Merge fa71050a07 into 5cd1113236

comfy/model_detection.py

@@ -237,6 +237,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         else:
             dit_config["vec_in_dim"] = None
 
+        dit_config["num_heads"] = dit_config["hidden_size"] // sum(dit_config["axes_dim"])
+
         dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
         dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
         if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma

comfy/model_management.py

@@ -368,7 +368,7 @@ try:
                     if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
                         ENABLE_PYTORCH_ATTENTION = True
                 if rocm_version >= (7, 0):
-                   if any((a in arch) for a in ["gfx1201"]):
+                   if any((a in arch) for a in ["gfx1200", "gfx1201"]):
                        ENABLE_PYTORCH_ATTENTION = True
         if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):
             if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx950"]):  # TODO: more arches, "gfx942" gives error on pytorch nightly 2.10 1013 rocm7.0

comfy/sd.py

@@ -1059,9 +1059,9 @@ def detect_te_model(sd):
         return TEModel.JINA_CLIP_2
     if "encoder.block.23.layer.1.DenseReluDense.wi_1.weight" in sd:
         weight = sd["encoder.block.23.layer.1.DenseReluDense.wi_1.weight"]
-        if weight.shape[-1] == 4096:
+        if weight.shape[0] == 10240:
             return TEModel.T5_XXL
-        elif weight.shape[-1] == 2048:
+        elif weight.shape[0] == 5120:
             return TEModel.T5_XL
     if 'encoder.block.23.layer.1.DenseReluDense.wi.weight' in sd:
         return TEModel.T5_XXL_OLD

comfy/text_encoders/cosmos.py

@@ -36,7 +36,7 @@ def te(dtype_t5=None, t5_quantization_metadata=None):
             if t5_quantization_metadata is not None:
                 model_options = model_options.copy()
                 model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
-            if dtype is None:
+            if dtype_t5 is not None:
                 dtype = dtype_t5
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return CosmosTEModel_

comfy/text_encoders/genmo.py

@@ -32,7 +32,7 @@ def mochi_te(dtype_t5=None, t5_quantization_metadata=None):
             if t5_quantization_metadata is not None:
                 model_options = model_options.copy()
                 model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
-            if dtype is None:
+            if dtype_t5 is not None:
                 dtype = dtype_t5
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return MochiTEModel_

comfy/text_encoders/pixart_t5.py

@@ -36,7 +36,7 @@ def pixart_te(dtype_t5=None, t5_quantization_metadata=None):
             if t5_quantization_metadata is not None:
                 model_options = model_options.copy()
                 model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
-            if dtype is None:
+            if dtype_t5 is not None:
                 dtype = dtype_t5
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return PixArtTEModel_

comfy_api_nodes/nodes_kling.py

@@ -567,7 +567,7 @@ async def execute_lipsync(
     # Upload the audio file to Comfy API and get download URL
     if audio:
         audio_url = await upload_audio_to_comfyapi(
-            cls, audio, container_format="mp3", codec_name="libmp3lame", mime_type="audio/mpeg", filename="output.mp3"
+            cls, audio, container_format="mp3", codec_name="libmp3lame", mime_type="audio/mpeg"
         )
         logging.info("Uploaded audio to Comfy API. URL: %s", audio_url)
     else:

comfy_api_nodes/util/conversions.py

@@ -55,7 +55,7 @@ def image_tensor_pair_to_batch(image1: torch.Tensor, image2: torch.Tensor) -> to
 
 def tensor_to_bytesio(
     image: torch.Tensor,
-    name: str | None = None,
+    *,
     total_pixels: int = 2048 * 2048,
     mime_type: str = "image/png",
 ) -> BytesIO:
@@ -75,7 +75,7 @@ def tensor_to_bytesio(
 
     pil_image = tensor_to_pil(image, total_pixels=total_pixels)
     img_binary = pil_to_bytesio(pil_image, mime_type=mime_type)
-    img_binary.name = f"{name if name else uuid.uuid4()}.{mimetype_to_extension(mime_type)}"
+    img_binary.name = f"{uuid.uuid4()}.{mimetype_to_extension(mime_type)}"
     return img_binary
 
 

comfy_api_nodes/util/upload_helpers.py

@@ -82,7 +82,6 @@ async def upload_audio_to_comfyapi(
     container_format: str = "mp4",
     codec_name: str = "aac",
     mime_type: str = "audio/mp4",
-    filename: str = "uploaded_audio.mp4",
 ) -> str:
     """
     Uploads a single audio input to ComfyUI API and returns its download URL.
@@ -92,7 +91,7 @@ async def upload_audio_to_comfyapi(
     waveform: torch.Tensor = audio["waveform"]
     audio_data_np = audio_tensor_to_contiguous_ndarray(waveform)
     audio_bytes_io = audio_ndarray_to_bytesio(audio_data_np, sample_rate, container_format, codec_name)
-    return await upload_file_to_comfyapi(cls, audio_bytes_io, filename, mime_type)
+    return await upload_file_to_comfyapi(cls, audio_bytes_io, f"{uuid.uuid4()}.{container_format}", mime_type)
 
 
 async def upload_video_to_comfyapi(