mirror of
https://github.com/modelscope/DiffSynth-Studio.git
synced 2026-03-18 22:08:13 +00:00
418 lines
17 KiB
Python
418 lines
17 KiB
Python
import math
|
|
import torch
|
|
import torch.nn as nn
|
|
from typing import Optional, Tuple
|
|
|
|
|
|
|
|
def rotate_half(x):
|
|
"""Rotates half the hidden dims of the input."""
|
|
x1 = x[..., : x.shape[-1] // 2]
|
|
x2 = x[..., x.shape[-1] // 2 :]
|
|
return torch.cat((-x2, x1), dim=-1)
|
|
|
|
|
|
def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
|
|
mrope_section = mrope_section * 2
|
|
cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
|
|
unsqueeze_dim
|
|
)
|
|
sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
|
|
unsqueeze_dim
|
|
)
|
|
|
|
q_embed = (q * cos) + (rotate_half(q) * sin)
|
|
k_embed = (k * cos) + (rotate_half(k) * sin)
|
|
return q_embed, k_embed
|
|
|
|
|
|
class Qwen2_5_VLRotaryEmbedding(nn.Module):
|
|
def __init__(self, config, device=None):
|
|
super().__init__()
|
|
# BC: "rope_type" was originally "type"
|
|
if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
|
|
self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
|
|
else:
|
|
self.rope_type = "default"
|
|
self.max_seq_len_cached = config.max_position_embeddings
|
|
self.original_max_seq_len = config.max_position_embeddings
|
|
|
|
self.config = config
|
|
from transformers.modeling_rope_utils import _compute_default_rope_parameters
|
|
self.rope_init_fn = _compute_default_rope_parameters
|
|
|
|
inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
|
|
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
|
self.original_inv_freq = self.inv_freq
|
|
|
|
|
|
def _dynamic_frequency_update(self, position_ids, device):
|
|
"""
|
|
dynamic RoPE layers should recompute `inv_freq` in the following situations:
|
|
1 - growing beyond the cached sequence length (allow scaling)
|
|
2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
|
|
"""
|
|
seq_len = torch.max(position_ids) + 1
|
|
if seq_len > self.max_seq_len_cached: # growth
|
|
inv_freq, self.attention_scaling = self.rope_init_fn(
|
|
self.config, device, seq_len=seq_len, **self.rope_kwargs
|
|
)
|
|
self.register_buffer("inv_freq", inv_freq, persistent=False) # TODO joao: may break with compilation
|
|
self.max_seq_len_cached = seq_len
|
|
|
|
if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len: # reset
|
|
self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
|
|
self.max_seq_len_cached = self.original_max_seq_len
|
|
|
|
|
|
@torch.no_grad()
|
|
def forward(self, x, position_ids):
|
|
if "dynamic" in self.rope_type:
|
|
self._dynamic_frequency_update(position_ids, device=x.device)
|
|
|
|
# Core RoPE block. In contrast to other models, Qwen2_5_VL has different position ids for the grids
|
|
# So we expand the inv_freq to shape (3, ...)
|
|
inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
|
|
position_ids_expanded = position_ids[:, :, None, :].float() # shape (3, bs, 1, positions)
|
|
# Force float32 (see https://github.com/huggingface/transformers/pull/29285)
|
|
device_type = x.device.type
|
|
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
|
|
with torch.autocast(device_type=device_type, enabled=False):
|
|
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
|
|
emb = torch.cat((freqs, freqs), dim=-1)
|
|
cos = emb.cos()
|
|
sin = emb.sin()
|
|
|
|
# Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
|
|
cos = cos * self.attention_scaling
|
|
sin = sin * self.attention_scaling
|
|
|
|
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
|
|
|
|
|
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
|
"""
|
|
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
|
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
|
"""
|
|
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
|
if n_rep == 1:
|
|
return hidden_states
|
|
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
|
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
|
|
|
|
|
class Qwen2_5_VLAttention(nn.Module):
|
|
def __init__(self, config, layer_idx: Optional[int] = None):
|
|
super().__init__()
|
|
self.config = config
|
|
self.layer_idx = layer_idx
|
|
|
|
self.hidden_size = config.hidden_size
|
|
self.num_heads = config.num_attention_heads
|
|
self.head_dim = self.hidden_size // self.num_heads
|
|
self.num_key_value_heads = config.num_key_value_heads
|
|
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
|
self.is_causal = True
|
|
self.attention_dropout = config.attention_dropout
|
|
self.rope_scaling = config.rope_scaling
|
|
|
|
if (self.head_dim * self.num_heads) != self.hidden_size:
|
|
raise ValueError(
|
|
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
|
|
f" and `num_heads`: {self.num_heads})."
