mirror of
https://git.datalinker.icu/ali-vilab/TeaCache
synced 2025-12-08 20:34:24 +08:00
272 lines
12 KiB
Python
272 lines
12 KiB
Python
import torchaudio
|
|
from tangoflux import TangoFluxInference
|
|
from typing import Any, Dict, Optional, Tuple, Union
|
|
from diffusers.models import FluxTransformer2DModel
|
|
from diffusers.models.modeling_outputs import Transformer2DModelOutput
|
|
from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
|
|
import torch
|
|
import numpy as np
|
|
import random
|
|
|
|
|
|
|
|
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
|
|
|
def teacache_forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
encoder_hidden_states: torch.Tensor = None,
|
|
pooled_projections: torch.Tensor = None,
|
|
timestep: torch.LongTensor = None,
|
|
img_ids: torch.Tensor = None,
|
|
txt_ids: torch.Tensor = None,
|
|
guidance: torch.Tensor = None,
|
|
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
|
|
return_dict: bool = True,
|
|
) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
|
|
"""
|
|
The [`FluxTransformer2DModel`] forward method.
|
|
|
|
Args:
|
|
hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
|
|
Input `hidden_states`.
|
|
encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
|
|
Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
|
|
pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
|
|
from the embeddings of input conditions.
|
|
timestep ( `torch.LongTensor`):
|
|
Used to indicate denoising step.
|
|
block_controlnet_hidden_states: (`list` of `torch.Tensor`):
|
|
A list of tensors that if specified are added to the residuals of transformer blocks.
|
|
joint_attention_kwargs (`dict`, *optional*):
|
|
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
|
`self.processor` in
|
|
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
|
return_dict (`bool`, *optional*, defaults to `True`):
|
|
Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
|
|
tuple.
|
|
|
|
Returns:
|
|
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
|
|
`tuple` where the first element is the sample tensor.
|
|
"""
|
|
if joint_attention_kwargs is not None:
|
|
joint_attention_kwargs = joint_attention_kwargs.copy()
|
|
lora_scale = joint_attention_kwargs.pop("scale", 1.0)
|
|
else:
|
|
lora_scale = 1.0
|
|
|
|
if USE_PEFT_BACKEND:
|
|
# weight the lora layers by setting `lora_scale` for each PEFT layer
|
|
scale_lora_layers(self, lora_scale)
|
|
else:
|
|
if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
|
|
logger.warning(
|
|
"Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
|
|
)
|
|
hidden_states = self.x_embedder(hidden_states)
|
|
|
|
timestep = timestep.to(hidden_states.dtype) * 1000
|
|
if guidance is not None:
|
|
guidance = guidance.to(hidden_states.dtype) * 1000
|
|
else:
|
|
guidance = None
|
|
temb = (
|
|
self.time_text_embed(timestep, pooled_projections)
|
|
if guidance is None
|
|
else self.time_text_embed(timestep, guidance, pooled_projections)
|
|
)
|
|
encoder_hidden_states = self.context_embedder(encoder_hidden_states)
|
|
|
|
ids = torch.cat((txt_ids, img_ids), dim=1)
|
|
image_rotary_emb = self.pos_embed(ids)
|
|
|
|
if self.enable_teacache:
|
|
inp = hidden_states.clone()
|
|
temb_ = temb.clone()
|
|
modulated_inp, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.transformer_blocks[0].norm1(inp, emb=temb_)
|
|
if self.cnt == 0 or self.cnt == self.num_steps-1:
|
|
should_calc = True
|
|
self.accumulated_rel_l1_distance = 0
|
|
else:
|
|
coefficients = [4.98651651e+02, -2.83781631e+02, 5.58554382e+01, -3.82021401e+00, 2.64230861e-01]
|
|
rescale_func = np.poly1d(coefficients)
|
|
self.accumulated_rel_l1_distance += rescale_func(((modulated_inp-self.previous_modulated_input).abs().mean() / self.previous_modulated_input.abs().mean()).cpu().item())
|
|
if self.accumulated_rel_l1_distance < self.rel_l1_thresh:
|
|
should_calc = False
|
|
else:
|
|
should_calc = True
|
|
self.accumulated_rel_l1_distance = 0
|
|
self.previous_modulated_input = modulated_inp
|
|
self.cnt += 1
|
|
if self.cnt == self.num_steps:
|
|
self.cnt = 0
|
|
|
|
if self.enable_teacache:
|
|
if not should_calc:
|
|
hidden_states += self.previous_residual
|
|
else:
|
|
ori_hidden_states = hidden_states.clone()
|
|
for index_block, block in enumerate(self.transformer_blocks):
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
def create_custom_forward(module, return_dict=None):
|
|
def custom_forward(*inputs):
|
|
if return_dict is not None:
|
|
