[{"data":1,"prerenderedAt":61},["Reactive",2],{"news-story-news-20261026-google-gemma-gemini-4-argon-vllm-gemma-4-e2b-2-85x-speedup":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":15,"content_zh":21,"content_en":22,"external_links":23,"source":33,"category":40,"impact_score":41,"views_count":42,"likes_count":43,"related_entities":44,"tags":54,"image_url":60},"news-20261026-google-gemma-gemini-4-argon-vllm-gemma-4-e2b-2-85x-speedup","Google Gemma 团队：用 Gemini 4 Argon 为 vLLM 提交 Gemma 4 优化 PR，官方称 gemma-4-E2B-it 在 A100 上提速 2.85 倍","Google's Gemma team says Gemini 4 Argon wrote vLLM optimizations that speed up gemma-4-E2B-it 2.85x on an A100","Google 官方账号 @googlegemma 10 月 26 日发帖称，借助 Gemini 4 Argon 完成的一组 vLLM 优化，让 gemma-4-E2B-it 在单张 A100 上的吞吐从约 37,900 提升到 108,000+ tokens/秒，约 2.85 倍，且输出 token 100% 一致。对应 PR 由 Google DeepMind 的 Luciano Martins 提交，大部分仍在评审中。","On Oct 26, Google's official @googlegemma account said a set of vLLM optimizations built with Gemini 4 Argon lifted gemma-4-E2B-it throughput on a single A100 from about 37,900 to 108,000+ tokens/sec, roughly 2.85x, with 100% token parity. The PRs were filed by Luciano Martins of Google DeepMind, and most are still under review.",[10,11,12,13,14],"官方称 gemma-4-E2B-it 在 A100 上吞吐约 2.85 倍（约 37.9K 到 108K+ tokens/秒），输出 token 100% 一致","PR 作者为 Google DeepMind 的 Luciano Martins，覆盖融合 Triton 内核、MTP 投机解码、KV 共享快速预填充等","2.85 倍是 Google 对整组优化的说法；单个 PR #60674 自测预填充最高 1.90 倍、解码约 1.06 倍","截至 10 月 27 日，链接中 10 个 PR 仅 3 个已合并，其余仍在评审","这是用前沿编码模型做推理内核性能工程的公开案例，结果尚待 vLLM 社区复现",[16,17,18,19,20],"Google says gemma-4-E2B-it throughput on an A100 rose about 2.85x (about 37.9K to 108K+ tokens/sec) with 100% token parity","The PRs come from Luciano Martins of Google DeepMind and cover fused Triton kernels, MTP speculative decoding and KV-sharing fast prefill","The 2.85x figure is Google's claim for the full set; PR #60674 alone self-reports up to 1.90x prefill and about 1.06x decode","As of Oct 27, only 3 of the 10 PRs in the linked list are merged; the rest are still in review","It is a public example of using a frontier coding model for inference-kernel performance work, still awaiting reproduction by the vLLM community","### 核心背景与行业痛点\n开源模型发布后，推理框架的专用内核常要很久才跟上。10 月 26 日 @googlegemma 称，团队用 Gemini 4 Argon 为 vLLM 做了一组 Gemma 4 优化，展示编码模型能做底层性能工程。\n\n### 架构亮点与底层机制\nPR 由 Google DeepMind 的 Luciano Martins 提交：#60674 为 norm/RoPE、PLE、MoE 路由加入融合 Triton 内核；#60689 用稀疏 gather-GEMV 优化 MTP 投机解码；#60445 为 KV 共享层自动开启快速预填充。\n\n### 权威 Benchmark 与实测跑分对比\nGoogle 称 gemma-4-E2B-it 在单张 A100 上从约 37,900 升到 108,000+ tokens/秒，约 2.85 倍，token 100% 一致，这是整组优化的数字。单看 #60674，E2B 预填充最高 1.90 倍、解码约 1.06 倍，更大模型仅 1.04 到 1.05 倍。均为作者自测，尚无第三方复现。\n\n### 开发者实战落地与开箱指南\n截至 10 月 27 日，链接中 10 个 PR 仅 3 个合并，正式版 vLLM 暂拿不到全部收益。可关注合并进度，或在测试环境拉分支，用自己的负载对比吞吐和输出一致性后再上生产。","### Background and pain points\nAfter an open model ships, specialized kernels in inference engines often lag by weeks or months, and small models suffer most from generic code paths. On Oct 26, Google's official @googlegemma account said the team used Gemini 4 Argon to build a set of Gemma 4 optimizations for vLLM, as a demonstration that a frontier coding model can take on low-level performance engineering.\n\n### Architecture and mechanism\nThe PRs were filed by Luciano Martins of Google DeepMind. PR #60674 adds fused Triton kernels for QKV norm/RoPE, pre/post-FFN norms, PLE and MoE routing. PR #60689 optimizes MTP speculative decoding with a sparse gather-GEMV kernel that cuts intermediate memory allocation. PR #60445 auto-enables KV-sharing fast prefill and adds piecewise CUDA Graph capture.\n\n### Benchmarks\nGoogle says gemma-4-E2B-it on a single A100 went from about 37,900 to 108,000+ tokens/sec, roughly 2.85x, with 100% token parity. That figure covers the whole set. PR #60674 alone self-reports up to 1.90x prefill and about 1.06x decode on E2B, and only 1.04x to 1.05x on larger Gemma 4 models. All numbers are the author's own measurements with no third-party reproduction yet.\n\n### Getting started\nAs of Oct 27, only 3 of the 10 PRs in the linked list are merged, so release builds of vLLM do not yet get the full gain. Track the PRs, or test the branches in a staging environment and compare throughput and output parity on your own workload before production.",[24,27,30],{"title":25,"url":26},"vLLM PRs by lucianommartins (Oct 5-10)","https://github.com/vllm-project/vllm/pulls?q=is:pr+author:lucianommartins+(created:%3E2026-10-05+created:%3C2026-10-10)",{"title":28,"url":29},"vLLM PR #60674: fused Triton kernels for Gemma 4","https://github.com/vllm-project/vllm/pull/60674",{"title":31,"url":32},"vLLM PR #60689: Gemma 4 MTP speculative decoding","https://github.com/vllm-project/vllm/pull/60689",{"platform":34,"author_name":35,"author_handle":36,"author_avatar":37,"tweet_url":38,"published_at":39},"x","Google Gemma","googlegemma","https://unavatar.io/x/googlegemma","https://x.com/googlegemma/status/2114757896563253305","2026-10-26T16:35:47Z","tooling",81,73,0,{"models":45,"agents":49,"vendors":50},[46,47,48],"Gemma 4","gemma-4-E2B-it","Gemini 4 Argon",[],[51,52,53],"Google","Google DeepMind","vLLM",[53,55,56,57,58,59],"Gemma","inference","Triton","performance","AI coding","https://pbs.twimg.com/amplify_video_thumb/2114757764908228608/img/h8e5tEPxAV8R9Xbn.jpg",1793718867659]