ComfyUI Node
DiffusionGemma Model Loader (Advanced)
A ComfyUI node in prompt/diffusiongemma with 8 inputs and 2 outputs.
DiffusionGemma Model Loader (Advanced)
- model_config
- runtime_status_json
◄model_path/tmp/ComfyUI/custom_nodes/models/LLM/diffusiongemma-26B-A4B-it►
◄backendtransformers_inprocess►
◄dtypeauto►
◄quantizationmodelopt_nvfp4►
◄local_files_onlytrue►
◄unload_policykeep_loaded►
◄max_memory_gb20.0 GB - safe practical►
◄temperature0.45►
Categoryprompt/diffusiongemma
Inputs (8)
| Name | Type | Default | Description |
|---|---|---|---|
| model_path | STRING | /tmp/ComfyUI/custom_nodes/models/LLM/diffusiongemma-26B-A4B-it | Whole Hugging Face repo folder for the in-process Transformers/NVFP4 runtime. |
| backend | COMBO | transformers_inprocess | 1 options: transformers_inprocess |
| dtype | COMBO | auto | 4 options: auto, bfloat16, float16, float32 |
| quantization | COMBO | modelopt_nvfp4 | 5 options: modelopt_nvfp4, none, bitsandbytes_4bit, quanto, nvfp4_metadata_only |
| local_files_only | BOOLEAN | true | — |
| unload_policy | COMBO | keep_loaded | 2 options: keep_loaded, unload_after_run |
| max_memory_gb | COMBO | 20.0 GB - safe practical | NVFP4 bridge memory profile. 20 GB is the safest practical default; 18 GB is the supported floor; 0 requests full GPU load. |
| temperatureopt | FLOAT | 0.450–2 | Generation sampling temperature. 0.2 is conservative; 0.45-0.8 is usually more creative. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| model_config | DG_MODEL_CONFIG | — |
| runtime_status_json | STRING | — |