943 lines
36 KiB
Python
943 lines
36 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
|
|
import math
|
|
from typing import Any
|
|
|
|
import numpy as np
|
|
import torch
|
|
import torch.nn as nn
|
|
from torchvision import transforms
|
|
|
|
from fastvideo.attention import DistributedAttention, LocalAttention
|
|
from fastvideo.configs.models.dits.cosmos2_5 import Cosmos25VideoConfig
|
|
from fastvideo.forward_context import get_forward_context
|
|
from fastvideo.layers.layernorm import RMSNorm
|
|
from fastvideo.layers.linear import ReplicatedLinear
|
|
from fastvideo.layers.mlp import MLP
|
|
from fastvideo.layers.rotary_embedding import apply_rotary_emb
|
|
from fastvideo.layers.visual_embedding import Timesteps
|
|
from fastvideo.models.dits.base import BaseDiT
|
|
from fastvideo.platforms import AttentionBackendEnum
|
|
|
|
|
|
class Cosmos25PatchEmbed(nn.Module):
|
|
"""
|
|
COSMOS 2.5 patch embedding - converts video (B, C, T, H, W) to patches (B, T', H', W', D).
|
|
Uses linear projection after rearranging patches.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
in_channels: int,
|
|
out_channels: int,
|
|
patch_size: tuple[int, int, int],
|
|
) -> None:
|
|
super().__init__()
|
|
self.patch_size = patch_size
|
|
self.dim = in_channels * patch_size[0] * patch_size[1] * patch_size[2]
|
|
|
|
self.proj = nn.Linear(self.dim, out_channels, bias=False)
|
|
|
|
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
|
"""
|
|
Args:
|
|
hidden_states: (B, C, T, H, W)
|
|
Returns:
|
|
(B, T', H', W', D) where T'=T//pt, H'=H//ph, W'=W//pw
|
|
"""
|
|
batch_size, num_channels, num_frames, height, width = hidden_states.shape
|
|
p_t, p_h, p_w = self.patch_size
|
|
|
|
# Rearrange: b c (t pt) (h ph) (w pw) -> b t h w (c pt ph pw)
|
|
hidden_states = hidden_states.reshape(batch_size, num_channels, num_frames // p_t, p_t, height // p_h, p_h,
|
|
width // p_w, p_w)
|
|
hidden_states = hidden_states.permute(0, 2, 4, 6, 1, 3, 5, 7)
|
|
hidden_states = hidden_states.flatten(4, 7) # Flatten patch dimensions
|
|
|
|
# Project to model dimension — cast to weight dtype to handle fp32/bf16 mismatch
|
|
# between training (fp32) and inference (bf16) contexts
|
|
hidden_states = self.proj(hidden_states.to(self.proj.weight.dtype))
|
|
return hidden_states
|
|
|
|
|
|
class Cosmos25TimestepEmbedding(nn.Module):
|
|
"""
|
|
COSMOS 2.5 timestep embedding with AdaLN-LoRA support.
|
|
Generates both standard embedding and AdaLN-LoRA parameters.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
in_features: int,
|
|
out_features: int,
|
|
use_adaln_lora: bool = True,
|
|
adaln_lora_dim: int = 256,
|
|
) -> None:
|
|
super().__init__()
|
|
self.use_adaln_lora = use_adaln_lora
|
|
|
|
self.linear_1 = nn.Linear(in_features, out_features, bias=False)
|
|
self.activation = nn.SiLU()
|
|
|
|
if use_adaln_lora:
|
|
self.linear_2 = nn.Linear(out_features, 3 * out_features, bias=False)
|
|
else:
|
|
self.linear_2 = nn.Linear(out_features, out_features, bias=False)
|
|
|
|
def forward(self, sample: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor | None]:
|
|
"""
|
|
Returns:
|
|
emb: Standard embedding (B, T, D)
|
|
adaln_lora: AdaLN-LoRA parameters (B, T, 3D) or None
|
|
"""
|
|
emb = self.linear_1(sample)
|
|
emb = self.activation(emb)
|
|
emb = self.linear_2(emb)
|
|
|
|
if self.use_adaln_lora:
|
|
adaln_lora = emb # (B, T, 3D)
|
|
emb_standard = sample # Use input as standard embedding
|
|
else:
|
|
emb_standard = emb
|
|
adaln_lora = None
|
|
|
|
return emb_standard, adaln_lora
|
|
|
|
|
|
class Cosmos25Embedding(nn.Module):
|
|
"""
|
|
COSMOS 2.5 timestep conditioning embedding.
