF_HUB_DISABLE_XET: "1"
	steps:
	- uses: actions/checkout@v6

	- uses: astral-sh/setup-uv@v7

	- name: Sync environment (cu128-train)
	run: uv sync --all-extras --group=cu128-train

	# Run the whole co-located suite (CPU + GPU). Tests that load internal
	# pretrained weights from S3 skip themselves when credentials/pretrained.secret
	# is absent (via RunIf / pytest.skip guards), so this is green without
	# internal credentials; provide the credential file on the runner to
	# exercise them. New tests are picked up automatically (no markers/lists).
	# Cache downloaded input assets in a persistent dir (outside the repo tree,
	# so the cleanup step keeps it) and reuse it across runs.
	- name: Unit tests
	run: \|
	export LD_LIBRARY_PATH=
	export COSMOS_DOWNLOAD_CACHE_DIR="$RUNNER_WORKSPACE/cosmos_input_cache"
	uv run --all-extras --group=cu128-train python -m pytest -v -s \
	cosmos_framework/ -o addopts=

	# The cfgp_ar / context_parallel tests call dist.init_process_group and
	# skip under plain pytest (world_size 1); they must be launched with
	# torchrun. They hardcode their world size via ParallelDims, so each file
	# needs the matching --nproc_per_node and they cannot share one launch:
	# * cfgp_ar -> cfgp=2, dp_shard=1 => world_size must be 2
	# * context_parallel -> cp=4 (and cp=world_size) => world_size must be 4
	# Over-provisioning (e.g. 8) makes ParallelDims' product != world_size and
	# fails, so these sizes are fixed, not "the more GPUs the better".
	- name: Distributed unit tests - cfgp_ar (torchrun, 2 ranks)
	run: \|
	export LD_LIBRARY_PATH=
	uv run --all-extras --group=cu128-train torchrun --nproc_per_node=2 -m pytest -v \
	cosmos_framework/model/generator/mot/cfgp_ar_test.py -o addopts=

	- name: Distributed unit tests - context_parallel (torchrun, 4 ranks)
	run: \|
	export LD_LIBRARY_PATH=
	uv run --all-extras --group=cu128-train torchrun --nproc_per_node=4 -m pytest -v \
	cosmos_framework/model/generator/mot/context_parallel_test.py -o addopts=

	# Clear everything the suite writes into the working tree (all gitignored
	# scratch): pytest tmp dirs (DCP checkpoint, logs), the script-test
	# `outputs/` dir, any `examples/checkpoints`, and the `schemas/` dir from
	# export_schemas_test. The HF cache lives outside the tree and is kept for
	# reuse across runs. Runs on success or failure.
	- name: Clean up run outputs
	if: always()
	run: \|
	rm -rf "${TMPDIR:-/tmp}"/pytest-of-* /tmp/pytest-of-* \
	outputs examples/checkpoints schemas \|\| true

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Release: sync from internal （clean run) #308

Workflow file

Release: sync from internal （clean run) #308

Uh oh!

Workflow file for this run