Nodes/MOSS LowVRAM Optimized/MOSS · 配置与准备
ComfyUI Node

MOSS · 配置与准备

Main MOSS configuration. Stored weight format and actual matrix computation are separate. Prepare/check inspects dependencies and offers explicit verified model downloads. Sampling parameters change only through Apply recommended sampling.

By Andy-linux·Created 5 days ago·Updated a day ago· 0
MOSS · 配置与准备
    • MOSS_SESSION
    ◄model_familySFT►
    ◄dit_storagePacked W5►
    ◄dit_computeINT8 × INT8 → INT32►
    ◄te_storageSumomo INT8 ConvRot v2►
    ◄te_computeW8A16►
    ◄profile6 GB►
    ◄precisionfp16_fixed►
    ◄attentionauto►
    ◄vae_precisionfp32►
    ◄vae_assetOriginal MOSS►
    ◄qkv_backend▾►
    ◄ffn_tile▾►
    ◄host_arenatrue►
    ◄adapter_cachefalse►
    ◄conditioning_memoryfalse►
    ◄attention_head_groupauto►
    ◄te_memoryauto►
    ◄vae_admissionfalse►
    ◄resident_blocks-1►
    ◄payload_pinningstaging►
    ◄te_resident_gib3.00►
    ◄dit_checkpointq5_ridge_gptq.safetensors►
    ◄te_checkpointmoss-image-qwen35-int8-convrot-v2.safetensors►
    ◄vae_checkpointmoss-image-vae.safetensors►
    CategoryMOSS/Optimized

    Inputs (24)

    NameTypeDefaultDescription
    model_familyCOMBOSFTSelected choices are checked against the server's environment policy. Changing family does not change prompt, seed or sampling steps.
    dit_storageCOMBOPacked W5Packed W5/W6 decode into INT8 weights for real W8A8 GEMM; storage bits are not multiplication bits. Base supports INT8/FP8 only.
    dit_computeCOMBOINT8 × INT8 → INT32Integer DiT supports W8A8 only. FP8 uses row-scaled FP8 weights with FP8 activations, or dequantized 16-bit matrix operations. Boundary precision is separate.
    te_storageCOMBOSumomo INT8 ConvRot v2Sumomo v2 and Comfy INT8 are distinct files, sizes and hashes. Original float downloads all four shards, index/config and tokenizer.
    te_computeCOMBOW8A16W8A16/W4A16 dequantize weights for 16-bit activation operations; W8A8/W4A8 execute real INT8 GEMM. Historical W4A16 preserves the old GPTQ numerical path.
    profileCOMBO6 GBSelected choices are checked against the server's environment policy. Changing family does not change prompt, seed or sampling steps.
    precisionCOMBOfp16_fixedBoundary/nonquantized DiT precision. Does not change stored integer codes or select a W8A16 DiT path.
    attentionCOMBOautoSelected choices are checked against the server's environment policy. Changing family does not change prompt, seed or sampling steps.
    vae_precisionCOMBOfp32Selected choices are checked against the server's environment policy. Changing family does not change prompt, seed or sampling steps.
    vae_assetCOMBOOriginal MOSSSelected choices are checked against the server's environment policy. Changing family does not change prompt, seed or sampling steps.
    qkv_backendoptCOMBOauto 按显卡和工作集选路径。shared 共享 A8 量化；native 合并投影；register 融合写回；register3 使用独立 Q/K/V 输出。W5/6G/Flash 实测 register3 小幅提速，其他组合未普遍验证；4G 长图请保持 auto，以启用完整容量策略。
    ffn_tileoptCOMBOauto 优先整段计算，容量紧张才分块；0 强制整段；2048 降低激活峰值，通常增加耗时。
    host_arenaoptBOOLEANtrue复用 W5/W6 packed 存储，或 W8/6G、8G 的原 INT8 CPU arena。W8 arena 约 4.43 GiB 系统内存，首次组装有成本；必要 GPU 上传仍保留。4G 流式路径总是复用原码。
    adapter_cacheoptBOOLEANfalse精确复用相同编码与尺寸的 adapter 输出；CPU 最多 64 MiB。重复抽图减少权重搬运，新提示词仍完整编码。
    conditioning_memoryoptBOOLEANfalse4 GB、Euler、1–64 步时预计算原调制结果，采样少占约 162 MiB；增加准备时间。容量可选项，其他配置自动走原路径。
    attention_head_groupoptCOMBOautoauto 为 4G、W8/6G 的 Sage/CK 选 8 头，其余不分组。0/24 不分组；8/12 保留完整 K/V，省临时显存但增加调度。Flash/auto 注意力仅支持不分组。
    te_memoryoptCOMBOautoauto 按已验证预算选择。ring 逐层搬运；partial 保留部分层；resident 保留约 6.46 GiB INT8 TE 主体，embedding 仍在 CPU。过多驻留可能使解码卸载模型，反而变慢。
    vae_admissionoptBOOLEANfalse使用同运行库、已测尺寸的 VAE 工作区估计，减少不必要的模型卸载。未知尺寸保留原估计；预算不增加，容量不足仍官方分块。
    resident_blocksoptINT-1-1–40高级：-1 自动；0–40 强制 DiT 常驻前缀并用双缓冲搬运其余层。更多常驻减少 PCIe 流量，也挤占激活空间；一般保持 -1。
    payload_pinningoptCOMBOstagingstaging 省锁页：流式用小双缓冲，W8 完整 arena 用普通内存整段上传。suffix 另缓存最多 2.5 GiB 活动锁页副本，4G 重复搬运有利；首次准备更重。卸载释放本会话缓存引用。
    te_resident_giboptFLOAT3.000.25–6仅 te_memory=partial 生效：TE 常驻前缀的权重上限，单位 GiB。另需后缀双缓冲与激活空间；不增加整个会话显存预算。仅 SM89 已适配。
    dit_checkpointoptCOMBOq5_ridge_gptq.safetensorsExplicit local checkpoint. Prepare shows pinned downloads only for the selected catalog asset; custom files have no automatic source link.
    te_checkpointoptCOMBOmoss-image-qwen35-int8-convrot-v2.safetensorsExplicit local checkpoint. Prepare shows pinned downloads only for the selected catalog asset; custom files have no automatic source link.
    vae_checkpointoptCOMBOmoss-image-vae.safetensorsExplicit local checkpoint. Prepare shows pinned downloads only for the selected catalog asset; custom files have no automatic source link.

    Outputs (1)

    NameTypeDescription
    MOSS_SESSIONMOSS_SESSION—