|
|
Generates sinusoidal embeddings and processes them through MLP.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
embedding_dim: int,
|
|
condition_dim: int,
|
|
use_adaln_lora: bool = True,
|
|
adaln_lora_dim: int = 256,
|
|
) -> None:
|
|
super().__init__()
|
|
|
|
self.time_proj = Timesteps(embedding_dim, flip_sin_to_cos=True, downscale_freq_shift=0.0)
|
|
self.t_embedder = Cosmos25TimestepEmbedding(
|
|
embedding_dim,
|
|
condition_dim,
|
|
use_adaln_lora=use_adaln_lora,
|
|
adaln_lora_dim=adaln_lora_dim,
|
|
)
|
|
self.norm = RMSNorm(embedding_dim, eps=1e-6)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
timestep: torch.Tensor,
|
|
) -> tuple[torch.Tensor, torch.Tensor | None]:
|
|
"""
|
|
Args:
|
|
timestep: (B, T) tensor of timesteps
|
|
|
|
Returns:
|
|
embedded_timestep: Normalized timestep embedding (B, T, D)
|
|
adaln_lora: AdaLN-LoRA parameters (B, T, 3D) or None
|
|
"""
|
|
# Handle 2D timestep input (B, T) like the official model
|
|
assert timestep.ndim == 2, f"Expected 2D timestep, got {timestep.ndim}D with shape {timestep.shape}"
|
|
B, T = timestep.shape
|
|
|
|
# Flatten for Timesteps layer which expects 1D, then reshape back
|
|
timestep_flat = timestep.flatten() # (B*T,)
|
|
timesteps_proj = self.time_proj(timestep_flat).type_as(hidden_states) # (B*T, D)
|
|
timesteps_proj = timesteps_proj.reshape(B, T, -1) # (B, T, D)
|
|
|
|
embedded_timestep, adaln_lora = self.t_embedder(timesteps_proj)
|
|
embedded_timestep = self.norm(embedded_timestep)
|
|
|
|
return embedded_timestep, adaln_lora
|
|
|
|
|
|
class Cosmos25AdaLayerNormZero(nn.Module):
|
|
"""
|
|
COSMOS 2.5 Adaptive Layer Normalization with zero initialization and gate.
|
|
This is a simplified version that expects pre-computed shift/scale/gate parameters.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
in_features: int,
|
|
) -> None:
|
|
super().__init__()
|
|
self.norm = nn.LayerNorm(in_features, elementwise_affine=False, eps=1e-6)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
shift: torch.Tensor,
|
|
scale: torch.Tensor,
|
|
) -> torch.Tensor:
|
|
"""
|
|
Args:
|
|
hidden_states: Input tensor
|
|
shift: Shift parameter for modulation
|
|
scale: Scale parameter for modulation
|
|
|
|
Returns:
|
|
normalized_hidden_states: Modulated normalized hidden states
|
|
"""
|
|
# Apply layer norm and modulation
|
|
hidden_states = self.norm(hidden_states)
|
|
hidden_states = hidden_states * (1 + scale) + shift
|
|
return hidden_states
|
|
|
|
|
|
class Cosmos25SelfAttention(nn.Module):
|
|
"""
|
|
COSMOS 2.5 self-attention with QK normalization and RoPE.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
dim: int,
|
|
num_heads: int,
|
|
qk_norm: bool = True,
|
|
eps: float = 1e-6,
|
|
supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None,
|
|
) -> None:
|
|
assert dim % num_heads == 0
|
|
super().__init__()
|
|
self.dim = dim
|
|
self.num_heads = num_heads
|
|
self.head_dim = dim // num_heads
|
|
|
|
self.to_q = ReplicatedLinear(dim, dim, bias=False)
|
|
self.to_k = ReplicatedLinear(dim, dim, bias=False)
|
|
self.to_v = ReplicatedLinear(dim, dim, bias=False)
|
|
self.to_out = ReplicatedLinear(dim, dim, bias=False)
|
|
|
|
self.norm_q = RMSNorm(self.head_dim, eps=eps) if qk_norm else nn.Identity()
|
|
self.norm_k = RMSNorm(self.head_dim, eps=eps) if qk_norm else nn.Identity()
|
|
|
|
# Use DistributedAttention for flexible backend support (torch SDPA / FlashAttention)
|
|
# For single-GPU (non-distributed), use LocalAttention to avoid distributed requirements
|
|
if supported_attention_backends is None:
|
|
supported_attention_backends = (AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA)
|
|
|
|
# Always use DistributedAttention (requires distributed environment to be initialized)
|
|
self.attn = DistributedAttention(num_heads=num_heads,
|
|
head_size=self.head_dim,
|
|
causal=False,
|
|
supported_attention_backends=supported_attention_backends,
|
|
prefix="self_attn")
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
rope_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
|
|
) -> torch.Tensor:
|
|
"""
|
|
Args:
|
|
hidden_states: (B, S, D) where S = T*H*W
|
|
rope_emb: Tuple of (cos, sin) for RoPE
|
|
"""
|
|
# Get QKV
|
|
query, _ = self.to_q(hidden_states)
|
|
key, _ = self.to_k(hidden_states)
|
|
value, _ = self.to_v(hidden_states)
|
|
|
|
# Reshape for multi-head attention: (B, S, D) -> (B, S, H, D_h) -> (B, H, S, D_h)
|
|
query = query.unflatten(-1, (self.num_heads, self.head_dim)).transpose(1, 2)
|
|
