vLLM runs DeepSeek v4.1 10-20% faster than SGLang on RTX Pro 6000
morgymcg · x · 2026-10-26
User morgymcg reports running DeepSeek v4.1 via YourLocalAILab's vLLM image on an RTX Pro 6000, with decode throughput about 10-20% faster than SGLang and no visible quality drop so far — evals are still finishing.
More from Infra
- Sam Altman: robots building robots and data centers building data centers form another AI flywheel — haider1 · 2026-10-26
- Middle schooler hand-writes Metal RMSNorm kernel for own inference engine on Apple Silicon — dosco · 2026-10-26
- Default-to-zero: a model-name mismatch silently wiped all LLM costs to $0 — Atm1n9 · 2026-10-26
- Dev: I regret buying software licenses, should have bought more compute — cephaloform · 2026-10-26
- Musk: 200+ neoclouds face 'guppy-level death' as debt markets shut for AI-only firms — Yuchenj_UW · 2026-10-26
- Tailscale VP of Engineering responds to SSH vuln concerns: no need to disable it yet — basedjensen · 2026-10-26