DeepSeek Runs Blackwell Inference Without Tensor Parallelism, Prefers Wide-EP for Decode
nrehiew_ · x · 2026-10-20
nrehiew observes that DeepSeek uses no tensor parallelism in production inference even on Blackwell (presumably NVL72), a style tracing back to DeepSeek v3 on interconnect-limited H800s — while vLLM and SGLang default recipes both recommend TP. A DeepSeek slide confirms: Prefill EP2–8/DP2–8 (smallest EP/DP HBM allows), Decode Wide-EP EP64–128/DP64–128 (largest GPUs allow), VE on a single cheaper GPU per instance, and no TP/CP for now. This also appears to confirm DeepSeek is using NVIDIA Blackwell/NVL72 for part of production inference.
More from Infra
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Jensen Huang lays out NVIDIA's full Vera Rubin agentic system: NVL72, MGX, CMX and RTX — rohanpaul_ai · 2026-10-23
- Jensen Huang details Vera Rubin agentic stack: NVL72, MGX, CMX and RTX — rohanpaul_ai · 2026-10-23
- ONNX to Transformers.js: running models directly in your browser — JosephJacks_ · 2026-10-23
- MiniMax M3.1 Flash free for 3 days on 0G, with 1M context and 512K output — MiniMax_AI · 2026-10-23
- "We bet on Cloudflare and it paid off": R2 and D1 at billion scale — dinasaur_404 · 2026-10-23