key = key.unflatten(-1, (self.num_heads, self.head_dim)).transpose(1, 2)
|
|
value = value.unflatten(-1, (self.num_heads, self.head_dim)).transpose(1, 2)
|
|
|
|
# Apply QK normalization
|
|
query = self.norm_q(query)
|
|
key = self.norm_k(key)
|
|
|
|
# Apply RoPE if provided (query/key are now in (B, H, S, D_h) format)
|
|
if rope_emb is not None:
|
|
cos, sin = rope_emb
|
|
query = apply_rotary_emb(query, (cos, sin), use_real=True, use_real_unbind_dim=-2)
|
|
key = apply_rotary_emb(key, (cos, sin), use_real=True, use_real_unbind_dim=-2)
|
|
|
|
# Attention computation using DistributedAttention or LocalAttention
|
|
# Both expect (B, S, H, D_h), so transpose first
|
|
query = query.transpose(1, 2) # (B, H, S, D_h) -> (B, S, H, D_h)
|
|
key = key.transpose(1, 2)
|
|
value = value.transpose(1, 2)
|
|
|
|
attn_output, _ = self.attn(query, key, value)
|
|
# Reshape back: (B, S, H, D_h) -> (B, S, H*D_h)
|
|
attn_output = attn_output.flatten(-2, -1)
|
|
|
|
# Output projection
|
|
attn_output, _ = self.to_out(attn_output)
|
|
return attn_output
|
|
|
|
|
|
class Cosmos25CrossAttention(nn.Module):
|
|
"""
|
|
COSMOS 2.5 cross-attention for text conditioning.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
dim: int,
|
|
cross_attention_dim: int,
|
|
num_heads: int,
|
|
qk_norm: bool = True,
|
|
eps: float = 1e-6,
|
|
supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None,
|
|
) -> None:
|
|
assert dim % num_heads == 0
|
|
super().__init__()
|
|
self.dim = dim
|
|
self.cross_attention_dim = cross_attention_dim
|
|
self.num_heads = num_heads
|
|
self.head_dim = dim // num_heads
|
|
|
|
self.to_q = ReplicatedLinear(dim, dim, bias=False)
|
|
self.to_k = ReplicatedLinear(cross_attention_dim, dim, bias=False)
|
|
self.to_v = ReplicatedLinear(cross_attention_dim, dim, bias=False)
|
|
self.to_out = ReplicatedLinear(dim, dim, bias=False)
|
|
|
|
self.norm_q = RMSNorm(self.head_dim, eps=eps) if qk_norm else nn.Identity()
|
|
self.norm_k = RMSNorm(self.head_dim, eps=eps) if qk_norm else nn.Identity()
|
|
|
|
if supported_attention_backends is None:
|
|
supported_attention_backends = (AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA)
|
|
|
|
# Use LocalAttention for cross-attention since text embeddings are not sharded
|
|
# in sequence parallelism (replicated across ranks)
|
|
self.attn = LocalAttention(
|
|
num_heads=num_heads,
|
|
head_size=self.head_dim,
|
|
causal=False,
|
|
supported_attention_backends=supported_attention_backends,
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
encoder_hidden_states: torch.Tensor,
|
|
attention_mask: torch.Tensor | None = None,
|
|
) -> torch.Tensor:
|
|
"""
|
|
Args:
|
|
hidden_states: (B, S, D)
|
|
encoder_hidden_states: (B, N, D_text)
|
|
"""
|
|
# Get QKV
|
|
query, _ = self.to_q(hidden_states)
|
|
key, _ = self.to_k(encoder_hidden_states)
|
|
value, _ = self.to_v(encoder_hidden_states)
|
|
|
|
# Reshape for multi-head attention
|
|
query = query.unflatten(-1, (self.num_heads, self.head_dim))
|
|
key = key.unflatten(-1, (self.num_heads, self.head_dim))
|
|
value = value.unflatten(-1, (self.num_heads, self.head_dim))
|
|
|
|
# Apply QK normalization
|
|
query = self.norm_q(query)
|
|
key = self.norm_k(key)
|
|
|
|
# LocalAttention expects (B, S, H, D_h), which is what we already have
|
|
attn_output = self.attn(query, key, value)
|
|
|
|
# Reshape back: (B, S, H, D_h) -> (B, S, H*D_h)
|
|
attn_output = attn_output.flatten(-2, -1)
|
|
|
|
# Output projection
|
|
attn_output, _ = self.to_out(attn_output)
|
|
return attn_output
|
|
|
|
|
|
class Cosmos25TransformerBlock(nn.Module):
|
|
"""
|
|
COSMOS 2.5 transformer block with self-attention, cross-attention, and MLP.
|
|
Uses AdaLN-LoRA for conditioning.
|
|
Matches the official architecture where modulation parameters are computed once per block.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
num_attention_heads: int,
|
|
attention_head_dim: int,
|
|
cross_attention_dim: int,
|
|
mlp_ratio: float = 4.0,
|
|
adaln_lora_dim: int = 256,
|
|
use_adaln_lora: bool = True,
|
|
qk_norm: bool = True,
|
|
supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None,
|
|
) -> None:
|
|
super().__init__()
|
|
|
|
hidden_size = num_attention_heads * attention_head_dim
|
|
self.use_adaln_lora = use_adaln_lora
|
|
|
|
# Layer norms (no modulation logic inside)
|
|
self.norm1 = Cosmos25AdaLayerNormZero(hidden_size)
|
|
self.norm2 = Cosmos25AdaLayerNormZero(hidden_size)
|
|
self.norm3 = Cosmos25AdaLayerNormZero(hidden_size)
|
|
|
|
# Attention and MLP layers
|
