UNET weights can now be stored in fp8.

--fp8_e4m3fn-unet and --fp8_e5m2-unet are the two different formats
supported by pytorch.

comfy/cldm/cldm.py

@@ -283,7 +283,7 @@ class ControlNet(nn.Module):
         return TimestepEmbedSequential(zero_module(operations.conv_nd(self.dims, channels, channels, 1, padding=0)))
 
     def forward(self, x, hint, timesteps, context, y=None, **kwargs):
-        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(self.dtype)
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
         emb = self.time_embed(t_emb)
 
         guided_hint = self.input_hint_block(hint, emb, context)
@@ -295,7 +295,7 @@ class ControlNet(nn.Module):
             assert y.shape[0] == x.shape[0]
             emb = emb + self.label_emb(y)
 
-        h = x.type(self.dtype)
+        h = x
         for module, zero_conv in zip(self.input_blocks, self.zero_convs):
             if guided_hint is not None:
                 h = module(h, emb, context)

comfy/cli_args.py

@@ -55,7 +55,10 @@ fp_group = parser.add_mutually_exclusive_group()
 fp_group.add_argument("--force-fp32", action="store_true", help="Force fp32 (If this makes your GPU work better please report it).")
 fp_group.add_argument("--force-fp16", action="store_true", help="Force fp16.")
 
-parser.add_argument("--bf16-unet", action="store_true", help="Run the UNET in bf16. This should only be used for testing stuff.")
+fpunet_group = parser.add_mutually_exclusive_group()
+fpunet_group.add_argument("--bf16-unet", action="store_true", help="Run the UNET in bf16. This should only be used for testing stuff.")
+fpunet_group.add_argument("--fp8_e4m3fn-unet", action="store_true", help="Store unet weights in fp8_e4m3fn.")
+fpunet_group.add_argument("--fp8_e5m2-unet", action="store_true", help="Store unet weights in fp8_e5m2.")
 
 fpvae_group = parser.add_mutually_exclusive_group()
 fpvae_group.add_argument("--fp16-vae", action="store_true", help="Run the VAE in fp16, might cause black images.")

comfy/controlnet.py

@@ -1,6 +1,7 @@
 import torch
 import math
 import os
+import contextlib
 import comfy.utils
 import comfy.model_management
 import comfy.model_detection
@@ -147,24 +148,31 @@ class ControlNet(ControlBase):
                 else:
                     return None
 
+        dtype = self.control_model.dtype
+        if comfy.model_management.supports_dtype(self.device, dtype):
+            precision_scope = lambda a: contextlib.nullcontext(a)
+        else:
+            precision_scope = torch.autocast
+            dtype = torch.float32
+
         output_dtype = x_noisy.dtype
         if self.cond_hint is None or x_noisy.shape[2] * 8 != self.cond_hint.shape[2] or x_noisy.shape[3] * 8 != self.cond_hint.shape[3]:
             if self.cond_hint is not None:
                 del self.cond_hint
             self.cond_hint = None
-            self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * 8, x_noisy.shape[2] * 8, 'nearest-exact', "center").to(self.control_model.dtype).to(self.device)
+            self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * 8, x_noisy.shape[2] * 8, 'nearest-exact', "center").to(dtype).to(self.device)
         if x_noisy.shape[0] != self.cond_hint.shape[0]:
             self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
 
-
         context = cond['c_crossattn']
         y = cond.get('y', None)
         if y is not None:
-            y = y.to(self.control_model.dtype)
+            y = y.to(dtype)
         timestep = self.model_sampling_current.timestep(t)
         x_noisy = self.model_sampling_current.calculate_input(t, x_noisy)
 
-        control = self.control_model(x=x_noisy.to(self.control_model.dtype), hint=self.cond_hint, timesteps=timestep.float(), context=context.to(self.control_model.dtype), y=y)
+        with precision_scope(comfy.model_management.get_autocast_device(self.device)):
+            control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.float(), context=context.to(dtype), y=y)
         return self.control_merge(None, control, control_prev, output_dtype)
 
     def copy(self):

comfy/ldm/modules/diffusionmodules/openaimodel.py

@@ -841,14 +841,14 @@ class UNetModel(nn.Module):
             self.num_classes is not None
         ), "must specify y if and only if the model is class-conditional"
         hs = []
-        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(self.dtype)
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
         emb = self.time_embed(t_emb)
 
         if self.num_classes is not None:
             assert y.shape[0] == x.shape[0]
             emb = emb + self.label_emb(y)
 
-        h = x.type(self.dtype)
+        h = x
         for id, module in enumerate(self.input_blocks):
             transformer_options["block"] = ("input", id)
             h = forward_timestep_embed(module, h, emb, context, transformer_options, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)

comfy/model_base.py

@@ -5,6 +5,7 @@ from comfy.ldm.modules.diffusionmodules.openaimodel import Timestep
 import comfy.model_management
 import comfy.conds
 from enum import Enum
+import contextlib
 from . import utils
 
 class ModelType(Enum):
@@ -61,6 +62,13 @@ class BaseModel(torch.nn.Module):
 
         context = c_crossattn
         dtype = self.get_dtype()
+
+        if comfy.model_management.supports_dtype(xc.device, dtype):
+            precision_scope = lambda a: contextlib.nullcontext(a)
+        else:
+            precision_scope = torch.autocast
+            dtype = torch.float32
+
         xc = xc.to(dtype)
         t = self.model_sampling.timestep(t).float()
         context = context.to(dtype)
@@ -70,7 +78,10 @@ class BaseModel(torch.nn.Module):
             if hasattr(extra, "to"):
                 extra = extra.to(dtype)
             extra_conds[o] = extra
-        model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float()
+
+        with precision_scope(comfy.model_management.get_autocast_device(xc.device)):
+            model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float()
+
         return self.model_sampling.calculate_denoised(sigma, model_output, x)
 
     def get_dtype(self):

comfy/model_management.py

@@ -459,6 +459,10 @@ def unet_inital_load_device(parameters, dtype):
 def unet_dtype(device=None, model_params=0):
     if args.bf16_unet:
         return torch.bfloat16
+    if args.fp8_e4m3fn_unet:
+        return torch.float8_e4m3fn
+    if args.fp8_e5m2_unet:
+        return torch.float8_e5m2
     if should_use_fp16(device=device, model_params=model_params):
         return torch.float16
     return torch.float32
@@ -515,6 +519,17 @@ def get_autocast_device(dev):
         return dev.type
     return "cuda"
 
+def supports_dtype(device, dtype): #TODO
+    if dtype == torch.float32:
+        return True
+    if torch.device("cpu") == device:
+        return False
+    if dtype == torch.float16:
+        return True
+    if dtype == torch.bfloat16:
+        return True
+    return False
+
 def cast_to_device(tensor, device, dtype, copy=False):
     device_supports_cast = False
     if tensor.dtype == torch.float32 or tensor.dtype == torch.float16: