Repository navigation

docs(kernels): add kernel framework docs and the LLVM interop design #57

Workflow file for this run

.github/workflows/llm-deploy.yml at 3f46529

	name: LLM Deploy v1.0

	# CI for the Qwen3-0.6B-on-RISC-V project (docs/llm-deploy-v1.0/).
	#
	# Separate from ci.yml on purpose: that one tests the ScratchV compiler, this
	# one tests the deployment project. A failure here should not read as "the
	# compiler is broken".
	#
	# The PR job executes small numeric gates. A separate manual/nightly job
	# downloads and validates the pinned full-model release with ORT.
	#
	# docs:interfaces docs:risks
	# probe:qemu-matmul probe:small-transformer probe:qwen3-onnx
	# numeric:ir-qwen3-small
	#
	# Later weeks' gates (numeric:ir-full-qwen3 at W3, numeric:qemu-full-qwen3 at
	# W4, e2e:qemu-paris at W5, release:gate at W8) are NOT stubbed here. They get
	# added when their week arrives, from the plan's §4.3 table.

	on:
	pull_request:
	paths:
	- 'docs/llm-deploy-v1.0/**'
	- 'probes/**'
	- 'scripts/check_llm_deploy_docs.py'
	- 'scripts/check_w1_repro_env.py'
	- 'export_qwen3_onnx.py'
	- 'scratchv/**'
	- 'requirements/**'
	- 'tests/test_qwen3_small*.py'
	- 'tests/test_tensor*.py'
	- 'tests/test_riscv_tensor*.py'
	- 'tests/test_optimizer*.py'
	- 'tests/qwen_artifacts.py'
	- 'tests/test_qwen_artifact_selection.py'
	- 'tests/test_w1_qwen3*.py'
	- 'tests/test_w1_repro_env.py'
	- '.github/workflows/llm-deploy.yml'
	schedule:
	# Nightly, off the hour: 19:23 UTC = 03:23 北京时间.
	- cron: '23 19 * * *'
	workflow_dispatch:

	permissions:
	contents: read

	env:
	# Delivered numeric gates are mandatory: missing scripts/tools are failures.
	PROBE_MATMUL: probes/w1_matmul_4x4/run.py
	PROBE_TRANSFORMER: probes/w1_tiny_transformer/run.py
	PROBE_QWEN3_EXPORT: probes/w1_qwen3_export/run.py
	PROBE_QWEN3_SMALL: probes/w2_qwen3_small/run.py
	QWEN3_PY: output/qwen3-probe-venv/bin/python

	jobs:
	llm-deploy:
	name: "LLM Deploy v1.0 / W1"
	runs-on: ubuntu-latest
	timeout-minutes: 60
	steps:
	- uses: actions/checkout@v4
	with:
	lfs: true

	- uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
	with:
	python-version: "3.12"

	# ── 前置：RISC-V 工具链 + Python 依赖 ──────────────────────────────
	- name: Install RISC-V toolchain and QEMU
	run: \|
	set -euo pipefail
	if ! command -v clang >/dev/null \|\| ! command -v ld.lld >/dev/null \|\| ! command -v qemu-riscv32 >/dev/null \|\| ! command -v qemu-system-riscv64 >/dev/null; then
	sudo apt-get update
	sudo apt-get install --no-install-recommends -y clang lld qemu-user qemu-system-misc
	fi
	missing=""
	for tool in clang ld.lld qemu-riscv32 qemu-system-riscv64; do
	command -v "$tool" >/dev/null 2>&1 \|\| missing="$missing $tool"
	done
	if [ -n "$missing" ]; then
	echo "::error::missing tools:$missing"
	exit 1
	fi
	clang --version \| head -1
	qemu-riscv32 --version \| head -1
	qemu-system-riscv64 --version \| head -1

