Greatly improve lowvram sampling speed by getting rid of accelerate.

Let me know if this breaks anything.

comfy/controlnet.py

@@ -283,7 +283,7 @@ class control_lora_ops(ControlLoraOps, comfy.ops.manual_cast):
         cm = self.control_model.state_dict()
 
         for k in sd:
-            weight = comfy.model_management.resolve_lowvram_weight(sd[k], diffusion_model, k)
+            weight = sd[k]
             try:
                 comfy.utils.set_attr(self.control_model, k, weight)
             except:

comfy/model_base.py

@@ -162,11 +162,7 @@ def process_latent_out(self, latent):
 
     def state_dict_for_saving(self, clip_state_dict, vae_state_dict):
         clip_state_dict = self.model_config.process_clip_state_dict_for_saving(clip_state_dict)
-        unet_sd = self.diffusion_model.state_dict()
-        unet_state_dict = {}
-        for k in unet_sd:
-            unet_state_dict[k] = comfy.model_management.resolve_lowvram_weight(unet_sd[k], self.diffusion_model, k)
-
+        unet_state_dict = self.diffusion_model.state_dict()
         unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict)
         vae_state_dict = self.model_config.process_vae_state_dict_for_saving(vae_state_dict)
         if self.get_dtype() == torch.float16:

comfy/model_management.py

@@ -218,15 +218,8 @@ def is_nvidia():
     FORCE_FP16 = True
 
 if lowvram_available:
-    try:
-        import accelerate
-        if set_vram_to in (VRAMState.LOW_VRAM, VRAMState.NO_VRAM):
-            vram_state = set_vram_to
-    except Exception as e:
-        import traceback
-        print(traceback.format_exc())
-        print("ERROR: LOW VRAM MODE NEEDS accelerate.")
-        lowvram_available = False
+    if set_vram_to in (VRAMState.LOW_VRAM, VRAMState.NO_VRAM):
+        vram_state = set_vram_to
 
 
 if cpu_state != CPUState.GPU:
@@ -298,8 +291,20 @@ def model_load(self, lowvram_model_memory=0):
 
         if lowvram_model_memory > 0:
             print("loading in lowvram mode", lowvram_model_memory/(1024 * 1024))
-            device_map = accelerate.infer_auto_device_map(self.real_model, max_memory={0: "{}MiB".format(lowvram_model_memory // (1024 * 1024)), "cpu": "16GiB"})
-            accelerate.dispatch_model(self.real_model, device_map=device_map, main_device=self.device)
+            mem_counter = 0
+            for m in self.real_model.modules():
+                if hasattr(m, "comfy_cast_weights"):
+                    m.prev_comfy_cast_weights = m.comfy_cast_weights
+                    m.comfy_cast_weights = True
+                    module_mem = 0
+                    sd = m.state_dict()
+                    for k in sd:
+                        t = sd[k]
+                        module_mem += t.nelement() * t.element_size()
+                    if mem_counter + module_mem < lowvram_model_memory:
+                        m.to(self.device)
+                        mem_counter += module_mem
+
             self.model_accelerated = True
 
         if is_intel_xpu() and not args.disable_ipex_optimize:
@@ -309,7 +314,11 @@ def model_load(self, lowvram_model_memory=0):
 
     def model_unload(self):
         if self.model_accelerated:
-            accelerate.hooks.remove_hook_from_submodules(self.real_model)
+            for m in self.real_model.modules():
+                if hasattr(m, "prev_comfy_cast_weights"):
+                    m.comfy_cast_weights = m.prev_comfy_cast_weights
+                    del m.prev_comfy_cast_weights
+
             self.model_accelerated = False
 
         self.model.unpatch_model(self.model.offload_device)
@@ -402,14 +411,14 @@ def load_models_gpu(models, memory_required=0):
         if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):
             model_size = loaded_model.model_memory_required(torch_dev)
             current_free_mem = get_free_memory(torch_dev)
-            lowvram_model_memory = int(max(256 * (1024 * 1024), (current_free_mem - 1024 * (1024 * 1024)) / 1.3 ))
+            lowvram_model_memory = int(max(64 * (1024 * 1024), (current_free_mem - 1024 * (1024 * 1024)) / 1.3 ))
             if model_size > (current_free_mem - inference_memory): #only switch to lowvram if really necessary
                 vram_set_state = VRAMState.LOW_VRAM
             else:
                 lowvram_model_memory = 0
 
         if vram_set_state == VRAMState.NO_VRAM:
-            lowvram_model_memory = 256 * 1024 * 1024
+            lowvram_model_memory = 64 * 1024 * 1024
 
         cur_loaded_model = loaded_model.model_load(lowvram_model_memory)
         current_loaded_models.insert(0, loaded_model)
@@ -566,6 +575,11 @@ def supports_dtype(device, dtype): #TODO
         return True
     return False
 
+def device_supports_non_blocking(device):
+    if is_device_mps(device):
+        return False #pytorch bug? mps doesn't support non blocking
+    return True
+
 def cast_to_device(tensor, device, dtype, copy=False):
     device_supports_cast = False
     if tensor.dtype == torch.float32 or tensor.dtype == torch.float16:
@@ -576,9 +590,7 @@ def cast_to_device(tensor, device, dtype, copy=False):
         elif is_intel_xpu():
             device_supports_cast = True
 
