LMSYS Blog · 实时热榜
- 01SpecForge v0.3.0: a Unified Disaggregated and Colocated Speculative Decoding Stack, and New Open SpecBundle Draft Models - LMSYS Org
SpecForge v0.3.0: a Unified Disaggregated and Colocated Speculative Decoding Stack, and New Open SpecBundle Draft Models LMSYS Org
最高第 1 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 02Full-Stack Performance Optimization of AR+DiT in SGL-Diffusion - LMSYS Org
Full-Stack Performance Optimization of AR+DiT in SGL-Diffusion LMSYS Org
最高第 1 名10:59 达到10:59 首次观测上榜当日结束时仍在榜累计约12小时6分 - 03RadixArk Joins Forces with Google to Bring Full SGLang Features to TPUs - LMSYS Org
RadixArk Joins Forces with Google to Bring Full SGLang Features to TPUs LMSYS Org
最高第 2 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 04Towards Blackwell-Native 8-bit and 4-bit RL: End-to-End MXFP8 and NVFP4 RL in Miles - LMSYS Org
Towards Blackwell-Native 8-bit and 4-bit RL: End-to-End MXFP8 and NVFP4 RL in Miles LMSYS Org
最高第 3 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 05Toward a Cleaner Quantization Stack in SGLang - LMSYS Org
Toward a Cleaner Quantization Stack in SGLang LMSYS Org
最高第 4 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 06SGLang and Miles Add Day-0 Support for Kimi K3 - LMSYS Org
SGLang and Miles Add Day-0 Support for Kimi K3 LMSYS Org
最高第 5 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 07OPD Support in Miles - LMSYS Org
OPD Support in Miles LMSYS Org
最高第 6 名00:00 达到当日首次采集时已在榜10:59 观测离榜累计约10小时59分 - 08Serving GLM5.2 NVFP4 Agentic Workload with SGLang: Reaching 500 TPS in 2 Weeks - LMSYS Org
Serving GLM5.2 NVFP4 Agentic Workload with SGLang: Reaching 500 TPS in 2 Weeks LMSYS Org
最高第 7 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 09Bringing DeepSeek-V4 Flash RL Training to AMD Instinct MI355X GPUs with Miles - LMSYS Org
Bringing DeepSeek-V4 Flash RL Training to AMD Instinct MI355X GPUs with Miles LMSYS Org
最高第 8 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 10Accelerating SGLang HiCache with Netpreme X-Mem™ MPU - LMSYS Org
Accelerating SGLang HiCache with Netpreme X-Mem™ MPU LMSYS Org
最高第 9 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 11DSpark in SGLang: Speculative Decoding with Confidence-Driven, Variable-Length Verification - LMSYS Org
DSpark in SGLang: Speculative Decoding with Confidence-Driven, Variable-Length Verification LMSYS Org
最高第 10 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 12Agent-Assisted SGLang Development: An Initial Exploration - LMSYS Org
Agent-Assisted SGLang Development: An Initial Exploration LMSYS Org
最高第 11 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 13Improving DeepEP MoE Load Balance in SGLang with Waterfill and LPLB - LMSYS Org
Improving DeepEP MoE Load Balance in SGLang with Waterfill and LPLB LMSYS Org
最高第 12 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 14MOSS-TTS Local Transformer v1.5 on SGLang-Omni: Serving Native-Streaming 48 kHz Speech - LMSYS Org
MOSS-TTS Local Transformer v1.5 on SGLang-Omni: Serving Native-Streaming 48 kHz Speech LMSYS Org
最高第 13 名00:00 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 15Optimizing Ling-2.6-1T on TPU with SGLang-JAX: Hiding MoE Data Movement Behind Compute with One Pallas Kernel - lmsys.org
Optimizing Ling-2.6-1T on TPU with SGLang-JAX: Hiding MoE Data Movement Behind Compute with One Pallas Kernel lmsys.org
最高第 14 名00:00 达到当日首次采集时已在榜23:05 观测离榜累计约23小时5分 - 16The next generation of speculative decoding: DFlash and Spec V2 - LMSYS Org
The next generation of speculative decoding: DFlash and Spec V2 LMSYS Org
最高第 14 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 17No Token Left Behind: Demystifying Token-In-Token-Out in Miles - LMSYS Org
No Token Left Behind: Demystifying Token-In-Token-Out in Miles LMSYS Org
最高第 15 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 18Higgs Audio v3 TTS on SGLang-Omni: Real-Time, Controllable Speech for Voice Agents - LMSYS Org
Higgs Audio v3 TTS on SGLang-Omni: Real-Time, Controllable Speech for Voice Agents LMSYS Org
最高第 16 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 19SGLang and Miles Add Day-0 Support for NVIDIA Nemotron 3 Ultra for Long-Running Autonomous Agents - LMSYS Org
SGLang and Miles Add Day-0 Support for NVIDIA Nemotron 3 Ultra for Long-Running Autonomous Agents LMSYS Org
最高第 17 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 20Heterogeneous CPU + GPU EPD Disaggregation to Boost VLM Serving - LMSYS Org
Heterogeneous CPU + GPU EPD Disaggregation to Boost VLM Serving LMSYS Org
最高第 18 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 21Win on TCO: How AMD Instinct™ MI355X Achieves Cost-Competitive Distributed Inference Through SGLang with MoRI - LMSYS Org
Win on TCO: How AMD Instinct™ MI355X Achieves Cost-Competitive Distributed Inference Through SGLang with MoRI LMSYS Org
最高第 19 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 22Updating 1T parameters in seconds — P2P weight transfer in Large Scale Distributed RL - LMSYS Org
Updating 1T parameters in seconds — P2P weight transfer in Large Scale Distributed RL LMSYS Org
最高第 20 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 23DeepSeek-V4 on Day 0: From Fast Inference to Verified RL with SGLang and Miles - LMSYS Org
DeepSeek-V4 on Day 0: From Fast Inference to Verified RL with SGLang and Miles LMSYS Org
最高第 21 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 24HiSparse: Turbocharging Sparse Attention with Hierarchical Memory - LMSYS Org
HiSparse: Turbocharging Sparse Attention with Hierarchical Memory LMSYS Org
最高第 22 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 25Highlights of SGLang at NVIDIA GTC 2026 - LMSYS Org
Highlights of SGLang at NVIDIA GTC 2026 LMSYS Org
最高第 23 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 26Elastic EP in SGLang: Achieving Partial Failure Tolerance for DeepSeek MoE Deployments - LMSYS Org
Elastic EP in SGLang: Achieving Partial Failure Tolerance for DeepSeek MoE Deployments LMSYS Org
最高第 24 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 27ROCm Support for Miles: Large-Scale RL Post-Training on AMD Instinct™ GPUs - LMSYS Org
ROCm Support for Miles: Large-Scale RL Post-Training on AMD Instinct™ GPUs LMSYS Org
最高第 25 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 28Unlocking 25x Inference Performance with SGLang on NVIDIA GB300 NVL72 - LMSYS Org
Unlocking 25x Inference Performance with SGLang on NVIDIA GB300 NVL72 LMSYS Org
最高第 26 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 29Deploying DeepSeek on GB300 NVL72: Big Wins in Long-Context Inference - LMSYS Org
Deploying DeepSeek on GB300 NVL72: Big Wins in Long-Context Inference LMSYS Org
最高第 27 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 30SGLang-Diffusion: Advanced Optimizations for Production-Ready Video Generation - LMSYS Org
SGLang-Diffusion: Advanced Optimizations for Production-Ready Video Generation LMSYS Org
最高第 28 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 31Unleashing Computational Power: Ultimate Latency Optimization of Qwen3 and Qwen3-VL on AMD MI300X Series - LMSYS Org
Unleashing Computational Power: Ultimate Latency Optimization of Qwen3 and Qwen3-VL on AMD MI300X Series LMSYS Org
最高第 29 名23:05 达到当日首次采集时已在榜当日结束时仍在榜累计约23小时5分 - 32Squeezing 1TB Model Rollout into a Single H200: INT4 QAT RL End-to-End Practice - LMSYS Org
Squeezing 1TB Model Rollout into a Single H200: INT4 QAT RL End-to-End Practice LMSYS Org
最高第 30 名23:05 达到23:05 首次观测上榜当日结束时仍在榜累计约0分钟