{
  "schemaVersion": 1,
  "checkedAt": "2026-11-09",
  "methodology": "核实固定配置、实现源码和逻辑缓存公式，未运行完整模型或测量GPU分配。证据状态只适用于逐项scope，不是跨硬件/后端的通用保证。",
  "statusDefinitions": {
    "verified": "实现已核实：列明配置的结构、dtype、布局与内容公式已有直接来源。",
    "reference": "含假设：存在明确的容量预算或间接配置证据，仍可按公开假设估算。",
    "pending": "信息不足：关键缓存维度或存储实现缺证据，不生成数值。"
  },
  "models": {
    "inklingsmall": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。 明确选择 Inkling-Small BF16 变体，而非现 recipe 默认 NVFP4 变体。",
      "finding": "42 层=7 full+35 SWA-512；KV8×128。每层四条 sconv stream 打包：raw width=2×128+2×(4096/8)=1280，向2幂补齐至2048；4槽×8 heads×2048×BF16，每层131072 B，共5505024 B。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/thinkingmachines/Inkling-Small/blob/8cc5877b44d343f88b92086aa1fb72897950f06a/config.json"
        },
        {
          "label": "固定 Inkling KV spec",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/models/inkling/nvidia/attention.py#L162-L198"
        },
        {
          "label": "固定 sconv state 布局及 BF16 断言",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/models/inkling/nvidia/sconv_swa_attn.py#L125-L192"
        },
        {
          "label": "每 decoder 层独立 conv owner",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/models/inkling/nvidia/model.py#L137-L170"
        },
        {
          "label": "官方卡指向 vLLM recipe",
          "url": "https://huggingface.co/thinkingmachines/Inkling-Small/blob/8cc5877b44d343f88b92086aa1fb72897950f06a/README.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。",
        "sconv 的2048宽度padding为所选实现内建状态布局，已计入；不是把逻辑卷积只按kernel-1估算。"
      ],
      "unresolved": [
        "固定实现对 TP 有 tp_size<=num_kv_heads=8 的断言；界面更高TP仅为数学外推，不可宣称可运行配置。"
      ]
    },
    "mimov25flash": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "官方full/SWA KV heads为4/8；已修正旧目录写反的8/4。9 full +39 SWA-128，K/V=192/128，BF16逻辑缓存1Mi为22.5238037109375 GiB。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5/blob/d6bf9db68673d1f903a53d8da417b293fdc97d1f/config.json"
        },
        {
          "label": "固定 MiMo Full/SWA 分支与不同 K/V 头维",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/mimo_v2.py#L326-L416"
        },
        {
          "label": "官方模型卡（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5/blob/d6bf9db68673d1f903a53d8da417b293fdc97d1f/README.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": [
        "此处不是所有引擎的统一缓存精度；官方部署镜像与不同 attention backend 的实际页分配尚未实机测量。"
      ]
    },
    "mimov25pro": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "10 full + 60 SWA-128；K=192/V=128；主缓存 BF16，无逐token量化 scale。 官方 full/SWA KV heads 均为 8。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/blob/98e90df48208087a2eb974866a736a20f8122d2a/config.json"
        },
        {
          "label": "固定 MiMo Full/SWA 分支与不同 K/V 头维",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/mimo_v2.py#L326-L416"
        },
        {
          "label": "官方模型卡（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/blob/98e90df48208087a2eb974866a736a20f8122d2a/README.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": [
        "此处不是所有引擎的统一缓存精度；官方部署镜像与不同 attention backend 的实际页分配尚未实机测量。"
      ]
    },
    "mimov26flashrl": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "9 full + 39 SWA-128；K=192/V=128；主缓存 BF16，无逐token量化 scale。 官方 full/SWA KV heads 为 4/8。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/blob/f657ba2f7d24c9adc5730675baaa2d34de878a49/config.json"
        },
        {
          "label": "固定 MiMo Full/SWA 分支与不同 K/V 头维",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/mimo_v2.py#L326-L416"
        },
        {
          "label": "官方模型卡（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/blob/f657ba2f7d24c9adc5730675baaa2d34de878a49/README.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": [
        "此处不是所有引擎的统一缓存精度；官方部署镜像与不同 attention backend 的实际页分配尚未实机测量。"
      ]
    },
    "mimov26flashmopd": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "9 full + 39 SWA-128；K=192/V=128；主缓存 BF16，无逐token量化 scale。 官方 full/SWA KV heads 为 4/8。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-MOPD/blob/5220c3f513065a0426be5e094571cbed0983386e/config.json"
        },
        {
          "label": "固定 MiMo Full/SWA 分支与不同 K/V 头维",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/mimo_v2.py#L326-L416"
        },
        {
          "label": "官方模型卡（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-MOPD/blob/5220c3f513065a0426be5e094571cbed0983386e/README.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": [
        "此处不是所有引擎的统一缓存精度；官方部署镜像与不同 attention backend 的实际页分配尚未实机测量。"
      ]
    },
    "mimov26prorl": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "10 full + 60 SWA-128；K=192/V=128；主缓存 BF16，无逐token量化 scale。 官方 full/SWA KV heads 均为 8。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/blob/a066ba0057192126bd5d255f37138265fc1a435e/config.json"
        },
        {
          "label": "固定 MiMo Full/SWA 分支与不同 K/V 头维",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/mimo_v2.py#L326-L416"
        },
        {
          "label": "官方模型卡（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/blob/a066ba0057192126bd5d255f37138265fc1a435e/README.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": [
        "此处不是所有引擎的统一缓存精度；官方部署镜像与不同 attention backend 的实际页分配尚未实机测量。"
      ]
    },
    "mimov26promopd": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "10 full + 60 SWA-128；K=192/V=128；主缓存 BF16，无逐token量化 scale。 官方 full/SWA KV heads 均为 8。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-MOPD/blob/da71ee15e76a8849bd1a9610195c452007a79678/config.json"
        },
        {
          "label": "固定 MiMo Full/SWA 分支与不同 K/V 头维",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/mimo_v2.py#L326-L416"
        },
        {
          "label": "官方模型卡（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-MOPD/blob/da71ee15e76a8849bd1a9610195c452007a79678/README.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": [
        "此处不是所有引擎的统一缓存精度；官方部署镜像与不同 attention backend 的实际页分配尚未实机测量。"
      ]
    },
    "mimov26distillqwen9b": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。 Qwen3_5ForCausalLM 实现；mamba_cache_dtype=auto、mamba_ssm_dtype=float32、无投机解码。",
      "finding": "32层=8 full GQA（KV4×256）+24 DeltaNet；conv 8192×3 BF16 + recurrent 32×128×128 FP32/层，固定状态51511296 B。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B/blob/1f09ddcc1a0ea45751ae0dfe2af20d533168c46f/config.json"
        },
        {
          "label": "固定 Qwen state dtype/shape 路径",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/qwen3_5.py#L385-L416"
        },
        {
          "label": "固定 DeltaNet dtype 与形状 helper",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/mamba/mamba_utils.py#L98-L130"
        },
        {
          "label": "固定 DeltaNet shape helper",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/mamba/mamba_utils.py#L280-L300"
        },
        {
          "label": "固定 Qwen recurrent dtype 覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/config.py#L853-L885"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": [
        "本 checkpoint 官方卡 quickstart 是 SGLang；本站数值是已明确选择的 vLLM 参考路径，不能称为厂商统一原生默认。建议标签加 vLLM 并加入 state-shape 直接证据。"
      ]
    },
    "minimaxm3": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "60层主 GQA KV4×128；前3层无 index，后57层各保存单个128维 BF16 index key=256B/token，不乘4个index query heads；无index scale。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/d95e30cb912fe3021bb9c2ba66a8ce60b6c937a9/config.json"
        },
        {
          "label": "固定 M3 主缓存 dtype / FullAttentionSpec",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/models/minimax_m3/nvidia/model.py#L493-L570"
        },
        {
          "label": "固定 M3 index key-only 单头 BF16 allocation",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/models/minimax_m3/common/indexer.py#L103-L161"
        },
        {
          "label": "固定 indexer auto->bf16 resolver",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/attention.py#L154-L158"
        },
        {
          "label": "官方 vLLM recipe，基础模型命令",
          "url": "https://github.com/vllm-project/recipes/blob/main/models/MiniMaxAI/MiniMax-M3.yaml#L343-L377"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": [
        "FP8/NVFP4/单独量化 checkpoint 不是本 profile。TP>4 的 KV 头复制与 index 副本不应按总payload/TP解释为真实每卡。"
      ]
    },