-    non_blocking = True
-    if is_device_mps(device):
-        non_blocking = False #pytorch bug? mps doesn't support non blocking
+    non_blocking = device_supports_non_blocking(device)
 
     if device_supports_cast:
         if copy:
@@ -742,11 +754,7 @@ def soft_empty_cache(force=False):
             torch.cuda.empty_cache()
             torch.cuda.ipc_collect()
 
-def resolve_lowvram_weight(weight, model, key):
-    if weight.device == torch.device("meta"): #lowvram NOTE: this depends on the inner working of the accelerate library so it might break.
-        key_split = key.split('.')              # I have no idea why they don't just leave the weight there instead of using the meta device.
-        op = comfy.utils.get_attr(model, '.'.join(key_split[:-1]))
-        weight = op._hf_hook.weights_map[key_split[-1]]
+def resolve_lowvram_weight(weight, model, key): #TODO: remove
     return weight
 
 #TODO: might be cleaner to put this somewhere else

comfy/ops.py

@@ -1,27 +1,93 @@
 import torch
 from contextlib import contextmanager
+import comfy.model_management
+
+def cast_bias_weight(s, input):
+    bias = None
+    non_blocking = comfy.model_management.device_supports_non_blocking(input.device)
+    if s.bias is not None:
+        bias = s.bias.to(device=input.device, dtype=input.dtype, non_blocking=non_blocking)
+    weight = s.weight.to(device=input.device, dtype=input.dtype, non_blocking=non_blocking)
+    return weight, bias
+
 
 class disable_weight_init:
     class Linear(torch.nn.Linear):
+        comfy_cast_weights = False
         def reset_parameters(self):
             return None
 
+        def forward_comfy_cast_weights(self, input):
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.linear(input, weight, bias)
+
+        def forward(self, *args, **kwargs):
+            if self.comfy_cast_weights:
+                return self.forward_comfy_cast_weights(*args, **kwargs)
+            else:
+                return super().forward(*args, **kwargs)
+
     class Conv2d(torch.nn.Conv2d):
+        comfy_cast_weights = False
         def reset_parameters(self):
             return None
 
+        def forward_comfy_cast_weights(self, input):
+            weight, bias = cast_bias_weight(self, input)
+            return self._conv_forward(input, weight, bias)
+
+        def forward(self, *args, **kwargs):
+            if self.comfy_cast_weights:
+                return self.forward_comfy_cast_weights(*args, **kwargs)
+            else:
+                return super().forward(*args, **kwargs)
+
     class Conv3d(torch.nn.Conv3d):
+        comfy_cast_weights = False
         def reset_parameters(self):
             return None
 
+        def forward_comfy_cast_weights(self, input):
+            weight, bias = cast_bias_weight(self, input)
+            return self._conv_forward(input, weight, bias)
+
+        def forward(self, *args, **kwargs):
+            if self.comfy_cast_weights:
+                return self.forward_comfy_cast_weights(*args, **kwargs)
+            else:
+                return super().forward(*args, **kwargs)
+
     class GroupNorm(torch.nn.GroupNorm):
+        comfy_cast_weights = False
         def reset_parameters(self):
             return None
 
+        def forward_comfy_cast_weights(self, input):
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
+
+        def forward(self, *args, **kwargs):
+            if self.comfy_cast_weights:
+                return self.forward_comfy_cast_weights(*args, **kwargs)
+            else:
+                return super().forward(*args, **kwargs)
+
+
     class LayerNorm(torch.nn.LayerNorm):
+        comfy_cast_weights = False
         def reset_parameters(self):
             return None
 
+        def forward_comfy_cast_weights(self, input):
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
+
+        def forward(self, *args, **kwargs):
+            if self.comfy_cast_weights:
+                return self.forward_comfy_cast_weights(*args, **kwargs)
+            else:
+                return super().forward(*args, **kwargs)
+
     @classmethod
     def conv_nd(s, dims, *args, **kwargs):
         if dims == 2:
@@ -31,35 +97,19 @@ def conv_nd(s, dims, *args, **kwargs):
         else:
             raise ValueError(f"unsupported dimensions: {dims}")
 
-def cast_bias_weight(s, input):
-    bias = None
-    if s.bias is not None:
-        bias = s.bias.to(device=input.device, dtype=input.dtype)
-    weight = s.weight.to(device=input.device, dtype=input.dtype)
-    return weight, bias
 
 class manual_cast(disable_weight_init):
     class Linear(disable_weight_init.Linear):
-        def forward(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.linear(input, weight, bias)
+        comfy_cast_weights = True
 
     class Conv2d(disable_weight_init.Conv2d):
-        def forward(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+        comfy_cast_weights = True
 
     class Conv3d(disable_weight_init.Conv3d):
-        def forward(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+        comfy_cast_weights = True
 
     class GroupNorm(disable_weight_init.GroupNorm):
-        def forward(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
+        comfy_cast_weights = True
 
     class LayerNorm(disable_weight_init.LayerNorm):
-        def forward(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
+        comfy_cast_weights = True

requirements.txt

@@ -4,7 +4,6 @@ einops
 transformers>=4.25.1
 safetensors>=0.3.0
 aiohttp
-accelerate
 pyyaml
 Pillow
 scipy