Merge pull request #324 from modelscope/vram_management

support vram management in flux

diffsynth/models/model_manager.py

@@ -80,7 +80,10 @@ def load_model_from_single_file(state_dict, model_names, model_classes, model_re
 def load_model_from_huggingface_folder(file_path, model_names, model_classes, torch_dtype, device):
     loaded_model_names, loaded_models = [], []
     for model_name, model_class in zip(model_names, model_classes):
-        model = model_class.from_pretrained(file_path, torch_dtype=torch_dtype).eval()
+        if torch_dtype in [torch.float32, torch.float16, torch.bfloat16]:
+            model = model_class.from_pretrained(file_path, torch_dtype=torch_dtype).eval()
+        else:
+            model = model_class.from_pretrained(file_path).eval().to(dtype=torch_dtype)
         if torch_dtype == torch.float16 and hasattr(model, "half"):
             model = model.half()
         try:

diffsynth/models/sd3_text_encoder.py

@@ -9,7 +9,8 @@ class SD3TextEncoder1(SDTextEncoder):
         super().__init__(vocab_size=vocab_size)
 
     def forward(self, input_ids, clip_skip=2, extra_mask=None):
-        embeds = self.token_embedding(input_ids) + self.position_embeds
+        embeds = self.token_embedding(input_ids)
+        embeds = embeds + self.position_embeds.to(dtype=embeds.dtype, device=input_ids.device)
         attn_mask = self.attn_mask.to(device=embeds.device, dtype=embeds.dtype)
         if extra_mask is not None:
             attn_mask[:, extra_mask[0]==0] = float("-inf")

diffsynth/pipelines/base.py

@@ -101,12 +101,22 @@ class BasePipeline(torch.nn.Module):
             if model_name not in loadmodel_names:
                 model = getattr(self, model_name)
                 if model is not None:
-                    model.cpu()
+                    if hasattr(model, "vram_management_enabled") and model.vram_management_enabled:
+                        for module in model.modules():
+                            if hasattr(module, "offload"):
+                                module.offload()
+                    else:
+                        model.cpu()
         # load the needed models to device
         for model_name in loadmodel_names:
             model = getattr(self, model_name)
             if model is not None:
-                model.to(self.device)
+                if hasattr(model, "vram_management_enabled") and model.vram_management_enabled:
+                    for module in model.modules():
+                        if hasattr(module, "onload"):
+                            module.onload()
+                else:
+                    model.to(self.device)
         # fresh the cuda cache
         torch.cuda.empty_cache()
 

diffsynth/pipelines/flux_image.py

@@ -11,6 +11,9 @@ from PIL import Image
 from ..models.tiler import FastTileWorker
 from transformers import SiglipVisionModel
 from copy import deepcopy
+from transformers.models.t5.modeling_t5 import T5LayerNorm, T5DenseActDense, T5DenseGatedActDense
+from ..models.flux_dit import RMSNorm
+from ..vram_management import enable_vram_management, AutoWrappedModule, AutoWrappedLinear
 
 
 class FluxImagePipeline(BasePipeline):
@@ -31,6 +34,105 @@ class FluxImagePipeline(BasePipeline):
         self.model_names = ['text_encoder_1', 'text_encoder_2', 'dit', 'vae_decoder', 'vae_encoder', 'controlnet', 'ipadapter', 'ipadapter_image_encoder']
 
