File size: 17,859 Bytes
bd9a568
 
 
 
 
60031f1
bd9a568
 
 
 
 
 
 
 
 
 
 
 
 
 
 
60031f1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5c5a7f4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
60031f1
5c5a7f4
 
 
60031f1
5c5a7f4
 
 
 
 
 
60031f1
 
5c5a7f4
 
 
 
 
 
60031f1
5c5a7f4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
60031f1
 
5c5a7f4
 
 
 
 
 
 
60031f1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bd9a568
 
 
 
 
60031f1
 
 
 
 
 
 
 
 
 
 
bd9a568
 
 
 
 
 
 
 
 
 
60031f1
bd9a568
 
 
53cf78b
60031f1
 
 
53cf78b
bd9a568
 
 
 
 
 
 
 
60031f1
 
 
 
 
 
bd9a568
 
 
 
 
60031f1
bd9a568
60031f1
bd9a568
60031f1
bd9a568
60031f1
bd9a568
 
 
 
60031f1
bd9a568
 
 
 
 
 
 
 
 
60031f1
 
 
 
 
bd9a568
 
 
 
60031f1
bd9a568
 
 
 
 
 
 
 
60031f1
 
 
 
 
53cf78b
60031f1
 
 
 
 
 
 
 
 
 
bd9a568
 
 
 
60031f1
bd9a568
 
 
53cf78b
 
60031f1
53cf78b
60031f1
 
 
 
 
 
bd9a568
60031f1
53cf78b
 
 
60031f1
bd9a568
 
60031f1
bd9a568
 
 
60031f1
 
 
 
 
53cf78b
60031f1
bd9a568
60031f1
 
 
 
 
 
 
 
 
53cf78b
bd9a568
 
 
 
60031f1
 
 
 
 
 
bd9a568
 
60031f1
 
 
 
 
 
 
 
bd9a568
 
60031f1
bd9a568
 
 
 
60031f1
 
 
 
bd9a568
 
 
60031f1
 
 
 
 
 
 
 
 
bd9a568
60031f1
bd9a568
60031f1
bd9a568
60031f1
 
 
 
 
 
 
 
 
 
 
 
 
bd9a568
60031f1
 
 
bd9a568
60031f1
bd9a568
 
 
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
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
import inspect
import math
import warnings
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Tuple, Union

import numpy as np

import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn

from transformers.activations import ACT2FN
from transformers.utils.import_utils import is_torch_fx_available

from torch.utils.checkpoint import checkpoint
from functools import partial



# Try to import flash-attn; if unavailable or fails to initialize on this device
# we will set a flag and provide a fallback implementation below.
try:
    from flash_attn import flash_attn_func as _flash_attn_func
    from flash_attn import flash_attn_varlen_func as _flash_attn_varlen_func
    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa: F401

    HAVE_FLASH_ATTN = True
except Exception:
    _flash_attn_func = None
    _flash_attn_varlen_func = None
    index_first_axis = None
    pad_input = None
    unpad_input = None
    HAVE_FLASH_ATTN = False



def _repeat_kv_for_gqa(x: torch.Tensor, repeat: int) -> torch.Tensor:
    # x: [B, S, Hk, D] -> [B, S, Hq, D], where Hq = Hk * repeat
    if repeat == 1:
        return x
    B, S, Hk, D = x.shape
    x = x.unsqueeze(2).expand(B, S, repeat, Hk, D)   # [B,S,repeat,Hk,D]
    return x.reshape(B, S, repeat * Hk, D)

@torch.no_grad()
def _build_window_mask(
    Sq: int, Sk: int, left: int, right: int, causal: bool, device: torch.device
) -> torch.Tensor:
    """
    FA2 window semantics:
      valid j for query i: j ∈ [ i + Sk - Sq - left, i + Sk - Sq + right ]
    FA2.1 causal alignment (bottom-right): additionally disallow j > i + Sk - Sq
    Return: float mask [1,1,Sq,Sk] with 0 for keep, -inf for mask.
    """
    i = torch.arange(Sq, device=device).view(-1, 1)  # [Sq,1]
    j = torch.arange(Sk, device=device).view(1, -1)  # [1,Sk]
    shift = Sk - Sq
    j_min = i + shift - left
    j_max = i + shift + right
    allowed = (j >= j_min) & (j <= j_max)
    if causal:
        # forbid looking ahead relative to FA2.1 alignment
        allowed &= (j <= (i + shift))
    masked = ~allowed
    m = torch.full((Sq, Sk), 0.0, device=device)
    m[masked] = -torch.finfo(m.dtype).max  # -inf
    return m.view(1, 1, Sq, Sk).contiguous()