	- name: Install Python dependencies
	run: \|
	python3 -m pip install --upgrade pip
	python3 -m pip install -e ".[all]"
	python3 -m pip install numpy "pytest==9.1.1"
	python3 -m pip install "ziglang==0.14.1"
	echo "SCRATCHV_CC=$(python3 -c 'from pathlib import Path; import ziglang; print(Path(ziglang.__file__).parent / "zig")')" >> "$GITHUB_ENV"

	# ── docs:interfaces / docs:risks ──────────────────────────────────
	- name: "docs:interfaces + docs:risks"
	id: docs
	run: python3 scripts/check_llm_deploy_docs.py --week W1

	# ── probe:qemu-matmul ─────────────────────────────────────────────
	- name: "probe:qemu-matmul"
	id: matmul
	run: \|
	set -euo pipefail
	python3 "$PROBE_MATMUL"

	- name: Tensor compiler and runtime regressions
	run: \|
	python3 -m pytest tests/test_tensor_c_codegen.py tests/test_tensor_compiler.py tests/test_riscv_tensor_runtime.py tests/test_optimizer_numeric_semantics.py tests/test_qwen_artifact_selection.py tests/test_w1_qwen3_export.py tests/test_qwen3_small_riscv_gate.py tests/test_w1_repro_env.py -q

	- name: Repeated Linux timeout cleanup regression
	timeout-minutes: 5
	run: \|
	set -euo pipefail
	mkdir -p output/process-cleanup-repeats
	for attempt in $(seq 1 20); do
	python3 -B -m pytest \
	tests/test_riscv_tensor_runtime.py::test_timeout_terminates_only_owned_process_tree \
	-q -p no:cacheprovider \
	--basetemp "output/process-cleanup-repeats/temp-$attempt" \
	--junit-xml="output/process-cleanup-repeats/attempt-$attempt.xml"
	done

	# ── probe:small-transformer ───────────────────────────────────────
	- name: "probe:small-transformer"
	id: transformer
	run: \|
	set -euo pipefail
	python3 "$PROBE_TRANSFORMER" --model probes/w1_tiny_transformer/out/tiny_transformer_2l.onnx

	# The real architecture probe uses random weights, needs no checkpoint
	# download, and runs on PRs. Keep its export stack isolated from W1 and
	# the generic compiler tests. Install the CPU wheel before requirements
	# so pip never resolves torch from PyPI with CUDA dependencies.
	- name: Install pinned CPU Qwen3 probe environment
	run: \|
	set -euo pipefail
	python3 -m venv output/qwen3-probe-venv
	"$QWEN3_PY" -m pip install --upgrade pip
	"$QWEN3_PY" -m pip install "torch==2.7.1" --index-url https://download.pytorch.org/whl/cpu
	"$QWEN3_PY" -m pip install -r requirements/qwen3-small-probe.txt
	"$QWEN3_PY" -m pip install --no-deps -e .
	"$QWEN3_PY" -m pip check
	"$QWEN3_PY" -c "import torch; assert torch.__version__ == '2.7.1+cpu' and torch.version.cuda is None, torch.__version__"

	- name: W1 reproduction environment preflight
	run: \|
	set -euo pipefail
	"$QWEN3_PY" -X utf8 scripts/check_w1_repro_env.py \
	--require-clean --output-dir output/w1-repro-env

	- name: "numeric:ir-qwen3-small"
	id: qwen3_small
	timeout-minutes: 15
	env:
	PYTHONHASHSEED: "0"
	OMP_NUM_THREADS: "1"
	OPENBLAS_NUM_THREADS: "1"
	MKL_NUM_THREADS: "1"
	HF_HUB_OFFLINE: "1"
	TRANSFORMERS_OFFLINE: "1"
	run: \|
	set -euo pipefail
	"$QWEN3_PY" "$PROBE_QWEN3_SMALL" --output-dir output/qwen3-small