|
self.attn1 = Cosmos25SelfAttention(
|
|
dim=hidden_size,
|
|
num_heads=num_attention_heads,
|
|
qk_norm=qk_norm,
|
|
supported_attention_backends=supported_attention_backends,
|
|
)
|
|
self.attn2 = Cosmos25CrossAttention(
|
|
dim=hidden_size,
|
|
cross_attention_dim=cross_attention_dim,
|
|
num_heads=num_attention_heads,
|
|
qk_norm=qk_norm,
|
|
supported_attention_backends=supported_attention_backends,
|
|
)
|
|
self.mlp = MLP(hidden_size, int(hidden_size * mlp_ratio), act_type="gelu", bias=False)
|
|
|
|
# AdaLN modulation layers (compute shift/scale/gate for each sub-layer)
|
|
# These match the official model's adaln_modulation_* layers
|
|
if use_adaln_lora:
|
|
self.adaln_modulation_self_attn = nn.Sequential(
|
|
nn.SiLU(),
|
|
nn.Linear(hidden_size, adaln_lora_dim, bias=False),
|
|
nn.Linear(adaln_lora_dim, 3 * hidden_size, bias=False),
|
|
)
|
|
self.adaln_modulation_cross_attn = nn.Sequential(
|
|
nn.SiLU(),
|
|
nn.Linear(hidden_size, adaln_lora_dim, bias=False),
|
|
nn.Linear(adaln_lora_dim, 3 * hidden_size, bias=False),
|
|
)
|
|
self.adaln_modulation_mlp = nn.Sequential(
|
|
nn.SiLU(),
|
|
nn.Linear(hidden_size, adaln_lora_dim, bias=False),
|
|
nn.Linear(adaln_lora_dim, 3 * hidden_size, bias=False),
|
|
)
|
|
else:
|
|
self.adaln_modulation_self_attn = nn.Sequential(nn.SiLU(),
|
|
nn.Linear(hidden_size, 3 * hidden_size, bias=False))
|
|
self.adaln_modulation_cross_attn = nn.Sequential(nn.SiLU(),
|
|
nn.Linear(hidden_size, 3 * hidden_size, bias=False))
|
|
self.adaln_modulation_mlp = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 3 * hidden_size, bias=False))
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
encoder_hidden_states: torch.Tensor,
|
|
embedded_timestep: torch.Tensor,
|
|
adaln_lora: torch.Tensor | None = None,
|
|
rope_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
|
|
extra_pos_emb: torch.Tensor | None = None,
|
|
attention_mask: torch.Tensor | None = None,
|
|
) -> torch.Tensor:
|
|
"""
|
|
Args:
|
|
hidden_states: (B, T, H, W, D)
|
|
encoder_hidden_states: (B, N, D_text)
|
|
embedded_timestep: (B, T, D)
|
|
adaln_lora: (B, T, 3D) AdaLN-LoRA parameters
|
|
rope_emb: Tuple of (cos, sin) for RoPE
|
|
extra_pos_emb: Optional learnable positional embeddings
|
|
"""
|
|
# Add extra positional embeddings if provided
|
|
if extra_pos_emb is not None:
|
|
hidden_states = hidden_states + extra_pos_emb
|
|
|
|
B, T, H, W, D = hidden_states.shape
|
|
|
|
# Step 1: Compute ALL modulation parameters once (matches official model)
|
|
if self.use_adaln_lora and adaln_lora is not None:
|
|
shift_self_attn, scale_self_attn, gate_self_attn = (self.adaln_modulation_self_attn(embedded_timestep) +
|
|
adaln_lora).chunk(3, dim=-1)
|
|
shift_cross_attn, scale_cross_attn, gate_cross_attn = (self.adaln_modulation_cross_attn(embedded_timestep) +
|
|
adaln_lora).chunk(3, dim=-1)
|
|
shift_mlp, scale_mlp, gate_mlp = (self.adaln_modulation_mlp(embedded_timestep) + adaln_lora).chunk(3,
|
|
dim=-1)
|
|
else:
|
|
shift_self_attn, scale_self_attn, gate_self_attn = self.adaln_modulation_self_attn(embedded_timestep).chunk(
|
|
3, dim=-1)
|
|
shift_cross_attn, scale_cross_attn, gate_cross_attn = self.adaln_modulation_cross_attn(
|
|
embedded_timestep).chunk(3, dim=-1)
|
|
shift_mlp, scale_mlp, gate_mlp = self.adaln_modulation_mlp(embedded_timestep).chunk(3, dim=-1)
|
|
|
|
# Reshape modulation parameters from (B, T, D) to (B, T, 1, 1, D) for broadcasting
|
|
shift_self_attn = shift_self_attn.unsqueeze(2).unsqueeze(2).type_as(hidden_states)
|
|
scale_self_attn = scale_self_attn.unsqueeze(2).unsqueeze(2).type_as(hidden_states)
|
|
gate_self_attn = gate_self_attn.unsqueeze(2).unsqueeze(2).type_as(hidden_states)
|
|
|
|
shift_cross_attn = shift_cross_attn.unsqueeze(2).unsqueeze(2).type_as(hidden_states)
|
|
scale_cross_attn = scale_cross_attn.unsqueeze(2).unsqueeze(2).type_as(hidden_states)
|
|
gate_cross_attn = gate_cross_attn.unsqueeze(2).unsqueeze(2).type_as(hidden_states)
|
|
|
|
shift_mlp = shift_mlp.unsqueeze(2).unsqueeze(2).type_as(hidden_states)
|
|
scale_mlp = scale_mlp.unsqueeze(2).unsqueeze(2).type_as(hidden_states)
|
|
gate_mlp = gate_mlp.unsqueeze(2).unsqueeze(2).type_as(hidden_states)
|
|
|
|
# Step 2: Self-attention block
|
|
norm_hidden_states = self.norm1(hidden_states, shift_self_attn, scale_self_attn)
|
|
# Flatten for attention: (B, T, H, W, D) -> (B, THW, D)
|
|
norm_hidden_states_flat = norm_hidden_states.flatten(1, 3)
|
|
|
|
attn_output = self.attn1(norm_hidden_states_flat, rope_emb=rope_emb)
