427 lines
14 KiB
Python
427 lines
14 KiB
Python
import torch
|
|
from torch import Tensor
|
|
from comfy.ldm.flux.layers import timestep_embedding
|
|
|
|
|
|
def has_affordable_memory(device: torch.device) -> bool:
|
|
free_memory, _ = torch.cuda.mem_get_info(device)
|
|
free_memory_gb = free_memory / (1024**3)
|
|
return free_memory_gb > 24
|
|
|
|
|
|
def is_newer_than_ada_lovelace(device: torch.device) -> int:
|
|
cc_major, cc_minor = torch.cuda.get_device_capability(device)
|
|
return cc_major * 10 + cc_minor >= 89
|
|
|
|
|
|
teacache_coef = [
|
|
4.98651651e02,
|
|
-2.83781631e02,
|
|
5.58554382e01,
|
|
-3.82021401e00,
|
|
2.64230861e-01,
|
|
]
|
|
|
|
|
|
def poly1d(x):
|
|
result = torch.zeros_like(x)
|
|
for i, coeff in enumerate(teacache_coef):
|
|
result += coeff * (x ** (len(teacache_coef) - 1 - i))
|
|
return result
|
|
|
|
|
|
def are_tensors_similar(tensor1: Tensor, tensor2: Tensor, threshold: float):
|
|
assert tensor1.shape == tensor2.shape
|
|
diff_abs_mean = (tensor1 - tensor2).abs().mean()
|
|
tensor1_abs_mean = tensor1.abs().mean()
|
|
diff = diff_abs_mean / tensor1_abs_mean
|
|
return diff.item() < threshold
|
|
|
|
|
|
def skip_forward_orig(
|
|
self,
|
|
img: Tensor,
|
|
img_ids: Tensor,
|
|
txt: Tensor,
|
|
txt_ids: Tensor,
|
|
timesteps: Tensor,
|
|
y: Tensor,
|
|
guidance: Tensor = None,
|
|
control=None,
|
|
transformer_options={},
|
|
attn_mask: Tensor = None,
|
|
) -> Tensor:
|
|
patches_replace = transformer_options.get("patches_replace", {})
|
|
ds_skip_blocks = transformer_options.get("ds_skip_blocks", [])
|
|
ss_skip_blocks = transformer_options.get("ss_skip_blocks", [])
|
|
|
|
if img.ndim != 3 or txt.ndim != 3:
|
|
raise ValueError("Input img and txt tensors must have 3 dimensions.")
|
|
|
|
# running on sequences img
|
|
img = self.img_in(img)
|
|
vec = self.time_in(timestep_embedding(timesteps, 256).to(img.dtype))
|
|
if self.params.guidance_embed:
|
|
if guidance is None:
|
|
raise ValueError(
|
|
"Didn't get guidance strength for guidance distilled model."