|
|
)
|
|
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
|
|
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
|
|
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
|
|
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
|
|
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC
|
|
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
|
bsz, q_len, _ = hidden_states.size()
|
|
|
|
query_states = self.q_proj(hidden_states)
|
|
key_states = self.k_proj(hidden_states)
|
|
value_states = self.v_proj(hidden_states)
|
|
|
|
query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
|
|
key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
|
|
value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
|
|
|
|
cos, sin = position_embeddings
|
|
query_states, key_states = apply_multimodal_rotary_pos_emb(
|
|
query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
|
|
)
|
|
|
|
# repeat k/v heads if n_kv_heads < n_heads
|
|
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
|
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
|
|
|
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
|
|
|
# Fix precision issues in Qwen2-VL float16 inference
|
|
# Replace inf values with zeros in attention weights to prevent NaN propagation
|
|
if query_states.dtype == torch.float16:
|
|
attn_weights = torch.where(torch.isinf(attn_weights), torch.zeros_like(attn_weights), attn_weights)
|
|
|
|
# upcast attention to fp32
|
|
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
|
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
|
|
attn_output = torch.matmul(attn_weights, value_states)
|
|
|
|
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
|
raise ValueError(
|
|
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
|
|
f" {attn_output.size()}"
|
|
)
|
|
|
|
attn_output = attn_output.transpose(1, 2).contiguous()
|
|
attn_output = attn_output.reshape(bsz, q_len, -1)
|
|
|
|
attn_output = self.o_proj(attn_output)
|
|
|
|
return attn_output
|
|
|
|
|
|
class Qwen2MLP(nn.Module):
|
|
def __init__(self, config):
|
|
super().__init__()
|
|
from transformers.activations import ACT2FN
|
|
self.config = config
|
|
self.hidden_size = config.hidden_size
|
|
self.intermediate_size = config.intermediate_size
|
|
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
|
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
|
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
|
self.act_fn = ACT2FN[config.hidden_act]
|
|
|
|
def forward(self, x):
|
|
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
|
return down_proj
|
|
|
|
|
|
class Qwen2RMSNorm(nn.Module):
|
|
def __init__(self, hidden_size, eps=1e-6):
|
|
"""
|
|
Qwen2RMSNorm is equivalent to T5LayerNorm
|
|
"""
|
|
super().__init__()
|
|
self.weight = nn.Parameter(torch.ones(hidden_size))
|
|
self.variance_epsilon = eps
|
|
|
|
def forward(self, hidden_states):
|
|
input_dtype = hidden_states.dtype
|
|
hidden_states = hidden_states.to(torch.float32)
|
|
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
|
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
|
return self.weight * hidden_states.to(input_dtype)
|
|
|
|
def extra_repr(self):
|
|
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
|
|
|
|
|
class Qwen2_5_VLDecoderLayer(nn.Module):
|
|
def __init__(self, config, layer_idx):
|
|
super().__init__()
|
|
self.hidden_size = config.hidden_size