|
|
|
|
# Reshape back and apply residual
|
|
attn_output = attn_output.unflatten(1, (T, H, W)) # (B, T, H, W, D)
|
|
hidden_states = hidden_states + gate_self_attn * attn_output
|
|
|
|
# Step 3: Cross-attention block
|
|
norm_hidden_states = self.norm2(hidden_states, shift_cross_attn, scale_cross_attn)
|
|
norm_hidden_states_flat = norm_hidden_states.flatten(1, 3)
|
|
|
|
attn_output = self.attn2(
|
|
norm_hidden_states_flat,
|
|
encoder_hidden_states=encoder_hidden_states,
|
|
attention_mask=attention_mask,
|
|
)
|
|
|
|
attn_output = attn_output.unflatten(1, (T, H, W))
|
|
hidden_states = hidden_states + gate_cross_attn * attn_output
|
|
|
|
# Step 4: MLP block
|
|
norm_hidden_states = self.norm3(hidden_states, shift_mlp, scale_mlp)
|
|
norm_hidden_states_flat = norm_hidden_states.flatten(1, 3)
|
|
|
|
mlp_output = self.mlp(norm_hidden_states_flat)
|
|
|
|
mlp_output = mlp_output.unflatten(1, (T, H, W))
|
|
hidden_states = hidden_states + gate_mlp * mlp_output
|
|
|
|
return hidden_states
|
|
|
|
|
|
class Cosmos25RotaryPosEmbed(nn.Module):
|
|
"""
|
|
COSMOS 2.5 3D Rotary Position Embedding with NTK-aware extrapolation.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
max_size: tuple[int, int, int] = (128, 240, 240),
|
|
patch_size: tuple[int, int, int] = (1, 2, 2),
|
|
base_fps: int = 24,
|
|
rope_scale: tuple[float, float, float] = (1.0, 1.0, 1.0),
|
|
enable_fps_modulation: bool = True,
|
|
) -> None:
|
|
super().__init__()
|
|
|
|
self.max_size = [size // patch for size, patch in zip(max_size, patch_size, strict=True)]
|
|
self.patch_size = patch_size
|
|
self.base_fps = base_fps
|
|
self.enable_fps_modulation = enable_fps_modulation
|
|
|
|
# Split dimensions: 1/3 for T, 1/3 for H, 1/3 for W
|
|
self.dim_h = hidden_size // 6 * 2
|
|
self.dim_w = hidden_size // 6 * 2
|
|
self.dim_t = hidden_size - self.dim_h - self.dim_w
|
|
|
|
# NTK-aware extrapolation factors
|
|
self.h_ntk_factor = rope_scale[1]**(self.dim_h / (self.dim_h - 2))
|
|
self.w_ntk_factor = rope_scale[2]**(self.dim_w / (self.dim_w - 2))
|
|
self.t_ntk_factor = rope_scale[0]**(self.dim_t / (self.dim_t - 2))
|
|
|
|
def forward(self, hidden_states: torch.Tensor, fps: int | None = None) -> tuple[torch.Tensor, torch.Tensor]:
|
|
"""
|
|
Generate 3D RoPE embeddings.
|
|
|
|
Args:
|
|
hidden_states: (B, T, H, W, D) - patch-embedded features
|
|
fps: Frames per second for temporal scaling
|
|
|
|
Returns:
|
|
cos, sin: RoPE embeddings (THW, D)
|
|
"""
|
|
batch_size, T, H, W, input_dim = hidden_states.shape
|
|
device = hidden_states.device
|
|
|
|
# T, H, W are already patch dimensions after patch_embed
|
|
# No need to divide by patch_size
|
|
|
|
# Generate frequency scales with NTK
|
|
h_theta = 10000.0 * self.h_ntk_factor
|
|
w_theta = 10000.0 * self.w_ntk_factor
|
|
t_theta = 10000.0 * self.t_ntk_factor
|
|
|
|
seq = torch.arange(max(self.max_size), device=device, dtype=torch.float32)
|
|
|
|
# Use self.dim_h/w/t which were set during initialization
|
|
dim_h_range = torch.arange(0, self.dim_h, 2, device=device,
|
|
dtype=torch.float32)[:(self.dim_h // 2)] / self.dim_h
|
|
dim_w_range = torch.arange(0, self.dim_w, 2, device=device,
|
|
dtype=torch.float32)[:(self.dim_w // 2)] / self.dim_w
|
|
dim_t_range = torch.arange(0, self.dim_t, 2, device=device,
|
|
dtype=torch.float32)[:(self.dim_t // 2)] / self.dim_t
|
|
|
|
h_spatial_freqs = 1.0 / (h_theta**dim_h_range)
|
|
w_spatial_freqs = 1.0 / (w_theta**dim_w_range)
|
|
temporal_freqs = 1.0 / (t_theta**dim_t_range)
|
|
|
|
# Generate positional embeddings
|
|
half_emb_h = torch.outer(seq[:H], h_spatial_freqs)
|
|
half_emb_w = torch.outer(seq[:W], w_spatial_freqs)
|
|
|
|
if self.enable_fps_modulation and fps is not None:
|
|
# Apply FPS scaling
|
|
half_emb_t = torch.outer(seq[:T] / fps * self.base_fps, temporal_freqs)
|
|
else:
|
|
half_emb_t = torch.outer(seq[:T], temporal_freqs)
|
|
|
|
# Broadcast and concatenate embeddings
|
|
emb_t = half_emb_t[:, None, None, :].repeat(1, H, W, 1)
|
|
emb_h = half_emb_h[None, :, None, :].repeat(T, 1, W, 1)
|
|
emb_w = half_emb_w[None, None, :, :].repeat(T, H, 1, 1)
|
|
|
|
# Concatenate [t, h, w, t, h, w] for sin/cos pairs
|
|
freqs = torch.cat([emb_t, emb_h, emb_w] * 2, dim=-1)
|
|
freqs = freqs.flatten(0, 2).float() # (THW, D)
|
|
|
|
cos = torch.cos(freqs) # (THW, D)
|
|
sin = torch.sin(freqs) # (THW, D)
|
|
|
|
return cos, sin
|
|
|
|
|
|
class Cosmos25LearnablePositionalEmbed(nn.Module):
|
|
"""
|
|
COSMOS 2.5 learnable absolute positional embeddings (optional).
