inference
End-to-end Alpamayo R1 multimodal driving inference, sample generation, and troubleshooting.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- inference
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- End-to-end Alpamayo R1 multimodal driving inference, sample generation, and troubleshooting.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/api-reference.md4.0 KB
- 参考文档references/data-formats.md3.6 KB
- 参考文档references/troubleshooting.md4.7 KB
- 参考文档references/workflows.md4.3 KB
- 脚本scripts/run_inference_smoke.py7.7 KB
03 SKILL.md 正文
Alpamayo R1 Inference
Use this sub-skill when the user wants to run, inspect, or adapt the Alpamayo R1 inference path on a PhysicalAI-AV clip.
What this covers
- Load a gated PhysicalAI-AV clip and egomotion history/future.
- Build the multimodal chat prompt from stacked camera frames.
- Load
nvidia/Alpamayo-R1-10Bon CUDA with the Alpamayo tokenizer and processor. - Sample future trajectories plus Chain-of-Causation text traces.
- Compare predicted trajectories with ground truth and visualize them in notebook style.
- Troubleshoot HF gating, CUDA OOM, flash-attn / SDPA fallback, device placement, prompt rank, and early
t0_usfailures.
What this does not cover
- Training, SFT, RL post-training, or repo maintenance.
- Import/export plumbing or package provenance.
- Any workflow that depends on the original checkout at runtime.
Fast route
- Read
references/api-reference.mdfor the public call contract. - Read
references/data-formats.mdfor shapes, frames, and output semantics. - Follow
references/workflows.mdfor the end-to-end inference flow and notebook-style visualization. - If anything fails, open
references/troubleshooting.md. - Run
scripts/run_inference_smoke.pyas the bundled smoke test.
Key defaults
- Default model id:
nvidia/Alpamayo-R1-10B - Default loader history / future / camera settings: 16 steps, 64 steps, 4 cameras
- Default attention path:
flash_attention_2 - SDPA is a fallback for flash-attn incompatibility, not a replacement for the CUDA path.
See also
../../SKILL.mdfor the root router../../references/repo-provenance.md../../references/repo-routing-metadata.json../../references/troubleshooting.mdfor shared Alpamayo R1 troubleshootingreferences/api-reference.mdreferences/data-formats.mdreferences/workflows.mdreferences/troubleshooting.mdscripts/run_inference_smoke.py
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).