alpasim
Use AlpaSim repository guidance for autonomous-driving simulation setup, Hydra wizard runs, runtime services, ego drivers and plugins, ASL evaluation, controller/physics/traffic components, gRPC contracts, and operational troubleshooting.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- alpasim
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use AlpaSim repository guidance for autonomous-driving simulation setup, Hydra wizard runs, runtime services, ego drivers and plugins, ASL evaluation, controller/physics/traffic components, gRPC contracts, and operational troubleshooting.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/coordinate-frames.md1.8 KB
- 参考文档references/package-and-backend-overview.md3.3 KB
- 参考文档references/repo-provenance.md2.9 KB
- 参考文档references/repo-routing-metadata.json428 B
- 参考文档references/troubleshooting.md3.9 KB
- 脚本scripts/check_env.py3.3 KB
03 SKILL.md 正文
AlpaSim
AlpaSim is a modular, data-driven autonomous-driving simulator. It orchestrates renderer, driver, controller, physics, traffic, runtime, and evaluation services through gRPC. Use this root as a router; read only the focused sub-skill needed for the task.
Route by intent
- Prepare or run a simulation, compose Hydra overrides, select scenes, Docker/Slurm deployment, tune GPUs/timing, or inspect telemetry: read simulation-wizard.
- Understand the event loop, service lifecycle, addresses, daemon mode, replay, timing validation, renderer caches, or video-model chunking: read runtime-services.
- Select or adapt an ego policy, manual/external driver, model input/output, camera rectification, or plugin entry points/config discovery: read drivers-and-plugins.
- Inspect ASL logs, extract frames, evaluate/aggregate/re-evaluate runs, interpret metrics/videos, or use trajectory/geometry utilities: read evaluation-and-logs.
- Choose MPC, inspect vehicle state, run ground physics, understand CATK traffic sessions/handover, or classify CUDA/PyG/Warp requirements: read control-physics-traffic.
- Change protobufs, compile generated stubs, use gRPC clients, inspect package entry points, or review safe developer/tool workflows: read grpc-and-developer-tools.
Cross-cutting details are in troubleshooting, package and backend notes, and coordinate frames. Run the read-only environment checker before optional backend work; the structured routing metadata is consumed by the specialized repo-skill importer. Check repository provenance before treating this graph as current for another checkout.
Installation orientation
AlpaSim is a uv workspace. The root project intentionally has no default
runtime dependencies; install only the workspace extra for the chosen workflow:
uv sync --extra wizard # wizard and its transitive core
uv sync --extra all # all core packages, no optional plugin
uv sync --extra all --extra transfuser # core plus the public Transfuser plugin
Use Python 3.11 for the core packages. alpasim_driver currently requires
Python 3.12, so use a compatible environment for driver/model inspection. A
full simulation additionally requires Docker Compose, NVIDIA Container
Toolkit/CUDA, scene artifacts, and often Hugging Face access; a package import
is not proof that those services or assets work.
After changing .proto definitions, compile stubs from the gRPC package with
uv run compile-protos; use the developer-tools route for the safe bundled
compiler and clean/build distinctions. Verify a selected environment with:
python -c "import alpasim_grpc, alpasim_utils, alpasim_runtime, alpasim_wizard"
python -c "from alpasim_grpc.v0 import runtime_pb2, runtime_pb2_grpc"
Operating boundaries
- Keep
HF_TOKEN, external endpoints, private paths, model weights, scene caches, and scheduler credentials out of prompts, generated YAML, logs, and skill content. Treat gated scene/model downloads as user-authorized actions. - Start with config generation and a one-scene/one-rollout plan. Prefer
wizard.run_method=NONEor a read-only checker before containers, network services, GPU inference, or scheduler submission. - Preserve resolved run files and the first causal service error. A rollout is
successful only when its
_completemarker and expected evaluation outputs exist; an ASL file alone is not completion. - Do not use CPU imports to claim CUDA renderer, Warp, CATK/PyG, or model inference coverage. The optional backend notes preserve those limits.
- Keep coordinate-frame names explicit:
local,rig,aabb,ecef, and the estimated/noised frame are not interchangeable. Read the coordinate reference before changing transforms or service messages.
Verification entry points
The generated graph is intended to support focused, reproducible checks rather than a default full simulation. Use the verification artifacts outside this runtime tree when reviewing coverage. Safe native candidates include wizard config/dispatch, runtime validation/replay, controller, evaluation/ASL, utilities, plugin registry, and protobuf compilation tests. Full Docker, NuRec/FlashDreams, gated HF assets, model inference, CATK compiled extensions, Slurm submission, and large benchmarks remain explicitly conditional.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).