    "minimax": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "62层全 GQA、8 KV heads×128；BF16 主缓存253952 B/token；官方基础模型部署命令未指定KV量化。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/0a6536713b0b03f3be97494e5cf8d22f668a3164/config.json"
        },
        {
          "label": "固定 MiniMax-M2 Attention",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/minimax_m2.py#L138-L214"
        },
        {
          "label": "官方部署说明（固定修订）",
          "url": "https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/0a6536713b0b03f3be97494e5cf8d22f668a3164/docs/vllm_deploy_guide.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": []
    },
    "lagunas21": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "48层=12 full+36 SWA-512；两类层 KV8×128。SWA query heads虽为72，对KV heads无影响；原始BF16 checkpoint，官方vLLM命令未覆盖KV。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/poolside/Laguna-S-2.1/blob/0f573140834b11cfac0c2af97a101a7a69a13e22/config.json"
        },
        {
          "label": "固定 Laguna per-layer window 与 KV头数",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/laguna.py#L279-L305"
        },
        {
          "label": "固定 Laguna Attention",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/laguna.py#L412-L424"
        },
        {
          "label": "官方模型卡及vLLM命令（固定修订）",
          "url": "https://huggingface.co/poolside/Laguna-S-2.1/blob/0f573140834b11cfac0c2af97a101a7a69a13e22/README.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": []
    },
    "mistral": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "88层 full GQA、KV8×128；BF16主缓存360448 B/token。FP8权重量化配置不含KV量化scheme，不能把主缓存改为FP8。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/blob/22b2b868a15677cfa6061277ed2f653d1349a9ab/config.json"
        },
        {
          "label": "固定 Mistral3 文本模型初始化",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/mistral3.py#L485-L491"
        },
        {
          "label": "官方vLLM命令（固定修订）",
          "url": "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/blob/22b2b868a15677cfa6061277ed2f653d1349a9ab/README.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": []
    },
    "mistrallarge4": {
      "scope": "Transformers b5e1cb8b015b6e17162a3efb9fc43dc6bf454612，BF16计算的非量化 DynamicCache 参考路径；官方FP8权重checkpoint；非通用vLLM默认。",
      "finding": "72层中63独立KV owners：54×(512+64)+9×512；另外9层复用已有KV。44独立BF16 index-key owner×128，另28层复用top-k；每token82688B。当前main config与原站点固定修订完全相同。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/af26e7869af601441d5cdf75a1272469ddc3d059/config.json"
        },
        {
          "label": "固定Transformers latent缓存与KV共享",
          "url": "https://github.com/huggingface/transformers/blob/b5e1cb8b015b6e17162a3efb9fc43dc6bf454612/src/transformers/models/mistral_large4/modeling_mistral_large4.py#L335-L445"
        },
        {
          "label": "固定DSA单头key及index更新",
          "url": "https://github.com/huggingface/transformers/blob/b5e1cb8b015b6e17162a3efb9fc43dc6bf454612/src/transformers/models/mistral_large4/modeling_mistral_large4.py#L193-L263"
        },
        {
          "label": "固定DynamicCache dtype/index存储",
          "url": "https://github.com/huggingface/transformers/blob/b5e1cb8b015b6e17162a3efb9fc43dc6bf454612/src/transformers/cache_utils.py#L315-L349"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": [
        "HF eager/SDPA会生成展开KV及attention工作区；这些不属于长期缓存但可造成很大运行峰值。跨设备to()会产生副本，未作为理想逻辑payload计算。"
      ]
    },
    "shieldstral10": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "26层 full GQA、KV8×128；BF16 主缓存106496 B/token。官方部署示例用32768上限，config max_position_embeddings为262144；保留两者区分。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/mistralai/Shieldstral-1.0-3B/blob/003ec7e2b0bab5f0e6307edbaf186fa5822b76f5/config.json"
        },
        {
          "label": "官方部署示例与BF16 Transformers示例（固定修订）",
          "url": "https://huggingface.co/mistralai/Shieldstral-1.0-3B/blob/003ec7e2b0bab5f0e6307edbaf186fa5822b76f5/README.md"
        },
        {
          "label": "固定 Mistral3 文本初始化",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/mistral3.py#L485-L491"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": []
    },
    "step37": {
      "scope": "vLLM 1dfe9fbdae734fc2ee8c0494c543c24d25b563d6，BF16-capable NVIDIA CUDA，原始 checkpoint，cache_dtype=auto，无 KV 量化覆盖。",
      "finding": "45层=12 full+33 SWA-512；KV8×128，BF16缓存。基础FP8权重模型的命令未覆盖KV；NVFP4命令明确另选FP8 KV，不应套用。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/blob/5f6244077ac62e04eec3f320501ff8c2b293373a/config.json"
        },
        {
          "label": "固定Step3.7委托文本模型",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/step3p7.py#L69-L76"
        },
        {
          "label": "固定Step3.5 full/SWA及Attention缓存",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/models/step3p5.py#L146-L253"
        },
        {
          "label": "官方部署三种变体（固定修订）",
          "url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash/blob/5f6244077ac62e04eec3f320501ff8c2b293373a/README.md"
        },
        {
          "label": "固定 vLLM CacheConfig：auto KV",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/config/cache.py#L122-L129"
        },
        {
          "label": "固定 vLLM auto dtype 解析",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/utils/torch_utils.py#L525-L531"
        },
        {
          "label": "固定 vLLM Attention cache dtype 与量化覆盖",
          "url": "https://github.com/vllm-project/vllm/blob/1dfe9fbdae734fc2ee8c0494c543c24d25b563d6/vllm/model_executor/layers/attention/attention.py#L274-L326"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。",
        "官方示例含3-token MTP；本profile只计算基础decoder，排除MTP，需继续显式标示。"
      ],
      "unresolved": []
    },
    "step5preview": {
      "scope": "StepFun官方 reference inference，HF固定修订364b3b03fed2477aa0953fe855a2c4c8c48ee3cc；allocate_cache(dtype=bfloat16)，非vLLM profile。",
      "finding": "92 decoder层=23 sparse/full+69 SWA-512；KV4×192 BF16。每8token完成一个256维无scale FP8 summary；每full层 pending_key与pending_z各8×256 BF16，合8192B/层、188416B总计。config layer_types共有95项，最后3项是MTP，不能算作基础decoder。",
      "sources": [
        {
          "label": "该 checkpoint 官方 config（固定修订）",
          "url": "https://huggingface.co/stepfun-ai/Step-5-Preview/blob/364b3b03fed2477aa0953fe855a2c4c8c48ee3cc/config.json"
        },
        {
          "label": "官方固定BF16 allocate_cache",
          "url": "https://huggingface.co/stepfun-ai/Step-5-Preview/blob/364b3b03fed2477aa0953fe855a2c4c8c48ee3cc/inference/model.py#L2226-L2231"
        },
        {
          "label": "官方固定summary及双pending buffer",
          "url": "https://huggingface.co/stepfun-ai/Step-5-Preview/blob/364b3b03fed2477aa0953fe855a2c4c8c48ee3cc/inference/kernel.py#L1858-L1911"
        },
        {
          "label": "官方固定SWA ring及KV分片",
          "url": "https://huggingface.co/stepfun-ai/Step-5-Preview/blob/364b3b03fed2477aa0953fe855a2c4c8c48ee3cc/inference/model.py#L978-L1060"
        }
      ],
      "status": "verified",
      "assumptions": [
        "单条序列、文本 decoder、T 为已保留有效 token；仅 KV/索引/列明固定状态。",
        "不计模型权重、视觉/音频编码器、MTP/投机解码、临时激活/工作区、分页空隙及空闲预分配。",
        "数值是独立逻辑 payload；TP 除法仅理想均分，不代表实际逐卡内存或集群总驻留副本。"
      ],
      "unresolved": [
        "实现按max_tokens预分配主cache/summary，SWA ring固定512；min(T,512)和floor(T/8)是已利用payload，不是torch.zeros物理容量。",
        "TP8时每rank实际缓存1个KV group且summary/pending复制；直接总量/8不是真实每卡值。block_table/lengths元数据及RoPE表未包含。"
      ]
    },
    "kimik25": {
      "scope": "moonshotai/Kimi-K2.5@70a8ad02d86dbb7459abd22b8a4feca294d2c993; vLLM@e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1; 原版 checkpoint，显式 BF16 模型 dtype 与 auto KV（默认配方未覆盖 KV dtype），MLA 压缩缓存",
      "finding": "本条 61 层 × (512+64) × BF16 两字节公式正确；限定该 profile 已按限定实现核实。不是任何 checkpoint/硬件的通用默认。",
      "sources": [
        {
          "label": "Moonshot 固定模型配置",
          "url": "https://huggingface.co/moonshotai/Kimi-K2.5/blob/70a8ad02d86dbb7459abd22b8a4feca294d2c993/config.json"
        },
        {
          "label": "固定官方原版配方（对照第三方 FP8 KV variants）",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/recipes/models/moonshotai/Kimi-K2.5.yaml"
        },
        {
          "label": "auto KV dtype 定义",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py"
        },
        {
          "label": "auto 解析为 model dtype",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/utils/torch_utils.py#L539-L545"
        },
        {
          "label": "MLA 单 KV head 的缓存 spec",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L1458-L1495"
