Spaces:
Running
on
L40S
Running
on
L40S
import torch | |
from .sd3_dit import TimestepEmbeddings, RMSNorm | |
from .utils import init_weights_on_device | |
from einops import rearrange, repeat | |
from tqdm import tqdm | |
from typing import Union, Tuple, List | |
from .utils import hash_state_dict_keys | |
def HunyuanVideoRope(latents): | |
def _to_tuple(x, dim=2): | |
if isinstance(x, int): | |
return (x,) * dim | |
elif len(x) == dim: | |
return x | |
else: | |
raise ValueError(f"Expected length {dim} or int, but got {x}") | |
def get_meshgrid_nd(start, *args, dim=2): | |
""" | |
Get n-D meshgrid with start, stop and num. | |
Args: | |
start (int or tuple): If len(args) == 0, start is num; If len(args) == 1, start is start, args[0] is stop, | |
step is 1; If len(args) == 2, start is start, args[0] is stop, args[1] is num. For n-dim, start/stop/num | |
should be int or n-tuple. If n-tuple is provided, the meshgrid will be stacked following the dim order in | |
n-tuples. | |
*args: See above. | |
dim (int): Dimension of the meshgrid. Defaults to 2. | |
Returns: | |
grid (np.ndarray): [dim, ...] | |
""" | |
if len(args) == 0: | |
# start is grid_size | |
num = _to_tuple(start, dim=dim) | |
start = (0,) * dim | |
stop = num | |
elif len(args) == 1: | |
# start is start, args[0] is stop, step is 1 | |
start = _to_tuple(start, dim=dim) | |
stop = _to_tuple(args[0], dim=dim) | |
num = [stop[i] - start[i] for i in range(dim)] | |
elif len(args) == 2: | |
# start is start, args[0] is stop, args[1] is num | |
start = _to_tuple(start, dim=dim) # Left-Top eg: 12,0 | |
stop = _to_tuple(args[0], dim=dim) # Right-Bottom eg: 20,32 | |
num = _to_tuple(args[1], dim=dim) # Target Size eg: 32,124 | |
else: | |
raise ValueError(f"len(args) should be 0, 1 or 2, but got {len(args)}") | |
# PyTorch implement of np.linspace(start[i], stop[i], num[i], endpoint=False) | |
axis_grid = [] | |
for i in range(dim): | |
a, b, n = start[i], stop[i], num[i] | |
g = torch.linspace(a, b, n + 1, dtype=torch.float32)[:n] | |
axis_grid.append(g) | |
grid = torch.meshgrid(*axis_grid, indexing="ij") # dim x [W, H, D] | |
grid = torch.stack(grid, dim=0) # [dim, W, H, D] | |
return grid | |
def get_1d_rotary_pos_embed( | |
dim: int, | |
pos: Union[torch.FloatTensor, int], | |
theta: float = 10000.0, | |
use_real: bool = False, | |
theta_rescale_factor: float = 1.0, | |
interpolation_factor: float = 1.0, | |
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]: | |
""" | |
Precompute the frequency tensor for complex exponential (cis) with given dimensions. | |
(Note: `cis` means `cos + i * sin`, where i is the imaginary unit.) | |
This function calculates a frequency tensor with complex exponential using the given dimension 'dim' | |
and the end index 'end'. The 'theta' parameter scales the frequencies. | |
The returned tensor contains complex values in complex64 data type. | |
Args: | |
dim (int): Dimension of the frequency tensor. | |
pos (int or torch.FloatTensor): Position indices for the frequency tensor. [S] or scalar | |
theta (float, optional): Scaling factor for frequency computation. Defaults to 10000.0. | |
use_real (bool, optional): If True, return real part and imaginary part separately. | |
Otherwise, return complex numbers. | |
theta_rescale_factor (float, optional): Rescale factor for theta. Defaults to 1.0. | |
Returns: | |
freqs_cis: Precomputed frequency tensor with complex exponential. [S, D/2] | |
freqs_cos, freqs_sin: Precomputed frequency tensor with real and imaginary parts separately. [S, D] | |
""" | |
if isinstance(pos, int): | |
pos = torch.arange(pos).float() | |
# proposed by reddit user bloc97, to rescale rotary embeddings to longer sequence length without fine-tuning | |
# has some connection to NTK literature | |
if theta_rescale_factor != 1.0: | |
theta *= theta_rescale_factor ** (dim / (dim - 2)) | |
freqs = 1.0 / ( | |
theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim) | |
) # [D/2] | |
# assert interpolation_factor == 1.0, f"interpolation_factor: {interpolation_factor}" | |
freqs = torch.outer(pos * interpolation_factor, freqs) # [S, D/2] | |
if use_real: | |
freqs_cos = freqs.cos().repeat_interleave(2, dim=1) # [S, D] | |
freqs_sin = freqs.sin().repeat_interleave(2, dim=1) # [S, D] | |
return freqs_cos, freqs_sin | |
else: | |
freqs_cis = torch.polar( | |
torch.ones_like(freqs), freqs | |
) # complex64 # [S, D/2] | |
return freqs_cis | |
def get_nd_rotary_pos_embed( | |
rope_dim_list, | |
start, | |
*args, | |
theta=10000.0, | |
use_real=False, | |