 
+    def enable_vram_management(self, num_persistent_param_in_dit=None):
+        dtype = next(iter(self.text_encoder_1.parameters())).dtype
+        enable_vram_management(
+            self.text_encoder_1,
+            module_map = {
+                torch.nn.Linear: AutoWrappedLinear,
+                torch.nn.Embedding: AutoWrappedModule,
+                torch.nn.LayerNorm: AutoWrappedModule,
+            },
+            module_config = dict(
+                offload_dtype=dtype,
+                offload_device="cpu",
+                onload_dtype=dtype,
+                onload_device="cpu",
+                computation_dtype=self.torch_dtype,
+                computation_device=self.device,
+            ),
+        )
+        dtype = next(iter(self.text_encoder_2.parameters())).dtype
+        enable_vram_management(
+            self.text_encoder_2,
+            module_map = {
+                torch.nn.Linear: AutoWrappedLinear,
+                torch.nn.Embedding: AutoWrappedModule,
+                T5LayerNorm: AutoWrappedModule,
+                T5DenseActDense: AutoWrappedModule,
+                T5DenseGatedActDense: AutoWrappedModule,
+            },
+            module_config = dict(
+                offload_dtype=dtype,
+                offload_device="cpu",
+                onload_dtype=dtype,
+                onload_device="cpu",
+                computation_dtype=self.torch_dtype,
+                computation_device=self.device,
+            ),
+        )
+        dtype = next(iter(self.dit.parameters())).dtype
+        enable_vram_management(
+            self.dit,
+            module_map = {
+                RMSNorm: AutoWrappedModule,
+                torch.nn.Linear: AutoWrappedLinear,
+            },
+            module_config = dict(
+                offload_dtype=dtype,
+                offload_device="cpu",
+                onload_dtype=dtype,
+                onload_device="cuda",
+                computation_dtype=self.torch_dtype,
+                computation_device=self.device,
+            ),
+            max_num_param=num_persistent_param_in_dit,
+            overflow_module_config = dict(
+                offload_dtype=dtype,
+                offload_device="cpu",
+                onload_dtype=dtype,
+                onload_device="cpu",
+                computation_dtype=self.torch_dtype,
+                computation_device=self.device,
+            ),
+        )
+        dtype = next(iter(self.vae_decoder.parameters())).dtype
+        enable_vram_management(
+            self.vae_decoder,
+            module_map = {
+                torch.nn.Linear: AutoWrappedLinear,
+                torch.nn.Conv2d: AutoWrappedModule,
+                torch.nn.GroupNorm: AutoWrappedModule,
+            },
+            module_config = dict(
+                offload_dtype=dtype,
+                offload_device="cpu",
+                onload_dtype=dtype,
+                onload_device="cpu",
+                computation_dtype=self.torch_dtype,
+                computation_device=self.device,
+            ),
+        )
+        dtype = next(iter(self.vae_encoder.parameters())).dtype
+        enable_vram_management(
+            self.vae_encoder,
+            module_map = {
+                torch.nn.Linear: AutoWrappedLinear,
+                torch.nn.Conv2d: AutoWrappedModule,
+                torch.nn.GroupNorm: AutoWrappedModule,
+            },
+            module_config = dict(
+                offload_dtype=dtype,
+                offload_device="cpu",
+                onload_dtype=dtype,
+                onload_device="cpu",
+                computation_dtype=self.torch_dtype,
+                computation_device=self.device,
+            ),
+        )
+        self.enable_cpu_offload()
+
+
     def denoising_model(self):
         return self.dit
 
@@ -62,10 +164,10 @@ class FluxImagePipeline(BasePipeline):
 
 
     @staticmethod
-    def from_model_manager(model_manager: ModelManager, controlnet_config_units: List[ControlNetConfigUnit]=[], prompt_refiner_classes=[], prompt_extender_classes=[], device=None):
+    def from_model_manager(model_manager: ModelManager, controlnet_config_units: List[ControlNetConfigUnit]=[], prompt_refiner_classes=[], prompt_extender_classes=[], device=None, torch_dtype=None):
         pipe = FluxImagePipeline(
             device=model_manager.device if device is None else device,
-            torch_dtype=model_manager.torch_dtype,
+            torch_dtype=model_manager.torch_dtype if torch_dtype is None else torch_dtype,
         )
         pipe.fetch_models(model_manager, controlnet_config_units, prompt_refiner_classes, prompt_extender_classes)
         return pipe

