diff options
Diffstat (limited to 'modules/sd_models.py')
| -rw-r--r-- | modules/sd_models.py | 49 |
1 files changed, 46 insertions, 3 deletions
diff --git a/modules/sd_models.py b/modules/sd_models.py index 841402e8..90437c87 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -339,10 +339,28 @@ class SkipWritingToConfig: SkipWritingToConfig.skip = self.previous
+def check_fp8(model):
+ if model is None:
+ return None
+ if devices.get_optimal_device_name() == "mps":
+ enable_fp8 = False
+ elif shared.opts.fp8_storage == "Enable":
+ enable_fp8 = True
+ elif getattr(model, "is_sdxl", False) and shared.opts.fp8_storage == "Enable for SDXL":
+ enable_fp8 = True
+ else:
+ enable_fp8 = False
+ return enable_fp8
+
+
def load_model_weights(model, checkpoint_info: CheckpointInfo, state_dict, timer):
sd_model_hash = checkpoint_info.calculate_shorthash()
timer.record("calculate hash")
+ if devices.fp8:
+ # prevent model to load state dict in fp8
+ model.half()
+
if not SkipWritingToConfig.skip:
shared.opts.data["sd_model_checkpoint"] = checkpoint_info.title
@@ -395,6 +413,28 @@ def load_model_weights(model, checkpoint_info: CheckpointInfo, state_dict, timer devices.dtype_unet = torch.float16
timer.record("apply half()")
+ for module in model.modules():
+ if hasattr(module, 'fp16_weight'):
+ del module.fp16_weight
+ if hasattr(module, 'fp16_bias'):
+ del module.fp16_bias
+
+ if check_fp8(model):
+ devices.fp8 = True
+ first_stage = model.first_stage_model
+ model.first_stage_model = None
+ for module in model.modules():
+ if isinstance(module, (torch.nn.Conv2d, torch.nn.Linear)):
+ if shared.opts.cache_fp16_weight:
+ module.fp16_weight = module.weight.clone().half()
+ if module.bias is not None:
+ module.fp16_bias = module.bias.clone().half()
+ module.to(torch.float8_e4m3fn)
+ model.first_stage_model = first_stage
+ timer.record("apply fp8")
+ else:
+ devices.fp8 = False
+
devices.unet_needs_upcast = shared.cmd_opts.upcast_sampling and devices.dtype == torch.float16 and devices.dtype_unet == torch.float16
model.first_stage_model.to(devices.dtype_vae)
@@ -737,7 +777,7 @@ def reuse_model_from_already_loaded(sd_model, checkpoint_info, timer): return None
-def reload_model_weights(sd_model=None, info=None):
+def reload_model_weights(sd_model=None, info=None, forced_reload=False):
checkpoint_info = info or select_checkpoint()
timer = Timer()
@@ -749,11 +789,14 @@ def reload_model_weights(sd_model=None, info=None): current_checkpoint_info = None
else:
current_checkpoint_info = sd_model.sd_checkpoint_info
- if sd_model.sd_model_checkpoint == checkpoint_info.filename:
+ if check_fp8(sd_model) != devices.fp8:
+ # load from state dict again to prevent extra numerical errors
+ forced_reload = True
+ elif sd_model.sd_model_checkpoint == checkpoint_info.filename:
return sd_model
sd_model = reuse_model_from_already_loaded(sd_model, checkpoint_info, timer)
- if sd_model is not None and sd_model.sd_checkpoint_info.filename == checkpoint_info.filename:
+ if not forced_reload and sd_model is not None and sd_model.sd_checkpoint_info.filename == checkpoint_info.filename:
return sd_model
if sd_model is not None:
|