theta_rescale_factor: Union[float, List[float]] = 1.0, | |
interpolation_factor: Union[float, List[float]] = 1.0, | |
): | |
""" | |
This is a n-d version of precompute_freqs_cis, which is a RoPE for tokens with n-d structure. | |
Args: | |
rope_dim_list (list of int): Dimension of each rope. len(rope_dim_list) should equal to n. | |
sum(rope_dim_list) should equal to head_dim of attention layer. | |
start (int | tuple of int | list of int): If len(args) == 0, start is num; If len(args) == 1, start is start, | |
args[0] is stop, step is 1; If len(args) == 2, start is start, args[0] is stop, args[1] is num. | |
*args: See above. | |
theta (float): Scaling factor for frequency computation. Defaults to 10000.0. | |
use_real (bool): If True, return real part and imaginary part separately. Otherwise, return complex numbers. | |
Some libraries such as TensorRT does not support complex64 data type. So it is useful to provide a real | |
part and an imaginary part separately. | |
theta_rescale_factor (float): Rescale factor for theta. Defaults to 1.0. | |
Returns: | |
pos_embed (torch.Tensor): [HW, D/2] | |
""" | |
grid = get_meshgrid_nd( | |
start, *args, dim=len(rope_dim_list) | |
) # [3, W, H, D] / [2, W, H] | |
if isinstance(theta_rescale_factor, int) or isinstance(theta_rescale_factor, float): | |
theta_rescale_factor = [theta_rescale_factor] * len(rope_dim_list) | |
elif isinstance(theta_rescale_factor, list) and len(theta_rescale_factor) == 1: | |
theta_rescale_factor = [theta_rescale_factor[0]] * len(rope_dim_list) | |
assert len(theta_rescale_factor) == len( | |
rope_dim_list | |
), "len(theta_rescale_factor) should equal to len(rope_dim_list)" | |
if isinstance(interpolation_factor, int) or isinstance(interpolation_factor, float): | |
interpolation_factor = [interpolation_factor] * len(rope_dim_list) | |
elif isinstance(interpolation_factor, list) and len(interpolation_factor) == 1: | |
interpolation_factor = [interpolation_factor[0]] * len(rope_dim_list) | |
assert len(interpolation_factor) == len( | |
rope_dim_list | |
), "len(interpolation_factor) should equal to len(rope_dim_list)" | |
# use 1/ndim of dimensions to encode grid_axis | |
embs = [] | |
for i in range(len(rope_dim_list)): | |
emb = get_1d_rotary_pos_embed( | |
rope_dim_list[i], | |
grid[i].reshape(-1), | |
theta, | |
use_real=use_real, | |
theta_rescale_factor=theta_rescale_factor[i], | |
interpolation_factor=interpolation_factor[i], | |
) # 2 x [WHD, rope_dim_list[i]] | |
embs.append(emb) | |
if use_real: | |
cos = torch.cat([emb[0] for emb in embs], dim=1) # (WHD, D/2) | |
sin = torch.cat([emb[1] for emb in embs], dim=1) # (WHD, D/2) | |
return cos, sin | |
else: | |
emb = torch.cat(embs, dim=1) # (WHD, D/2) | |
return emb | |
freqs_cos, freqs_sin = get_nd_rotary_pos_embed( | |
[16, 56, 56], | |
[latents.shape[2], latents.shape[3] // 2, latents.shape[4] // 2], | |
theta=256, | |
use_real=True, | |
theta_rescale_factor=1, | |
) | |
return freqs_cos, freqs_sin | |
class PatchEmbed(torch.nn.Module): | |
def __init__(self, patch_size=(1, 2, 2), in_channels=16, embed_dim=3072): | |
super().__init__() | |
self.proj = torch.nn.Conv3d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size) | |
def forward(self, x): | |
x = self.proj(x) | |
x = x.flatten(2).transpose(1, 2) | |
return x | |
class IndividualTokenRefinerBlock(torch.nn.Module): | |
def __init__(self, hidden_size=3072, num_heads=24): | |
super().__init__() | |
self.num_heads = num_heads | |
self.norm1 = torch.nn.LayerNorm(hidden_size, elementwise_affine=True, eps=1e-6) | |
self.self_attn_qkv = torch.nn.Linear(hidden_size, hidden_size * 3) | |
self.self_attn_proj = torch.nn.Linear(hidden_size, hidden_size) | |
self.norm2 = torch.nn.LayerNorm(hidden_size, elementwise_affine=True, eps=1e-6) | |
self.mlp = torch.nn.Sequential( | |
torch.nn.Linear(hidden_size, hidden_size * 4), | |
torch.nn.SiLU(), | |
torch.nn.Linear(hidden_size * 4, hidden_size) | |
) | |
self.adaLN_modulation = torch.nn.Sequential( | |
torch.nn.SiLU(), | |
torch.nn.Linear(hidden_size, hidden_size * 2, device="cuda", dtype=torch.bfloat16), | |
) | |
def forward(self, x, c, attn_mask=None): | |
gate_msa, gate_mlp = self.adaLN_modulation(c).chunk(2, dim=1) | |
norm_x = self.norm1(x) | |
qkv = self.self_attn_qkv(norm_x) | |
q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads) | |
attn = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask) | |
attn = rearrange(attn, "B H L D -> B L (H D)") | |
x = x + self.self_attn_proj(attn) * gate_msa.unsqueeze(1) | |
