[pull] master from comfyanonymous:master

comfy/sd.py

@@ -193,6 +193,7 @@ def encode_from_tokens_scheduled(self, tokens, unprojected=False, add_dict: dict
                 self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
             self.load_model()
+            self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
             all_hooks.reset()
             self.patcher.patch_hooks(None)
             if show_pbar:
@@ -240,6 +241,7 @@ def encode_from_tokens(self, tokens, return_pooled=False, return_dict=False):
             self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
         self.load_model()
+        self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
         o = self.cond_stage_model.encode_token_weights(tokens)
         cond, pooled = o[:2]
         if return_dict:

comfy/sd1_clip.py

@@ -147,6 +147,7 @@ def __init__(self, device="cpu", max_length=77,
         self.layer_norm_hidden_state = layer_norm_hidden_state
         self.return_projected_pooled = return_projected_pooled
         self.return_attention_masks = return_attention_masks
+        self.execution_device = None
 
         if layer == "hidden":
             assert layer_idx is not None
@@ -163,6 +164,7 @@ def freeze(self):
     def set_clip_options(self, options):
         layer_idx = options.get("layer", self.layer_idx)
         self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled)
+        self.execution_device = options.get("execution_device", self.execution_device)
         if isinstance(self.layer, list) or self.layer == "all":
             pass
         elif layer_idx is None or abs(layer_idx) > self.num_layers:
@@ -175,6 +177,7 @@ def reset_clip_options(self):
         self.layer = self.options_default[0]
         self.layer_idx = self.options_default[1]
         self.return_projected_pooled = self.options_default[2]
+        self.execution_device = None
 
     def process_tokens(self, tokens, device):
         end_token = self.special_tokens.get("end", None)
@@ -258,7 +261,11 @@ def process_tokens(self, tokens, device):
         return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens, embeds_info
 
     def forward(self, tokens):
-        device = self.transformer.get_input_embeddings().weight.device
+        if self.execution_device is None:
+            device = self.transformer.get_input_embeddings().weight.device
+        else:
+            device = self.execution_device
+
         embeds, attention_mask, num_tokens, embeds_info = self.process_tokens(tokens, device)
 
         attention_mask_model = None

comfy_api/latest/_ui.py

@@ -319,9 +319,10 @@ def save_audio(
             for key, value in metadata.items():
                 output_container.metadata[key] = value
 
+            layout = "mono" if waveform.shape[0] == 1 else "stereo"
             # Set up the output stream with appropriate properties
             if format == "opus":
-                out_stream = output_container.add_stream("libopus", rate=sample_rate)
+                out_stream = output_container.add_stream("libopus", rate=sample_rate, layout=layout)
                 if quality == "64k":
                     out_stream.bit_rate = 64000
                 elif quality == "96k":
@@ -333,7 +334,7 @@ def save_audio(
                 elif quality == "320k":
                     out_stream.bit_rate = 320000
             elif format == "mp3":
-                out_stream = output_container.add_stream("libmp3lame", rate=sample_rate)
+                out_stream = output_container.add_stream("libmp3lame", rate=sample_rate, layout=layout)
                 if quality == "V0":
                     # TODO i would really love to support V3 and V5 but there doesn't seem to be a way to set the qscale level, the property below is a bool
                     out_stream.codec_context.qscale = 1
@@ -342,12 +343,12 @@ def save_audio(
                 elif quality == "320k":
                     out_stream.bit_rate = 320000
             else:  # format == "flac":
-                out_stream = output_container.add_stream("flac", rate=sample_rate)
+                out_stream = output_container.add_stream("flac", rate=sample_rate, layout=layout)
 
             frame = av.AudioFrame.from_ndarray(
                 waveform.movedim(0, 1).reshape(1, -1).float().numpy(),
                 format="flt",
-                layout="mono" if waveform.shape[0] == 1 else "stereo",
+                layout=layout,
             )
             frame.sample_rate = sample_rate
             frame.pts = 0