lmdeploy-runtime-debugging

Use when adding verified layer guides or compact GPU display lanes to an existing Torch Profiler trace.

Use when measuring LMDeploy serving throughput, latency, or capacity without a profiler.

Use when an LMDeploy production incident needs evidence preservation and replay.

Use when optimizing or validating an identified LMDeploy CUDA/Triton kernel or dispatch path.

Use when reviewing LMDeploy changes against backend contracts or maintainer conventions.

Use when adding or reviewing an LLM or VLM architecture in LMDeploy's PyTorch backend.

Use when tracing LMDeploy attention dispatch, KV layout, or quantization contracts.

Use when measuring LMDeploy model or API accuracy with deterministic requests or real datasets.

Use when capturing or analyzing LMDeploy, vLLM, or SGLang serving traces.

Use when committing, publishing LMDeploy PRs, or resolving branch and remote divergence.

Use when a session or requested retrospective yields a recurring lesson worth preserving in skills.

Use when LMDeploy commands fail because of environment, tool-path, or sandbox setup.