        }
      ],
      "status": "verified",
      "assumptions": [
        "数值是指定实现下的缓存内容字节，不是实测总显存，也不是 allocator 预留池容量。",
        "未启用 DCP/PCP、推测解码、缓存卸载或第三方量化 checkpoint；S 是保留的历史 token 总数，不是仅本轮新增 token。",
        "MLA 缓存为压缩 latent，不另乘 2(K/V) 或 attention/query heads；无 DCP 时 latent 在 TP ranks 复制。",
        "只核实固定源码与内容公式，未执行GPU实测；Docker标签未视为固定镜像digest。"
      ],
      "unresolved": [],
      "formula": "B × 70272 × T bytes；界面TP>1只作理想除法，不表示真实每卡。",
      "parameters": {
        "modelRevision": "70a8ad02d86dbb7459abd22b8a4feca294d2c993",
        "engine": "vllm",
        "engineRevision": "e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1",
        "layers": 61,
        "kvLoraRank": 512,
        "qkRopeHeadDim": 64,
        "mainBytes": 2,
        "latentRowBytes": 1152,
        "bytesPerToken": 70272,
        "maxContext": 262144
      }
    },
    "kimik26": {
      "scope": "moonshotai/Kimi-K2.6@81bcaaa7947338ce2641983d98947bca0cc1a4d4; vLLM@e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1; 原版 checkpoint，显式 BF16 模型 dtype 与 auto KV（默认配方未覆盖 KV dtype），MLA 压缩缓存",
      "finding": "本条 61 层 × (512+64) × BF16 两字节公式正确；限定该 profile 已按限定实现核实。不是任何 checkpoint/硬件的通用默认。",
      "sources": [
        {
          "label": "Moonshot 固定模型配置",
          "url": "https://huggingface.co/moonshotai/Kimi-K2.6/blob/81bcaaa7947338ce2641983d98947bca0cc1a4d4/config.json"
        },
        {
          "label": "固定官方原版配方（对照第三方 FP8 KV variants）",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/recipes/models/moonshotai/Kimi-K2.6.yaml"
        },
        {
          "label": "auto KV dtype 定义",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py"
        },
        {
          "label": "auto 解析为 model dtype",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/utils/torch_utils.py#L539-L545"
        },
        {
          "label": "MLA 单 KV head 的缓存 spec",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L1458-L1495"
        }
      ],
      "status": "verified",
      "assumptions": [
        "数值是指定实现下的缓存内容字节，不是实测总显存，也不是 allocator 预留池容量。",
        "未启用 DCP/PCP、推测解码、缓存卸载或第三方量化 checkpoint；S 是保留的历史 token 总数，不是仅本轮新增 token。",
        "MLA 缓存为压缩 latent，不另乘 2(K/V) 或 attention/query heads；无 DCP 时 latent 在 TP ranks 复制。",
        "只核实固定源码与内容公式，未执行GPU实测；Docker标签未视为固定镜像digest。"
      ],
      "unresolved": [],
      "formula": "B × 70272 × T bytes；界面TP>1只作理想除法，不表示真实每卡。",
      "parameters": {
        "modelRevision": "81bcaaa7947338ce2641983d98947bca0cc1a4d4",
        "engine": "vllm",
        "engineRevision": "e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1",
        "layers": 61,
        "kvLoraRank": 512,
        "qkRopeHeadDim": 64,
        "mainBytes": 2,
        "latentRowBytes": 1152,
        "bytesPerToken": 70272,
        "maxContext": 262144
      }
    },
    "kimik27code": {
      "scope": "moonshotai/Kimi-K2.7-Code@da26fe4d239c5c9d14507f1f909894201f911ede; vLLM@e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1; 原版 checkpoint，显式 BF16 模型 dtype 与 auto KV（默认配方未覆盖 KV dtype），MLA 压缩缓存",
      "finding": "本条 61 层 × (512+64) × BF16 两字节公式正确；限定该 profile 已按限定实现核实。不是任何 checkpoint/硬件的通用默认。",
      "sources": [
        {
          "label": "Moonshot 固定模型配置",
          "url": "https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/da26fe4d239c5c9d14507f1f909894201f911ede/config.json"
        },
        {
          "label": "固定官方原版配方（对照第三方 FP8 KV variants）",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/recipes/models/moonshotai/Kimi-K2.7-Code.yaml"
        },
        {
          "label": "auto KV dtype 定义",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py"
        },
        {
          "label": "auto 解析为 model dtype",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/utils/torch_utils.py#L539-L545"
        },
        {
          "label": "MLA 单 KV head 的缓存 spec",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L1458-L1495"
        }
      ],
      "status": "verified",
      "assumptions": [
        "数值是指定实现下的缓存内容字节，不是实测总显存，也不是 allocator 预留池容量。",
        "未启用 DCP/PCP、推测解码、缓存卸载或第三方量化 checkpoint；S 是保留的历史 token 总数，不是仅本轮新增 token。",
        "MLA 缓存为压缩 latent，不另乘 2(K/V) 或 attention/query heads；无 DCP 时 latent 在 TP ranks 复制。",
        "只核实固定源码与内容公式，未执行GPU实测；Docker标签未视为固定镜像digest。"
      ],
      "unresolved": [],
      "formula": "B × 70272 × T bytes；界面TP>1只作理想除法，不表示真实每卡。",
      "parameters": {
        "modelRevision": "da26fe4d239c5c9d14507f1f909894201f911ede",
        "engine": "vllm",
        "engineRevision": "e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1",
        "layers": 61,
        "kvLoraRank": 512,
        "qkRopeHeadDim": 64,
        "mainBytes": 2,
        "latentRowBytes": 1152,
        "bytesPerToken": 70272,
        "maxContext": 262144
      }
    },
    "kimik3": {
      "scope": "moonshotai/Kimi-K3@c5d1dd4c428bd1ce8b88c5044f3b6ccde9e3b721; vLLM@e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1; NVIDIA Blackwell, TOKENSPEED_MLA, FP8 latent + BF16 conv + FP32 KDA",
      "finding": "本条 FP8 latent 576 B × 24 层与 69 层 KDA 固定状态公式正确。限定此 profile 已按限定实现核实。KDA 是固定单状态内容，不是与 S 线性增长的全历史 KV，也不是整个运行时全部状态副本。",
      "sources": [
        {
          "label": "Moonshot 固定模型配置",
          "url": "https://huggingface.co/moonshotai/Kimi-K3/blob/c5d1dd4c428bd1ce8b88c5044f3b6ccde9e3b721/config.json"
        },
        {
          "label": "固定 Blackwell 配方",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/recipes/models/moonshotai/Kimi-K3.yaml"
        },
        {
          "label": "MLA 实际 cache spec",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/kimi_k3/nvidia/mla.py#L426-L443"
        },
        {
          "label": "KDA dtype 与 shape",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py"
        },
        {
          "label": "TokenSpeed KDA packed recipe（独立实现佐证）",
          "url": "https://github.com/lightseekorg/tokenspeed/blob/16889c2c00d72f6cd6dcde2bde5bfeaaaacd5e26/python/tokenspeed/runtime/layers/attention/kv_cache/recipes/kimi_k3.py"
        }
      ],
      "status": "verified",
      "assumptions": [
        "数值是指定实现下的缓存内容字节，不是实测总显存，也不是 allocator 预留池容量。",
        "未启用 DCP/PCP、推测解码、缓存卸载或第三方量化 checkpoint；S 是保留的历史 token 总数，不是仅本轮新增 token。",
        "MLA 缓存为压缩 latent，不另乘 2(K/V) 或 attention/query heads；无 DCP 时 latent 在 TP ranks 复制。",
        "MLA 的额外 64 通道在 NoPE 下未旋转但仍保存。",
        "不含 FP8 每层 scalar calibration、KDA 前缀/checkpoint/rollback 副本、spec state 和草稿缓存。",
        "只核实固定源码与内容公式，未执行GPU实测；Docker标签未视为固定镜像digest。"
      ],
      "unresolved": [],
      "formula": "T>0: B × (13824 × T + 449372160); T=0: 0 bytes；界面TP>1只作理想除法，不表示真实每卡。",
      "parameters": {
        "modelRevision": "c5d1dd4c428bd1ce8b88c5044f3b6ccde9e3b721",
        "engine": "vllm",
        "engineRevision": "e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1",
        "layers": 93,
        "fullAttentionLayers": 24,
        "kdaLayers": 69,
        "kvLoraRank": 512,
        "qkRopeHeadDim": 64,
        "mainBytes": 1,
        "latentRowBytes": 576,
        "bytesPerToken": 13824,
        "maxContext": 1048576,
        "kda": {
          "layers": 69,
          "heads": 96,
          "headDim": 128,
          "convKernel": 4,
          "recurrentBytes": 4,
          "convBytes": 2,
          "unshardedBytesPerLiveSequence": 449372160,
          "perTP8BytesPerLiveSequence": 56171520,
          "formula": "69 * [96*128*128*4 + 3*96*128*(4-1)*2] / TP"
        }
      }
    },
    "kimik31pro": {
      "scope": "moonshotai/Kimi-K3.1-Pro@6d0a0202956796543364d666ab19c6ad1a494ee1; vLLM@e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1; NVIDIA Blackwell SM10x, FLASHINFER_MLA_SPARSE, 显式 FP8 latent + FP8 indexer(auto) + BF16 conv + FP32 KDA; TokenSpeed@16889c2c00d72f6cd6dcde2bde5bfeaaaacd5e26 独立证据链",
      "finding": "Kimi 专属 config→indexer format→每层独立 cache wiring→packed recipe/test 的证据链已闭合，此前信息不足 中的 132 B/index token 现可确认。已为这个已指定profile开放内容公式；仍不得称运行时实测或通用物理分配。",
      "sources": [
        {
          "label": "Moonshot 固定模型配置",
          "url": "https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/6d0a0202956796543364d666ab19c6ad1a494ee1/config.json"
        },
        {
          "label": "vLLM Kimi 专属 FP8/MXFP4 index row 与 packed size",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/kimi_k3/common/dsa_cache.py"
        },
        {