x = x + self.mlp(self.norm2(x)) * gate_mlp.unsqueeze(1) | |
return x | |
class SingleTokenRefiner(torch.nn.Module): | |
def __init__(self, in_channels=4096, hidden_size=3072, depth=2): | |
super().__init__() | |
self.input_embedder = torch.nn.Linear(in_channels, hidden_size, bias=True) | |
self.t_embedder = TimestepEmbeddings(256, hidden_size, computation_device="cpu") | |
self.c_embedder = torch.nn.Sequential( | |
torch.nn.Linear(in_channels, hidden_size), | |
torch.nn.SiLU(), | |
torch.nn.Linear(hidden_size, hidden_size) | |
) | |
self.blocks = torch.nn.ModuleList([IndividualTokenRefinerBlock(hidden_size=hidden_size) for _ in range(depth)]) | |
def forward(self, x, t, mask=None): | |
timestep_aware_representations = self.t_embedder(t, dtype=torch.float32) | |
mask_float = mask.float().unsqueeze(-1) | |
context_aware_representations = (x * mask_float).sum(dim=1) / mask_float.sum(dim=1) | |
context_aware_representations = self.c_embedder(context_aware_representations) | |
c = timestep_aware_representations + context_aware_representations | |
x = self.input_embedder(x) | |
mask = mask.to(device=x.device, dtype=torch.bool) | |
mask = repeat(mask, "B L -> B 1 D L", D=mask.shape[-1]) | |
mask = mask & mask.transpose(2, 3) | |
mask[:, :, :, 0] = True | |
for block in self.blocks: | |
x = block(x, c, mask) | |
return x | |
class ModulateDiT(torch.nn.Module): | |
def __init__(self, hidden_size, factor=6): | |
super().__init__() | |
self.act = torch.nn.SiLU() | |
self.linear = torch.nn.Linear(hidden_size, factor * hidden_size) | |
def forward(self, x): | |
return self.linear(self.act(x)) | |
def modulate(x, shift=None, scale=None, tr_shift=None, tr_scale=None, tr_token=None): | |
if tr_shift is not None: | |
x_zero = x[:, :tr_token] * (1 + tr_scale.unsqueeze(1)) + tr_shift.unsqueeze(1) | |
x_orig = x[:, tr_token:] * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1) | |
x = torch.concat((x_zero, x_orig), dim=1) | |
return x | |
if scale is None and shift is None: | |
return x | |
elif shift is None: | |
return x * (1 + scale.unsqueeze(1)) | |
elif scale is None: | |
return x + shift.unsqueeze(1) | |
else: | |
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1) | |
def reshape_for_broadcast( | |
freqs_cis, | |
x: torch.Tensor, | |
head_first=False, | |
): | |
ndim = x.ndim | |
assert 0 <= 1 < ndim | |
if isinstance(freqs_cis, tuple): | |
# freqs_cis: (cos, sin) in real space | |
if head_first: | |
assert freqs_cis[0].shape == ( | |
x.shape[-2], | |
x.shape[-1], | |
), f"freqs_cis shape {freqs_cis[0].shape} does not match x shape {x.shape}" | |
shape = [ | |
d if i == ndim - 2 or i == ndim - 1 else 1 | |
for i, d in enumerate(x.shape) | |
] | |
else: | |
assert freqs_cis[0].shape == ( | |
x.shape[1], | |
x.shape[-1], | |
), f"freqs_cis shape {freqs_cis[0].shape} does not match x shape {x.shape}" | |
shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)] | |
return freqs_cis[0].view(*shape), freqs_cis[1].view(*shape) | |
else: | |
# freqs_cis: values in complex space | |
if head_first: | |
assert freqs_cis.shape == ( | |
x.shape[-2], | |
x.shape[-1], | |
), f"freqs_cis shape {freqs_cis.shape} does not match x shape {x.shape}" | |
shape = [ | |
d if i == ndim - 2 or i == ndim - 1 else 1 | |
for i, d in enumerate(x.shape) | |
] | |
else: | |
assert freqs_cis.shape == ( | |
x.shape[1], | |
x.shape[-1], | |
), f"freqs_cis shape {freqs_cis.shape} does not match x shape {x.shape}" | |
shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)] | |
return freqs_cis.view(*shape) | |
def rotate_half(x): | |
x_real, x_imag = ( | |
x.float().reshape(*x.shape[:-1], -1, 2).unbind(-1) | |
) # [B, S, H, D//2] | |
return torch.stack([-x_imag, x_real], dim=-1).flatten(3) | |
def apply_rotary_emb( | |
xq: torch.Tensor, | |
xk: torch.Tensor, | |
freqs_cis, | |
head_first: bool = False, | |
): | |
xk_out = None | |
if isinstance(freqs_cis, tuple): | |
cos, sin = reshape_for_broadcast(freqs_cis, xq, head_first) # [S, D] | |
cos, sin = cos.to(xq.device), sin.to(xq.device) | |
# real * cos - imag * sin | |
# imag * cos + real * sin | |
xq_out = (xq.float() * cos + rotate_half(xq.float()) * sin).type_as(xq) | |
xk_out = (xk.float() * cos + rotate_half(xk.float()) * sin).type_as(xk) | |
else: | |
# view_as_complex will pack [..., D/2, 2](real) to [..., D/2](complex) | |
xq_ = torch.view_as_complex( | |
xq.float().reshape(*xq.shape[:-1], -1, 2) | |
) # [B, S, H, D//2] | |
freqs_cis = reshape_for_broadcast(freqs_cis, xq_, head_first).to( | |
xq.device | |
) # [S, D//2] --> [1, S, 1, D//2] | |
