Merge branch 'master' into seedvr2

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]

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/model_patcher.py

@@ -718,6 +718,7 @@ class ModelPatcher:
                             continue
 
                 cast_weight = self.force_cast_weights
+                m.comfy_force_cast_weights = self.force_cast_weights
                 if lowvram_weight:
                     if hasattr(m, "comfy_cast_weights"):
                         m.weight_function = []
@@ -790,11 +791,12 @@ class ModelPatcher:
                 for param in params:
                     self.pin_weight_to_device("{}.{}".format(n, param))
 
+            usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
-                logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
+                logging.info("loaded partially; {} {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(usable_stat, mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
                 self.model.model_lowvram = True
             else:
-                logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
+                logging.info("loaded completely; {} {:.2f} MB loaded, full load: {}".format(usable_stat, mem_counter / (1024 * 1024), full_load))
                 self.model.model_lowvram = False
                 if full_load:
                     self.model.to(device_to)

comfy/ops.py

@@ -654,29 +654,29 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 run_every_op()
 
                 input_shape = input.shape
-                tensor_3d = input.ndim == 3
-
-                if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
-                    return self.forward_comfy_cast_weights(input, *args, **kwargs)
+                reshaped_3d = False
 
                 if (getattr(self, 'layout_type', None) is not None and
-                    not isinstance(input, QuantizedTensor)):
+                    not isinstance(input, QuantizedTensor) and not self._full_precision_mm and
+                    not getattr(self, 'comfy_force_cast_weights', False) and
+                    len(self.weight_function) == 0 and len(self.bias_function) == 0):
 
                     # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others)
-                    if tensor_3d:
-                        input = input.reshape(-1, input_shape[2])
+                    input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input
 
-                    if input.ndim != 2:
-                        # Fall back to comfy_cast_weights for non-2D tensors
-                        return self.forward_comfy_cast_weights(input.reshape(input_shape), *args, **kwargs)
+                    # Fall back to non-quantized for non-2D tensors
+                    if input_reshaped.ndim == 2:
+                        reshaped_3d = input.ndim == 3
+                        # dtype is now implicit in the layout class
+                        scale = getattr(self, 'input_scale', None)
+                        if scale is not None:
+                            scale = comfy.model_management.cast_to_device(scale, input.device, None)
+                        input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale)
 
-                    # dtype is now implicit in the layout class
-                    input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None))
-
-                output = self._forward(input, self.weight, self.bias)
+                output = self.forward_comfy_cast_weights(input)
 
                 # Reshape output back to 3D if input was 3D
-                if tensor_3d:
+                if reshaped_3d:
                     output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0]))
 
                 return output

comfy/quant_ops.py

@@ -19,6 +19,7 @@ try:
         cuda_version = tuple(map(int, str(torch.version.cuda).split('.')))
         if cuda_version < (13,):
             ck.registry.disable("cuda")
+            logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
     ck.registry.disable("triton")
     for k, v in ck.list_backends().items():

comfy/sd.py

@@ -219,7 +219,7 @@ class CLIP:
             if unprojected:
                 self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-            self.load_model()
+            self.load_model(tokens)
             self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
             all_hooks.reset()
             self.patcher.patch_hooks(None)
@@ -267,7 +267,7 @@ class CLIP:
         if return_pooled == "unprojected":
             self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-        self.load_model()
+        self.load_model(tokens)
         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]
@@ -300,8 +300,11 @@ class CLIP:
             sd_clip[k] = sd_tokenizer[k]
         return sd_clip
 
-    def load_model(self):
-        model_management.load_model_gpu(self.patcher)
+    def load_model(self, tokens={}):
+        memory_used = 0
+        if hasattr(self.cond_stage_model, "memory_estimation_function"):
+            memory_used = self.cond_stage_model.memory_estimation_function(tokens, device=self.patcher.load_device)
+        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
         return self.patcher
 
     def get_key_patches(self):
@@ -491,8 +494,8 @@ 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)

comfy/supported_models.py

@@ -845,7 +845,7 @@ class LTXAV(LTXV):
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
-        self.memory_usage_factor = 0.055  # TODO
+        self.memory_usage_factor = 0.061  # TODO
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.LTXAV(self, device=device)

comfy/text_encoders/lt.py

@@ -36,10 +36,10 @@ class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer):
 
 class Gemma3_12BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
-        llama_scaled_fp8 = model_options.get("gemma_scaled_fp8", None)
-        if llama_scaled_fp8 is not None:
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
             model_options = model_options.copy()
-            model_options["scaled_fp8"] = llama_scaled_fp8
+            model_options["quantization_metadata"] = llama_quantization_metadata
 
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
@@ -98,10 +98,13 @@ class LTXAVTEModel(torch.nn.Module):
 
         out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs)
         out_device = out.device
+        if comfy.model_management.should_use_bf16(self.execution_device):
+            out = out.to(device=self.execution_device, dtype=torch.bfloat16)
         out = out.movedim(1, -1).to(self.execution_device)
         out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6)
         out = out.reshape((out.shape[0], out.shape[1], -1))
         out = self.text_embedding_projection(out)
+        out = out.float()
         out_vid = self.video_embeddings_connector(out)[0]
         out_audio = self.audio_embeddings_connector(out)[0]
         out = torch.concat((out_vid, out_audio), dim=-1)
@@ -118,13 +121,21 @@ class LTXAVTEModel(torch.nn.Module):
 
             return self.load_state_dict(sdo, strict=False)
 
+    def memory_estimation_function(self, token_weight_pairs, device=None):
+        constant = 6.0
+        if comfy.model_management.should_use_bf16(device):
+            constant /= 2.0
 
-def ltxav_te(dtype_llama=None, llama_scaled_fp8=None):
+        token_weight_pairs = token_weight_pairs.get("gemma3_12b", [])
+        num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
+        return num_tokens * constant * 1024 * 1024
+
+def ltxav_te(dtype_llama=None, llama_quantization_metadata=None):
     class LTXAVTEModel_(LTXAVTEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
+            if llama_quantization_metadata is not None:
                 model_options = model_options.copy()
-                model_options["llama_scaled_fp8"] = llama_scaled_fp8
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
             if dtype_llama is not None:
                 dtype = dtype_llama
             super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)

comfy_extras/nodes_lt_audio.py

@@ -185,6 +185,10 @@ class LTXAVTextEncoderLoader(io.ComfyNode):
                 io.Combo.Input(
                     "ckpt_name",
                     options=folder_paths.get_filename_list("checkpoints"),
+                ),
+                io.Combo.Input(
+                    "device",
+                    options=["default", "cpu"],
                 )
             ],
             outputs=[io.Clip.Output()],
@@ -197,7 +201,11 @@ class LTXAVTextEncoderLoader(io.ComfyNode):
         clip_path1 = folder_paths.get_full_path_or_raise("text_encoders", text_encoder)
         clip_path2 = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name)
 
-        clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type)
+        model_options = {}
+        if device == "cpu":
+            model_options["load_device"] = model_options["offload_device"] = torch.device("cpu")
+
+        clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type, model_options=model_options)
         return io.NodeOutput(clip)
 
 

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"

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.35.9
-comfyui-workflow-templates==0.7.67
+comfyui-workflow-templates==0.7.69
 comfyui-embedded-docs==0.3.1
 torch
 torchsde
@@ -21,7 +21,7 @@ psutil
 alembic
 SQLAlchemy
 av>=14.2.0
-comfy-kitchen>=0.2.3
+comfy-kitchen>=0.2.5
 
 #non essential dependencies:
 kornia>=0.7.1