comfyui-nunchaku-rocm 1.0.2
Basename is uppercase README.mdPython JSON parsing or serialization reference
SHA-25600544ce96f053edcba0bfaa03bf7e4893f7b485919f1afadb98a4c5d5aef62ab
MaleculeK(Am)H₄(CmDb₄Os)Md₆(In₂NePa₃Pt)
Evidence
1# ComfyUI-nunchaku-rocm
2
3ComfyUI nodes that run [Nunchaku](https://github.com/nunchaku-ai/nunchaku) 4-bit SVDQuant models on AMD GPUs, on
4[nunchaku-rocm](https://github.com/TwelveNights/nunchaku-rocm) (Trit …
23:46… b.com/TwelveNights/nunchaku-rocm/blob/main/comfyui/README.md).
24
25## Nodes
26
27Category *Nunchaku (ROCm)*:
28
29- **Nunchaku DiT Loader (ROCm)** loads a Nunchaku checkpoint from `models/diffusion_models` as a `MODEL`. The model
30 type is read from the checkpoint. Qwen-Image only: `cpu_offload` (`auto` / ` …
40:37… UI Manager, or clone it into `ComfyUI/custom_nodes`:
41
42 ```
43 git clone https://github.com/TwelveNights/ComfyUI-nunchaku-rocm.git
44 ```
45
463. Restart ComfyUI and open a workflow from [`example_workflows`](example_workflows).
47
48The first image of each model co …
1"""
2ComfyUI nodes for Nunchaku SVDQuant (4-bit) models on AMD GPUs, running on nunchaku-rocm.
3
4Needs nunchaku-rocm installed in ComfyUI's Python environment: https://github.com/TwelveNights/nunchaku-rocm
5"""
6
7import importlib.util
8import os
9
10if importlib.util.find_spec("nunchaku") is None:
11 raise ImportError(
12 "ComfyUI-nunchaku-rocm needs nunchaku-rocm in ComfyUI's Python environment, installed on top of a ROCm "
13 "PyTorch: see https://github.com/TwelveNights/nunchaku-rocm#install"
14 )
15
16# select the Triton backend before nunchaku is first imported; NUNCHAKU_BACKEND=cuda can still override
17os.environ.setdefault("NUNCHAKU_BACKEND", "triton")
18
19from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS # noqa: E402
20
21__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]
50:51… == "flux2"
51 self.transformer = transformer
52 self._warned = set()
53 # LoRAs are requested per sampling call through transformer_options (see nodes.NunchakuLoraLoader), so
54 # MODEL branches with and without LoRAs share this transformer; applied lazily when the request changes
55 self._applied_loras: tuple = ()
56 …
No evidence locations were recorded for this file. Raw result