@torch.no_grad()
def _build_causal_mask_fa21(
    Sq: int, Sk: int, device: torch.device
) -> torch.Tensor:
    """
    FA2.1 causal only (no window): mask positions with j > i + (Sk - Sq).
    Returns float mask [1,1,Sq,Sk] with 0 keep, -inf mask.
    """
    i = torch.arange(Sq, device=device).view(-1, 1)
    j = torch.arange(Sk, device=device).view(1, -1)
    shift = Sk - Sq
    allowed = (j <= (i + shift))
    masked = ~allowed
    m = torch.full((Sq, Sk), 0.0, device=device)
    m[masked] = -torch.finfo(m.dtype).max
    return m.view(1, 1, Sq, Sk).contiguous()

def _sdpa_flash_attn_compat(
    q: torch.Tensor,  # [B,Sq,Hq,D]
    k: torch.Tensor,  # [B,Sk,Hk,D]
    v: torch.Tensor,  # [B,Sk,Hk,D]
    *,
    dropout_p: float = 0.0,
    softmax_scale: Optional[float] = None,    # default 1/sqrt(D) if None
    causal: bool = False,
    window_size: Tuple[int, int] = (-1, -1),  # (-1,-1) == no window
    alibi_slopes: Optional[torch.Tensor] = None,  # (Hq,) or (B,Hq)
    training: Optional[bool] = None,
) -> torch.Tensor:
    """
    SDPA path emulating flash_attn_func semantics (v2):
      - supports GQA (Hq divisible by Hk)
      - FA2.1 causal alignment when Sq != Sk
      - sliding window: j in [i + Sk - Sq - left, i + Sk - Sq + right]
      - ALiBi additive bias
    Returns: [B,Sq,Hq,D] with original dtype.
    """
    assert q.dim() == k.dim() == v.dim() == 4, "Expect [B,S,H,D] tensors"
    B, Sq, Hq, D = q.shape
    Bk, Sk, Hk, Dk = k.shape
    assert (Bk, Sk, Dk) == (B, k.shape[1], D), "Batch/Dim mismatch"
    assert v.shape[:3] == k.shape[:3] and v.shape[3] == D, "K/V mismatch"
    assert Hq % Hk == 0, "Hq must be divisible by Hk for GQA/MQA"
    repeat = Hq // Hk

    # GQA: expand K,V heads to match Q heads so SDPA sees [B,Hq,*,D]
    k_exp = _repeat_kv_for_gqa(k, repeat)  # [B,Sk,Hq,D]
    v_exp = _repeat_kv_for_gqa(v, repeat)  # [B,Sk,Hq,D]

    # layout for SDPA: [B,H,S,D]
    qh = q.permute(0, 2, 1, 3).to(torch.float32)     # [B,Hq,Sq,D]
    kh = k_exp.permute(0, 2, 1, 3).to(torch.float32) # [B,Hq,Sk,D]
    vh = v_exp.permute(0, 2, 1, 3).to(torch.float32) # [B,Hq,Sk,D]
    in_dtype = q.dtype
    device = q.device

    # softmax scale: default 1/sqrt(D); emulate custom s by scaling Q by s*sqrt(D)
    if softmax_scale is None:
        softmax_scale = 1.0 / math.sqrt(D)
    qh = qh * (softmax_scale * math.sqrt(D))

    # Build float mask (+ALiBi) as additive bias; pass is_causal=False to SDPA.
    left, right = window_size
    use_window = (left, right) != (-1, -1)
    attn_bias = None  # [B,Hq,Sq,Sk] float, 0 for keep, -inf for mask, +ALiBi

    if use_window:
        # Per FA2 semantics; also clamp look-ahead under causal
        if causal and right > 0:
            right = 0
        base = _build_window_mask(Sq, Sk, left, right, causal, device)  # [1,1,Sq,Sk]
        attn_bias = base.expand(B, Hq, Sq, Sk)
        is_causal = False
    elif causal:
        base = _build_causal_mask_fa21(Sq, Sk, device)  # [1,1,Sq,Sk]
        attn_bias = base.expand(B, Hq, Sq, Sk)
        is_causal = False
    else:
        is_causal = False
        attn_bias = None  # fastest path