# (real, imag) * (cos, sin) = (real * cos - imag * sin, imag * cos + real * sin) | |
# view_as_real will expand [..., D/2](complex) to [..., D/2, 2](real) | |
xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3).type_as(xq) | |
xk_ = torch.view_as_complex( | |
xk.float().reshape(*xk.shape[:-1], -1, 2) | |
) # [B, S, H, D//2] | |
xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3).type_as(xk) | |
return xq_out, xk_out | |
def attention(q, k, v): | |
q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2) | |
x = torch.nn.functional.scaled_dot_product_attention(q, k, v) | |
x = x.transpose(1, 2).flatten(2, 3) | |
return x | |
def apply_gate(x, gate, tr_gate=None, tr_token=None): | |
if tr_gate is not None: | |
x_zero = x[:, :tr_token] * tr_gate.unsqueeze(1) | |
x_orig = x[:, tr_token:] * gate.unsqueeze(1) | |
return torch.concat((x_zero, x_orig), dim=1) | |
else: | |
return x * gate.unsqueeze(1) | |
class MMDoubleStreamBlockComponent(torch.nn.Module): | |
def __init__(self, hidden_size=3072, heads_num=24, mlp_width_ratio=4): | |
super().__init__() | |
self.heads_num = heads_num | |
self.mod = ModulateDiT(hidden_size) | |
self.norm1 = torch.nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) | |
self.to_qkv = torch.nn.Linear(hidden_size, hidden_size * 3) | |
self.norm_q = RMSNorm(dim=hidden_size // heads_num, eps=1e-6) | |
self.norm_k = RMSNorm(dim=hidden_size // heads_num, eps=1e-6) | |
self.to_out = torch.nn.Linear(hidden_size, hidden_size) | |
self.norm2 = torch.nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) | |
self.ff = torch.nn.Sequential( | |
torch.nn.Linear(hidden_size, hidden_size * mlp_width_ratio), | |
torch.nn.GELU(approximate="tanh"), | |
torch.nn.Linear(hidden_size * mlp_width_ratio, hidden_size) | |
) | |
def forward(self, hidden_states, conditioning, freqs_cis=None, token_replace_vec=None, tr_token=None): | |
mod1_shift, mod1_scale, mod1_gate, mod2_shift, mod2_scale, mod2_gate = self.mod(conditioning).chunk(6, dim=-1) | |
if token_replace_vec is not None: | |
assert tr_token is not None | |
tr_mod1_shift, tr_mod1_scale, tr_mod1_gate, tr_mod2_shift, tr_mod2_scale, tr_mod2_gate = self.mod(token_replace_vec).chunk(6, dim=-1) | |
else: | |
tr_mod1_shift, tr_mod1_scale, tr_mod1_gate, tr_mod2_shift, tr_mod2_scale, tr_mod2_gate = None, None, None, None, None, None | |
norm_hidden_states = self.norm1(hidden_states) | |
norm_hidden_states = modulate(norm_hidden_states, shift=mod1_shift, scale=mod1_scale, | |
tr_shift=tr_mod1_shift, tr_scale=tr_mod1_scale, tr_token=tr_token) | |
qkv = self.to_qkv(norm_hidden_states) | |
q, k, v = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.heads_num) | |
q = self.norm_q(q) | |
k = self.norm_k(k) | |
if freqs_cis is not None: | |
q, k = apply_rotary_emb(q, k, freqs_cis, head_first=False) | |
return (q, k, v), (mod1_gate, mod2_shift, mod2_scale, mod2_gate), (tr_mod1_gate, tr_mod2_shift, tr_mod2_scale, tr_mod2_gate) | |
def process_ff(self, hidden_states, attn_output, mod, mod_tr=None, tr_token=None): | |
mod1_gate, mod2_shift, mod2_scale, mod2_gate = mod | |
if mod_tr is not None: | |
tr_mod1_gate, tr_mod2_shift, tr_mod2_scale, tr_mod2_gate = mod_tr | |
else: | |
tr_mod1_gate, tr_mod2_shift, tr_mod2_scale, tr_mod2_gate = None, None, None, None | |
hidden_states = hidden_states + apply_gate(self.to_out(attn_output), mod1_gate, tr_mod1_gate, tr_token) | |
x = self.ff(modulate(self.norm2(hidden_states), shift=mod2_shift, scale=mod2_scale, tr_shift=tr_mod2_shift, tr_scale=tr_mod2_scale, tr_token=tr_token)) | |
hidden_states = hidden_states + apply_gate(x, mod2_gate, tr_mod2_gate, tr_token) | |
return hidden_states | |
class MMDoubleStreamBlock(torch.nn.Module): | |
def __init__(self, hidden_size=3072, heads_num=24, mlp_width_ratio=4): | |
super().__init__() | |
self.component_a = MMDoubleStreamBlockComponent(hidden_size, heads_num, mlp_width_ratio) | |
self.component_b = MMDoubleStreamBlockComponent(hidden_size, heads_num, mlp_width_ratio) | |
def forward(self, hidden_states_a, hidden_states_b, conditioning, freqs_cis, token_replace_vec=None, tr_token=None, split_token=71): | |
(q_a, k_a, v_a), mod_a, mod_tr = self.component_a(hidden_states_a, conditioning, freqs_cis, token_replace_vec, tr_token) | |
(q_b, k_b, v_b), mod_b, _ = self.component_b(hidden_states_b, conditioning, freqs_cis=None) | |
q_a, q_b = torch.concat([q_a, q_b[:, :split_token]], dim=1), q_b[:, split_token:].contiguous() | |
k_a, k_b = torch.concat([k_a, k_b[:, :split_token]], dim=1), k_b[:, split_token:].contiguous() | |
