Merge 3e34344eeb into b48d6a83d4

comfy/ldm/hunyuan_video/upsampler.py

@@ -3,8 +3,8 @@ import torch.nn as nn
 import torch.nn.functional as F
 from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d
 from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm
-import model_management
-import model_patcher
+import comfy.model_management
+import comfy.model_patcher
 
 class SRResidualCausalBlock3D(nn.Module):
     def __init__(self, channels: int):
@@ -103,13 +103,13 @@ UPSAMPLERS = {
 
 class HunyuanVideo15SRModel():
     def __init__(self, model_type, config):
-        self.load_device = model_management.vae_device()
-        offload_device = model_management.vae_offload_device()
-        self.dtype = model_management.vae_dtype(self.load_device)
+        self.load_device = comfy.model_management.vae_device()
+        offload_device = comfy.model_management.vae_offload_device()
+        self.dtype = comfy.model_management.vae_dtype(self.load_device)
         self.model_class = UPSAMPLERS.get(model_type)
         self.model = self.model_class(**config).eval()
 
-        self.patcher = model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
 
     def load_sd(self, sd):
         return self.model.load_state_dict(sd, strict=True)
@@ -118,5 +118,5 @@ class HunyuanVideo15SRModel():
         return self.model.state_dict()
 
     def resample_latent(self, latent):
-        model_management.load_model_gpu(self.patcher)
+        comfy.model_management.load_model_gpu(self.patcher)
         return self.model(latent.to(self.load_device))

comfy/model_management.py

@@ -22,7 +22,6 @@ from enum import Enum
 from comfy.cli_args import args, PerformanceFeature
 import torch
 import sys
-import importlib
 import platform
 import weakref
 import gc
@@ -349,10 +348,22 @@ try:
         except:
             rocm_version = (6, -1)
 
+        def aotriton_supported(gpu_arch):
+            path = torch.__path__[0]
+            path = os.path.join(os.path.join(path, "lib"), "aotriton.images")
+            gfx = set(map(lambda a: a[4:], filter(lambda a: a.startswith("amd-gfx"), os.listdir(path))))
+            if gpu_arch in gfx:
+                return True
+            if "{}x".format(gpu_arch[:-1]) in gfx:
+                return True
+            if "{}xx".format(gpu_arch[:-2]) in gfx:
+                return True
+            return False
+
         logging.info("AMD arch: {}".format(arch))
         logging.info("ROCm version: {}".format(rocm_version))
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
-            if importlib.util.find_spec('triton') is not None:  # AMD efficient attention implementation depends on triton. TODO: better way of detecting if it's compiled in or not.
+            if aotriton_supported(arch):  # AMD efficient attention implementation depends on aotriton.
                 if torch_version_numeric >= (2, 7):  # works on 2.6 but doesn't actually seem to improve much
                     if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
                         ENABLE_PYTORCH_ATTENTION = True

comfy_extras/nodes_audio.py

@@ -399,6 +399,58 @@ class SplitAudioChannels(IO.ComfyNode):
 
     separate = execute  # TODO: remove
 
+class JoinAudioChannels(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="JoinAudioChannels",
+            display_name="Join Audio Channels",
+            description="Joins left and right mono audio channels into a stereo audio.",
+            category="audio",
+            inputs=[
+                IO.Audio.Input("audio_left"),
+                IO.Audio.Input("audio_right"),
+            ],
+            outputs=[
+                IO.Audio.Output(display_name="audio"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, audio_left, audio_right) -> IO.NodeOutput:
+        waveform_left = audio_left["waveform"]
+        sample_rate_left = audio_left["sample_rate"]
+        waveform_right = audio_right["waveform"]
+        sample_rate_right = audio_right["sample_rate"]
+
+        if waveform_left.shape[1] != 1 or waveform_right.shape[1] != 1:
+            raise ValueError("AudioJoin: Both input audios must be mono.")
+
+        # Handle different sample rates by resampling to the higher rate
+        waveform_left, waveform_right, output_sample_rate = match_audio_sample_rates(
+            waveform_left, sample_rate_left, waveform_right, sample_rate_right
+        )
+
+        # Handle different lengths by trimming to the shorter length
+        length_left = waveform_left.shape[-1]
+        length_right = waveform_right.shape[-1]
+
+        if length_left != length_right:
+            min_length = min(length_left, length_right)
+            if length_left > min_length:
+                logging.info(f"JoinAudioChannels: Trimming left channel from {length_left} to {min_length} samples.")
+                waveform_left = waveform_left[..., :min_length]
+            if length_right > min_length:
+                logging.info(f"JoinAudioChannels: Trimming right channel from {length_right} to {min_length} samples.")
+                waveform_right = waveform_right[..., :min_length]
+
+        # Join the channels into stereo
+        left_channel = waveform_left[..., 0:1, :]
+        right_channel = waveform_right[..., 0:1, :]
+        stereo_waveform = torch.cat([left_channel, right_channel], dim=1)
+
+        return IO.NodeOutput({"waveform": stereo_waveform, "sample_rate": output_sample_rate})
+
 
 def match_audio_sample_rates(waveform_1, sample_rate_1, waveform_2, sample_rate_2):
     if sample_rate_1 != sample_rate_2:
@@ -616,6 +668,7 @@ class AudioExtension(ComfyExtension):
             RecordAudio,
             TrimAudioDuration,
             SplitAudioChannels,
+            JoinAudioChannels,
             AudioConcat,
             AudioMerge,
             AudioAdjustVolume,

requirements.txt

@@ -5,7 +5,7 @@ torch
 torchsde
 torchvision
 torchaudio
-numpy>=1.25.0
+numpy==2.2.6
 einops
 transformers>=4.50.3
 tokenizers>=0.13.3

server.py

@@ -184,7 +184,7 @@ def create_block_external_middleware():
         else:
             response = await handler(request)
 
-        response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' 'unsafe-eval' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self'; frame-src 'self'; object-src 'self';"
+        response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' 'unsafe-eval' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self' data:; frame-src 'self'; object-src 'self';"
         return response
 
     return block_external_middleware