{
  "sources": [
    {
      "id": "windows-baseline",
      "url": "https://github.com/fakoli/anvil-serving/blob/8abcc5dc75a1b2389210b553120abb76e5864d3a/docs/findings/2026-09-02-glm53-sglang-sm120-393k-promotion.md",
      "published_or_observed": "2026-09-02",
      "observed_at": "2026-09-08",
      "age_class": "recent",
      "evidence_type": "local historical Windows/WSL benchmark",
      "hardware_runtime_relevance": "Same physical GPU pair, model and image pins; historical OS/runtime",
      "decision_impact": "Baseline only; retained raw request metrics determine comparisons."
    },
    {
      "id": "nccl-gpu-guidance",
      "url": "https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/troubleshooting/gpu_troubleshooting.html",
      "published_or_observed": "2026-09-08 observed; publication date not stated",
      "observed_at": "2026-09-08",
      "age_class": "current-observed",
      "evidence_type": "official documentation, follow-up planning only",
      "hardware_runtime_relevance": "Linux multi-GPU guidance; exact installed-version applicability still requires verification",
      "decision_impact": "P2P correctness/bandwidth and platform checks before changing the existing workaround"
    },
    {
      "id": "nccl-environment",
      "url": "https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html",
      "published_or_observed": "2026-09-08 observed; publication date not stated",
      "observed_at": "2026-09-08",
      "age_class": "current-observed",
      "evidence_type": "official documentation, follow-up planning only",
      "hardware_runtime_relevance": "Linux multi-GPU guidance; exact installed-version applicability still requires verification",
      "decision_impact": "Separate P2P and cuMem controls; verify installed version"
    },
    {
      "id": "sglang-server-controls",
      "url": "https://docs.sglang.io/docs/advanced_features/server_arguments",
      "published_or_observed": "2026-09-08 observed; publication date not stated",
      "observed_at": "2026-09-08",
      "age_class": "current-observed",
      "evidence_type": "official documentation, follow-up planning only",
      "hardware_runtime_relevance": "Linux multi-GPU guidance; exact installed-version applicability still requires verification",
      "decision_impact": "Custom all-reduce and prefill candidates; no unconditional enabling"
    },
    {
      "id": "linux-iommu-domain-abi",
      "url": "https://raw.githubusercontent.com/torvalds/linux/v7.0/Documentation/ABI/testing/sysfs-kernel-iommu_groups",
      "published_or_observed": "Kernel v7.0 retained source; observed2026-09-08",
      "observed_at": "2026-09-08",
      "age_class": "version-pinned",
      "evidence_type": "official kernel ABI definition",
      "hardware_runtime_relevance": "Explains live DMA-FQ sysfs values on the native Linux GPU groups",
      "decision_impact": "Translated DMA is a platform prerequisite to review before P2P qualification; not proof of hang cause"
    }
  ]
}