	- name: Explicit Qwen3 artifact optimization integration
	id: qwen3_artifact_tests
	env:
	SCRATCHV_QWEN_ARTIFACT_DIR: output/qwen3-small
	OMP_NUM_THREADS: "1"
	OPENBLAS_NUM_THREADS: "1"
	MKL_NUM_THREADS: "1"
	run: \|
	set -euo pipefail
	"$QWEN3_PY" -B -m pytest \
	tests/test_optimizer_numeric_semantics.py::test_real_qwen_artifact_optimization_matches_ort \
	-q -rs -o junit_family=xunit1 --junit-xml=output/qwen3-small/optimization-tests.xml
	"$QWEN3_PY" - <<'PY'
	import xml.etree.ElementTree as ET
	cases = ET.parse('output/qwen3-small/optimization-tests.xml').getroot().findall('.//testcase')
	expected = {f'test_real_qwen_artifact_optimization_matches_ort[{level}]' for level in ('none', 'basic', 'all')}
	assert len(cases) == 3 and {case.get('name') for case in cases} == expected
	assert not any(case.find(tag) is not None for case in cases for tag in ('skipped', 'failure', 'error'))
	print('Qwen artifact integration: 3 passed, 0 skipped')
	PY

	- name: Qwen3 small probe regressions
	id: qwen3_small_tests
	env:
	HF_HUB_OFFLINE: "1"
	TRANSFORMERS_OFFLINE: "1"
	run: \|
	set -euo pipefail
	"$QWEN3_PY" -m pytest \
	tests/test_qwen3_small_probe.py \
	tests/test_qwen3_small_model.py \
	tests/test_qwen3_small_gate.py -q \
	--junit-xml=output/qwen3-small/tests.xml

	- name: "numeric:qemu-qwen3-small"
	id: qwen3_riscv
	timeout-minutes: 25
	env:
	OMP_NUM_THREADS: "1"
	OPENBLAS_NUM_THREADS: "1"
	run: \|
	set -euo pipefail
	"$QWEN3_PY" probes/w2_qwen3_small/riscv.py \
	--model-dir output/qwen3-small --output-dir output/qwen3-riscv

	- name: Upload RISC-V tensor evidence
	if: always()
	uses: actions/upload-artifact@v4
	with:
	name: riscv-tensor-probe-output
	path: \|
	output/qemu-matmul/
	output/qwen3-riscv/
	output/process-cleanup-repeats/*.xml
	if-no-files-found: ignore

	- name: Summarize RISC-V numeric gate
	if: always()
	env:
	PROBE_OUTCOME: ${{ steps.qwen3_riscv.outcome }}
	run: \|
	{
	echo "## numeric:qemu-qwen3-small"
	echo "Outcome: $PROBE_OUTCOME"
	if [ -f output/qwen3-riscv/report.md ]; then
	cat output/qwen3-riscv/report.md
	fi
	} >> "$GITHUB_STEP_SUMMARY"

	- name: Summarize Qwen3 small probe
	if: always()
	env:
	PROBE_OUTCOME: ${{ steps.qwen3_small.outcome }}
	TEST_OUTCOME: ${{ steps.qwen3_small_tests.outcome }}
	ARTIFACT_TEST_OUTCOME: ${{ steps.qwen3_artifact_tests.outcome }}
	run: \|
	{
	echo "## numeric:ir-qwen3-small"
	echo "Probe: $PROBE_OUTCOME; regression tests: $TEST_OUTCOME"
	echo "Explicit artifact integration (requires 3 passed / 0 skipped): $ARTIFACT_TEST_OUTCOME"
	echo ""
	if [ -f output/qwen3-small/report.md ]; then
	cat output/qwen3-small/report.md
	else
	echo "No report was produced; inspect the dependency/probe step logs."
	fi
	} >> "$GITHUB_STEP_SUMMARY"

	- name: Upload Qwen3 small probe artifacts
	if: always()
	uses: actions/upload-artifact@v4
	with:
	name: qwen3-small-probe-output
	path: output/qwen3-small/
	if-no-files-found: ignore

