Merge ca6da3d9fb into 1a20656448

This reverts commit ce0000c4f2.

.github/workflows/test-ci.yml

@@ -20,7 +20,6 @@ jobs:
   test-stable:
     strategy:
       fail-fast: false
-      max-parallel: 1  # This forces sequential execution
       matrix:
         # os: [macos, linux, windows]
         # os: [macos, linux]
@@ -75,7 +74,6 @@ jobs:
   test-unix-nightly:
     strategy:
       fail-fast: false
-      max-parallel: 1  # This forces sequential execution
       matrix:
         # os: [macos, linux]
         os: [linux]

README.md

@@ -69,6 +69,8 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Hunyuan Image 2.1](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_image/)
    - [Flux 2](https://comfyanonymous.github.io/ComfyUI_examples/flux2/)
    - [Z Image](https://comfyanonymous.github.io/ComfyUI_examples/z_image/)
+   - [Chroma](https://comfyanonymous.github.io/ComfyUI_examples/chroma/)
+   - [Ovis Image](https://docs.comfy.org/tutorials/image/ovis/ovis-image)
 - Image Editing Models
    - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
    - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)
@@ -82,6 +84,9 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/)
    - [Wan 2.2](https://comfyanonymous.github.io/ComfyUI_examples/wan22/)
    - [Hunyuan Video 1.5](https://docs.comfy.org/tutorials/video/hunyuan/hunyuan-video-1-5)
+   - [Kandinsky 5.0](https://docs.comfy.org/tutorials/video/kandinsky/kandinsky-5)
+   - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos)
+   - [Nvidia Cosmos Predict2](https://comfyanonymous.github.io/ComfyUI_examples/cosmos_predict2)
 - Audio Models
    - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
    - [ACE Step](https://comfyanonymous.github.io/ComfyUI_examples/audio/)

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/sd.py

@@ -479,8 +479,8 @@ class VAE:
                 self.first_stage_model = comfy.ldm.lightricks.vae.causal_video_autoencoder.VideoVAE(version=version, config=vae_config)
                 self.latent_channels = 128
                 self.latent_dim = 3
-                self.memory_used_decode = lambda shape, dtype: (900 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
-                self.memory_used_encode = lambda shape, dtype: (70 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (1200 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
+                self.memory_used_encode = lambda shape, dtype: (80 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
                 self.upscale_ratio = (lambda a: max(0, a * 8 - 7), 32, 32)
                 self.upscale_index_formula = (8, 32, 32)
                 self.downscale_ratio = (lambda a: max(0, math.floor((a + 7) / 8)), 32, 32)

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.8.0"
+__version__ = "0.8.2"

manager_requirements.txt

@@ -1 +1 @@
-comfyui_manager==4.0.4
+comfyui_manager==4.0.5

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.8.0"
+version = "0.8.2"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"