|
|
)
|
|
vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
|
|
|
|
vec = vec + self.vector_in(y[:, : self.params.vec_in_dim])
|
|
txt = self.txt_in(txt)
|
|
|
|
ids = torch.cat((txt_ids, img_ids), dim=1)
|
|
pe = self.pe_embedder(ids)
|
|
|
|
blocks_replace = patches_replace.get("dit", {})
|
|
for i, block in enumerate(self.double_blocks):
|
|
#### skip blocks
|
|
if i in ds_skip_blocks:
|
|
continue
|
|
|
|
if ("double_block", i) in blocks_replace:
|
|
|
|
def block_wrap(args):
|
|
out = {}
|
|
out["img"], out["txt"] = block(
|
|
img=args["img"],
|
|
txt=args["txt"],
|
|
vec=args["vec"],
|
|
pe=args["pe"],
|
|
attn_mask=args.get("attn_mask"),
|
|
)
|
|
return out
|
|
|
|
out = blocks_replace[("double_block", i)](
|
|
{"img": img, "txt": txt, "vec": vec, "pe": pe, "attn_mask": attn_mask},
|
|
{"original_block": block_wrap},
|
|
)
|
|
txt = out["txt"]
|
|
img = out["img"]
|
|
else:
|
|
img, txt = block(img=img, txt=txt, vec=vec, pe=pe, attn_mask=attn_mask)
|
|
|
|
if control is not None: # Controlnet
|
|
control_i = control.get("input")
|
|
if i < len(control_i):
|
|
add = control_i[i]
|
|
if add is not None:
|
|
img += add
|
|
|
|
img = torch.cat((txt, img), 1)
|
|
|
|
for i, block in enumerate(self.single_blocks):
|
|
if ("single_block", i) in blocks_replace:
|
|
#### skip blocks
|
|
if i in ss_skip_blocks:
|
|
continue
|
|
|
|
def block_wrap(args):
|
|
out = {}
|
|
out["img"] = block(
|
|
args["img"],
|
|
vec=args["vec"],
|
|
pe=args["pe"],
|
|
attn_mask=args.get("attn_mask"),
|
|
)
|
|
return out
|
|
|
|
out = blocks_replace[("single_block", i)](
|
|
{"img": img, "vec": vec, "pe": pe, "attn_mask": attn_mask},
|
|
{"original_block": block_wrap},
|
|
)
|
|
img = out["img"]
|
|
else:
|
|
img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
|
|
|
|
if control is not None: # Controlnet
|
|
control_o = control.get("output")
|
|
if i < len(control_o):
|
|
add = control_o[i]
|
|
if add is not None:
|
|
img[:, txt.shape[1] :, ...] += add
|
|
|
|
img = img[:, txt.shape[1] :, ...]
|
|
|
|
img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels)
|
|
return img
|
|
|
|
|
|
def teacache_skip_forward_orig(
|
|
self,
|
|
img: Tensor,
|
|
img_ids: Tensor,
|
|
txt: Tensor,
|
|
txt_ids: Tensor,
|
|
timesteps: Tensor,
|
|
y: Tensor,
|
|
guidance: Tensor = None,
|
|
control=None,
|
|
transformer_options={},
|
|
attn_mask: Tensor = None,
|
|
) -> Tensor:
|
|
patches_replace = transformer_options.get("patches_replace", {})
|
|
rel_l1_threshold = transformer_options.get("rel_l1_threshold")
|
|
ds_skip_blocks = transformer_options.get("ds_skip_blocks")
|
|
ss_skip_blocks = transformer_options.get("ss_skip_blocks")
|
|
|
|
if img.ndim != 3 or txt.ndim != 3:
|
|
raise ValueError("Input img and txt tensors must have 3 dimensions.")
|
|
|
|
# running on sequences img
|
|
img = self.img_in(img)
|
|
vec = self.time_in(timestep_embedding(timesteps, 256).to(img.dtype))
|
|
if self.params.guidance_embed:
|
|
if guidance is None:
|
|
raise ValueError(
|
|
"Didn't get guidance strength for guidance distilled model."