|
|
|
|
self.self_attn = Qwen2_5_VLAttention(config, layer_idx)
|
|
|
|
self.mlp = Qwen2MLP(config)
|
|
self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC
|
|
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
|
|
|
residual = hidden_states
|
|
|
|
hidden_states = self.input_layernorm(hidden_states)
|
|
|
|
# Self Attention
|
|
hidden_states = self.self_attn(
|
|
hidden_states=hidden_states,
|
|
position_embeddings=position_embeddings,
|
|
)
|
|
hidden_states = residual + hidden_states
|
|
|
|
# Fully Connected
|
|
residual = hidden_states
|
|
hidden_states = self.post_attention_layernorm(hidden_states)
|
|
hidden_states = self.mlp(hidden_states)
|
|
hidden_states = residual + hidden_states
|
|
|
|
return hidden_states
|
|
|
|
|
|
class NexusGenImageEmbeddingMerger(nn.Module):
|
|
def __init__(self, num_layers=1, out_channel=4096, expand_ratio=4, device='cpu'):
|
|
super().__init__()
|
|
from transformers import Qwen2_5_VLConfig
|
|
from transformers.activations import ACT2FN
|
|
config = Qwen2_5_VLConfig(**{
|
|
"_name_or_path": "DiffSynth-Studio/Nexus-GenV2",
|
|
"architectures": [
|
|
"Qwen2_5_VLForConditionalGeneration"
|
|
],
|
|
"attention_dropout": 0.0,
|
|
"auto_map": {
|
|
"AutoConfig": "configuration_qwen2_5_vl.Qwen2_5_VLConfig",
|
|
"AutoModel": "modeling_qwen2_5_vl.Qwen2_5_VLModel",
|
|
"AutoModelForCausalLM": "modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration"
|
|
},
|
|
"bos_token_id": 151643,
|
|
"eos_token_id": 151645,
|
|
"hidden_act": "silu",
|
|
"hidden_size": 3584,
|
|
"image_token_id": 151655,
|
|
"initializer_range": 0.02,
|
|
"intermediate_size": 18944,
|
|
"max_position_embeddings": 128000,
|
|
"max_window_layers": 28,
|
|
"model_type": "qwen2_5_vl",
|
|
"num_attention_heads": 28,
|
|
"num_hidden_layers": 28,
|
|
"num_key_value_heads": 4,
|
|
"pad_token_id": 151643,
|
|
"rms_norm_eps": 1e-06,
|
|
"rope_scaling": {
|
|
"mrope_section": [
|
|
16,
|
|
24,
|
|
24
|
|
],
|
|
"rope_type": "default",
|
|
"type": "default"
|
|
},
|
|
"rope_theta": 1000000.0,
|
|
"sliding_window": 32768,
|
|
"tie_word_embeddings": False,
|
|
"torch_dtype": "bfloat16",
|
|
"transformers_version": "4.49.0",
|
|
"use_cache": False,
|
|
"use_sliding_window": False,
|
|
"video_token_id": 151656,
|
|
"vision_config": {
|
|
"hidden_size": 1280,
|
|
"in_chans": 3,
|
|
"model_type": "qwen2_5_vl",
|
|
"spatial_patch_size": 14,
|
|
"tokens_per_second": 2,
|
|
"torch_dtype": "bfloat16"
|
|
},
|
|
"vision_end_token_id": 151653,
|
|
"vision_start_token_id": 151652,
|
|
"vision_token_id": 151654,
|
|
"vocab_size": 152064
|
|
})
|
|
self.config = config
|
|
self.num_layers = num_layers
|
|
self.layers = nn.ModuleList([Qwen2_5_VLDecoderLayer(config, layer_idx) for layer_idx in range(num_layers)])
|
|
self.projector = nn.Sequential(Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps),
|
|
nn.Linear(config.hidden_size, out_channel * expand_ratio),
|
|
Qwen2RMSNorm(out_channel * expand_ratio, eps=config.rms_norm_eps),
|
|
ACT2FN[config.hidden_act], nn.Linear(out_channel * expand_ratio, out_channel),
|
|
Qwen2RMSNorm(out_channel, eps=config.rms_norm_eps))
|
|
self.base_grid = torch.tensor([[1, 72, 72]], device=device)
|
|
self.rotary_emb = Qwen2_5_VLRotaryEmbedding(config=config, device=device)
|
|
|
|
def get_position_ids(self, image_grid_thw):
|
|
"""
|
|
Generates position ids for the input embeddings grid.