diffsynth/vram_management/__init__.py

@@ -0,0 +1 @@
+from .layers import *

diffsynth/vram_management/layers.py

@@ -0,0 +1,95 @@
+import torch, copy
+from ..models.utils import init_weights_on_device
+
+
+def cast_to(weight, dtype, device):
+    r = torch.empty_like(weight, dtype=dtype, device=device)
+    r.copy_(weight)
+    return r
+
+
+class AutoWrappedModule(torch.nn.Module):
+    def __init__(self, module: torch.nn.Module, offload_dtype, offload_device, onload_dtype, onload_device, computation_dtype, computation_device):
+        super().__init__()
+        self.module = module.to(dtype=offload_dtype, device=offload_device)
+        self.offload_dtype = offload_dtype
+        self.offload_device = offload_device
+        self.onload_dtype = onload_dtype
+        self.onload_device = onload_device
+        self.computation_dtype = computation_dtype
+        self.computation_device = computation_device
+        self.state = 0
+
+    def offload(self):
+        if self.state == 1 and (self.offload_dtype != self.onload_dtype or self.offload_device != self.onload_device):
+            self.module.to(dtype=self.offload_dtype, device=self.offload_device)
+            self.state = 0
+
+    def onload(self):
+        if self.state == 0 and (self.offload_dtype != self.onload_dtype or self.offload_device != self.onload_device):
+            self.module.to(dtype=self.onload_dtype, device=self.onload_device)
+            self.state = 1
+
+    def forward(self, *args, **kwargs):
+        if self.onload_dtype == self.computation_dtype and self.onload_device == self.computation_device:
+            module = self.module
+        else:
+            module = copy.deepcopy(self.module).to(dtype=self.computation_dtype, device=self.computation_device)
+        return module(*args, **kwargs)
+    
+
+class AutoWrappedLinear(torch.nn.Linear):
+    def __init__(self, module: torch.nn.Linear, offload_dtype, offload_device, onload_dtype, onload_device, computation_dtype, computation_device):
+        with init_weights_on_device(device=torch.device("meta")):
+            super().__init__(in_features=module.in_features, out_features=module.out_features, bias=module.bias is not None, dtype=offload_dtype, device=offload_device)
+        self.weight = module.weight
+        self.bias = module.bias
+        self.offload_dtype = offload_dtype
+        self.offload_device = offload_device
+        self.onload_dtype = onload_dtype
+        self.onload_device = onload_device
+        self.computation_dtype = computation_dtype
+        self.computation_device = computation_device
+        self.state = 0
+
+    def offload(self):
+        if self.state == 1 and (self.offload_dtype != self.onload_dtype or self.offload_device != self.onload_device):
+            self.to(dtype=self.offload_dtype, device=self.offload_device)
+            self.state = 0
+
+    def onload(self):
+        if self.state == 0 and (self.offload_dtype != self.onload_dtype or self.offload_device != self.onload_device):
+            self.to(dtype=self.onload_dtype, device=self.onload_device)
+            self.state = 1
+
+    def forward(self, x, *args, **kwargs):
+        if self.onload_dtype == self.computation_dtype and self.onload_device == self.computation_device:
+            weight, bias = self.weight, self.bias
+        else:
+            weight = cast_to(self.weight, self.computation_dtype, self.computation_device)
+            bias = None if self.bias is None else cast_to(self.bias, self.computation_dtype, self.computation_device)
+        return torch.nn.functional.linear(x, weight, bias)
+
+
+def enable_vram_management_recursively(model: torch.nn.Module, module_map: dict, module_config: dict, max_num_param=None, overflow_module_config: dict = None, total_num_param=0):
+    for name, module in model.named_children():
+        for source_module, target_module in module_map.items():
+            if isinstance(module, source_module):
+                num_param = sum(p.numel() for p in module.parameters())
+                if max_num_param is not None and total_num_param + num_param > max_num_param:
+                    module_config_ = overflow_module_config
+                else:
+                    module_config_ = module_config
+                module_ = target_module(module, **module_config_)
+                setattr(model, name, module_)
+                total_num_param += num_param
+                break
+        else:
+            total_num_param = enable_vram_management_recursively(module, module_map, module_config, max_num_param, overflow_module_config, total_num_param)
+    return total_num_param
+
+
+def enable_vram_management(model: torch.nn.Module, module_map: dict, module_config: dict, max_num_param=None, overflow_module_config: dict = None):
+    enable_vram_management_recursively(model, module_map, module_config, max_num_param, overflow_module_config, total_num_param=0)
+    model.vram_management_enabled = True
+

examples/vram_management/README.md

@@ -0,0 +1,3 @@
+# VRAM Management
+
+Experimental feature. Still under development.

examples/vram_management/flux_text_to_image.py

@@ -0,0 +1,25 @@
+import torch
+from diffsynth import ModelManager, FluxImagePipeline
+
+
+model_manager = ModelManager(
+    file_path_list=[
+        "models/FLUX/FLUX.1-dev/text_encoder/model.safetensors",
+        "models/FLUX/FLUX.1-dev/text_encoder_2",
+        "models/FLUX/FLUX.1-dev/flux1-dev.safetensors",
+        "models/FLUX/FLUX.1-dev/ae.safetensors",
+    ],
+    torch_dtype=torch.float8_e4m3fn,
+    device="cpu"
+)
+pipe = FluxImagePipeline.from_model_manager(model_manager, torch_dtype=torch.bfloat16, device="cuda")
+
+# Enable VRAM management
+# `num_persistent_param_in_dit` indicates the number of parameters that reside persistently in VRAM within the DiT model.
+# When `num_persistent_param_in_dit=None`, it means all parameters reside persistently in memory.
+# When `num_persistent_param_in_dit=7*10**9`, it indicates that 7 billion parameters reside persistently in memory.
+# When `num_persistent_param_in_dit=0`, it means no parameters reside persistently in memory, and they are loaded layer by layer during inference.
+pipe.enable_vram_management(num_persistent_param_in_dit=None)
+
+image = pipe(prompt="a beautiful orange cat", seed=0)
+image.save("image.jpg")