DeepSeek

• DeepSeek V4 Attention: Code Reading Map — Navigable implementation map of DeepSeek V4’s hybrid compressed attention across vLLM (NVIDIA/AMD/XPU) and vllm-ascend (Ascend NPU), covering CSA/HCA compressors, sparse MLA backends, heterogeneous KV cache, multi-stream overlap, and platform-specific kernel dispatch.
• DeepSeek-V4.1-Flash on NVIDIA vLLM: GPU Inference Path — Pinned worker round trip through the V4.1 wrapper, configuration-driven cache/indexer ownership, NVIDIA sparse-MLA backends, bounded SWA replay, and sampled-token return.
• DeepSeek-V4 Inference on Ascend: The DSA Serving Stack — End-to-end DeepSeek-V4 serving on Ascend NPUs: the model override with mHC hyper-connections, hybrid c4/c128 compressor layers, the AscendDSA prefill/decode flow, the five-type heterogeneous KV cache, the sparse-attention custom operator, and the MTP draft model.

Related: DSpark — DeepSeek’s speculative decoding framework.