Use comfy.quant_ops NVFP4 layout and exact MM3_NVFP4=1 gate

- Import TensorCoreNVFP4Layout from comfy.quant_ops so ComfyUI's quantize override and Params contract apply; comfy_kitchen stays lazy for the dequantize helpers.

- Gate the conversion on the exact documented value MM3_NVFP4=1 in both load paths.

comfy/sd.py

@@ -2267,7 +2267,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None, disable
     parameters = comfy.utils.calculate_parameters(sd)
     weight_dtype = comfy.utils.weight_dtype(sd)
 
-    if os.environ.get("MM3_NVFP4") and metadata is not None and metadata.get("comfy_model") == "minimax_music3_dit":
+    if os.environ.get("MM3_NVFP4") == "1" and metadata is not None and metadata.get("comfy_model") == "minimax_music3_dit":
         comfy.text_encoders.minimax_music._to_nvfp4(sd)
 
     load_device = model_options.get("load_device", model_management.get_torch_device())

comfy/text_encoders/minimax_music.py

@@ -16,7 +16,7 @@ def _to_nvfp4(sd):
     bytes, so the AR model's resident set fits VRAM and aimdo stops paging.
     """
     import comfy_kitchen as ck
-    from comfy_kitchen.tensor.nvfp4 import TensorCoreNVFP4Layout
+    from comfy.quant_ops import TensorCoreNVFP4Layout
 
     dtype_code = ck.DTYPE_TO_CODE[torch.float32]
     for key in list(sd):
@@ -132,7 +132,7 @@ class MiniMaxMusic3TEModel(MiniMaxMusic3AR):
         return hidden.unsqueeze(0), None, {}
 
     def load_state_dict(self, state_dict, strict=True, assign=False):
-        if os.environ.get("MM3_NVFP4"):
+        if os.environ.get("MM3_NVFP4") == "1":
             _to_nvfp4(state_dict)
         if self.model.pruned_embedding is None:
             self.model.pruned_embedding = "model.embed_tokens_prefill.weight" in state_dict