Skill Claude CodeCodex
Validate PD-disaggregation Prefill/Decode launch commands and environment for Path C. Rewrite only network fields plus computed kvport/engineid/dpsize; never rename kvconnector. Renders mooncakemaster + dual proxy (fetched from vllm-ascend). Use for PD ratio / PD deploy check.
Skill Claude CodeCodex
Deploy Mooncake master then Prefill/Decode/Proxy from Path C rendered scripts. Use after pd-config-env-check and aisbench-install. On failure diagnose and rollback to Phase 1.
Skill Claude CodeCodex
Run AISBench tests to measure Prefill-side and Decode-side QPS under SLO, then compute best PD instance ratio (Path C Phase 4). Use after aisbench-install.
Skill Claude CodeCodex
A tool that reads service startup logs and extracts non-default settings such as model paths, quantization, parallel-processing choices, configuration, and feature switches into an Excel report.
Skill Claude CodeCodex
Estimates available KV cache capacity and memory-bound max concurrency for each DP×TP×EP parallel combination under vLLM-Ascend. Use after find-possible-parallel-strategy or when evaluating KV headroom vs context length.
Skill Claude CodeCodex
Orchestrates vLLM-Ascend single-node parallel strategy tuning from baseline config: resolve SLO constraints, clone vllm-ascend/msmodeling, then run find-possible-parallel-strategy, serving-kv-cache-capacity, and serving-slo-concurrency. Use in Phase 2 serving tuning.
Skill Claude CodeCodex
A guide to measuring the runtime performance of a software service, including its required calls, CSV output format, 11 extracted measurements, and supported log formats.
Skill Claude CodeCodex
Estimates max concurrency under TTFT/TPOT SLO for arbitrary vLLM-Ascend models (Qwen any size, GLM, DeepSeek, MiniMax, …) by analyzing vllm-ascend attention dispatch (usemla/usesparse) and msmodeling profilingdatabase. Requires local clones of both repos. Use as sub-skill of serving-parallel-strategy-tuning.
Skill Claude CodeCodex
Extract and compare configuration switches between vLLM and vLLM-Ascend repos. Invoke when user needs to audit, compare, or document config options across vLLM and vLLM-Ascend.
Skill Claude CodeCodex
A guide to tuning vLLM-Ascend, a system for running language models on Ascend hardware. It covers standalone use, pipeline use, and quantization tuning, which reduces model precision to change performance or resource use.