          "label": "vLLM 逐 DSA 层独立 indexer 及 MLA 连接",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/kimi_k3/nvidia/dsa.py#L600-L670"
        },
        {
          "label": "MLA backend 特定 FP8 layout canonicalization",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L362-L385"
        },
        {
          "label": "KDA dtype 与 shape",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py"
        },
        {
          "label": "固定 K3.1 配方与发布限制",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/recipes/models/moonshotai/Kimi-K3.1-Pro.yaml"
        },
        {
          "label": "TokenSpeed FP8 index row 定义",
          "url": "https://github.com/lightseekorg/tokenspeed/blob/16889c2c00d72f6cd6dcde2bde5bfeaaaacd5e26/python/tokenspeed/runtime/layers/attention/configs/dsa.py"
        },
        {
          "label": "TokenSpeed Kimi-K3.1 cache plane 测试",
          "url": "https://github.com/lightseekorg/tokenspeed/blob/16889c2c00d72f6cd6dcde2bde5bfeaaaacd5e26/test/runtime/test_kimi_k31.py#L355-L403"
        },
        {
          "label": "TokenSpeed K3.1 NoPE 仍保留 64 维",
          "url": "https://github.com/lightseekorg/tokenspeed/blob/16889c2c00d72f6cd6dcde2bde5bfeaaaacd5e26/python/tokenspeed/runtime/models/kimi_k31.py#L213-L221"
        },
        {
          "label": "vLLM 发布说明的独立数值校验",
          "url": "https://vllm.ai/blog/2026-11-04-kimi-k31-pro"
        }
      ],
      "status": "verified",
      "assumptions": [
        "数值是指定实现下的缓存内容字节，不是实测总显存，也不是 allocator 预留池容量。",
        "未启用 DCP/PCP、推测解码、缓存卸载或第三方量化 checkpoint；S 是保留的历史 token 总数，不是仅本轮新增 token。",
        "MLA 缓存为压缩 latent，不另乘 2(K/V) 或 attention/query heads；无 DCP 时 latent 在 TP ranks 复制。",
        "每个 DSA 层有一个 128 维 index K，不乘 32 个 query heads；每 token 一个 FP32 scale，合计 132 B。",
        "top-k=2048 只限制每次稀疏读取/计算，不把全历史缓存截成 2048 token。",
        "NoPE 不删除 64 维额外通道，因此 latent 仍是 576。",
        "不含 FP8 每层 scalar calibration、KDA 前缀/checkpoint/rollback 副本、spec state 和草稿缓存。",
        "强制限定 FLASHINFER_MLA_SPARSE / SM10x；FLASHMLA_SPARSE 与 SM120 路径会规范化到 fp8_ds_mla（656 B latent），不能套用 576 B。",
        "只核实固定源码与内容公式，未执行GPU实测；Docker标签未视为固定镜像digest。"
      ],
      "unresolved": [],
      "formula": "T>0: B × (16992 × T + 449372160); T=0: 0 bytes；界面TP>1只作理想除法，不表示真实每卡。",
      "parameters": {
        "modelRevision": "6d0a0202956796543364d666ab19c6ad1a494ee1",
        "engine": "vllm",
        "engineRevision": "e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1",
        "corroboratingEngine": "tokenspeed",
        "corroboratingEngineRevision": "16889c2c00d72f6cd6dcde2bde5bfeaaaacd5e26",
        "layers": 93,
        "fullAttentionLayers": 24,
        "indexLayers": 24,
        "kdaLayers": 69,
        "kvLoraRank": 512,
        "qkRopeHeadDim": 64,
        "mainBytes": 1,
        "latentRowBytes": 576,
        "indexHeadDim": 128,
        "indexQueryHeads": 32,
        "indexTopK": 2048,
        "indexBytes": 1,
        "indexScaleBytes": 4,
        "indexRowBytes": 132,
        "bytesPerToken": 16992,
        "maxContext": 1048576,
        "kda": {
          "layers": 69,
          "heads": 96,
          "headDim": 128,
          "convKernel": 4,
          "recurrentBytes": 4,
          "convBytes": 2,
          "unshardedBytesPerLiveSequence": 449372160,
          "perTP8BytesPerLiveSequence": 56171520,
          "formula": "69 * [96*128*128*4 + 3*96*128*(4-1)*2] / TP"
        }
      }
    },
    "hy3": {
      "scope": "vLLM cache_dtype=auto；BF16-capable CUDA；no KV override",
      "finding": "核对固定模型配置及所选vLLM默认缓存实现，未发现该限定配置的数值错误；保留BF16主KV/conv与独立FP32 recurrent（适用时）的区别。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/Hy3/blob/a960ebc3da325ba167f069f76c41eb62c9280d22/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "327680 * T"
    },
    "hy4preview": {
      "scope": "vLLM NVIDIA；BF16 main latent MLA + 128 FP8 bytes and 4 FP32 scale bytes/index row；21 owners",
      "finding": "主latent MLA 576维BF16，21独立index owner各132 B/行的实现链通过。官方快速示例为FP8权重+MTP变体，不等同本站所选原始checkpoint、无MTP内容口径。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/Hy4-preview/blob/705d81ee51566a186d645b74c974d642ef2828fe/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/hy_v4/nvidia/attention.py#L133-L146"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/hy_v4/nvidia/attention.py#L282-L307"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "92628*T = (78*576*2 + 21*132)*T"
    },
    "uimate9b": {
      "scope": "vLLM auto；BF16 KV/conv、FP32 recurrent；no speculative/MTP",
      "finding": "核对固定模型配置及所选vLLM默认缓存实现，未发现该限定配置的数值错误；保留BF16主KV/conv与独立FP32 recurrent（适用时）的区别。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/UI-Mate-9B/blob/05dd5f2975195a5bb03d4363e8767f12158c8421/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L1251-L1320"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L104-L139"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L280-L302"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "32768 * T + 51511296 (T > 0; T=0 => 0)"
    },
    "uimate27b": {
      "scope": "vLLM auto；BF16 KV/conv、FP32 recurrent；no speculative/MTP",
      "finding": "核对固定模型配置及所选vLLM默认缓存实现，未发现该限定配置的数值错误；保留BF16主KV/conv与独立FP32 recurrent（适用时）的区别。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/UI-Mate-27B/blob/3ade2378fc84032d5017c1a9c93c4eaa77d65e57/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L1251-L1320"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L104-L139"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L280-L302"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "65536 * T + 153944064 (T > 0; T=0 => 0)"
    },
    "uimatedemocua27b": {
      "scope": "vLLM auto；BF16 KV/conv、FP32 recurrent；no speculative/MTP",
      "finding": "核对固定模型配置及所选vLLM默认缓存实现，未发现该限定配置的数值错误；保留BF16主KV/conv与独立FP32 recurrent（适用时）的区别。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/UI-Mate-democua-27B/blob/a76300b510ee8c9b509cdd933ff5e31a994a3400/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L1251-L1320"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L104-L139"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L280-L302"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "65536 * T + 153944064 (T > 0; T=0 => 0)"
    },
    "contextpilot8b": {
      "scope": "Official Transformers torch_dtype=auto；BF16 config；DynamicCache inherits projected KV dtype",
      "finding": "官方Transformers示例配合BF16 config，DynamicCache沿用投影后K/V dtype；未发现该明确实现下的数值错误。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/ContextPilot-8B/blob/595abeaed72e4a4d9da94d04917f0618989a5553/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/huggingface/transformers/blob/de35c2b1bff085ac422e45f2dc63618c7789fe52/src/transformers/cache_utils.py#L117-L143"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "147456 * T"
    },
    "contextpilot14b": {
      "scope": "Official Transformers torch_dtype=auto；BF16 config；DynamicCache inherits projected KV dtype",
      "finding": "官方Transformers示例配合BF16 config，DynamicCache沿用投影后K/V dtype；未发现该明确实现下的数值错误。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/ContextPilot-14B/blob/8eafd8356c5c5e713e92e623c8f1d346de6d0f95/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/huggingface/transformers/blob/de35c2b1bff085ac422e45f2dc63618c7789fe52/src/transformers/cache_utils.py#L117-L143"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "163840 * T"
    },
    "contextpilote4b": {
      "scope": "Transformers DynamicCache；BF16；4 global + 20 sliding owners; 18 reuse layers",
      "finding": "修正更新后持久缓存长度为min(T,511)，attention窗口仍512；4全局+20滑窗owner，18层共享。T≥512时比旧值减少40,960 B/序列。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/ContextPilot-E4B/blob/a7ba41c18ddcd334f23f63ef7a608ca46bfe3aa4/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/huggingface/transformers/blob/de35c2b1bff085ac422e45f2dc63618c7789fe52/src/transformers/cache_utils.py#L117-L143"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/huggingface/transformers/blob/de35c2b1bff085ac422e45f2dc63618c7789fe52/src/transformers/cache_utils.py#L241-L258"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/huggingface/transformers/blob/de35c2b1bff085ac422e45f2dc63618c7789fe52/src/transformers/models/gemma4/configuration_gemma4.py#L228-L253"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/huggingface/transformers/blob/de35c2b1bff085ac422e45f2dc63618c7789fe52/src/transformers/cache_utils.py#L1930-L1938"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "16384*T + 40960*min(T,511) [HF DynamicCache retained tensor shape]"
