vllm-metax-upgrade
Skill192 starsChanged 12 days agoApache-2.0View on GitHub
Coordinate a complete vLLM-MetaX source upgrade in ordered framework, model and patch phases, then run final model e2e and batched validation with a resumable full-file inventory. Use a domain skill for standalone component work.
In this repository10 other files it gives its agents.
- AGENTS.mdAGENTS.md
- attention-upgradeSkill
- common-senseSkill
- csrc-upgradeSkill
- environment-checkSkill
- model-adaptationSkill
- model-upgradeSkill
- patch-upgradeSkill
- registry-upgradeSkill
- vllm-metax-model-trimSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 1 is distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 28
Before entering platform, record analysis-put with the global dependency graph and read-only contracts for every baseline model.
- Very specific: line 44
Read orchestration.md for domain boundaries and handoffs.
- Very specific: line 45
Use phase-next only after every file through the current phase has current source adaptation and required local evidence.
- Distinctive: line 68
Before any model execution, require current source readiness for the entire inventory, including every model spec/import closure and all …
- Very specific: line 78
Run pre-commit run --all-files after source adaptation and again on the final code before completion.
- Very specific: line 85
Run finish only when every inventoried file is marked 适配完成并验证完毕, mappings and all required evidence are current, every supported row has …
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
vLLM-MetaX Full Upgrade
First apply environment-check: approve the preparation plan once and successfully install/verify both frameworks before any adaptation. An installation failure blocks adaptation; source repairs cannot bypass this prerequisite. Then read and apply common-sense. It owns upstream-first source style, dependency closure, file ownership, evidence and recovery. This coordinator owns the full inventory, phase order and whole-upgrade acceptance.
Prepare the complete upgrade
Resume the existing full run, or initialize one before adaptation. Inventory every
vllm_metax/ Python file and production/build/requirements file before edits. Assign
phase, owning skill, exact upstream/prior mapping and required acceptance. Keep added,
deleted and no-edit files in the ledger. Freeze the confirmed target revision and the
current-chip support matrix; inspect all target implementation paths, including branches
unavailable for runtime tests on the current hardware.
Assign csrc-upgrade every native source/header/build contract, regardless of whether production uses mcoplib. Place required native dependencies with their earliest consumer and finish the remaining native sweep in framework.
Before entering platform, record analysis-put with the global dependency graph and
read-only contracts for every baseline model. Resolve cycles/ownership in the plan;
this planning step does not authorize early model construction or adaptation.
Enforce the phase order
| Phase | Required source work |
1. platform | Review platform.py against target CUDA/base platform APIs and MACA contracts. |
2. registry | Upgrade all registrations, CustomOps, quantization configs and their local transitive kernels/experts. |
3. attention | Upgrade attention, quantized KV/indexer cache contracts and split prefill/decode execution. |
4. framework | Scan every remaining non-model/non-patch Python file, initializers and build/runtime dependencies, including a full csrc upgrade under either kernel provider. |
5. models | Upgrade every local model and model dependency; maintain each model's spec and prohibit dependency on upstream concrete model implementations. |
6. patch | Review every patch with removal as the objective; remove, replace via supported extension points, or minimize and justify. |
7. validation | Only now construct/load models and run model smoke, generation, e2e and batched inference correctness. |
Read orchestration.md for domain boundaries and handoffs.
Use phase-next only after every file through the current phase has current source
adaptation and required local evidence. All four framework phases must finish before
models begin. Domain completion or one working model never substitutes for the full
framework upgrade. Do not schedule an early minimal model vertical slice.
Framework/operation unit tests and component numerical checks may run while adapting. A model-free plugin/dispatch closure is allowed; complete-model construction, real weight loading, dummy/trimmed smoke and generation all wait until phase 7.
Handle discoveries without hiding unfinished work
The user permits required framework repairs discovered during models. Register a separate earlier-phase dependency task with its owner and reason; invalidate affected source/evidence, complete the repair and revisit consumers. The full framework gate must be current before continuing model work or final model validation.
The user also permits an early blocking patch repair. Record its specific blocker
with --exception-note, scope it to resolving that blocker, and keep the comprehensive
patch review in phase 6. Early repair does not exempt it from the removal-first review;
record a fresh patch-phase review before counting its source adaptation complete.
Final model validation and acceptance
Before any model execution, require current source readiness for the entire inventory,
including every model spec/import closure and all patches. Execute normal-entry model
e2e and tools/batched_test inference correctness for every supported matrix row and
at least every sample in tools/models/. Bind samples with model-bind. GPU-count/VRAM
shortages may be explicitly skipped using model-skip; all other missing coverage blocks
completion. Report skipped samples and required/available resources separately from passes.
Performance runs are additional only when requested; a benchmark alone is not correctness
acceptance. Read common validation for actual
child/worker origins, numerical contracts and evidence limits.
Run pre-commit run --all-files after source adaptation and again on the final code
before completion. Fix every current hook failure, including pre-existing formatting
issues, and repeat until it passes. Review auto-fixes, reconcile source changes and
rerun affected required validation; preserve the final successful log with the run.
The common validation procedure defines this
mandatory repository-wide acceptance check.
Run finish only when every inventoried file is marked 适配完成并验证完毕, mappings and
all required evidence are current, every supported row has required e2e/batched
results, and the final full pre-commit run passes. Preserve explicit rejection evidence for unsupported rows. Record unavailable
chip quantization branches as unverified; report completion only within the agreed
current-chip scope, never as support for those branches. Required failures or blockers
within that scope prevent completion.
At handoff provide phase, source-ready and final-verified file counts, owner/blocker, run/TODO/FILES paths, remaining matrix rows and the next executable action. A partial framework or model subset cannot satisfy full completion.
For baseline samples using dummy weights, acceptance is explicitly dummy_smoke:
startup, valid prefill/decode generation and concurrent request/batch stability, without
semantic accuracy claims. Real-weight configurations use accuracy and their semantic
cases. Declare configuration.validation_mode in model rows and load_format: dummy
where applicable. A dummy smoke pass cannot satisfy real-weight accuracy acceptance.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

