Merge d1bf1c5a7a into e84a200a3c

If a subclass BYO _load_from_state_dict and doesnt call the super() the
needed default init of these weights is missed and can lead to problems
for uninitialized weights.

comfy/cli_args.py

@@ -83,6 +83,8 @@ fpte_group.add_argument("--fp16-text-enc", action="store_true", help="Store text
 fpte_group.add_argument("--fp32-text-enc", action="store_true", help="Store text encoder weights in fp32.")
 fpte_group.add_argument("--bf16-text-enc", action="store_true", help="Store text encoder weights in bf16.")
 
+parser.add_argument("--fp16-intermediates", action="store_true", help="Experimental: Use fp16 for intermediate tensors between nodes instead of fp32.")
+
 parser.add_argument("--force-channels-last", action="store_true", help="Force channels last format when inferencing the models.")
 
 parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")
@@ -90,6 +92,7 @@ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE"
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
 parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.")
 parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
+parser.add_argument("--enable-triton-backend", action="store_true", help="ComfyUI will enable the use of Triton backend in comfy-kitchen. Is disabled at launch by default.")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -11,6 +11,7 @@ from .causal_conv3d import CausalConv3d
 from .pixel_norm import PixelNorm
 from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings
 import comfy.ops
+import comfy.model_management
 from comfy.ldm.modules.diffusionmodules.model import torch_cat_if_needed
 
 ops = comfy.ops.disable_weight_init
@@ -536,7 +537,7 @@ class Decoder(nn.Module):
                     mark_conv3d_ended(self.conv_out)
                 sample = self.conv_out(sample, causal=self.causal)
                 if sample is not None and sample.shape[2] > 0:
-                    output.append(sample)
+                    output.append(sample.to(comfy.model_management.intermediate_device()))
                 return
 
             up_block = self.up_blocks[idx]

comfy/model_management.py

@@ -1050,6 +1050,12 @@ def intermediate_device():
     else:
         return torch.device("cpu")
 
+def intermediate_dtype():
+    if args.fp16_intermediates:
+        return torch.float16
+    else:
+        return torch.float32
+
 def vae_device():
     if args.cpu_vae:
         return torch.device("cpu")

comfy/ops.py

@@ -336,7 +336,10 @@ class disable_weight_init:
     class Linear(torch.nn.Linear, CastWeightBiasOp):
 
         def __init__(self, in_features, out_features, bias=True, device=None, dtype=None):
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            # don't trust subclasses that BYO state dict loader to call us.
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Linear._load_from_state_dict):
                 super().__init__(in_features, out_features, bias, device, dtype)
                 return
 
@@ -357,7 +360,9 @@ class disable_weight_init:
         def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                                 strict, missing_keys, unexpected_keys, error_msgs):
 
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Linear._load_from_state_dict):
                 return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
                                                      missing_keys, unexpected_keys, error_msgs)
             disable_weight_init._lazy_load_from_state_dict(
@@ -564,7 +569,10 @@ class disable_weight_init:
         def __init__(self, num_embeddings, embedding_dim, padding_idx=None, max_norm=None,
                      norm_type=2.0, scale_grad_by_freq=False, sparse=False, _weight=None,
                      _freeze=False, device=None, dtype=None):
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            # don't trust subclasses that BYO state dict loader to call us.
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Embedding._load_from_state_dict):
                 super().__init__(num_embeddings, embedding_dim, padding_idx, max_norm,
                                  norm_type, scale_grad_by_freq, sparse, _weight,
                                  _freeze, device, dtype)
@@ -590,7 +598,9 @@ class disable_weight_init:
         def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                                 strict, missing_keys, unexpected_keys, error_msgs):
 
-            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+            if (not comfy.model_management.WINDOWS
+                or not comfy.memory_management.aimdo_enabled
+                or type(self)._load_from_state_dict is not disable_weight_init.Embedding._load_from_state_dict):
                 return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
                                                      missing_keys, unexpected_keys, error_msgs)
             disable_weight_init._lazy_load_from_state_dict(

comfy/quant_ops.py

@@ -1,6 +1,8 @@
 import torch
 import logging
 
+from comfy.cli_args import args
+
 try:
     import comfy_kitchen as ck
     from comfy_kitchen.tensor import (
@@ -21,7 +23,15 @@ try:
             ck.registry.disable("cuda")
             logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
-    ck.registry.disable("triton")
+    if args.enable_triton_backend:
+        try:
+            import triton
+            logging.info("Found triton %s. Enabling comfy-kitchen triton backend.", triton.__version__)
+        except ImportError as e:
+            logging.error(f"Failed to import triton, Error: {e}, the comfy-kitchen triton backend will not be available.")
+            ck.registry.disable("triton")
+    else:
+        ck.registry.disable("triton")
     for k, v in ck.list_backends().items():
         logging.info(f"Found comfy_kitchen backend {k}: {v}")
 except ImportError as e:

comfy/sd.py

@@ -871,13 +871,16 @@ class VAE:
                 pixels = torch.nn.functional.pad(pixels, (0, self.output_channels - pixels.shape[-1]), mode=mode, value=value)
         return pixels
 