	- name: Upload W1 reproduction environment
	if: always()
	uses: actions/upload-artifact@v4
	with:
	name: w1-reproduction-environment
	path: output/w1-repro-env/
	if-no-files-found: ignore

	- name: Upload probe artifacts
	if: always()
	uses: actions/upload-artifact@v4
	with:
	name: w1-probe-output
	path: probes/**/out/
	if-no-files-found: ignore

	# ── 状态汇总 ──────────────────────────────────────────────────────
	- name: W1 交付物状态
	if: always()
	env:
	DOCS_OUTCOME: ${{ steps.docs.outcome }}
	MATMUL_OUTCOME: ${{ steps.matmul.outcome }}
	TRANSFORMER_OUTCOME: ${{ steps.transformer.outcome }}
	QWEN_IR_OUTCOME: ${{ steps.qwen3_small.outcome }}
	QWEN_RISCV_OUTCOME: ${{ steps.qwen3_riscv.outcome }}
	run: \|
	{
	echo "## W1 / small Qwen3 execution status"
	echo ""
	echo "\| Gate \| Actual step outcome \|"
	echo "\|---\|---\|"
	echo "\| Document structure \| $DOCS_OUTCOME \|"
	echo "\| QEMU MatMul \| $MATMUL_OUTCOME \|"
	echo "\| Synthetic Transformer IR \| $TRANSFORMER_OUTCOME \|"
	echo "\| Real Qwen3 structure IR \| $QWEN_IR_OUTCOME \|"
	echo "\| Real Qwen3 structure QEMU \| $QWEN_RISCV_OUTCOME \|"
	echo ""
	echo "Full-model ORT validation runs in the separate manual/nightly job."
	echo "Document checks do not certify team interface approval or second-person reproduction."
	} >> "$GITHUB_STEP_SUMMARY"

	full-qwen3-onnx:
	name: "W1 / full Qwen3 ONNX release validation"
	# This downloads 1.24 GB and executes the 2.38 GB FP32 model. It is not
	# counted as a passing gate on PRs where it did not execute.
	if: github.event_name != 'pull_request'
	runs-on: ubuntu-latest
	timeout-minutes: 60
	steps:
	- uses: actions/checkout@v4
	with:
	persist-credentials: false
	- uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
	with:
	python-version: "3.12"
	- name: Install pinned full-model validation dependencies
	run: \|
	python3 -m pip install "numpy==2.2.6" "onnx==1.18.0" "onnxruntime==1.22.1" "protobuf==5.29.5"
	python3 -m pip check
	- name: "probe:qwen3-onnx"
	id: full_model
	env:
	OMP_NUM_THREADS: "2"
	OPENBLAS_NUM_THREADS: "2"
	run: \|
	set -euo pipefail
	python3 "$PROBE_QWEN3_EXPORT" --mode download \
	--model-dir output/qwen3-full-model \
	--output-dir output/qwen3-full-probe --threads 2
	- name: Summarize full-model gate
	if: always()
	env:
	PROBE_OUTCOME: ${{ steps.full_model.outcome }}
	run: \|
	{
	echo "## probe:qwen3-onnx (complete FP32 release, ORT CPU)"
	echo "Outcome: $PROBE_OUTCOME"
	if [ -f output/qwen3-full-probe/report.md ]; then
	cat output/qwen3-full-probe/report.md
	else
	echo "No report produced; inspect the failed step."
	fi
	echo "This validates the complete ONNX artifact, not ScratchV full-model execution."
	} >> "$GITHUB_STEP_SUMMARY"
	- name: Upload full-model validation evidence
	if: always()
	uses: actions/upload-artifact@v4
	with:
	name: qwen3-full-onnx-probe-output
	path: output/qwen3-full-probe/
	if-no-files-found: ignore

Navigation Menu

Repository navigation

docs(kernels): add kernel framework docs and the LLVM interop design #57

Workflow file

docs(kernels): add kernel framework docs and the LLVM interop design #57

Uh oh!

Workflow file for this run