Nodes/comfyui-gpu-resident-loader/Diffusion Model Loader Resident
ComfyUI Node

Diffusion Model Loader Resident

KJ-compatible diffusion-model loader with GPU-resident ingest and residency tracking. It mirrors the KJ node's weight dtype, compute dtype, cublas_ops, SageAttention, fp16 accumulation, and optional extra-state-dict features.

By xmarre·Created 4 months ago·Updated 4 months ago· 2
Diffusion Model Loader Resident
    • MODEL
    model_name
    weight_dtype
    compute_dtypedefault
    patch_cublaslinearfalse
    sage_attentiondisabled
    enable_fp16_accumulationfalse
    extra_state_dict
    policy_override
    CategoryGPU Resident Loader/loaders

    Inputs (8)

    NameTypeDefaultDescription
    model_nameCOMBOThe diffusion model file to load.
    weight_dtypeCOMBO7 options: default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2, fp16, bf16, +1
    compute_dtypeCOMBOdefaultCompute dtype to apply after model creation.
    patch_cublaslinearBOOLEANfalseToggle ComfyUI's cublas_ops performance feature.
    sage_attentionCOMBOdisabledPatch optimized attention override to a SageAttention variant.
    enable_fp16_accumulationBOOLEANfalseSet torch.backends.cuda.matmul.allow_fp16_accumulation.
    extra_state_dictoptSTRINGOptional absolute path to a second state dict merged into the main diffusion state dict before model detection.
    policy_overrideoptSTRINGOptional residency policy override. Connect Set Global Residency Policy here, not to extra_state_dict.

    Outputs (1)

    NameTypeDescription
    MODELMODEL