|
|
modified from the qwen2_vl mrope.
|
|
"""
|
|
batch_size = image_grid_thw.shape[0]
|
|
spatial_merge_size = self.config.vision_config.spatial_merge_size
|
|
t, h, w = (
|
|
image_grid_thw[0][0],
|
|
image_grid_thw[0][1],
|
|
image_grid_thw[0][2],
|
|
)
|
|
llm_grid_t, llm_grid_h, llm_grid_w = (
|
|
t.item(),
|
|
h.item() // spatial_merge_size,
|
|
w.item() // spatial_merge_size,
|
|
)
|
|
scale_h = self.base_grid[0][1].item() / h.item()
|
|
scale_w = self.base_grid[0][2].item() / w.item()
|
|
|
|
range_tensor = torch.arange(llm_grid_t).view(-1, 1)
|
|
expanded_range = range_tensor.expand(-1, llm_grid_h * llm_grid_w)
|
|
time_tensor = expanded_range * self.config.vision_config.tokens_per_second
|
|
t_index = time_tensor.long().flatten().to(image_grid_thw.device)
|
|
h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten().to(image_grid_thw.device) * scale_h
|
|
w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten().to(image_grid_thw.device) * scale_w
|
|
# 3, B, L
|
|
position_ids = torch.stack([t_index, h_index, w_index]).unsqueeze(0).repeat(batch_size, 1, 1).permute(1, 0, 2)
|
|
return position_ids
|
|
|
|
def forward(self, embeds, embeds_grid, ref_embeds=None, ref_embeds_grid=None):
|
|
position_ids = self.get_position_ids(embeds_grid)
|
|
hidden_states = embeds
|
|
if ref_embeds is not None:
|
|
position_ids_ref_embeds = self.get_position_ids(ref_embeds_grid)
|
|
position_ids = torch.cat((position_ids, position_ids_ref_embeds), dim=-1)
|
|
hidden_states = torch.cat((embeds, ref_embeds), dim=1)
|
|
|
|
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
|
for layer in self.layers:
|
|
hidden_states = layer(hidden_states, position_embeddings)
|
|
|
|
hidden_states = self.projector(hidden_states)
|
|
return hidden_states
|
|
|
|
@staticmethod
|
|
def state_dict_converter():
|
|
return NexusGenMergerStateDictConverter()
|
|
|
|
|
|
class NexusGenMergerStateDictConverter:
|
|
def __init__(self):
|
|
pass
|
|
|
|
def from_diffusers(self, state_dict):
|
|
return state_dict
|
|
|
|
def from_civitai(self, state_dict):
|
|
merger_state_dict = {key.replace("embedding_merger.", ""): value for key, value in state_dict.items() if key.startswith('embedding_merger.')}
|
|
return merger_state_dict
|
|
|
|
|
|
class NexusGenAdapter(nn.Module):
|
|
"""
|
|
Adapter for Nexus-Gen generation decoder.
|
|
"""
|
|
def __init__(self, input_dim=3584, output_dim=4096):
|
|
super(NexusGenAdapter, self).__init__()
|
|
self.adapter = nn.Sequential(nn.Linear(input_dim, output_dim),
|
|
nn.LayerNorm(output_dim), nn.ReLU(),
|
|
nn.Linear(output_dim, output_dim),
|
|
nn.LayerNorm(output_dim))
|
|
|
|
def forward(self, x):
|
|
return self.adapter(x)
|
|
|
|
@staticmethod
|
|
def state_dict_converter():
|
|
return NexusGenAdapterStateDictConverter()
|
|
|
|
|
|
class NexusGenAdapterStateDictConverter:
|
|
def __init__(self):
|
|
pass
|
|
|
|
def from_diffusers(self, state_dict):
|
|
return state_dict
|
|
|
|
def from_civitai(self, state_dict):
|
|
adapter_state_dict = {key: value for key, value in state_dict.items() if key.startswith('adapter.')}
|
|
return adapter_state_dict
|