Review and upgrade MetaX model support against target vLLM and installed MACA components, including model-dependent attention, weight loading and kernels. Use for model adaptations and shared-upstream-bug attribution, not standalone patch or registry audits.
中文简介正在准备中。当前显示原文说明。
实测评分 未实测
查看安装方法安装前可以查看交给 AI 代理的指令内容。
Fully read and apply vllm-metax-upgrade-common before this workflow.
Cover the requested models under vllm_metax/models/, their registration/configuration,
and required attention, cache, compressor, indexer, projection, MTP/draft, MoE and
kernel wrappers. Model-private attention remains in scope wherever its files live.
Build the matrix for each requested model family using the effective configuration after platform and speculative-decoding rewrites, not only the raw HF config.
config.json for layer
layout. Record backbone and MTP/draft counts separately, count layer-indexed arrays
such as layer_types, rope_theta and partial_rotary_factors, and map zero-based
checkpoint IDs to each part. Config defaults and the highest layer ID do not supply
the layer count; identify missing official configuration as an evidence gap.| Contract | Record for both upstream and MACA |
|---|---|
| Weights | Checkpoint/runtime dtype, per-channel/block/MX grouping, scale encoding, packing, post-load conversion and excluded modules. |
| Activations and projections | Compute/output/accumulation dtype, fused/separate projections, grouped GEMM, padding, quantization and inverse RoPE order. |
| KV and indexer caches | Logical/physical dtype, quantization scales, compressed/SWA separation, page layout/stride and index units. |
| Attention | Actual backend/model-private path, prefill/decode/MTP, head dimensions, sinks, causal behavior, compression and capability restrictions. |
| MoE and parallelism | Actual experts implementation, routing/scaling, shared experts, TP/PP/SP/EP/DCP collective ownership and supported combinations. |
| Optional features | LoRA, MTP/DSpark, graph capture and each feature's actual dispatch and restrictions. |
| Components | Selected DeepGEMM, FlashAttention/FlashMLA, MCOPLIB and other callable contracts and verified semantics. |
A model's FP8/FP4 checkpoint label does not replace required MACA BF16/INT8 compute, weight or cache paths. Inspect the model's executable selection and conversion rather than deriving these paths from its checkpoint label.
Trace model registration, config rewriting, layer construction, weight loading, execution selection and kernels. Include model-private attention/compressor/FlashMLA implementations such as DeepSeek V4 even outside a generic attention directory.
Check config transforms and model/draft class selection alongside parameter mappings; identically named imported classes can bind to different local/upstream implementations. Verify binary API contracts at the installed boundary. Read relevant review-cases.md for quantization/loading, private attention, parallelism and model-identity cases; its historical examples are investigation prompts.
When a defect also affects or may affect target upstream:
logger.warning or logger.warning_once with trigger, consequence and known workaround,
and report the unresolved repair plan. Prefer construction/configuration time to
repeated token-time logging.Verify registration and construction, real checkpoint loading, transformed parameter state, selected model/kernel paths and the affected numerical behavior. Derive boundary and rejection cases from the difference matrix, including loader ordering and cache scale/packing cases.
For parallel changes, verify collective ownership and residual semantics through the actual distributed path. LoRA/MTP support requires their model execution paths, and wrapper-only or kernel-only results cannot establish that support. State unresolved shared-bug warnings alongside the model-specific difference and dependency coverage.
还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
Review and adapt MetaX attention backends, MLA, sparse indexers, cache layouts, and kernel wrappers against target vLLM and installed MetaX component contracts. Use for attention work outside vllm_metax/patch/.
中文简介正在准备中。当前显示原文说明。
Trim large MetaX model directories for dummy smoke tests or real-checkpoint loading on limited GPUs. Preserve target execution paths while reducing depth; use the real-weight workflow when a loadable checkpoint is requested. Not for accuracy evaluation or model compatibility upgrades.
中文简介正在准备中。当前显示原文说明。
Audit and adapt monkey patches in vllm_metax/patch/ against a target upstream revision. Decide whether to retain, update, migrate, or remove each patch; maintain patch headers and explanations; validate changes and record evidence. Use only for compatibility work on this directory, not standalone attention backend, model, kernel, or other adaptations elsewhere in vllm_metax.
中文简介正在准备中。当前显示原文说明。
Review and adapt vllm_metax/registry registrations, quantization configurations, CustomOps and kernel dispatch against a target vLLM revision and installed MetaX APIs. Verify registration, inherited state, weight layouts and actual backend selection. Use for registry compatibility work, not standalone attention algorithms or monkey patches.
中文简介正在准备中。当前显示原文说明。
Orchestrate a repository-wide vLLM-MetaX compatibility upgrade or coordinated multi-component adaptation through a support matrix, dependency graph and scoped specialists. Use a domain skill for standalone component work.
中文简介正在准备中。当前显示原文说明。
Shared workflow for vLLM-MetaX compatibility adaptations and reviews, including confirmed environments, upstream-first code, dependency contracts, validation and durable progress. Required by standalone domain skills and the full-upgrade coordinator; collecting environment facts may precede confirmation.
中文简介正在准备中。当前显示原文说明。