return module(*inputs, return_dict=return_dict)
|
|
else:
|
|
return module(*inputs)
|
|
|
|
return custom_forward
|
|
|
|
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
|
encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(block),
|
|
hidden_states,
|
|
encoder_hidden_states,
|
|
temb,
|
|
image_rotary_emb,
|
|
**ckpt_kwargs,
|
|
)
|
|
|
|
else:
|
|
encoder_hidden_states, hidden_states = block(
|
|
hidden_states=hidden_states,
|
|
encoder_hidden_states=encoder_hidden_states,
|
|
temb=temb,
|
|
image_rotary_emb=image_rotary_emb,
|
|
)
|
|
|
|
hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
|
|
|
|
for index_block, block in enumerate(self.single_transformer_blocks):
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
def create_custom_forward(module, return_dict=None):
|
|
def custom_forward(*inputs):
|
|
if return_dict is not None:
|
|
return module(*inputs, return_dict=return_dict)
|
|
else:
|
|
return module(*inputs)
|
|
|
|
return custom_forward
|
|
|
|
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
|
hidden_states = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(block),
|
|
hidden_states,
|
|
temb,
|
|
image_rotary_emb,
|
|
**ckpt_kwargs,
|
|
)
|
|
|
|
else:
|
|
hidden_states = block(
|
|
hidden_states=hidden_states,
|
|
temb=temb,
|
|
image_rotary_emb=image_rotary_emb,
|
|
)
|
|
|
|
hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]
|
|
self.previous_residual = hidden_states - ori_hidden_states
|
|
else:
|
|
for index_block, block in enumerate(self.transformer_blocks):
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
def create_custom_forward(module, return_dict=None):
|
|
def custom_forward(*inputs):
|
|
if return_dict is not None:
|
|
return module(*inputs, return_dict=return_dict)
|
|
else:
|
|
return module(*inputs)
|
|
|
|
return custom_forward
|
|
|
|
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
|
encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(block),
|
|
hidden_states,
|
|
encoder_hidden_states,
|
|
temb,
|
|
image_rotary_emb,
|
|
**ckpt_kwargs,
|
|
)
|
|
|
|
else:
|
|
encoder_hidden_states, hidden_states = block(
|
|
hidden_states=hidden_states,
|
|
encoder_hidden_states=encoder_hidden_states,
|
|
temb=temb,
|
|
image_rotary_emb=image_rotary_emb,
|
|
)
|
|
|
|
hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
|
|
|
|
for index_block, block in enumerate(self.single_transformer_blocks):
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
def create_custom_forward(module, return_dict=None):
|
|
def custom_forward(*inputs):
|
|
if return_dict is not None:
|
|
return module(*inputs, return_dict=return_dict)
|
|
else:
|
|
return module(*inputs)
|
|
|
|
return custom_forward
|
|
|
|
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
|
hidden_states = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(block),
|
|
hidden_states,
|
|
temb,
|
|
image_rotary_emb,
|
|
**ckpt_kwargs,
|
|
)
|
|
|
|
else:
|
|
hidden_states = block(
|
|
hidden_states=hidden_states,
|
|
temb=temb,
|
|
image_rotary_emb=image_rotary_emb,
|
|
)
|
|
|
|
hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]
|
|
|
|
hidden_states = self.norm_out(hidden_states, temb)
|
|
output = self.proj_out(hidden_states)
|
|
|
|
if USE_PEFT_BACKEND:
|
|
# remove `lora_scale` from each PEFT layer
|
|
unscale_lora_layers(self, lora_scale)
|
|
|
|
if not return_dict:
|
|
return (output,)
|
|
|
|
return Transformer2DModelOutput(sample=output)
|
|
|
|
FluxTransformer2DModel.forward = teacache_forward
|
|
seed = 42
|
|
random.seed(seed)
|
|
np.random.seed(seed)
|
|
torch.manual_seed(seed)
|
|
if torch.cuda.is_available():
|
|
torch.cuda.manual_seed(seed)
|
|
torch.cuda.manual_seed_all(seed)
|
|
torch.backends.cudnn.deterministic = True
|
|
|
|
prompt = 'Hammer slowly hitting the wooden table'
|
|
steps = 50
|
|
|
|
|
|
model = TangoFluxInference(name='declare-lab/TangoFlux')
|
|
# TeaCache
|
|
model.model.transformer.__class__.enable_teacache = True
|
|
model.model.transformer.__class__.cnt = 0
|
|
model.model.transformer.__class__.num_steps = steps
|
|
model.model.transformer.__class__.rel_l1_thresh = 0.25 # 0.25 for 1.7x speedup, 0.4 for 2.1x speedup
|
|
model.model.transformer.__class__.accumulated_rel_l1_distance = 0
|
|
model.model.transformer.__class__.previous_modulated_input = None
|
|
model.model.transformer.__class__.previous_residual = None
|
|
|
|
audio = model.generate(prompt, steps=steps, duration=10)
|
|
torchaudio.save('teacache_tango_flux_{}.wav'.format(prompt), audio, 44100) |