|
|
)
|
|
vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
|
|
|
|
vec = vec + self.vector_in(y[:, : self.params.vec_in_dim])
|
|
txt = self.txt_in(txt)
|
|
|
|
ids = torch.cat((txt_ids, img_ids), dim=1)
|
|
pe = self.pe_embedder(ids)
|
|
|
|
blocks_replace = patches_replace.get("dit", {})
|
|
|
|
### TeaCache
|
|
img_, vec_ = img.clone(), vec.clone()
|
|
img_mod1, _ = self.double_blocks[0].img_mod(vec_)
|
|
curr_modulated_img = self.double_blocks[0].img_norm1(img_)
|
|
curr_modulated_img = (1 + img_mod1.scale) * curr_modulated_img + img_mod1.shift
|
|
|
|
try:
|
|
self.accum_rel_l1_distance += poly1d(
|
|
(curr_modulated_img - self.prev_modulated_img).abs().mean()
|
|
/ self.prev_modulated_img.abs().mean()
|
|
)
|
|
except:
|
|
torch._dynamo.graph_break() ### speeding up compiling
|
|
self.accum_rel_l1_distance = rel_l1_threshold
|
|
|
|
self.prev_modulated_img = curr_modulated_img
|
|
|
|
if self.accum_rel_l1_distance < rel_l1_threshold:
|
|
img += self.previous_residual
|
|
else:
|
|
self.accum_rel_l1_distance = 0
|
|
orig_img = img.clone()
|
|
|
|
for i, block in enumerate(self.double_blocks):
|
|
### skip blocks
|
|
if i in ds_skip_blocks:
|
|
continue
|
|
|
|
if ("double_block", i) in blocks_replace:
|
|
|
|
def block_wrap(args):
|
|
out = {}
|
|
out["img"], out["txt"] = block(
|
|
img=args["img"],
|
|
txt=args["txt"],
|
|
vec=args["vec"],
|
|
pe=args["pe"],
|
|
attn_mask=args.get("attn_mask"),
|
|
)
|
|
return out
|
|
|
|
out = blocks_replace[("double_block", i)](
|
|
{
|
|
"img": img,
|
|
"txt": txt,
|
|
"vec": vec,
|
|
"pe": pe,
|
|
"attn_mask": attn_mask,
|
|
},
|
|
{"original_block": block_wrap},
|
|
)
|
|
txt = out["txt"]
|
|
img = out["img"]
|
|
else:
|
|
img, txt = block(img=img, txt=txt, vec=vec, pe=pe, attn_mask=attn_mask)
|
|
|
|
if control is not None: # Controlnet
|
|
control_i = control.get("input")
|
|
if i < len(control_i):
|
|
add = control_i[i]
|
|
if add is not None:
|
|
img += add
|
|
|
|
img = torch.cat((txt, img), 1)
|
|
|
|
for i, block in enumerate(self.single_blocks):
|
|
if ("single_block", i) in blocks_replace:
|
|
### skip blocks
|
|
if i in ss_skip_blocks:
|
|
continue
|
|
|
|
def block_wrap(args):
|
|
out = {}
|
|
out["img"] = block(
|
|
args["img"],
|
|
vec=args["vec"],
|
|
pe=args["pe"],
|
|
attn_mask=args.get("attn_mask"),
|
|
)
|
|
return out
|
|
|
|
out = blocks_replace[("single_block", i)](
|
|
{"img": img, "vec": vec, "pe": pe, "attn_mask": attn_mask},
|
|
{"original_block": block_wrap},
|
|
)
|
|
img = out["img"]
|
|
else:
|
|
img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
|
|
|
|
if control is not None: # Controlnet
|
|
control_o = control.get("output")
|
|
if i < len(control_o):
|
|
add = control_o[i]
|
|
if add is not None:
|
|
img[:, txt.shape[1] :, ...] += add
|
|
|
|
img = img[:, txt.shape[1] :, ...]
|
|
|
|
self.previous_residual = img - orig_img
|
|
|
|
img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels)
|
|
return img
|
|
|
|
|
|
def fbcache_skip_forward_orig(
|
|
self,
|
|
img: Tensor,
|
|
img_ids: Tensor,
|
|
txt: Tensor,
|
|
txt_ids: Tensor,
|
|
timesteps: Tensor,
|
|
y: Tensor,
|
|
guidance: Tensor = None,
|
|
control=None,
|
|
transformer_options={},
|
|
attn_mask: Tensor = None,
|
|
) -> Tensor:
|
|
patches_replace = transformer_options.get("patches_replace", {})
|
|
fbcache_threshold = transformer_options.get("fbcache_threshold")
|
|
validate_use_cached = transformer_options.get("validator")
|
|
ds_skip_blocks = transformer_options.get("ds_skip_blocks")
|
|
ss_skip_blocks = transformer_options.get("ss_skip_blocks")
|
|
|
|
if img.ndim != 3 or txt.ndim != 3:
|
|
raise ValueError("Input img and txt tensors must have 3 dimensions.")