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
max_size: tuple[int, int, int],
|
|
patch_size: tuple[int, int, int],
|
|
eps: float = 1e-6,
|
|
) -> None:
|
|
super().__init__()
|
|
|
|
self.max_size = [size // patch for size, patch in zip(max_size, patch_size, strict=True)]
|
|
self.patch_size = patch_size
|
|
self.eps = eps
|
|
|
|
self.pos_emb_t = nn.Parameter(torch.zeros(self.max_size[0], hidden_size))
|
|
self.pos_emb_h = nn.Parameter(torch.zeros(self.max_size[1], hidden_size))
|
|
self.pos_emb_w = nn.Parameter(torch.zeros(self.max_size[2], hidden_size))
|
|
|
|
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
|
"""
|
|
Args:
|
|
hidden_states: (B, T, H, W, D)
|
|
Returns:
|
|
pos_emb: (B, T, H, W, D)
|
|
"""
|
|
B, T, H, W, D = hidden_states.shape
|
|
|
|
emb_t = self.pos_emb_t[:T][None, :, None, None, :].repeat(B, 1, H, W, 1)
|
|
emb_h = self.pos_emb_h[:H][None, None, :, None, :].repeat(B, T, 1, W, 1)
|
|
emb_w = self.pos_emb_w[:W][None, None, None, :, :].repeat(B, T, H, 1, 1)
|
|
|
|
emb = emb_t + emb_h + emb_w
|
|
|
|
# Normalize
|
|
norm = torch.linalg.vector_norm(emb, dim=-1, keepdim=True, dtype=torch.float32)
|
|
norm = torch.add(self.eps, norm, alpha=np.sqrt(norm.numel() / emb.numel()))
|
|
return (emb / norm).type_as(hidden_states)
|
|
|
|
|
|
class Cosmos25FinalLayer(nn.Module):
|
|
"""
|
|
COSMOS 2.5 final layer with AdaLN modulation and unpatchification.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
out_channels: int,
|
|
patch_size: tuple[int, int, int],
|
|
adaln_lora_dim: int = 256,
|
|
use_adaln_lora: bool = True,
|
|
) -> None:
|
|
super().__init__()
|
|
self.hidden_size = hidden_size
|
|
self.use_adaln_lora = use_adaln_lora
|
|
|
|
self.norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
|
self.activation = nn.SiLU()
|
|
|
|
if use_adaln_lora:
|
|
self.linear_1 = nn.Linear(hidden_size, adaln_lora_dim, bias=False)
|
|
self.linear_2 = nn.Linear(adaln_lora_dim, 2 * hidden_size, bias=False)
|
|
else:
|
|
self.linear_1 = nn.Identity()
|
|
self.linear_2 = nn.Linear(hidden_size, 2 * hidden_size, bias=False)
|
|
|
|
# Output projection
|
|
output_dim = out_channels * patch_size[0] * patch_size[1] * patch_size[2]
|
|
self.proj_out = nn.Linear(hidden_size, output_dim, bias=False)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
embedded_timestep: torch.Tensor,
|
|
adaln_lora: torch.Tensor | None = None,
|
|
) -> torch.Tensor:
|
|
"""
|
|
Args:
|
|
hidden_states: (B, T, H, W, D)
|
|
embedded_timestep: (B, T, D) or (B, D)
|
|
adaln_lora: (B, T, 3D) or None
|
|
"""
|
|
# Generate modulation parameters
|
|
embedded_timestep = self.activation(embedded_timestep)
|
|
embedded_timestep = self.linear_1(embedded_timestep)
|
|
embedded_timestep = self.linear_2(embedded_timestep)
|
|
|
|
if self.use_adaln_lora and adaln_lora is not None:
|
|
# Use first 2*hidden_size elements for shift/scale
|
|
embedded_timestep = embedded_timestep + adaln_lora[..., :2 * self.hidden_size]
|
|
|
|
shift, scale = embedded_timestep.chunk(2, dim=-1)
|
|
|
|
# Apply normalization and modulation
|
|
hidden_states = self.norm(hidden_states)
|
|
|
|
# Reshape for broadcasting if needed
|
|
if embedded_timestep.ndim == 2:
|
|
shift, scale = (x.unsqueeze(1) for x in (shift, scale))
|
|
elif embedded_timestep.ndim == 3 and hidden_states.ndim == 5:
|
|
shift, scale = (x.unsqueeze(2).unsqueeze(2) for x in (shift, scale))
|
|
|
|
hidden_states = hidden_states * (1 + scale) + shift
|
|
|
|
# Project to output
|
|
hidden_states = self.proj_out(hidden_states)
|
|
|
|
return hidden_states
|
|
|
|
|
|
class Cosmos25Transformer3DModel(BaseDiT):
|
|
"""
|
|
COSMOS 2.5 DiT - MiniTrainDIT architecture adapted for FastVideo.