v_a, v_b = torch.concat([v_a, v_b[:, :split_token]], dim=1), v_b[:, split_token:].contiguous() | |
attn_output_a = attention(q_a, k_a, v_a) | |
attn_output_b = attention(q_b, k_b, v_b) | |
attn_output_a, attn_output_b = attn_output_a[:, :-split_token].contiguous(), torch.concat([attn_output_a[:, -split_token:], attn_output_b], dim=1) | |
hidden_states_a = self.component_a.process_ff(hidden_states_a, attn_output_a, mod_a, mod_tr, tr_token) | |
hidden_states_b = self.component_b.process_ff(hidden_states_b, attn_output_b, mod_b) | |
return hidden_states_a, hidden_states_b | |
class MMSingleStreamBlockOriginal(torch.nn.Module): | |
def __init__(self, hidden_size=3072, heads_num=24, mlp_width_ratio=4): | |
super().__init__() | |
self.hidden_size = hidden_size | |
self.heads_num = heads_num | |
self.mlp_hidden_dim = hidden_size * mlp_width_ratio | |
self.linear1 = torch.nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim) | |
self.linear2 = torch.nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size) | |
self.q_norm = RMSNorm(dim=hidden_size // heads_num, eps=1e-6) | |
self.k_norm = RMSNorm(dim=hidden_size // heads_num, eps=1e-6) | |
self.pre_norm = torch.nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) | |
self.mlp_act = torch.nn.GELU(approximate="tanh") | |
self.modulation = ModulateDiT(hidden_size, factor=3) | |
def forward(self, x, vec, freqs_cis=None, txt_len=256): | |
mod_shift, mod_scale, mod_gate = self.modulation(vec).chunk(3, dim=-1) | |
x_mod = modulate(self.pre_norm(x), shift=mod_shift, scale=mod_scale) | |
qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) | |
q, k, v = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.heads_num) | |
q = self.q_norm(q) | |
k = self.k_norm(k) | |
q_a, q_b = q[:, :-txt_len, :, :], q[:, -txt_len:, :, :] | |
k_a, k_b = k[:, :-txt_len, :, :], k[:, -txt_len:, :, :] | |
q_a, k_a = apply_rotary_emb(q_a, k_a, freqs_cis, head_first=False) | |
q = torch.cat((q_a, q_b), dim=1) | |
k = torch.cat((k_a, k_b), dim=1) | |
attn_output_a = attention(q[:, :-185].contiguous(), k[:, :-185].contiguous(), v[:, :-185].contiguous()) | |
attn_output_b = attention(q[:, -185:].contiguous(), k[:, -185:].contiguous(), v[:, -185:].contiguous()) | |
attn_output = torch.concat([attn_output_a, attn_output_b], dim=1) | |
output = self.linear2(torch.cat((attn_output, self.mlp_act(mlp)), 2)) | |
return x + output * mod_gate.unsqueeze(1) | |
class MMSingleStreamBlock(torch.nn.Module): | |
def __init__(self, hidden_size=3072, heads_num=24, mlp_width_ratio=4): | |
super().__init__() | |
self.heads_num = heads_num | |
self.mod = ModulateDiT(hidden_size, factor=3) | |
self.norm = torch.nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) | |
self.to_qkv = torch.nn.Linear(hidden_size, hidden_size * 3) | |
self.norm_q = RMSNorm(dim=hidden_size // heads_num, eps=1e-6) | |
self.norm_k = RMSNorm(dim=hidden_size // heads_num, eps=1e-6) | |
self.to_out = torch.nn.Linear(hidden_size, hidden_size) | |
self.ff = torch.nn.Sequential( | |
torch.nn.Linear(hidden_size, hidden_size * mlp_width_ratio), | |
torch.nn.GELU(approximate="tanh"), | |
torch.nn.Linear(hidden_size * mlp_width_ratio, hidden_size, bias=False) | |
) | |
def forward(self, hidden_states, conditioning, freqs_cis=None, txt_len=256, token_replace_vec=None, tr_token=None, split_token=71): | |
mod_shift, mod_scale, mod_gate = self.mod(conditioning).chunk(3, dim=-1) | |
if token_replace_vec is not None: | |
assert tr_token is not None | |
tr_mod_shift, tr_mod_scale, tr_mod_gate = self.mod(token_replace_vec).chunk(3, dim=-1) | |
else: | |
tr_mod_shift, tr_mod_scale, tr_mod_gate = None, None, None | |
norm_hidden_states = self.norm(hidden_states) | |
norm_hidden_states = modulate(norm_hidden_states, shift=mod_shift, scale=mod_scale, | |
tr_shift=tr_mod_shift, tr_scale=tr_mod_scale, tr_token=tr_token) | |
qkv = self.to_qkv(norm_hidden_states) | |
q, k, v = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.heads_num) | |
q = self.norm_q(q) | |
k = self.norm_k(k) | |
q_a, q_b = q[:, :-txt_len, :, :], q[:, -txt_len:, :, :] | |
k_a, k_b = k[:, :-txt_len, :, :], k[:, -txt_len:, :, :] | |
q_a, k_a = apply_rotary_emb(q_a, k_a, freqs_cis, head_first=False) | |
v_len = txt_len - split_token | |
q_a, q_b = torch.concat([q_a, q_b[:, :split_token]], dim=1), q_b[:, split_token:].contiguous() | |
k_a, k_b = torch.concat([k_a, k_b[:, :split_token]], dim=1), k_b[:, split_token:].contiguous() | |
v_a, v_b = v[:, :-v_len].contiguous(), v[:, -v_len:].contiguous() | |
attn_output_a = attention(q_a, k_a, v_a) | |