|
|
|
|
# running on sequences img
|
|
img = self.img_in(img)
|
|
vec = self.time_in(timestep_embedding(timesteps, 256).to(img.dtype))
|
|
if self.params.guidance_embed:
|
|
if guidance is None:
|
|
raise ValueError(
|
|
"Didn't get guidance strength for guidance distilled model."
|
|
)
|
|
vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
|
|
|
|
vec = vec + self.vector_in(y[:, : self.params.vec_in_dim])
|
|
txt = self.txt_in(txt)
|
|
|
|
ids = torch.cat((txt_ids, img_ids), dim=1)
|
|
pe = self.pe_embedder(ids)
|
|
|
|
blocks_replace = patches_replace.get("dit", {})
|
|
|
|
for i, block in enumerate(self.double_blocks):
|
|
#### skip blocks
|
|
if i in ds_skip_blocks:
|
|
if i == 0:
|
|
use_cached = False
|
|
continue
|
|
|
|
if ("double_block", i) in blocks_replace:
|
|
|
|
def block_wrap(args):
|
|
out = {}
|
|
out["img"], out["txt"] = block(
|
|
img=args["img"],
|
|
txt=args["txt"],
|
|
vec=args["vec"],
|
|
pe=args["pe"],
|
|
attn_mask=args.get("attn_mask"),
|
|
)
|
|
return out
|
|
|
|
out = blocks_replace[("double_block", i)](
|
|
{"img": img, "txt": txt, "vec": vec, "pe": pe, "attn_mask": attn_mask},
|
|
{"original_block": block_wrap},
|
|
)
|
|
txt = out["txt"]
|
|
img = out["img"]
|
|
else:
|
|
img, txt = block(img=img, txt=txt, vec=vec, pe=pe, attn_mask=attn_mask)
|
|
|
|
if control is not None: # Controlnet
|
|
control_i = control.get("input")
|
|
if i < len(control_i):
|
|
add = control_i[i]
|
|
if add is not None:
|
|
img += add
|
|
|
|
### FBCache
|
|
if i == 0:
|
|
try:
|
|
use_cached = are_tensors_similar(
|
|
img, self.prev_first_block_output, fbcache_threshold
|
|
)
|
|
use_cached = validate_use_cached(use_cached, timesteps.item())
|
|
except:
|
|
torch._dynamo.graph_break() ### speeding up compiling
|
|
use_cached = False
|
|
self.prev_first_block_output = img
|
|
|
|
if use_cached:
|
|
break
|
|
|
|
if use_cached:
|
|
img += self.prev_residual
|
|
else:
|
|
img = torch.cat((txt, img), 1)
|
|
|
|
for i, block in enumerate(self.single_blocks):
|
|
if ("single_block", i) in blocks_replace:
|
|
#### skip blocks
|
|
if i in ss_skip_blocks:
|
|
continue
|
|
|
|
def block_wrap(args):
|
|
out = {}
|
|
out["img"] = block(
|
|
args["img"],
|
|
vec=args["vec"],
|
|
pe=args["pe"],
|
|
attn_mask=args.get("attn_mask"),
|
|
)
|
|
return out
|
|
|
|
out = blocks_replace[("single_block", i)](
|
|
{"img": img, "vec": vec, "pe": pe, "attn_mask": attn_mask},
|
|
{"original_block": block_wrap},
|
|
)
|
|
img = out["img"]
|
|
else:
|
|
img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
|
|
|
|
if control is not None: # Controlnet
|
|
control_o = control.get("output")
|
|
if i < len(control_o):
|
|
add = control_o[i]
|
|
if add is not None:
|
|
img[:, txt.shape[1] :, ...] += add
|
|
|
|
img = img[:, txt.shape[1] :, ...]
|
|
|
|
self.prev_residual = img - self.prev_first_block_output
|
|
|
|
img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels)
|
|
return img
|