rpc-deployment
Guides TVM RPC tracker/server/proxy setup, cross-compilation, remote module upload/load/run, and deployment troubleshooting.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- rpc-deployment
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Guides TVM RPC tracker/server/proxy setup, cross-compilation, remote module upload/load/run, and deployment troubleshooting.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/cli-and-runtime.md2.1 KB
- 参考文档references/cross-compilation.md2.4 KB
- 参考文档references/troubleshooting.md1.9 KB
- 脚本scripts/rpc_cli_help_probe.py2.0 KB
03 SKILL.md 正文
TVM RPC and Deployment
Use this route when a task involves TVM RPC, cross-compilation, remote device execution, tracker/server/proxy commands, module upload/load, target triples, remote timing, or RPC-backed meta-schedule runners.
Route
- Confirm local TVM import and target compile support. Use
../install-build/SKILL.mdfor build/import failures. - Decide whether the task needs direct RPC server, tracker-managed devices, or
a proxy. Read
references/cli-and-runtime.md. - For cross-compilation and remote execution, read
references/cross-compilation.mdand keep target triple, device runtime, export artifact, upload path, and run device explicit. - Run
scripts/rpc_cli_help_probe.pyfor safe CLI/API availability. It prints help text and does not start persistent services. - Use
references/troubleshooting.mdfor port conflicts, keys, timeouts, tracker query failures, target-host mismatch, module load errors, and remote device availability.
API anchors
from tvm import rpc
remote = rpc.connect(url, port, key="", session_timeout=0)
tracker = rpc.connect_tracker(url, port)
remote.upload("lib.tar")
func = remote.load_module("lib.tar")
dev = remote.cpu(0)
Use rpc.LocalSession() for a local smoke that exercises upload/load/run
without opening a network listener.
CLI anchors
python -m tvm.exec.rpc_server --host 0.0.0.0 --port 9090
python -m tvm.exec.rpc_tracker --host 0.0.0.0 --port 9190 --port-end 9191
python -m tvm.exec.query_rpc_tracker --host TRACKER_HOST --port 9190
python -m tvm.exec.rpc_proxy --host PROXY_HOST --port 9090 --tracker TRACKER_HOST:9190
Start long-lived services only after the user approves host/port/key choices and knows how to stop the processes.
Boundaries
- Relax compilation before export:
../relax-compile/SKILL.md. - S-TIR/meta-schedule runner selection:
../s-tir-tuning/SKILL.md. - TIRx CUDA kernel codegen:
../tirx-kernels/SKILL.md.
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).