attn_output_b = attention(q_b, k_b, v_b) | |
attn_output = torch.concat([attn_output_a, attn_output_b], dim=1) | |
hidden_states = hidden_states + apply_gate(self.to_out(attn_output), mod_gate, tr_mod_gate, tr_token) | |
hidden_states = hidden_states + apply_gate(self.ff(norm_hidden_states), mod_gate, tr_mod_gate, tr_token) | |
return hidden_states | |
class FinalLayer(torch.nn.Module): | |
def __init__(self, hidden_size=3072, patch_size=(1, 2, 2), out_channels=16): | |
super().__init__() | |
self.norm_final = torch.nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) | |
self.linear = torch.nn.Linear(hidden_size, patch_size[0] * patch_size[1] * patch_size[2] * out_channels) | |
self.adaLN_modulation = torch.nn.Sequential(torch.nn.SiLU(), torch.nn.Linear(hidden_size, 2 * hidden_size)) | |
def forward(self, x, c): | |
shift, scale = self.adaLN_modulation(c).chunk(2, dim=1) | |
x = modulate(self.norm_final(x), shift=shift, scale=scale) | |
x = self.linear(x) | |
return x | |
class HunyuanVideoDiT(torch.nn.Module): | |
def __init__(self, in_channels=16, hidden_size=3072, text_dim=4096, num_double_blocks=20, num_single_blocks=40, guidance_embed=True): | |
super().__init__() | |
self.img_in = PatchEmbed(in_channels=in_channels, embed_dim=hidden_size) | |
self.txt_in = SingleTokenRefiner(in_channels=text_dim, hidden_size=hidden_size) | |
self.time_in = TimestepEmbeddings(256, hidden_size, computation_device="cpu") | |
self.vector_in = torch.nn.Sequential( | |
torch.nn.Linear(768, hidden_size), | |
torch.nn.SiLU(), | |
torch.nn.Linear(hidden_size, hidden_size) | |
) | |
self.guidance_in = TimestepEmbeddings(256, hidden_size, computation_device="cpu") if guidance_embed else None | |
self.double_blocks = torch.nn.ModuleList([MMDoubleStreamBlock(hidden_size) for _ in range(num_double_blocks)]) | |
self.single_blocks = torch.nn.ModuleList([MMSingleStreamBlock(hidden_size) for _ in range(num_single_blocks)]) | |
self.final_layer = FinalLayer(hidden_size) | |
# TODO: remove these parameters | |
self.dtype = torch.bfloat16 | |
self.patch_size = [1, 2, 2] | |
self.hidden_size = 3072 | |
self.heads_num = 24 | |
self.rope_dim_list = [16, 56, 56] | |
def unpatchify(self, x, T, H, W): | |
x = rearrange(x, "B (T H W) (C pT pH pW) -> B C (T pT) (H pH) (W pW)", H=H, W=W, pT=1, pH=2, pW=2) | |
return x | |
def enable_block_wise_offload(self, warm_device="cuda", cold_device="cpu"): | |
self.warm_device = warm_device | |
self.cold_device = cold_device | |
self.to(self.cold_device) | |
def load_models_to_device(self, loadmodel_names=[], device="cpu"): | |
for model_name in loadmodel_names: | |
model = getattr(self, model_name) | |
if model is not None: | |
model.to(device) | |
torch.cuda.empty_cache() | |
def prepare_freqs(self, latents): | |
return HunyuanVideoRope(latents) | |
def forward( | |
self, | |
x: torch.Tensor, | |
t: torch.Tensor, | |
prompt_emb: torch.Tensor = None, | |
text_mask: torch.Tensor = None, | |
pooled_prompt_emb: torch.Tensor = None, | |
freqs_cos: torch.Tensor = None, | |
freqs_sin: torch.Tensor = None, | |
guidance: torch.Tensor = None, | |
**kwargs | |
): | |
B, C, T, H, W = x.shape | |
vec = self.time_in(t, dtype=torch.float32) + self.vector_in(pooled_prompt_emb) | |
if self.guidance_in is not None: | |
vec += self.guidance_in(guidance * 1000, dtype=torch.float32) | |
img = self.img_in(x) | |
txt = self.txt_in(prompt_emb, t, text_mask) | |
for block in tqdm(self.double_blocks, desc="Double stream blocks"): | |
img, txt = block(img, txt, vec, (freqs_cos, freqs_sin)) | |
x = torch.concat([img, txt], dim=1) | |
for block in tqdm(self.single_blocks, desc="Single stream blocks"): | |
x = block(x, vec, (freqs_cos, freqs_sin)) | |
img = x[:, :-256] | |
img = self.final_layer(img, vec) | |
img = self.unpatchify(img, T=T//1, H=H//2, W=W//2) | |
return img | |
def enable_auto_offload(self, dtype=torch.bfloat16, device="cuda"): | |
def cast_to(weight, dtype=None, device=None, copy=False): | |
if device is None or weight.device == device: | |
if not copy: | |
if dtype is None or weight.dtype == dtype: | |
return weight | |
return weight.to(dtype=dtype, copy=copy) | |
r = torch.empty_like(weight, dtype=dtype, device=device) | |
r.copy_(weight) | |
return r | |
def cast_weight(s, input=None, dtype=None, device=None): | |
if input is not None: | |
if dtype is None: | |
dtype = input.dtype | |
if device is None: | |
device = input.device | |
weight = cast_to(s.weight, dtype, device) | |
return weight | |
def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None): | |