    # ALiBi: add -(slope * |(i + Sk - Sq) - j|) to logits (i=0..Sq-1, j=0..Sk-1)
    if alibi_slopes is not None:
        # make slopes shape [B,Hq,1,1]
        if alibi_slopes.dim() == 1:
            # [Hq] -> [1,Hq,1,1]
            alibi = alibi_slopes.view(1, Hq, 1, 1).to(dtype=torch.float32, device=device)
            alibi = alibi.expand(B, Hq, 1, 1)
        elif alibi_slopes.dim() == 2:
            # [B,Hq] -> [B,Hq,1,1]
            alibi = alibi_slopes.view(B, Hq, 1, 1).to(dtype=torch.float32, device=device)
        else:
            raise ValueError("alibi_slopes must be (Hq,) or (B,Hq)")
        i = torch.arange(Sq, device=device).view(1, 1, -1, 1)
        j = torch.arange(Sk, device=device).view(1, 1, 1, -1)
        shift = Sk - Sq
        dist = (i + shift - j).abs().to(torch.float32)  # [1,1,Sq,Sk]
        alibi_term = -(alibi * dist)                     # [B,Hq,Sq,Sk]
        if attn_bias is None:
            attn_bias = alibi_term
        else:
            attn_bias = attn_bias + alibi_term

    # Dropout (train) vs eval
    if training is None:
        training = (dropout_p > 0.0) and any(t.requires_grad for t in (q, k, v))
    dp = dropout_p if training else 0.0

    out = F.scaled_dot_product_attention(
        qh, kh, vh,
        attn_mask=attn_bias,   # float additive mask/bias or None
        dropout_p=dp,
        is_causal=is_causal,   # we encode causal via mask/bias when needed
    )  # [B,Hq,Sq,D] fp32

    return out.permute(0, 2, 1, 3).to(in_dtype).contiguous()  # [B,Sq,Hq,D]



def _attn_dispatch(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    *,
    causal: bool = True,
    window_size: Optional[Tuple[int, int]] = None,
) -> torch.Tensor:
    """
    Dispatches to either flash attention or the SDPA fallback.  This function
    accepts and returns tensors shaped ``[batch, seq_len, num_heads, head_dim]``.
    """
    if HAVE_FLASH_ATTN:
        # If flash attention is available we use it directly.  Note that
        # ``flash_attn_func`` accepts the same tensor layout and returns a
        # tensor with identical shape.  Additional keyword arguments such as
        # ``softmax_scale`` and ``dropout_p`` will use default values.
        return _flash_attn_func(
            q,
            k,
            v,
            causal=causal,
            window_size=window_size,
        )
    # Otherwise use the fallback implementation.
    return _sdpa_flash_attn_compat(
        q,
        k,
        v,
        causal=causal,
        window_size=window_size,
    )


def rotate_half(x: torch.Tensor) -> torch.Tensor:
    """Rotate half the hidden dimensions of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(
    q: Optional[torch.Tensor],
    k: Optional[torch.Tensor],
    cos: torch.Tensor,
    sin: torch.Tensor,
    position_ids: Optional[torch.Tensor] = None,
    unsqueeze_dim: int = 1,
) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
    """
    Applies rotary position embeddings to the query and key tensors.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    if q is not None:
        q_embed = (q * cos) + (rotate_half(q) * sin)
    else:
        q_embed = None
    if k is not None:
        k_embed = (k * cos) + (rotate_half(k) * sin)
    else:
        k_embed = None
    return q_embed, k_embed


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    Equivalent to ``torch.repeat_interleave(x, dim=1, repeats=n_rep)``.  Converts
    hidden states from shape (batch, num_key_value_heads, seq_len, head_dim) to
    (batch, num_attention_heads, seq_len, 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 RotaryEmbedding(nn.Module):
    """
    Computes rotary position embeddings.  See
    https://arxiv.org/abs/2104.09864 for details.
    """

    def __init__(self, dim: int, base: int = 10000):
        super().__init__()
        self.dim = dim
        self.base = base
        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float() / self.dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    @torch.no_grad()
    def forward(self, x: torch.Tensor, position_ids: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, torch.Tensor]:
        if position_ids is None:
            # position_ids shape: [batch, seq_len]
            position_ids = torch.arange(x.shape[2], device=x.device, dtype=torch.int64).unsqueeze(0).expand(x.shape[0], -1)
        # x shape: [batch, num_heads, seq_len, head_size]
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        # Force float32 for numerical stability on long contexts.
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


class RMSNorm(nn.Module):
    """
    Root Mean Square layer normalization.  Equivalent to T5LayerNorm.
    """

    def __init__(self, hidden_size: int, eps: float = 1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        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)


class Attention(nn.Module):
    """
    Multi‑head attention module with optional rotary positional embeddings and
    windowed attention.  Uses flash attention when available, otherwise falls
    back to PyTorch's scaled dot product attention.
    """

