maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Profile and attribute HBM memory usage on Ascend NPU for vLLM serving scenarios. Breaks down memory into fixed overhead, model weights, KV cache, HCCL buffers, activations, and runtime, with traceable evidence chains. Use for requests like "分析显存占用", "显存 profiling", "HBM 用了多少", "内存各部分拆分". Do not use for performance…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Reduce an Ascend model-level failure to one torchnpu, ACLNN, or custom operator call, then validate explicit dtype, shape, layout, and eager/compile/graph cases against a reference implementation. Use for operator crashes, unsupported dtype or layout errors, shape-dependent numerical mismatches, or workspace API…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Analyze Ascend NPU torch profiler output (kerneldetails.csv / traceview.json / opsummary / communication.json) for one or many profiling roots and produce a traceable report (rank/step/layer/operator summary, cross-rank alignment, diagnosis findings, report.md / report.xlsx / report.html with single-step inspectors…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Collect one Ascend torch-profiler case end-to-end on a workspace-managed remote NPU container. Starts a profiled vLLM service, brackets a workload with /startprofile and /stopprofile, runs analyse(), verifies kerneldetails.csv landed, and writes a manifest the analysis skill can consume. Use for requests like "采集…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Profile and iteratively optimize a correctness-passed Ascend Triton kernel with explicit NPU baselines, per-shape measurements, UB live-set and physical-core reasoning, MTE/Vector/Scalar bottleneck attribution, one-hypothesis rounds, noise-aware KEEP/DISCARD decisions, and Run Manifest evidence. Use for single-kernel…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Validate one Ascend Triton kernel against a trusted reference across an explicit shape, dtype, layout, stride, scalar-option, and execution-mode case matrix, with static detection of missing kernel launches and PyTorch computation fallback plus Run Manifest evidence. Use before any performance claim, after migration…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Develop a first correct Ascend Triton operator from a PyTorch reference or migrate an existing GPU Triton kernel to Ascend, including semantic audit, explicit task contracts, hardware-aware grid and tiling design, implementation, and handoff to correctness validation. Use for new kernel implementation, CUDA/GPU Triton…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Orchestrate an end-to-end Ascend Triton operator effort across task definition, GPU-to-NPU migration or direct development, explicit correctness validation, profiler-driven optimization, and evidence aggregation with Run Manifest v1. Use when the request spans two or more lifecycle stages or asks for a complete…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Review, deduplicate, promote, merge, reject, or deprecate verified vLLM Ascend workspace knowledge candidates. Use only when the user explicitly asks to curate, persist, review, merge, promote, or deprecate project knowledge (沉淀、整理、复盘、合并、提升、废弃), or explicitly invokes this Skill to review…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Add, verify, repair, or remove a managed remote NPU host for this workspace. Use for requests like “配置服务器”, “加一台机器”, “检查 ready”, “修容器 SSH”, or “移除机器”. Do not use for code sync, rebuilds, serving, or benchmarking.
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Download, resume, status-check, and SHA256-verify ModelScope model weights. Use for $modelscope download/status/verify/check, Chinese requests to 下载/续传/补全/查看进度/校验 ModelScope 权重, and tasks that need durable background ModelScope downloads under explicit local directories.
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Bootstrap or locate the standalone vaws-top worktree and provide its basic CLI/MCP query entrypoints. Use when vaws-top is not yet available, for basic fleet discovery and server inspection, or to deploy, inspect, restart, or stop its local service. Detailed fleet-query guidance lives on the vaws-top branch.
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Ensure a ready remote runtime runs the exact current local workspace state before any remote smoke, service launch, or benchmark. Use automatically immediately before remote execution when direct local-to-container SSH already works and local uncommitted changes must be reflected remotely. Do not use for initial…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Resolve, probe, execute, observe jobs, sync, manage service lifecycle, transfer artifacts, and clean VAWS remote Ascend session containers through structured agent-facing CLI entrypoints.
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Initialize this workspace after clone. Use for requests like “初始化仓库”, “配置 gh / GitHub 登录”, “初始化子模块”, or “把 vllm / vllm-ascend remotes 改成我的 fork”. Do not use for ordinary coding, serving, benchmarking, or unrelated Git tasks.
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Create, list, inspect, remove, garbage-collect, and group isolated VAWS agent sessions, and optionally coordinate NPU task intent across independent agents on one host. Use before remote execution when tasks must not share worktrees, containers, serving state, or resource leases, when cooperative NPU queueing is…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Run vLLM online-serving benchmarks on a workspace-managed remote container. Use for requests like "跑个 benchmark", "对比性能", "压测一下", "测下吞吐", or "看下有没有性能回退". Do not use for accuracy tests, nightly CI matrix runs, offline inference, or service-only lifecycle.
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Analyze vLLM and vllm-ascend diffs, map affected components to the minimum sufficient correctness, build, performance, graph, operator, distributed, and profiling evidence, link downstream Run Manifest results, and produce a PR-ready validation report. Use for PR validation, workspace-diff risk analysis, deciding what…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Plan, execute, normalize, and compare vLLM Ascend inference correctness across baseline and candidate code states, eager and graph modes, offline generate or chat, online chat completions, and AISBench task metrics. Use for accuracy validation, token-output comparison, graph-versus-eager checks, deterministic…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Diagnose vLLM Ascend multi-rank and multi-node startup, rank mapping, process-group, collective, HCCL, Ray, scheduler, connector, and distributed hang failures from structured topology and per-rank evidence. Use when a failure depends on rank count, parallel topology, nodes, collectives, or distributed endpoints. Do…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Diagnose vLLM Ascend cudagraph and ACL Graph compile, capture, replay, hang, and graph-versus-eager correctness problems. Use when eager passes but graph mode fails, hangs, or diverges, or when graph/eager intermediate tensors must be aligned. Do not use to plan a correctness matrix, after the failure is reduced to…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Plan, start, inspect, smoke-test, and stop a multi-session vLLM Ascend prefill/decode deployment with explicit connector configuration, role ordering, proxy endpoint health, rollback, and KV-transfer request evidence. Use for PD disaggregation with NIXL, Mooncake, or another KV connector. Do not use for one colocated…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Plan, record, and analyze controlled baseline-versus-candidate vLLM Ascend serving performance experiments with isolated sessions, identical non-code configuration, alternating A/B order, warmup exclusion, variance and outlier reporting, and metric-specific regression thresholds. Use for throughput, TTFT, TPOT, ITL…
maoxx241/vllm-ascend-workspace
Skill Claude CodeCodex
Start, check, or stop a single-node vLLM Ascend online service on a workspace-managed ready remote container. Use for requests like "拉服务", "在远端起个服务", "重启服务", "看服务状态", "停掉服务". Do not use for machine attach, environment bootstrap, code sync, benchmark orchestration, or offline inference.