if input is not None: | |
if dtype is None: | |
dtype = input.dtype | |
if bias_dtype is None: | |
bias_dtype = dtype | |
if device is None: | |
device = input.device | |
weight = cast_to(s.weight, dtype, device) | |
bias = cast_to(s.bias, bias_dtype, device) if s.bias is not None else None | |
return weight, bias | |
class quantized_layer: | |
class Linear(torch.nn.Linear): | |
def __init__(self, *args, dtype=torch.bfloat16, device="cuda", **kwargs): | |
super().__init__(*args, **kwargs) | |
self.dtype = dtype | |
self.device = device | |
def block_forward_(self, x, i, j, dtype, device): | |
weight_ = cast_to( | |
self.weight[j * self.block_size: (j + 1) * self.block_size, i * self.block_size: (i + 1) * self.block_size], | |
dtype=dtype, device=device | |
) | |
if self.bias is None or i > 0: | |
bias_ = None | |
else: | |
bias_ = cast_to(self.bias[j * self.block_size: (j + 1) * self.block_size], dtype=dtype, device=device) | |
x_ = x[..., i * self.block_size: (i + 1) * self.block_size] | |
y_ = torch.nn.functional.linear(x_, weight_, bias_) | |
del x_, weight_, bias_ | |
torch.cuda.empty_cache() | |
return y_ | |
def block_forward(self, x, **kwargs): | |
# This feature can only reduce 2GB VRAM, so we disable it. | |
y = torch.zeros(x.shape[:-1] + (self.out_features,), dtype=x.dtype, device=x.device) | |
for i in range((self.in_features + self.block_size - 1) // self.block_size): | |
for j in range((self.out_features + self.block_size - 1) // self.block_size): | |
y[..., j * self.block_size: (j + 1) * self.block_size] += self.block_forward_(x, i, j, dtype=x.dtype, device=x.device) | |
return y | |
def forward(self, x, **kwargs): | |
weight, bias = cast_bias_weight(self, x, dtype=self.dtype, device=self.device) | |
return torch.nn.functional.linear(x, weight, bias) | |
class RMSNorm(torch.nn.Module): | |
def __init__(self, module, dtype=torch.bfloat16, device="cuda"): | |
super().__init__() | |
self.module = module | |
self.dtype = dtype | |
self.device = device | |
def forward(self, hidden_states, **kwargs): | |
input_dtype = hidden_states.dtype | |
variance = hidden_states.to(torch.float32).square().mean(-1, keepdim=True) | |
hidden_states = hidden_states * torch.rsqrt(variance + self.module.eps) | |
hidden_states = hidden_states.to(input_dtype) | |
if self.module.weight is not None: | |
weight = cast_weight(self.module, hidden_states, dtype=torch.bfloat16, device="cuda") | |
hidden_states = hidden_states * weight | |
return hidden_states | |
class Conv3d(torch.nn.Conv3d): | |
def __init__(self, *args, dtype=torch.bfloat16, device="cuda", **kwargs): | |
super().__init__(*args, **kwargs) | |
self.dtype = dtype | |
self.device = device | |
def forward(self, x): | |
weight, bias = cast_bias_weight(self, x, dtype=self.dtype, device=self.device) | |
return torch.nn.functional.conv3d(x, weight, bias, self.stride, self.padding, self.dilation, self.groups) | |
class LayerNorm(torch.nn.LayerNorm): | |
def __init__(self, *args, dtype=torch.bfloat16, device="cuda", **kwargs): | |
super().__init__(*args, **kwargs) | |
self.dtype = dtype | |
self.device = device | |
def forward(self, x): | |
if self.weight is not None and self.bias is not None: | |
weight, bias = cast_bias_weight(self, x, dtype=self.dtype, device=self.device) | |
return torch.nn.functional.layer_norm(x, self.normalized_shape, weight, bias, self.eps) | |
else: | |
return torch.nn.functional.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps) | |
def replace_layer(model, dtype=torch.bfloat16, device="cuda"): | |
for name, module in model.named_children(): | |
if isinstance(module, torch.nn.Linear): | |
with init_weights_on_device(): | |
new_layer = quantized_layer.Linear( | |
module.in_features, module.out_features, bias=module.bias is not None, | |
dtype=dtype, device=device | |
) | |
new_layer.load_state_dict(module.state_dict(), assign=True) | |
setattr(model, name, new_layer) | |
elif isinstance(module, torch.nn.Conv3d): | |
with init_weights_on_device(): | |
new_layer = quantized_layer.Conv3d( | |
module.in_channels, module.out_channels, kernel_size=module.kernel_size, stride=module.stride, | |
dtype=dtype, device=device | |
) | |
new_layer.load_state_dict(module.state_dict(), assign=True) | |
setattr(model, name, new_layer) | |
elif isinstance(module, RMSNorm): | |
new_layer = quantized_layer.RMSNorm( | |
module, | |
dtype=dtype, device=device | |
) | |
setattr(model, name, new_layer) | |
elif isinstance(module, torch.nn.LayerNorm): | |
with init_weights_on_device(): | |
