xLLM AI Coding Workflow

Languages: English | 简体中文

Evidence-driven workflows, skills, scripts, and reference knowledge for large-model serving on Ascend NPUs. xLLM is the complete primary path. vLLM-Ascend and SGLang are limited to the experimental adapter or artifact-analysis scopes declared in the skill map.

Choose your path

Goal	Start with
Create, resume, finalize, or archive an experiment	xllm-experiment-lifecycle
Optimize xLLM performance through multiple iterations	xllm-npu-sota-loop
Run a mixed performance and accuracy suite	xllm-npu-eval-runner
Run one performance or accuracy workload	xllm-npu-perf-runner or xllm-npu-accuracy-runner
Launch, verify, smoke-test, and clean up a service	xllm-npu-server-manager
Review fairness or a publishable performance claim	xllm-npu-benchmark
Collect or analyze profiling evidence	xllm-npu-profiler or xllm-npu-pipeline-analysis
Diagnose wrong output or a runtime/build incident	xllm-npu-accuracy-debug or xllm-npu-incident-triage
Query historical model and PR lessons	model-pr-optimization-history

For an optimization campaign, xllm-npu-sota-loop owns the optimization decisions while xllm-experiment-lifecycle owns the resumable run and evidence state.

Five-minute start

Initialize or reuse the xLLM checkout and install the canonical skills:

python scripts/init_xllm_workspace.py

Start Codex from this repository:

codex

Then choose a copy-ready prompt or create an experiment from experiment.example.yaml. Follow the getting-started guide for a synthetic lifecycle walkthrough.

How the pieces fit together

flowchart LR
  U["User intent"] --> S["Skills: decide and delegate"]
  S --> D["Scripts: deterministic execution"]
  D --> E["Run evidence and ledgers"]
  R["Reference and IO contracts"] --> S
  R --> D
  E --> H["Promoted durable lessons"]

Loading

Lifecycle is the control plane. Orchestrators delegate bounded work to runners, analyzers, planners, and gates. Deterministic scripts produce auditable artifacts.

Browse capabilities

• Skill capability index — grouped by role and domain.
• Detailed skill architecture — dependencies, scopes, intents, and outputs.
• Agent routing contract — exact task-to-skill and repository constraints.

Documentation

Use the documentation hub to find current user, architecture, workflow, and maintainer guides. The standard NPU AI coding workflow explains the evidence-driven phases. Historical PR design records are preserved separately and are not the normal operating entry point.

Repository map

skills/       procedural decisions and delegation
scripts/      deterministic execution
reference/    stable knowledge and IO contracts
runs/         per-task evidence; local and gitignored
humanize/     promoted validated lessons
docs/         current guides, architecture, and design history
tests/        catalog, navigation, schema, and workflow validation

See the repository and evidence map for ownership boundaries.

Contribution and support boundaries

Reusable decision workflows belong in skills; deterministic operations belong in scripts; schemas and stable knowledge belong in reference; only validated durable lessons are promoted to humanize. Read Adding or changing a skill before changing cataloged capabilities.

Do not commit local paths, private hosts, credentials, or non-public logs. The repository does not currently include a license; add one before broad external reuse.