vllm-deepseek-v3-r1-optimization

What problem does it solve?

This skill provides a structured, PR-backed guide to optimize and verify DeepSeek V3 and R1 implementations within vLLM, enabling engineers to audit changes, extend support, and document best practices for MLA, MoE, packed-module loading, LoRA, MTP/Eagle, and ROCm/CUDA validation paths.

Core Features & Use Cases

• PR-diff driven optimization guidance for DeepSeek V3/R1 changes across the vLLM runtime surfaces.
• Maps to concrete model files and tooling, including vllm/model_executor/models/deepseek_v2.py, deepseek_eagle.py, and deepseek_mtp.py.
• Use cases include validating new LoRA support, packed-module loading improvements, and end-to-end validation across BF16/FP8 and ROCm paths.

Quick Start

Review the PR diffs and canonical notes in references/pr-history.md to compose a production-ready optimization dossier for a given PR.