+    def vae_output_dtype(self):
+        return model_management.intermediate_dtype()
+
     def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16):
         steps = samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x, tile_y, overlap)
         steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x // 2, tile_y * 2, overlap)
         steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x * 2, tile_y // 2, overlap)
         pbar = comfy.utils.ProgressBar(steps)
 
-        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         output = self.process_output(
             (comfy.utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar) +
             comfy.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar) +
@@ -887,16 +890,16 @@ class VAE:
 
     def decode_tiled_1d(self, samples, tile_x=256, overlap=32):
         if samples.ndim == 3:
-            decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+            decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         else:
             og_shape = samples.shape
             samples = samples.reshape((og_shape[0], og_shape[1] * og_shape[2], -1))
-            decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).float()
+            decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
 
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, output_device=self.output_device))
 
     def decode_tiled_3d(self, samples, tile_t=999, tile_x=32, tile_y=32, overlap=(1, 8, 8)):
-        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, index_formulas=self.upscale_index_formula, output_device=self.output_device))
 
     def encode_tiled_(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
@@ -905,7 +908,7 @@ class VAE:
         steps += pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x * 2, tile_y // 2, overlap)
         pbar = comfy.utils.ProgressBar(steps)
 
-        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         samples = comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
         samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
         samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
@@ -914,7 +917,7 @@ class VAE:
 
     def encode_tiled_1d(self, samples, tile_x=256 * 2048, overlap=64 * 2048):
         if self.latent_dim == 1:
-            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
             out_channels = self.latent_channels
             upscale_amount = 1 / self.downscale_ratio
         else:
@@ -923,7 +926,7 @@ class VAE:
             tile_x = tile_x // extra_channel_size
             overlap = overlap // extra_channel_size
             upscale_amount = 1 / self.downscale_ratio
-            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).float()
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).to(dtype=self.vae_output_dtype())
 
         out = comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=upscale_amount, out_channels=out_channels, output_device=self.output_device)
         if self.latent_dim == 1:
@@ -932,7 +935,7 @@ class VAE:
             return out.reshape(samples.shape[0], self.latent_channels, extra_channel_size, -1)
 
     def encode_tiled_3d(self, samples, tile_t=9999, tile_x=512, tile_y=512, overlap=(1, 64, 64)):
-        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         return comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.downscale_ratio, out_channels=self.latent_channels, downscale=True, index_formulas=self.downscale_index_formula, output_device=self.output_device)
 
     def decode(self, samples_in, vae_options={}):
@@ -950,9 +953,9 @@ class VAE:
 
             for x in range(0, samples_in.shape[0], batch_number):
                 samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device)
-                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).float())
+                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).to(dtype=self.vae_output_dtype()))
                 if pixel_samples is None:
-                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
+                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 pixel_samples[x:x+batch_number] = out
         except Exception as e:
             model_management.raise_non_oom(e)
@@ -1025,9 +1028,9 @@ class VAE:
             samples = None
             for x in range(0, pixel_samples.shape[0], batch_number):
                 pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype).to(self.device)
-                out = self.first_stage_model.encode(pixels_in).to(self.output_device).float()
+                out = self.first_stage_model.encode(pixels_in).to(self.output_device).to(dtype=self.vae_output_dtype())
                 if samples is None:
-                    samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
+                    samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 samples[x:x + batch_number] = out
 
         except Exception as e:

manager_requirements.txt

@@ -1 +1 @@
-comfyui_manager==4.1b4
+comfyui_manager==4.1b5

nodes.py

@@ -1724,6 +1724,8 @@ class LoadImage:
         output_masks = []
         w, h = None, None
 
+        dtype = comfy.model_management.intermediate_dtype()
+
         for i in ImageSequence.Iterator(img):
             i = node_helpers.pillow(ImageOps.exif_transpose, i)
 
@@ -1748,8 +1750,8 @@ class LoadImage:
                 mask = 1. - torch.from_numpy(mask)
             else:
                 mask = torch.zeros((64,64), dtype=torch.float32, device="cpu")
-            output_images.append(image)
-            output_masks.append(mask.unsqueeze(0))
+            output_images.append(image.to(dtype=dtype))
+            output_masks.append(mask.unsqueeze(0).to(dtype=dtype))
 
             if img.format == "MPO":
                 break  # ignore all frames except the first one for MPO format

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.41.19
+comfyui-frontend-package==1.41.20
 comfyui-workflow-templates==0.9.21
 comfyui-embedded-docs==0.4.3
 torch