File size: 12,023 Bytes
7e53c9a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 | from typing import Any, Optional, Tuple, Union
import torch
from torch import Tensor, nn
from torch.nn.functional import scaled_dot_product_attention
from transformers import WhisperConfig
from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions
from transformers.models.whisper.modeling_whisper import WhisperEncoder, WhisperEncoderLayer
from transformers.utils import logging
logger = logging.get_logger(__name__)
# ==========================================
# 1. Core Rotary Embedding components
# ==========================================
class RotaryEmbedding(nn.Module):
def __init__(self, dim, rope_ratio=1):
super().__init__()
self.dim = dim
self.rope_ratio = rope_ratio
@torch.no_grad()
def get_emb(self, seq_len: int, dtype: torch.dtype, device: torch.device, base: int = 10000):
"""Generate the cached RoPE table."""
base = base * self.rope_ratio
# Compute the theta frequencies.
inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, dtype=torch.float, device=device) / self.dim))
# Build the position indices.
t = torch.arange(seq_len, device=device, dtype=torch.float)
freqs = torch.outer(t, inv_freq) # [seq_len, dim/2]
# Construct the cos/sin cache.
# Shape: [seq_len, dim/2, 2]
emb = torch.stack([torch.cos(freqs), torch.sin(freqs)], dim=-1)
if dtype in (torch.float16, torch.bfloat16):
emb = emb.to(dtype)
return emb
def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
"""
x: [batch, num_heads, seq_len, head_dim]
rope_cache: [1, seq_len, dim/2, 2]
"""
b, nh, sq, hd = x.shape
rot_dim = rope_cache.shape[-2] * 2
# Split x into the rotated and pass-through portions.
x_rot, x_pass = x[..., :rot_dim], x[..., rot_dim:]
# Reshape x_rot to match rope_cache: [b, nh, sq, rot_dim/2, 2]
x_shaped = x_rot.reshape(b, nh, sq, rot_dim // 2, 2)
# Apply the complex rotation: (a+bi)(c+di) = (ac-bd) + (ad+bc)i
cos = rope_cache[..., 0] # [1, sq, rot_dim/2]
sin = rope_cache[..., 1] # [1, sq, rot_dim/2]
# Add the head dimension.
cos = cos.unsqueeze(1) # [1, 1, sq, rot_dim/2]
sin = sin.unsqueeze(1) # [1, 1, sq, rot_dim/2]
x_out = torch.stack([
x_shaped[..., 0] * cos - x_shaped[..., 1] * sin,
x_shaped[..., 1] * cos + x_shaped[..., 0] * sin
], dim=-1)
x_out = x_out.flatten(3) # Merge the final two dimensions into rot_dim.
return torch.cat([x_out, x_pass], dim=-1)
# ==========================================
# 2. RoPE attention built on SDPA
# ==========================================
class WhisperRoPESdpaAttention(nn.Module):
"""
Replace WhisperFlashAttention2 with PyTorch's native scaled_dot_product_attention.
"""
def __init__(self, config: WhisperConfig, embed_dim: int, num_heads: int, dropout: float = 0.0):
super().__init__()
self.config = config
self.embed_dim = embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
# Standard Whisper projection layers.
self.q_proj = nn.Linear(embed_dim, embed_dim)
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=False)
self.v_proj = nn.Linear(embed_dim, embed_dim)
self.out_proj = nn.Linear(embed_dim, embed_dim)
self.is_causal = False
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
layer_head_mask: Optional[torch.Tensor] = None,
output_attentions: bool = False,
rotary_pos_emb: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], None]:
bsz, q_len, _ = hidden_states.size()