|
|
|
|
Key features:
|
|
- AdaLN-LoRA conditioning
|
|
- 3D RoPE with NTK-aware extrapolation
|
|
- Optional learnable positional embeddings
|
|
- QK normalization
|
|
- Cross-attention projection (optional)
|
|
"""
|
|
|
|
_fsdp_shard_conditions = Cosmos25VideoConfig()._fsdp_shard_conditions
|
|
_compile_conditions = Cosmos25VideoConfig()._compile_conditions
|
|
param_names_mapping = Cosmos25VideoConfig().param_names_mapping
|
|
lora_param_names_mapping = Cosmos25VideoConfig().lora_param_names_mapping
|
|
|
|
def __init__(self, config: Cosmos25VideoConfig, hf_config: dict[str, Any]) -> None:
|
|
super().__init__(config=config, hf_config=hf_config)
|
|
|
|
inner_dim = config.num_attention_heads * config.attention_head_dim
|
|
self.hidden_size = inner_dim
|
|
self.num_attention_heads = config.num_attention_heads
|
|
self.in_channels = config.in_channels
|
|
self.out_channels = config.out_channels
|
|
self.num_channels_latents = config.num_channels_latents
|
|
self.patch_size = config.patch_size
|
|
self.max_size = config.max_size
|
|
self.rope_scale = config.rope_scale
|
|
self.concat_padding_mask = config.concat_padding_mask
|
|
self.use_adaln_lora = getattr(config, "use_adaln_lora", True)
|
|
self.adaln_lora_dim = getattr(config, "adaln_lora_dim", 256)
|
|
self.extra_pos_embed_type = getattr(config, "extra_pos_embed_type", None)
|
|
self.use_crossattn_projection = getattr(config, "use_crossattn_projection", False)
|
|
|
|
# 1. Patch Embedding
|
|
# Account for: VAE channels + condition_mask (1) + padding_mask (1 if concat_padding_mask)
|
|
patch_embed_in_channels = config.in_channels # Base VAE channels (16)
|
|
patch_embed_in_channels += 1 # Always add 1 for condition_mask
|
|
if config.concat_padding_mask:
|
|
patch_embed_in_channels += 1 # Add 1 for padding_mask
|
|
# Total: 16 + 1 + 1 = 18 (with concat_padding_mask=True)
|
|
|
|
self.patch_embed = Cosmos25PatchEmbed(patch_embed_in_channels, inner_dim, config.patch_size)
|
|
|
|
# 2. Positional Embeddings
|
|
self.rope = Cosmos25RotaryPosEmbed(
|
|
hidden_size=config.attention_head_dim,
|
|
max_size=config.max_size,
|
|
patch_size=config.patch_size,
|
|
rope_scale=config.rope_scale,
|
|
enable_fps_modulation=getattr(config, "rope_enable_fps_modulation", True),
|
|
)
|
|
|
|
self.learnable_pos_embed = None
|
|
if self.extra_pos_embed_type == "learnable":
|
|
self.learnable_pos_embed = Cosmos25LearnablePositionalEmbed(
|
|
hidden_size=inner_dim,
|
|
max_size=config.max_size,
|
|
patch_size=config.patch_size,
|
|
)
|
|
|
|
# 3. Time Embedding
|
|
self.time_embed = Cosmos25Embedding(
|
|
inner_dim,
|
|
inner_dim,
|
|
use_adaln_lora=self.use_adaln_lora,
|
|
adaln_lora_dim=self.adaln_lora_dim,
|
|
)
|
|
|
|
# 4. Cross-attention projection (optional)
|
|
if self.use_crossattn_projection:
|
|
crossattn_proj_in_channels = getattr(config, "crossattn_proj_in_channels", config.text_embed_dim)
|
|
self.crossattn_proj = nn.Sequential(
|
|
nn.Linear(crossattn_proj_in_channels, config.text_embed_dim, bias=True),
|
|
nn.GELU(),
|
|
)
|
|
|
|
# 5. Transformer Blocks
|
|
self.transformer_blocks = nn.ModuleList([
|
|
Cosmos25TransformerBlock(
|
|
num_attention_heads=config.num_attention_heads,
|
|
attention_head_dim=config.attention_head_dim,
|
|
cross_attention_dim=config.text_embed_dim,
|
|
mlp_ratio=config.mlp_ratio,
|
|
adaln_lora_dim=self.adaln_lora_dim,
|
|
use_adaln_lora=self.use_adaln_lora,
|
|
qk_norm=(config.qk_norm == "rms_norm"),
|
|
supported_attention_backends=config._supported_attention_backends,
|
|
) for i in range(config.num_layers)
|
|
])
|
|
|
|
# 6. Final Layer
|
|
self.final_layer = Cosmos25FinalLayer(
|
|
hidden_size=inner_dim,
|
|
out_channels=config.out_channels,
|
|
patch_size=config.patch_size,
|
|
adaln_lora_dim=self.adaln_lora_dim,
|
|
use_adaln_lora=self.use_adaln_lora,
|
|
)
|
|
|
|
self.gradient_checkpointing = False
|
|
self.__post_init__()
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
timestep: torch.Tensor,
|
|
encoder_hidden_states: torch.Tensor | list[torch.Tensor],
|
|
attention_mask: torch.Tensor | None = None,
|
|
fps: int | None = None,