    },
    "wevisdoc2b": {
      "scope": "vLLM cache_dtype=auto；BF16-capable CUDA；no KV override",
      "finding": "核对固定模型配置及所选vLLM默认缓存实现，未发现该限定配置的数值错误；保留BF16主KV/conv与独立FP32 recurrent（适用时）的区别。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/WeVisDoc-2B/blob/86ecb7330566296b695d4f91aea807a260cc84a7/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "114688 * T"
    },
    "wevisdoc4b": {
      "scope": "vLLM cache_dtype=auto；BF16-capable CUDA；no KV override",
      "finding": "核对固定模型配置及所选vLLM默认缓存实现，未发现该限定配置的数值错误；保留BF16主KV/conv与独立FP32 recurrent（适用时）的区别。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/WeVisDoc-4B/blob/1754bfa79ba1a6e24aacc001da2e7bb39b9fbc34/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "147456 * T"
    },
    "youtuparsingomni": {
      "scope": "Official vLLM 0.19.0 + Youtu plugin；explicit BF16；DeepseekV2 latent MLA",
      "finding": "官方vLLM0.19与Youtu插件显式BF16，采用DeepseekV2 latent MLA；维持40层576维公式，不混入HF expanded K/V路径。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/Youtu-Parsing-Omni/blob/77d243bfd75c3e79b9e322f0668654ca22b8b187/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/TencentCloudADP/youtu-parsing/blob/c6b57f755b5bb2c6d4ed2887c8aaf983a322c50b/youtu_parsing_omni/scripts/vllm.sh#L45-L51"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/TencentCloudADP/youtu-parsing/blob/c6b57f755b5bb2c6d4ed2887c8aaf983a322c50b/youtu_parsing_omni/vllm-plugin-vita-omni/vita_omni/vita_for_causal_lm.py#L36-L42"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "46080*T = 40*(512+64)*2*T"
    },
    "hymemory30ba3b": {
      "scope": "Official-tested vLLM 0.31.0；CUDA Ampere+；dtype=auto resolves to BF16, cache=auto follows it; config has no dtype field",
      "finding": "config未声明dtype；限定官方测试vLLM0.31、Ampere+ CUDA，权重元数据与运行时auto解析为BF16，KV随运行dtype。其它硬件/参数另算。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/tencent/Hy-Memory-30B-A3B/blob/64a460a94d5498b7316fb02395c72d01c47a7347/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/v0.31.0/vllm/config/model.py#L2355-L2438"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/v0.31.0/vllm/transformers_utils/model_arch_config_convertor.py#L220-L258"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/v0.31.0/vllm/platforms/cuda.py#L255-L264"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/v0.31.0/vllm/model_executor/models/hy_v3.py#L282-L300"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "98304 * T"
    },
    "qwen": {
      "scope": "vLLM auto；BF16 KV/conv、FP32 recurrent；no speculative/MTP",
      "finding": "核对固定模型配置及所选vLLM默认缓存实现，未发现该限定配置的数值错误；保留BF16主KV/conv与独立FP32 recurrent（适用时）的区别。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B/blob/8472618112abcbd45acbcdc58436aff4233c23f7/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L1251-L1320"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L104-L139"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L280-L302"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "30720 * T + 192061440 (T > 0; T=0 => 0)"
    },
    "qwen3827b": {
      "scope": "vLLM auto；BF16 KV/conv、FP32 recurrent；no speculative/MTP",
      "finding": "核对固定模型配置及所选vLLM默认缓存实现，未发现该限定配置的数值错误；保留BF16主KV/conv与独立FP32 recurrent（适用时）的区别。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/Qwen/Qwen3.8-27B/blob/37561816e66b5b9153b1b64c10968c18c1e6f4a0/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L1251-L1320"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L104-L139"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L280-L302"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "65536 * T + 153944064 (T > 0; T=0 => 0)"
    },
    "qwen38flashnext": {
      "scope": "vLLM NVIDIA default BF16 QSA；multimodal MRoPE enabled；num_speculative_tokens=0；BF16 main/index/ring/conv/PLE + FP32 recurrent",
      "finding": "QSA、MRoPE、PLE与recurrent存储公式通过；更新旧失效链接。限定MRoPE启用、无推测解码：ring为1120 B/层；language-model-only会变为1024 B，MTP可能扩容。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/812d96e10a0dd82b69d749e68693698cf6c36daa/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/qwen4_exp/nvidia/qsa_indexer.py#L149-L181"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/qwen4_exp/common/qsa_indexer.py#L817-L881"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/qwen4_exp/nvidia/qsa.py#L281-L426"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/qwen4_exp/nvidia/ple_layer.py#L98-L105"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L1251-L1320"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L280-L302"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。",
        "启用多模态MRoPE、num_speculative_tokens=0；不能与language-model-only或推测解码配置混用。"
      ],
      "unresolved": [],
      "formula": "24576*T + 3072*floor(T/4) + 115655808 (T > 0; T=0 => 0)"
    },
    "qwen3824ta95b": {
      "scope": "vLLM auto；BF16 KV/conv、FP32 recurrent；no speculative/MTP",
      "finding": "核对固定模型配置及所选vLLM默认缓存实现，未发现该限定配置的数值错误；保留BF16主KV/conv与独立FP32 recurrent（适用时）的区别。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/4f885e51d6bbfee5866446d56d3f96f336c39723/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/config/cache.py#L238-L250"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L1251-L1320"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L104-L139"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L280-L302"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "94208 * T + 587292672 (T > 0; T=0 => 0)"
    },
    "qwen38omniflash": {
      "scope": "历史API产品占位；公开高层架构，不是已核实公开权重checkpoint。",
      "finding": "已公开Gated DeltaNet + QSA / Thinker–Talker架构；完整缓存维度、dtype与实际实现证据仍不足，因此不生成曲线。",
      "sources": [
        {
          "label": "实现/运行证据",
          "url": "https://arxiv.org/html/2609.25611v1#S2.SS1"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/QwenLM/Qwen-Live-Harness/blob/main/README.md"
        },
        {
          "label": "实现/运行证据",
          "url": "https://huggingface.co/api/models?author=Qwen&search=Qwen3.8-Omni&limit=100"
        }
      ],
      "status": "pending",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [
        "未核实公开checkpoint的层数/heads/维度与缓存dtype、量化scale/packing、完整状态结构。"
      ],
      "formula": null
    },
    "qwen38omniflashrealtime": {
      "scope": "历史API产品占位；公开高层架构，不是已核实公开权重checkpoint。",
      "finding": "已公开Gated DeltaNet + QSA / Thinker–Talker架构；完整缓存维度、dtype与实际实现证据仍不足，因此不生成曲线。",
      "sources": [
        {
          "label": "实现/运行证据",
          "url": "https://arxiv.org/html/2609.25611v1#S2.SS1"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/QwenLM/Qwen-Live-Harness/blob/main/README.md"
        },
        {
          "label": "实现/运行证据",
          "url": "https://huggingface.co/api/models?author=Qwen&search=Qwen3.8-Omni&limit=100"
        }
      ],
      "status": "pending",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [
        "未核实公开checkpoint的层数/heads/维度与缓存dtype、量化scale/packing、完整状态结构。",
        "精确发布日期有来源差异，目前只保留2026年9月。"
      ],
      "formula": null
    },
    "qwendrive104b": {
      "scope": "Official Qwen-Drive VQA code；Transformers 5.14.0 + FLA 0.5.1；BF16 KV/conv and FP32 recurrent；not planning/perception system",
      "finding": "补齐官方VQA→Transformers5.14→FLA0.5.1证据链，解除信息不足。主KV/conv为BF16、recurrent为FP32；HF conv保留4槽，不能套用vLLM的3槽。",
      "sources": [
        {
          "label": "官方配置（固定版本）",
          "url": "https://huggingface.co/Qwen/Qwen-Drive-1.0-4B/blob/28484089a7cc8c335cf5089fb0745cf7c49b6eaa/config.json"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/QwenLM/Qwen-Drive-1.0/blob/28091c1532e869bc7aee91fc0aef6b3e6fd0b2e0/scripts/run_vqa.py#L47-L57"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/QwenLM/Qwen-Drive-1.0/blob/28091c1532e869bc7aee91fc0aef6b3e6fd0b2e0/src/qwen_drive/modeling_qwen_drive.py#L102-L109"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/QwenLM/Qwen-Drive-1.0/blob/28091c1532e869bc7aee91fc0aef6b3e6fd0b2e0/src/qwen_drive/modeling_qwen_drive.py#L394-L421"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/QwenLM/Qwen-Drive-1.0/blob/28091c1532e869bc7aee91fc0aef6b3e6fd0b2e0/pyproject.toml#L10-L23"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/huggingface/transformers/blob/v5.14.0/src/transformers/models/qwen3_5/modeling_qwen3_5.py#L485-L491"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/huggingface/transformers/blob/v5.14.0/src/transformers/cache_utils.py#L942-L1026"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/fla-org/flash-linear-attention/blob/v0.5.1/fla/ops/common/chunk_delta_h.py#L626-L632"
