SemiAnalysis 批评 AMD 领导层将计算资源从其内部 vLLM 团队转移，导致 vLLM 自动化测试的开发出现倒退。这种集群资源的转移正在阻碍 AMD 在实现与 CUDA 的 vLLM 测试能力相匹配方面取得进展。尽管面临这些挑战，AMD 的 DI CI 系统在最初的请求和会谈后得以实施，已成功识别出 DeepSeek V4 和 Kimi 等模型中的重大错误，提高了整体代码质量。 AI
影响 AMD 的内部资源分配决策可能会减缓其在开发具有竞争力的 AI 模型和测试基础设施方面的进展。
排序理由 该集群包含 SemiAnalysis 关于 AMD 内部资源分配及其对 AI 发展影响的关键推文，而不是 AMD 的官方公告或发布。
AI 生成摘要 · Google Gemini · 来自 7 个来源。 我们如何撰写摘要 →
@sgl_project @AnushElangovan @LisaSu We hope AMD leadership can reprioritize providing its internal vLLM team with stable clusters so that AMD’s hardcore internal vLLM engineers can do the work required to reach 90%+ gating parity with CUDA vLLM. 7/7🧵
@sgl_project @AnushElangovan @LisaSu Gating/blocking tests mean that the tests are of the highest quality, since PRs cannot merge unless the tests pass. While AMD leadership may distract non-technical folks by showing its non-gating pass rate, gating parity and the gating pass ra…
@sgl_project @AnushElangovan @LisaSu On the vLLM side, progress on automated vLLM gating tests has massively regressed due to AMD cluster infrastructure stability issues. AMD’s hardcore engineers had been making good progress on vLLM gating over the past couple of weeks until AMD…
@sgl_project @AnushElangovan @LisaSu In order for AMD DI CI to reach parity with CUDA SGLang, it also needs to implement nightly tests for the WideEP decode optimization, on which it is making good progress in 31500. 4/7🧵
@sgl_project @AnushElangovan @LisaSu The first bug it caught was an AMD MoRI buffer MR error in #30336 on DeepSeek V4. The second was a Kimi K2.6 error caused by an AMD AITER kernel regression in #30433 that made Kimi’s math accuracy massively lower. Through AMD’s nightly CI, the…
@sgl_project The first AMD DI CI PR, #29084, was implemented three months after SemiAnalysis’s initial request and multiple meetings with @AnushElangovan and Vamsi (Head of AI), as well as a meeting with @LisaSu early in the year regarding improvements to code quality. We will wa…
Great work by the AMD @sgl_project team on enabling nightly disaggregated serving CI to improve code quality! It has already caught and prevented 2 massive bugs from reaching customers, as we explained before 👇️ 1/7🧵 https://t.co/cUCr6niKEt