|
|
condition_mask: torch.Tensor | None = None,
|
|
padding_mask: torch.Tensor | None = None,
|
|
**kwargs,
|
|
) -> torch.Tensor:
|
|
"""
|
|
Args:
|
|
hidden_states: (B, C, T, H, W) latent video
|
|
timestep: (B,) or (B, T) diffusion timesteps
|
|
encoder_hidden_states: (B, N, D_text) text embeddings
|
|
attention_mask: Optional attention mask
|
|
fps: Frames per second
|
|
condition_mask: (B, 1, T, H, W) conditioning mask
|
|
padding_mask: (B, 1, H, W) padding mask
|
|
"""
|
|
if not isinstance(encoder_hidden_states, torch.Tensor):
|
|
encoder_hidden_states = encoder_hidden_states[0]
|
|
|
|
batch_size, num_channels, num_frames, height, width = hidden_states.shape
|
|
|
|
# 1. Concatenate condition mask if provided
|
|
if condition_mask is not None:
|
|
hidden_states = torch.cat([hidden_states, condition_mask], dim=1)
|
|
|
|
# 2. Concatenate padding mask if needed
|
|
if self.concat_padding_mask and padding_mask is not None:
|
|
padding_mask = transforms.functional.resize(
|
|
padding_mask,
|
|
list(hidden_states.shape[-2:]),
|
|
interpolation=transforms.InterpolationMode.NEAREST,
|
|
)
|
|
hidden_states = torch.cat(
|
|
[hidden_states, padding_mask.unsqueeze(2).repeat(1, 1, num_frames, 1, 1)],
|
|
dim=1,
|
|
)
|
|
|
|
# 3. Patchify input
|
|
p_t, p_h, p_w = self.patch_size
|
|
post_patch_num_frames = num_frames // p_t
|
|
post_patch_height = height // p_h
|
|
post_patch_width = width // p_w
|
|
|
|
hidden_states = self.patch_embed(hidden_states) # (B, T', H', W', D)
|
|
|
|
# 4. Generate RoPE embeddings (after patchify, using patch dimensions)
|
|
rope_emb = self.rope(hidden_states, fps=fps)
|
|
|
|
# 5. Generate learnable positional embeddings (if used)
|
|
extra_pos_emb = None
|
|
if self.learnable_pos_embed is not None:
|
|
extra_pos_emb = self.learnable_pos_embed(hidden_states)
|
|
|
|
# 6. Timestep embeddings
|
|
# Official model expects timestep in (B, T) format, so ensure it has 2D shape
|
|
if timestep.ndim == 1:
|
|
# Scalar timestep per sample: (B,) -> (B, 1)
|
|
timestep = timestep.unsqueeze(1)
|
|
elif timestep.ndim == 2:
|
|
# Already in (B, T) format
|
|
pass
|
|
else:
|
|
raise ValueError(f"Unsupported timestep shape: {timestep.shape}")
|
|
|
|
# Now timestep is always (B, T), pass directly to time_embed
|
|
embedded_timestep, adaln_lora = self.time_embed(hidden_states, timestep)
|
|
|
|
# 7. Apply cross-attention projection (if used)
|
|
if self.use_crossattn_projection:
|
|
encoder_hidden_states = self.crossattn_proj(encoder_hidden_states.to(self.crossattn_proj[0].weight.dtype))
|
|
|
|
# Prepare attention mask
|
|
if attention_mask is not None:
|
|
attention_mask = attention_mask.unsqueeze(1).unsqueeze(1) # (B, 1, 1, N)
|
|
|
|
# 8. Transformer blocks
|
|
if torch.is_grad_enabled() and self.gradient_checkpointing:
|
|
for block in self.transformer_blocks:
|
|
hidden_states = self._gradient_checkpointing_func(
|
|
block,
|
|
hidden_states,
|
|
encoder_hidden_states,
|
|
embedded_timestep,
|
|
adaln_lora,
|
|
rope_emb,
|
|
extra_pos_emb,
|
|
attention_mask,
|
|
)
|
|
else:
|
|
for i, block in enumerate(self.transformer_blocks):
|
|
hidden_states = block(
|
|
hidden_states=hidden_states,
|
|
encoder_hidden_states=encoder_hidden_states,
|
|
embedded_timestep=embedded_timestep,
|
|
adaln_lora=adaln_lora,
|
|
rope_emb=rope_emb,
|
|
extra_pos_emb=extra_pos_emb,
|
|
attention_mask=attention_mask,
|
|
)
|
|
|
|
# 9. Final layer - output norm & projection
|
|
hidden_states = self.final_layer(hidden_states, embedded_timestep, adaln_lora)
|
|
|
|
# 10. Unpatchify: (B, T', H', W', P) -> (B, C, T, H, W)
|
|
# After unflatten: (B, T', H', W', p_t, p_h, p_w, C) with dims [0,1,2,3,4,5,6,7]
|
|
hidden_states = hidden_states.unflatten(-1, (p_t, p_h, p_w, self.out_channels))
|
|
# Permute to: (B, C, T', p_t, H', p_h, W', p_w)
|
|
hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
|
|
# Flatten pairs to get (B, C, T, H, W)
|
|
hidden_states = hidden_states.flatten(2, 3).flatten(3, 4).flatten(4, 5)
|
|
|
|
return hidden_states
|
|
|
|
|
|
# Entry point for model registry
|
|
EntryClass = Cosmos25Transformer3DModel
|