        },
        {
          "label": "实现/运行证据",
          "url": "https://github.com/fla-org/flash-linear-attention/blob/v0.5.1/fla/ops/gated_delta_rule/fused_recurrent.py#L208-L215"
        }
      ],
      "status": "verified",
      "assumptions": [
        "仅计算所选实现的一份逻辑持久缓存，不是GPU实测。",
        "排除权重、工作区、分配器/分页空闲和推测解码/MTP额外状态；TP仅为理想均分。"
      ],
      "unresolved": [],
      "formula": "32768*T + 51904512 (T > 0; T=0 => 0)"
    },
    "glm51": {
      "scope": "vLLM e5f2ea9；SM100；显式FLASHMLA_SPARSE + fp8_ds_mla主行656 B；FP8 index132 B；未启用推测解码。",
      "finding": "未发现该明确后端基线的几何或公式错误。656 B 不等于所有 Blackwell 引擎的FP8行宽，FlashInfer plain FP8 是不同布局。；本轮已固定实现与配方来源，明确attention backend和kv_cache_dtype。；1Mi 结果60.0234 GiB只是公式外推，超过原生202,752 tokens；原生上限约11.6061 GiB。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/zai-org/GLM-5.1/blob/main/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/recipes/blob/2adff3774c56906c8292568eda004c4d0e00e6a8/models/zai-org/GLM-5.1.yaml"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L121-L158"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/v1/attention/backends/mla/flashmla_sparse.py#L77-L108"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L362-L378"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L1467-L1488"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v32/attention.py#L170-L194"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v32/attention.py#L93-L105"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/utils.py#L1019-L1050"
        }
      ],
      "status": "verified",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "SM100；FLASHMLA_SPARSE；量化KV解析为fp8_ds_mla；index132B"
      ],
      "unresolved": [],
      "formula": "61464 × T = (78×656 + 78×132)×T"
    },
    "glm52": {
      "scope": "vLLM e5f2ea9；SM100；显式FLASHMLA_SPARSE + fp8_ds_mla主行656 B；FP8 index132 B；未启用推测解码。",
      "finding": "未发现该明确后端基线的几何或公式错误。656 B 不等于所有 Blackwell 引擎的FP8行宽，FlashInfer plain FP8 是不同布局。；本轮已固定实现与配方来源，明确attention backend和kv_cache_dtype。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/zai-org/GLM-5.2/blob/main/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/recipes/blob/2adff3774c56906c8292568eda004c4d0e00e6a8/models/zai-org/GLM-5.2.yaml"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L121-L158"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/v1/attention/backends/mla/flashmla_sparse.py#L77-L108"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L362-L378"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L1467-L1488"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v32/attention.py#L170-L194"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v32/attention.py#L93-L105"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/utils.py#L1019-L1050"
        }
      ],
      "status": "verified",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "SM100；FLASHMLA_SPARSE；量化KV解析为fp8_ds_mla；index132B"
      ],
      "unresolved": [],
      "formula": "53940 × T = (78×656 + 21×132)×T"
    },
    "glm53flash": {
      "scope": "vLLM e5f2ea9；SM100 FlashInfer MLA sparse plain FP8；无推测解码，raw K+gate tail=4槽，conv=3槽。",
      "finding": "主MLA 512 B、pooled index132 B、FP32 recurrent、BF16 conv均有实现依据。；2,048 B tail由4 slots×2向量×128×2 B得出，仅num_speculative_tokens=0时成立。官方MTP示例启用5 draft tokens后ring扩为16 slots，每层tail8,192 B；KDA conv宽度从3扩为8。；明确排除推测解码带来的主干状态扩容，不仅草稿层。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/zai-org/GLM-5.3-Flash/blob/main/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/recipes/blob/2adff3774c56906c8292568eda004c4d0e00e6a8/models/zai-org/GLM-5.3-Flash.yaml"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L121-L158"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/v1/attention/backends/mla/flashinfer_mla_sparse.py#L137-L155"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L1467-L1488"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/glm5next/common/attention.py#L125-L180"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/glm5next/common/attention.py#L252-L276"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/glm5next/common/kda.py#L212-L236"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L140-L158"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py#L304-L327"
        }
      ],
      "status": "verified",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "SM100 FlashInfer MLA sparse plain FP8 main512B；index132B；无推测解码；recurrent自动FP32，conv自动BF16；tail含raw K+gate score"
      ],
      "unresolved": [],
      "formula": "11×512×T + 11×132×floor(T/4) + 11×2048 + 34×[64×128×128×4 + 3×64×128×3×2] (T>0)"
    },
    "glm53": {
      "scope": "vLLM e5f2ea9；SM100；显式FLASHMLA_SPARSE + fp8_ds_mla主行656 B；FP8 index132 B；未启用推测解码。",
      "finding": "未发现该明确后端基线的几何或公式错误。656 B 不等于所有 Blackwell 引擎的FP8行宽，FlashInfer plain FP8 是不同布局。；本轮已固定实现与配方来源，明确attention backend和kv_cache_dtype。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/zai-org/GLM-5.3/blob/main/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/recipes/blob/2adff3774c56906c8292568eda004c4d0e00e6a8/models/zai-org/GLM-5.3.yaml"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/config.py#L121-L158"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/v1/attention/backends/mla/flashmla_sparse.py#L77-L108"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L362-L378"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/attention/mla_attention.py#L1467-L1488"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v32/attention.py#L170-L194"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v32/attention.py#L93-L105"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/utils.py#L1019-L1050"
        }
      ],
      "status": "verified",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "SM100；FLASHMLA_SPARSE；量化KV解析为fp8_ds_mla；index132B"
      ],
      "unresolved": [],
      "formula": "53940 × T = (78×656 + 21×132)×T"
    },
    "deepseek": {
      "scope": "vLLM e5f2ea9；FLASHMLA_SPARSE_DSV4 / fp8_ds_mla + FP8 index；主干KV/index逻辑payload加满窗口压缩器容量预算。",
      "finding": "584 B主/SWA行与132 B FP8 index行在选定FlashMLA packed路径正确；只c4层建index，SWA与压缩行并存，不扣窗口，floor边界正确。；“CUDA默认”太宽：当前CUDA选择器在SM12默认FlashInfer；plain-row后端auto为BF16、显式fp8为普通FP8，均不等于584 B packed。；每c4 81,920 B=8×2×2×(512+128)×4、每c128 524,288 B=128×2×512×4，是压缩器满窗口状态预算。当前T=1即加全部状态常数，不能称短序列已占用payload或真实分配。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/blob/main/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/v1/attention/backends/mla/flashmla_sparse.py#L77-L108"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/attention.py"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/compressor.py#L138-L192"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/compressor.py#L247-L301"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/nvidia/model.py#L539-L570"
        }
      ],
      "status": "reference",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "FLASHMLA_SPARSE_DSV4；fp8_ds_mla；FP8 index（非可选MXFP4）",
        "T>0加入满窗口压缩器容量预算，不声称短序列实际已用全部状态容量",
        "T>0使用满窗口压缩状态容量预算；不是短序列已用状态的精确值。"
      ],
      "unresolved": [
        "短序列实际活跃压缩状态及物理分配量未建模；当前明确使用满窗口容量预算。"
      ],
      "formula": "584×[43×min(T,128)+21×floor(T/4)+20×floor(T/128)] + 132×21×floor(T/4) + 12206080 (T>0)"
    },
    "deepseekpro": {
      "scope": "vLLM e5f2ea9；FLASHMLA_SPARSE_DSV4 / fp8_ds_mla + FP8 index；主干KV/index逻辑payload加满窗口压缩器容量预算。",
      "finding": "584 B主/SWA行与132 B FP8 index行在选定FlashMLA packed路径正确；只c4层建index，SWA与压缩行并存，不扣窗口，floor边界正确。；“CUDA默认”太宽：当前CUDA选择器在SM12默认FlashInfer；plain-row后端auto为BF16、显式fp8为普通FP8，均不等于584 B packed。；每c4 81,920 B=8×2×2×(512+128)×4、每c128 524,288 B=128×2×512×4，是压缩器满窗口状态预算。当前T=1即加全部状态常数，不能称短序列已占用payload或真实分配。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/v1/attention/backends/mla/flashmla_sparse.py#L77-L108"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/attention.py"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/compressor.py#L138-L192"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/compressor.py#L247-L301"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/nvidia/model.py#L539-L570"
        }
      ],
      "status": "reference",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "FLASHMLA_SPARSE_DSV4；fp8_ds_mla；FP8 index（非可选MXFP4）",