new_layer = quantized_layer.LayerNorm( | |
module.normalized_shape, elementwise_affine=module.elementwise_affine, eps=module.eps, | |
dtype=dtype, device=device | |
) | |
new_layer.load_state_dict(module.state_dict(), assign=True) | |
setattr(model, name, new_layer) | |
else: | |
replace_layer(module, dtype=dtype, device=device) | |
replace_layer(self, dtype=dtype, device=device) | |
def state_dict_converter(): | |
return HunyuanVideoDiTStateDictConverter() | |
class HunyuanVideoDiTStateDictConverter: | |
def __init__(self): | |
pass | |
def from_civitai(self, state_dict): | |
origin_hash_key = hash_state_dict_keys(state_dict, with_shape=True) | |
if "module" in state_dict: | |
state_dict = state_dict["module"] | |
direct_dict = { | |
"img_in.proj": "img_in.proj", | |
"time_in.mlp.0": "time_in.timestep_embedder.0", | |
"time_in.mlp.2": "time_in.timestep_embedder.2", | |
"vector_in.in_layer": "vector_in.0", | |
"vector_in.out_layer": "vector_in.2", | |
"guidance_in.mlp.0": "guidance_in.timestep_embedder.0", | |
"guidance_in.mlp.2": "guidance_in.timestep_embedder.2", | |
"txt_in.input_embedder": "txt_in.input_embedder", | |
"txt_in.t_embedder.mlp.0": "txt_in.t_embedder.timestep_embedder.0", | |
"txt_in.t_embedder.mlp.2": "txt_in.t_embedder.timestep_embedder.2", | |
"txt_in.c_embedder.linear_1": "txt_in.c_embedder.0", | |
"txt_in.c_embedder.linear_2": "txt_in.c_embedder.2", | |
"final_layer.linear": "final_layer.linear", | |
"final_layer.adaLN_modulation.1": "final_layer.adaLN_modulation.1", | |
} | |
txt_suffix_dict = { | |
"norm1": "norm1", | |
"self_attn_qkv": "self_attn_qkv", | |
"self_attn_proj": "self_attn_proj", | |
"norm2": "norm2", | |
"mlp.fc1": "mlp.0", | |
"mlp.fc2": "mlp.2", | |
"adaLN_modulation.1": "adaLN_modulation.1", | |
} | |
double_suffix_dict = { | |
"img_mod.linear": "component_a.mod.linear", | |
"img_attn_qkv": "component_a.to_qkv", | |
"img_attn_q_norm": "component_a.norm_q", | |
"img_attn_k_norm": "component_a.norm_k", | |
"img_attn_proj": "component_a.to_out", | |
"img_mlp.fc1": "component_a.ff.0", | |
"img_mlp.fc2": "component_a.ff.2", | |
"txt_mod.linear": "component_b.mod.linear", | |
"txt_attn_qkv": "component_b.to_qkv", | |
"txt_attn_q_norm": "component_b.norm_q", | |
"txt_attn_k_norm": "component_b.norm_k", | |
"txt_attn_proj": "component_b.to_out", | |
"txt_mlp.fc1": "component_b.ff.0", | |
"txt_mlp.fc2": "component_b.ff.2", | |
} | |
single_suffix_dict = { | |
"linear1": ["to_qkv", "ff.0"], | |
"linear2": ["to_out", "ff.2"], | |
"q_norm": "norm_q", | |
"k_norm": "norm_k", | |
"modulation.linear": "mod.linear", | |
} | |
# single_suffix_dict = { | |
# "linear1": "linear1", | |
# "linear2": "linear2", | |
# "q_norm": "q_norm", | |
# "k_norm": "k_norm", | |
# "modulation.linear": "modulation.linear", | |
# } | |
state_dict_ = {} | |
for name, param in state_dict.items(): | |
names = name.split(".") | |
direct_name = ".".join(names[:-1]) | |
if direct_name in direct_dict: | |
name_ = direct_dict[direct_name] + "." + names[-1] | |
state_dict_[name_] = param | |
elif names[0] == "double_blocks": | |
prefix = ".".join(names[:2]) | |
suffix = ".".join(names[2:-1]) | |
name_ = prefix + "." + double_suffix_dict[suffix] + "." + names[-1] | |
state_dict_[name_] = param | |
elif names[0] == "single_blocks": | |
prefix = ".".join(names[:2]) | |
suffix = ".".join(names[2:-1]) | |
if isinstance(single_suffix_dict[suffix], list): | |
if suffix == "linear1": | |
name_a, name_b = single_suffix_dict[suffix] | |
param_a, param_b = torch.split(param, (3072*3, 3072*4), dim=0) | |
state_dict_[prefix + "." + name_a + "." + names[-1]] = param_a | |
state_dict_[prefix + "." + name_b + "." + names[-1]] = param_b | |
elif suffix == "linear2": | |
if names[-1] == "weight": | |
name_a, name_b = single_suffix_dict[suffix] | |
param_a, param_b = torch.split(param, (3072*1, 3072*4), dim=-1) | |
state_dict_[prefix + "." + name_a + "." + names[-1]] = param_a | |
state_dict_[prefix + "." + name_b + "." + names[-1]] = param_b | |
else: | |
name_a, name_b = single_suffix_dict[suffix] | |
state_dict_[prefix + "." + name_a + "." + names[-1]] = param | |
else: | |
pass | |
else: | |
name_ = prefix + "." + single_suffix_dict[suffix] + "." + names[-1] | |
state_dict_[name_] = param | |
elif names[0] == "txt_in": | |
prefix = ".".join(names[:4]).replace(".individual_token_refiner.", ".") | |
suffix = ".".join(names[4:-1]) | |
name_ = prefix + "." + txt_suffix_dict[suffix] + "." + names[-1] | |
state_dict_[name_] = param | |
else: | |
pass | |
return state_dict_ | |