# 1. Project to queries, keys, and values.
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
# 2. Reshape to [batch, heads, seq, dim] and keep memory contiguous.
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
key_states = key_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
value_states = value_states.view(bsz, -1, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
# 3. Apply RoPE.
if rotary_pos_emb is not None:
query_states = apply_rotary_pos_emb(query_states, rotary_pos_emb)
key_states = apply_rotary_pos_emb(key_states, rotary_pos_emb)
# 4. Align dtypes to avoid mismatches introduced by fp32 LayerNorm.
target_dtype = self.q_proj.weight.dtype
query_states = query_states.to(target_dtype)
key_states = key_states.to(target_dtype)
value_states = value_states.to(target_dtype)
# 5. Run SDPA. Do not apply manual scaling; SDPA handles it internally.
# If a 4D attention_mask is provided, SDPA applies it correctly.
attn_output = scaled_dot_product_attention(
query_states,
key_states,
value_states,
attn_mask=attention_mask,
dropout_p=self.dropout if self.training else 0.0,
is_causal=self.is_causal,
)
# 6. Restore shape and apply the output projection.
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, None, None
# ==========================================
# 3. Wrapped encoder layer and encoder
# ==========================================
class WhisperSpecialEncoderLayer(WhisperEncoderLayer):
def __init__(self, config: WhisperConfig):
super().__init__(config)
# Replace self-attention with the RoPE + SDPA implementation.
self.self_attn = WhisperRoPESdpaAttention(
config=config,
embed_dim=self.embed_dim,
num_heads=config.encoder_attention_heads,
dropout=config.attention_dropout,
)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
layer_head_mask: Optional[torch.Tensor] = None,
output_attentions: bool = False,
rotary_pos_emb: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Any]:
residual = hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
hidden_states, attn_weights, _ = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
layer_head_mask=layer_head_mask,
output_attentions=output_attentions,
rotary_pos_emb=rotary_pos_emb,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
residual = hidden_states
hidden_states = self.final_layer_norm(hidden_states)
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
hidden_states = self.fc2(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
if hidden_states.dtype == torch.float16:
clamp_value = torch.finfo(hidden_states.dtype).max - 1000
hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
return (hidden_states, None) # Keep the tuple length aligned with the Whisper interface.
class WhisperSpecialEncoder(WhisperEncoder):
def __init__(self, config: WhisperConfig, use_rope=True, rope_ratio=1):
super().__init__(config)
self.use_rope = use_rope
# Override the parent layer stack.
self.layers = nn.ModuleList(
[WhisperSpecialEncoderLayer(config) for _ in range(config.encoder_layers)]
)
if use_rope:
# Compute the RoPE dimension, typically a subset of head_dim.
head_dim = config.d_model // config.encoder_attention_heads
self.rotary_embedding = RotaryEmbedding(head_dim // 2, rope_ratio)
def forward(
self,
input_features,
attention_mask=None,
head_mask=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
position_ids=None,
):
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# Whisper convolutional feature extraction.
inputs_embeds = nn.functional.gelu(self.conv1(input_features))
inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))
inputs_embeds = inputs_embeds.permute(0, 2, 1) # [B, T_down, D]
if self.use_rope:
# Build the rotary embedding cache.
rotary_embs = self.rotary_embedding.get_emb(
seq_len=inputs_embeds.shape[1],
dtype=inputs_embeds.dtype,
device=inputs_embeds.device
)
# Reshape to [1, seq_len, dim/2, 2] for broadcasting.
rotary_embs = rotary_embs.unsqueeze(0)
hidden_states = inputs_embeds
else:
rotary_embs = None
# Fall back to absolute positional embeddings.
embed_pos = self.embed_positions.weight[:inputs_embeds.shape[1]]
hidden_states = inputs_embeds + embed_pos
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
encoder_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
for idx, encoder_layer in enumerate(self.layers):
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
encoder_layer.__call__,
hidden_states,
None, # attention_mask
(head_mask[idx] if head_mask is not None else None),
output_attentions,
rotary_embs,
position_ids,
)
else:
layer_outputs = encoder_layer(
hidden_states,
attention_mask=None,
layer_head_mask=(head_mask[idx] if head_mask is not None else None),
output_attentions=output_attentions,
rotary_pos_emb=rotary_embs,
position_ids=position_ids,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[2],)
hidden_states = self.layer_norm(hidden_states)
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
hidden_states=encoder_states,
attentions=all_attentions,
) |