theluyuan/DiffSynth-Studio
1
0
Fork 0
mirror of https://github.com/modelscope/DiffSynth-Studio.git synced 2026-03-18 22:08:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

support qwen-image fp8 lora training

Browse Source
This commit is contained in:
Artiprocher
2025-08-25 20:32:36 +08:00
parent c795e35142
commit ce0b948655
3 changed files with 46 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
diffsynth/pipelines/qwen_image.py
View File

@@ -150,6 +150,35 @@ class QwenImagePipeline(BasePipeline):
return loss
def _enable_fp8_lora_training(self, dtype):
from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLRotaryEmbedding, Qwen2RMSNorm, Qwen2_5_VisionPatchEmbed, Qwen2_5_VisionRotaryEmbedding
from ..models.qwen_image_dit import RMSNorm
from ..models.qwen_image_vae import QwenImageRMS_norm
module_map = {
RMSNorm: AutoWrappedModule,
torch.nn.Linear: AutoWrappedLinear,
torch.nn.Conv3d: AutoWrappedModule,
torch.nn.Conv2d: AutoWrappedModule,
torch.nn.Embedding: AutoWrappedModule,
Qwen2_5_VLRotaryEmbedding: AutoWrappedModule,
Qwen2RMSNorm: AutoWrappedModule,
Qwen2_5_VisionPatchEmbed: AutoWrappedModule,
Qwen2_5_VisionRotaryEmbedding: AutoWrappedModule,
QwenImageRMS_norm: AutoWrappedModule,
}
model_config = dict(
offload_dtype=dtype,
offload_device="cuda",
onload_dtype=dtype,
onload_device="cuda",
computation_dtype=self.torch_dtype,
computation_device="cuda",
)
enable_vram_management(self.text_encoder, module_map=module_map, module_config=model_config)
enable_vram_management(self.dit, module_map=module_map, module_config=model_config)
enable_vram_management(self.vae, module_map=module_map, module_config=model_config)
def enable_vram_management(self, num_persistent_param_in_dit=None, vram_limit=None, vram_buffer=0.5, enable_dit_fp8_computation=False):
self.vram_management_enabled = True
if vram_limit is None:

7
diffsynth/trainers/utils.py
View File

@@ -338,11 +338,15 @@ class DiffusionTrainingModule(torch.nn.Module):
return trainable_param_names
def add_lora_to_model(self, model, target_modules, lora_rank, lora_alpha=None):
def add_lora_to_model(self, model, target_modules, lora_rank, lora_alpha=None, upcast_dtype=None):
if lora_alpha is None:
lora_alpha = lora_rank
lora_config = LoraConfig(r=lora_rank, lora_alpha=lora_alpha, target_modules=target_modules)
model = inject_adapter_in_model(lora_config, model)
if upcast_dtype is not None:
for param in model.parameters():
if param.requires_grad:
param.data = param.to(upcast_dtype)
return model
@@ -555,4 +559,5 @@ def qwen_image_parser():
parser.add_argument("--dataset_num_workers", type=int, default=0, help="Number of workers for data loading.")
parser.add_argument("--weight_decay", type=float, default=0.01, help="Weight decay.")
parser.add_argument("--processor_path", type=str, default=None, help="Path to the processor. If provided, the processor will be used for image editing.")
parser.add_argument("--enable_fp8_training", default=False, action="store_true", help="Whether to enable FP8 training. Only available for LoRA training on a single GPU.")
return parser