feat: support Qwen-Image in ComfyUI-nunchaku (#632)

* update

* add parameter of act unsigned

* upgrade the diffusers to v0.35.1

* bump the comfyui version to 0.3.51

* update version

* revert the test comfyui-version back to 0.3.44

.github/workflows/pr-test.yaml

@@ -67,8 +67,8 @@ jobs:
           which python
           echo "Installing dependencies"
           pip install torch==2.7 torchvision==0.22 torchaudio==2.7 --index-url https://download.pytorch.org/whl/cu128
-          pip install git+https://github.com/huggingface/diffusers
-          pip install ninja wheel transformers==4.51 accelerate==1.7 sentencepiece==0.2 protobuf==6.31 huggingface_hub==0.34
+          pip install diffusers==0.35.1
+          pip install ninja wheel transformers==4.55 accelerate==1.10 sentencepiece==0.2 protobuf==6.31 huggingface_hub==0.34
       - name: Build
         run: |
           source $(conda info --base)/etc/profile.d/conda.sh

nunchaku/models/linear.py

@@ -14,10 +14,13 @@ class SVDQW4A4Linear(nn.Module):
         rank: int = 32,
         bias: bool = True,
         precision: str = "int4",
+        act_unsigned: bool = False,
         torch_dtype: torch.dtype = torch.bfloat16,
-        device: str | torch.device = "cpu",
+        device: str | torch.device | None = None,
     ):
         super(SVDQW4A4Linear, self).__init__()
+        if device is None:
+            device = torch.device("cpu")
         self.in_features = in_features
         self.out_features = out_features
         self.rank = rank
@@ -69,7 +72,7 @@ class SVDQW4A4Linear(nn.Module):
                 torch.ones(out_features, dtype=torch_dtype, device=device), requires_grad=False
             )
 
-        self.act_unsigned = False
+        self.act_unsigned = act_unsigned
 
     @classmethod
     def from_linear(cls, linear: nn.Linear, **kwargs):
@@ -140,9 +143,11 @@ class AWQW4A16Linear(nn.Module):
         bias: bool = True,
         group_size: int = 64,
         torch_dtype: torch.dtype = torch.bfloat16,
-        device: str | torch.device = "cuda",
+        device: str | torch.device | None = None,
     ):
         super(AWQW4A16Linear, self).__init__()
+        if device is None:
+            device = torch.device("cpu")
         self.in_features = in_features
         self.out_features = out_features
 

nunchaku/utils.py

@@ -280,3 +280,18 @@ def check_hardware_compatibility(quantization_config: dict, device: str | torch.
             f"Unsupported GPU architecture {sm} due to the lack of 4-bit tensorcores. "
             "Please use a Turing, Ampere, Ada or Blackwell GPU for this quantization configuration."
         )
+
+
+def get_precision_from_quantization_config(quantization_config: dict) -> str:
+    """
+    Get the precision from the quantization configuration.
+    """
+    if quantization_config["weight"]["dtype"] == "fp4_e2m1_all":
+        if quantization_config["weight"]["group_size"] == 16:
+            return "nvfp4"
+        else:
+            raise ValueError("Currently, nunchaku only supports nvfp4.")
+    elif quantization_config["weight"]["dtype"] == "int4":
+        return "int4"
+    else:
+        raise ValueError(f"Unsupported quantization dtype: {quantization_config['weight']['dtype']}")