        "T>0加入满窗口压缩器容量预算，不声称短序列实际已用全部状态容量",
        "T>0使用满窗口压缩状态容量预算；不是短序列已用状态的精确值。"
      ],
      "unresolved": [
        "短序列实际活跃压缩状态及物理分配量未建模；当前明确使用满窗口容量预算。"
      ],
      "formula": "584×[61×min(T,128)+30×floor(T/4)+31×floor(T/128)] + 132×30×floor(T/4) + 18710528 (T>0)"
    },
    "deepseekv4flashvisionexp": {
      "scope": "vLLM e5f2ea9；FLASHMLA_SPARSE_DSV4 / fp8_ds_mla + FP8 index；主干KV/index逻辑payload加满窗口压缩器容量预算。",
      "finding": "584 B主/SWA行与132 B FP8 index行在选定FlashMLA packed路径正确；只c4层建index，SWA与压缩行并存，不扣窗口，floor边界正确。；“CUDA默认”太宽：当前CUDA选择器在SM12默认FlashInfer；plain-row后端auto为BF16、显式fp8为普通FP8，均不等于584 B packed。；每c4 81,920 B=8×2×2×(512+128)×4、每c128 524,288 B=128×2×512×4，是压缩器满窗口状态预算。当前T=1即加全部状态常数，不能称短序列已占用payload或真实分配。；这是decoder-token缓存口径；T必须包括注入decoder的图像token，不包含vision encoder、图像预处理和双向image-span prefill工作区。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp/blob/main/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/v1/attention/backends/mla/flashmla_sparse.py#L77-L108"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/attention.py"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/compressor.py#L138-L192"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/compressor.py#L247-L301"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/nvidia/model.py#L539-L570"
        }
      ],
      "status": "reference",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "FLASHMLA_SPARSE_DSV4；fp8_ds_mla；FP8 index（非可选MXFP4）",
        "T>0加入满窗口压缩器容量预算，不声称短序列实际已用全部状态容量",
        "T>0使用满窗口压缩状态容量预算；不是短序列已用状态的精确值。"
      ],
      "unresolved": [
        "短序列实际活跃压缩状态及物理分配量未建模；当前明确使用满窗口容量预算。"
      ],
      "formula": "584×[43×min(T,128)+21×floor(T/4)+20×floor(T/128)] + 132×21×floor(T/4) + 12206080 (T>0)"
    },
    "deepseekv4pro0813": {
      "scope": "vLLM e5f2ea9；FLASHMLA_SPARSE_DSV4 / fp8_ds_mla + FP8 index；主干KV/index逻辑payload加满窗口压缩器容量预算。",
      "finding": "584 B主/SWA行与132 B FP8 index行在选定FlashMLA packed路径正确；只c4层建index，SWA与压缩行并存，不扣窗口，floor边界正确。；“CUDA默认”太宽：当前CUDA选择器在SM12默认FlashInfer；plain-row后端auto为BF16、显式fp8为普通FP8，均不等于584 B packed。；每c4 81,920 B=8×2×2×(512+128)×4、每c128 524,288 B=128×2×512×4，是压缩器满窗口状态预算。当前T=1即加全部状态常数，不能称短序列已占用payload或真实分配。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813/blob/72e1d3230f6c080a530b0a1d46f8eb4602340597/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/v1/attention/backends/mla/flashmla_sparse.py#L77-L108"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/attention.py"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/compressor.py#L138-L192"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/compressor.py#L247-L301"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/nvidia/model.py#L539-L570"
        }
      ],
      "status": "reference",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "FLASHMLA_SPARSE_DSV4；fp8_ds_mla；FP8 index（非可选MXFP4）",
        "T>0加入满窗口压缩器容量预算，不声称短序列实际已用全部状态容量",
        "T>0使用满窗口压缩状态容量预算；不是短序列已用状态的精确值。"
      ],
      "unresolved": [
        "短序列实际活跃压缩状态及物理分配量未建模；当前明确使用满窗口容量预算。"
      ],
      "formula": "584×[61×min(T,128)+30×floor(T/4)+31×floor(T/128)] + 132×30×floor(T/4) + 18710528 (T>0)"
    },
    "deepseekv4flash0731": {
      "scope": "vLLM e5f2ea9；FLASHMLA_SPARSE_DSV4 / fp8_ds_mla + FP8 index；主干KV/index逻辑payload加满窗口压缩器容量预算。",
      "finding": "584 B主/SWA行与132 B FP8 index行在选定FlashMLA packed路径正确；只c4层建index，SWA与压缩行并存，不扣窗口，floor边界正确。；“CUDA默认”太宽：当前CUDA选择器在SM12默认FlashInfer；plain-row后端auto为BF16、显式fp8为普通FP8，均不等于584 B packed。；每c4 81,920 B=8×2×2×(512+128)×4、每c128 524,288 B=128×2×512×4，是压缩器满窗口状态预算。当前T=1即加全部状态常数，不能称短序列已占用payload或真实分配。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731/blob/main/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/v1/attention/backends/mla/flashmla_sparse.py#L77-L108"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/attention.py"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/compressor.py#L138-L192"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/compressor.py#L247-L301"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v4/nvidia/model.py#L539-L570"
        }
      ],
      "status": "reference",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "FLASHMLA_SPARSE_DSV4；fp8_ds_mla；FP8 index（非可选MXFP4）",
        "T>0加入满窗口压缩器容量预算，不声称短序列实际已用全部状态容量",
        "T>0使用满窗口压缩状态容量预算；不是短序列已用状态的精确值。"
      ],
      "unresolved": [
        "短序列实际活跃压缩状态及物理分配量未建模；当前明确使用满窗口容量预算。"
      ],
      "formula": "584×[43×min(T,128)+21×floor(T/4)+20×floor(T/128)] + 132×21×floor(T/4) + 12206080 (T>0)"
    },
    "deepseekv41flash": {
      "scope": "vLLM e5f2ea9；SM100 Mega/DEP FLASHMLA_MEGA_ATTN_DSV41；nvfp4_ds_mla global + MXFP4 index；compressor ring=8槽，无额外推测扩容。",
      "finding": "未发现该限定参考路径的几何或公式错误。288 B global、68 B index、528 B SWA、三份32 KiB compressor ring均有实现依据。；98,304 B为3×max(8,next_power_of_two(num_speculative_tokens+2))×1024×4在ring=8时的预算；大量draft tokens会继续扩容。；非SM100、split FlashMLA、FlashInfer、显式fp8_ds_mla等可选路径不保证同一行格式；并非通用默认。；Engram embedding表属于权重，hash实现当前由runner提供lookback输入而无独立序列状态缓存；不应把Engram权重错误加入本KV公式。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/recipes/blob/2adff3774c56906c8292568eda004c4d0e00e6a8/models/deepseek-ai/DeepSeek-V4.1-Flash.yaml"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/attention.py#L570-L590"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/attention.py#L375-L417"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/attention.py#L502-L544"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/attention.py#L181-L188"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/nvidia/flash_mla_mega_attn.py#L245-L248"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/compressor.py#L238-L265"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/compressor.py#L309-L342"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/common/engram.py#L352-L405"
        }
      ],
      "status": "verified",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "SM100 FLASHMLA_MEGA_ATTN_DSV41；auto/fp8解析nvfp4_ds_mla；indexer_kv_dtype=mxfp4；compressor ring=8行"
      ],
      "unresolved": [],
      "formula": "356×[3×floor(T/2)+T] + 40×528×min(T,128) + 98,304 (T>0)"
    },
    "deepseekv41flash1017": {
      "scope": "vLLM e5f2ea9；SM100 Mega/DEP FLASHMLA_MEGA_ATTN_DSV41；nvfp4_ds_mla global + MXFP4 index；compressor ring=8槽，无额外推测扩容。",
      "finding": "未发现该限定参考路径的几何或公式错误。288 B global、68 B index、528 B SWA、三份32 KiB compressor ring均有实现依据。；98,304 B为3×max(8,next_power_of_two(num_speculative_tokens+2))×1024×4在ring=8时的预算；大量draft tokens会继续扩容。；非SM100、split FlashMLA、FlashInfer、显式fp8_ds_mla等可选路径不保证同一行格式；并非通用默认。；Engram embedding表属于权重，hash实现当前由runner提供lookback输入而无独立序列状态缓存；不应把Engram权重错误加入本KV公式。",
      "sources": [
        {
          "label": "官方配置",
          "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/1bae0e1088993b9b82de08f59de5be5a34e42170/config.json"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/recipes/blob/2adff3774c56906c8292568eda004c4d0e00e6a8/models/deepseek-ai/DeepSeek-V4.1-Flash.yaml"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/attention.py#L570-L590"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/attention.py#L375-L417"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/attention.py#L502-L544"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/attention.py#L181-L188"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/nvidia/flash_mla_mega_attn.py#L245-L248"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/compressor.py#L238-L265"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/compressor.py#L309-L342"
        },
        {
          "label": "核验直接来源",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/models/deepseek_v41/common/engram.py#L352-L405"
        }
      ],
      "status": "verified",
      "assumptions": [
        "Batch=1，TP=1基线；TP>1仅理想均分，不代表实际每卡",
        "不含权重、工作区、分页对齐和空闲预留",
        "明确选择下述后端和精度；不是由权重dtype推断",
        "SM100 FLASHMLA_MEGA_ATTN_DSV41；auto/fp8解析nvfp4_ds_mla；indexer_kv_dtype=mxfp4；compressor ring=8行"
      ],
      "unresolved": [],
      "formula": "356×[3×floor(T/2)+T] + 40×528×min(T,128) + 98,304 (T>0)"
    },
    "nemotron35supervl": {
      "scope": "官方NVFP4 checkpoint da5b2650a835292c678994a1fa3fb653250fb508；vLLM参考实现 e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1；FP8 KV、BF16 conv、FP16 SSM。",
      "finding": "88blocks=8GQA+40Mamba+40MoE已逐项计数。T>0为4096T+86343680B；1Mi=4.080413818GiB，主公式通过。",
      "sources": [