    def __init__(
        self,
        num_attention_heads: int,
        num_key_value_heads: int,
        attention_head_size: int,
        attention_window_size: Optional[int] = None,
        seq_length: Optional[int] = None,
        use_positional_embedding: bool = False,
        rope_base: Optional[int] = None,
    ):
        super().__init__()
        self.num_attention_heads = num_attention_heads
        self.num_key_value_heads = num_key_value_heads
        self.attention_head_size = attention_head_size
        self.num_key_value_groups = self.num_attention_heads // self.num_key_value_heads
        self.attention_window_size = attention_window_size
        self.seq_length = seq_length
        self.use_positional_embedding = use_positional_embedding
        self.rope_base = rope_base
        if self.use_positional_embedding:
            self.rotary_emb = RotaryEmbedding(dim=self.attention_head_size, base=self.rope_base)

    def forward(
        self,
        query_states: torch.Tensor,
        key_states: torch.Tensor,
        value_states: torch.Tensor,
    ) -> torch.Tensor:
        bsz, q_len, _ = query_states.size()
        # Reshape to [batch, seq_len, num_heads, head_dim] and bring heads to axis 2.
        query_states = query_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_size).transpose(1, 2).contiguous()
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.attention_head_size).transpose(1, 2).contiguous()
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.attention_head_size).transpose(1, 2).contiguous()
        # Repeat keys/values if there are more query heads than key/value heads.
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
        # Apply rotary positional embeddings if requested.
        if self.use_positional_embedding:
            cos, sin = self.rotary_emb(query_states)
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
        # Move the seq_len dimension back to axis 1: [B, S, H, D].
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)
        # Compute attention.  Window size is specified as a tuple when present.
        if self.attention_window_size is not None:
            ws = (self.attention_window_size, self.attention_window_size)
        else:
            ws = None
        attn_outputs = _attn_dispatch(
            query_states,
            key_states,
            value_states,
            causal=True,
            window_size=ws,
        )
        # Merge heads back: [B, S, H*D].
        attn_outputs = attn_outputs.reshape(bsz, q_len, int(self.num_attention_heads * self.attention_head_size)).contiguous()
        return attn_outputs


class Block(nn.Module):
    """
    Basic transformer block consisting of an input projection into query/key/value
    and residual channels, a single attention layer, layer normalization and an
    output projection.
    """

    def __init__(
        self,
        hidden_size: int = 768,
        num_attention_heads: int = 12,
        num_key_value_heads: int = 4,
        attention_window_size: Optional[int] = None,
        seq_length: Optional[int] = None,
        use_positional_embedding: bool = False,
        rope_base: Optional[int] = None,
    ):
        super().__init__()
        self.hidden_size = hidden_size
        # In this architecture the intermediate size equals the hidden size.
        self.intermediate_size = self.hidden_size
        self.num_attention_heads = num_attention_heads
        self.num_key_value_heads = num_key_value_heads
        self.attention_head_size = int(self.intermediate_size / self.num_attention_heads)
        # The latent dimension contains the residual channel (intermediate_size)
        # plus separate query/key/value projections.  The factor of 2 accounts
        # for concatenated key and value tensors.
        self.latent_dim = self.intermediate_size + self.attention_head_size * self.num_key_value_heads * 2
        self.pre_avg_layernorm = RMSNorm(self.intermediate_size)
        self.in_proj = nn.Linear(self.hidden_size, self.latent_dim, bias=True)
        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=True)
        self.self_attn = Attention(
            self.num_attention_heads,
            self.num_key_value_heads,
            self.attention_head_size,
            attention_window_size,
            seq_length,
            use_positional_embedding,
            rope_base,
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, seq_len, hidden_size = hidden_states.shape
        # Project to queries, keys, values, and residuals.
        hidden_states = self.in_proj(hidden_states).transpose(1, 2)
        # Split into (q,k,v,residual).  Note: tensor_split returns views.
        q, k, v, residual = hidden_states.tensor_split(
            (
                self.intermediate_size,
                self.intermediate_size + self.attention_head_size * self.num_key_value_heads,
                self.intermediate_size + self.attention_head_size * self.num_key_value_heads * 2,
            ),
            dim=1,
        )
        q = q.transpose(1, 2)
        k = k.transpose(1, 2)
        v = v.transpose(1, 2)
        # Apply self attention.
        attn_outputs = self.self_attn(
            query_states=q,
            key_states=k,
            value_states=v,
        )
        # Normalize and project back to hidden size.
        hidden_states = self.pre_avg_layernorm(attn_outputs)
        contextualized_states = self.out_proj(hidden_states)
        return contextualized_states