        {
          "label": "官方config固定修订",
          "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4/blob/da5b2650a835292c678994a1fa3fb653250fb508/config.json"
        },
        {
          "label": "官方配方固定修订",
          "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4/blob/da5b2650a835292c678994a1fa3fb653250fb508/README.md"
        },
        {
          "label": "固定vLLM Mamba状态形状",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/nemotron_h.py"
        },
        {
          "label": "固定vLLM conv/SSM维度与dtype",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py"
        },
        {
          "label": "固定vLLM KV-specific auto解析",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/utils/torch_utils.py"
        }
      ],
      "status": "verified",
      "assumptions": [
        "只计目标文本主干一份当前状态。",
        "排除prefix快照、MTP/LiLiCorr、speculative conv扩展、分页/预留容量。",
        "TP仅理想均分；2KV heads在较大TP下可能复制。"
      ],
      "unresolved": []
    },
    "nemotron35bolt24gb": {
      "scope": "官方24GB checkpoint e0590ed2f3a564fd99725d9d891cb09deb7cc201；vLLM参考实现 e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1；FP8 KV、BF16 conv/SSM。",
      "finding": "88blocks=8GQA+40Mamba+40MoE。官方配方明确三类缓存dtype。T>0为4096T+86343680B；原生256Ki=1.080413818GiB，主公式通过。",
      "sources": [
        {
          "label": "官方config固定修订",
          "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB/blob/e0590ed2f3a564fd99725d9d891cb09deb7cc201/config.json"
        },
        {
          "label": "官方24GB配方固定修订",
          "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB/blob/e0590ed2f3a564fd99725d9d891cb09deb7cc201/README.md"
        },
        {
          "label": "固定vLLM Mamba状态形状",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/models/nemotron_h.py"
        },
        {
          "label": "固定vLLM conv/SSM维度与dtype",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/model_executor/layers/mamba/mamba_utils.py"
        },
        {
          "label": "固定vLLM KV-specific auto解析",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/vllm/utils/torch_utils.py"
        }
      ],
      "status": "verified",
      "assumptions": [
        "24GB为型号/部署目标，不是KV容量。",
        "只计一份当前状态，排除prefix/speculative等额外容量。",
        "大于262144token为外推；TP仅理想均分。"
      ],
      "unresolved": []
    },
    "musesparkopen": {
      "scope": "公开vLLM day-zero AMD实现6cc692b5a87a3f33c7fc5431d8add0640cb032f6；配方冻结e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1的BF16 AMD variant；BF16 KV。",
      "finding": "公开配置/配方支持62层=16full+46SWA2048，KV16×64。4096×[16T+46min(T,2048)]B；1Mi=64.359375GiB，公式通过。",
      "sources": [
        {
          "label": "公开配方与结构",
          "url": "https://recipes.vllm.ai/meta-models/Muse-Spark-Open-816B-A42B"
        },
        {
          "label": "固定AMD BF16配方",
          "url": "https://github.com/vllm-project/vllm/blob/e5f2ea9be8b2ca62ac6e049effe2dc438b78c8f1/recipes/models/meta-models/Muse-Spark-Open-816B-A42B.yaml"
        },
        {
          "label": "固定day-zero AMD Attention",
          "url": "https://github.com/vllm-project/vllm/blob/6cc692b5a87a3f33c7fc5431d8add0640cb032f6/vllm/models/muse_spark/amd/model.py#L349-L531"
        },
        {
          "label": "公开Transformers配置（动态来源）",
          "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/muse_spark/configuration_muse_spark.py"
        }
      ],
      "status": "reference",
      "assumptions": [
        "仅文本decoder。",
        "BF16 activation与auto KV参考路径。",
        "滑窗逻辑payload，不含物理页、预留、vision和DSpark。"
      ],
      "unresolved": [
        "原始checkpoint配置尚无本轮直接核实证据；当前用公开官方框架配置与固定实现交叉验证。",
        "所选AMD BF16配方仅核实32K；更长长度为几何公式外推。"
      ]
    },
    "llama4scout": {
      "scope": "Meta llama-models原生参考0e0b8c519242d5833d8c11bffc1232b77ad7f301；所有48层保留全历史，BF16/FP16均2B。",
      "finding": "缓存分配及dtype已核验，48层KV8×128由公开Transformers Scout配置交叉验证；196608T B，1Mi=192GiB，10Mi=1920GiB。",
      "sources": [
        {
          "label": "Meta原生K/V全历史分配",
          "url": "https://github.com/meta-llama/llama-models/blob/0e0b8c519242d5833d8c11bffc1232b77ad7f301/models/llama4/model.py"
        },
        {
          "label": "Meta默认BF16/FP16推理",
          "url": "https://github.com/meta-llama/llama-models/blob/0e0b8c519242d5833d8c11bffc1232b77ad7f301/models/llama4/generation.py"
        },
        {
          "label": "Meta Instruct上下文上限",
          "url": "https://github.com/meta-llama/llama-models/blob/0e0b8c519242d5833d8c11bffc1232b77ad7f301/models/sku_types.py"
        },
        {
          "label": "公开Scout框架配置（动态来源）",
          "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama4/configuration_llama4.py"
        },
        {
          "label": "固定vLLM chunk回收备选",
          "url": "https://github.com/vllm-project/vllm/blob/7b77ed581d92fa14e34ab43344837bb1ca1b4257/vllm/v1/core/single_type_kv_cache_manager.py"
        }
      ],
      "status": "reference",
      "assumptions": [
        "曲线为已用token payload；Meta实际初始化按max_seq_len预分配。",
        "不把chunk mask误当旧KV自动回收。",
        "10Mi为Instruct原生上限，不与Base的256Ki混用。"
      ],
      "unresolved": [
        "原始checkpoint配置尚无本轮直接核实证据；几何由公开官方Transformers Scout配置交叉验证。"
      ]
    },
    "llama4maverick": {
      "scope": "Meta llama-models原生参考0e0b8c519242d5833d8c11bffc1232b77ad7f301；所有48层保留全历史，BF16/FP16均2B。",
      "finding": "Meta分配/dtype和1Mi Instruct上限可验证；RedHat固定转换config核对48层KV8×128、chunk8192。196608T B，1Mi=192GiB。",
      "sources": [
        {
          "label": "Meta原生K/V全历史分配",
          "url": "https://github.com/meta-llama/llama-models/blob/0e0b8c519242d5833d8c11bffc1232b77ad7f301/models/llama4/model.py"
        },
        {
          "label": "Meta默认BF16/FP16推理",
          "url": "https://github.com/meta-llama/llama-models/blob/0e0b8c519242d5833d8c11bffc1232b77ad7f301/models/llama4/generation.py"
        },
        {
          "label": "Meta Instruct上下文上限",
          "url": "https://github.com/meta-llama/llama-models/blob/0e0b8c519242d5833d8c11bffc1232b77ad7f301/models/sku_types.py"
        },
        {
          "label": "RedHat转换配置固定修订（仅几何）",
          "url": "https://huggingface.co/RedHatAI/Llama-4-Maverick-17B-128E-Instruct-FP8/blob/59006a935b9ace6b220b8894bb88f6120d98cbeb/config.json"
        }
      ],
      "status": "reference",
      "assumptions": [
        "已用payload，不是初始化预分配。",
        "RedHat FP8只用来交叉核对几何，不能据权重量化推KV dtype。",
        "大于1Mi为外推。"
      ],
      "unresolved": [
        "原始checkpoint配置尚无本轮直接核实证据；几何由明确标注的RedHat转换配置交叉验证。"
      ]
    },
    "ernie55flashbase": {
      "scope": "官方config028544b2ffceea189b2d8768ed2b9d09b06f2343；Baidu vLLM-ERNIE 34cceda13419ae3ed0e554561f01e4be0390a6b5；vLLM打包实现5df63a7203cdb9328d824a27b682e77c7b9928be；FP8主缓存/index+FP32压缩状态预算。",
      "finding": "43主干=37HCA+6DSA，44项compress_ratios的末项MTP已正确排除。37×584×[min(T,128)+floor(T/128)]+6×788T+19398656B，T>0；1Mi=4.802685738GiB。",
      "sources": [
        {
          "label": "官方config固定修订",
          "url": "https://huggingface.co/baidu/ERNIE-5.5-Flash-Base/blob/028544b2ffceea189b2d8768ed2b9d09b06f2343/config.json"
        },
        {
          "label": "官方FP8命令固定提交",
          "url": "https://github.com/baidu/vLLM-ERNIE/blob/34cceda13419ae3ed0e554561f01e4be0390a6b5/README.md"
        },
        {
          "label": "HCA/DSA/index实现",
          "url": "https://github.com/baidu/vLLM-ERNIE/blob/34cceda13419ae3ed0e554561f01e4be0390a6b5/src/vllm_ernie/models/ernie5_5/nvidia/attention.py"
        },
        {
          "label": "584/656B打包行",
          "url": "https://github.com/vllm-project/vllm/blob/5df63a7203cdb9328d824a27b682e77c7b9928be/vllm/v1/attention/backends/mla/flashmla_sparse.py"
        },
        {
          "label": "FP32压缩状态容量",
          "url": "https://github.com/vllm-project/vllm/blob/5df63a7203cdb9328d824a27b682e77c7b9928be/vllm/models/deepseek_v4/compressor.py"
        },
        {
          "label": "C128完成边界",
          "url": "https://github.com/vllm-project/vllm/blob/5df63a7203cdb9328d824a27b682e77c7b9928be/vllm/models/deepseek_v4/nvidia/ops/sparse_attn_compress_cutedsl.py"
        }
      ],
      "status": "reference",
      "assumptions": [
        "18.5MiB是满窗口compressor容量预算，不是短序列精确live状态。",
        "T=0按已用payload约定为0。",
        "排除MTP/ECHO、prefix额外状态、物理分页/对齐、预留和TP复制。",
        "T>0使用满窗口压缩状态容量预算；不是短序列已用状态的精确值。"
      ],
      "unresolved": [
        "短序列实际活跃压缩状态及物理分配量未建模；当前明确使用满窗口容量预算。"
      ]
    }
  }
}
