File size: 13,205 Bytes
c43aac0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from dataclasses import dataclass
from typing import Optional


@dataclass
class ModelConfig:
    vocab_size: int
    hidden_size: int
    n_heads: int
    n_kv_heads: int
    n_kv_groups: int
    head_dim: int
    n_layers: int
    attention_bias: bool
    intermediate_size: int
    mlp_bias: bool
    eps: float
    dropout: float
    max_position_embeddings: int
    pre_norm: bool
    tie_weights: bool
    max_seq_len: int


class RMSNorm(nn.Module):
    def __init__(self, config: ModelConfig):
        super().__init__()
        self.eps = config.eps
        self.weight = nn.Parameter(torch.ones(config.hidden_size))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        rms = torch.sqrt(torch.mean(x * x, dim=-1, keepdim=True) + self.eps)
        return (x / rms) * self.weight


class RotaryEmbedding(nn.Module):
    def __init__(self, head_dim, max_position_embeddings=2048):
        super().__init__()
        inv_freq = 1.0 / (10000 ** (torch.arange(0, head_dim, 2).float() / head_dim))
        t = torch.arange(max_position_embeddings, dtype=torch.float32)
        freqs = torch.einsum("i,j->ij", t, inv_freq)
        emb = torch.cat((freqs, freqs), dim=-1)
        self.register_buffer("cos", emb.cos()[None, None, :, :], persistent=False)
        self.register_buffer("sin", emb.sin()[None, None, :, :], persistent=False)

    def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype):
        cos = self.cos[:, :, :seq_len, :].to(device=device, dtype=dtype)
        sin = self.sin[:, :, :seq_len, :].to(device=device, dtype=dtype)
        return cos, sin


def apply_rotary(x, cos, sin):
    x1, x2 = x[..., ::2], x[..., 1::2]
    x_rot = torch.stack([-x2, x1], dim=-1).reshape_as(x)
    return (x * cos) + (x_rot * sin)


class GroupedMultiQueryAttention(nn.Module):
    def __init__(self, config: ModelConfig):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.n_heads = config.n_heads
        self.n_kv_heads = config.n_kv_heads
        self.head_dim = config.head_dim
        self.attention_bias = config.attention_bias
        self.dropout = nn.Dropout(config.dropout)

        if self.n_heads * self.head_dim != self.hidden_size:
            raise ValueError("hidden_size must equal n_heads * head_dim")

        # derive n_kv_groups if None
        if config.n_kv_groups is None:
            if self.n_kv_heads == 0:
                raise ValueError("n_kv_heads must be > 0")
            self.n_kv_groups = self.n_heads // self.n_kv_heads
            if self.n_heads % self.n_kv_heads != 0:
                raise ValueError("n_heads must be divisible by n_kv_heads to derive groups")
        else:
            self.n_kv_groups = config.n_kv_groups
            if self.n_kv_heads * self.n_kv_groups != self.n_heads:
                raise ValueError("n_heads must equal n_kv_heads * n_kv_groups")

        self.q_proj = nn.Linear(self.hidden_size, self.n_heads * self.head_dim, bias=self.attention_bias)
        self.k_proj = nn.Linear(self.hidden_size, self.n_kv_heads * self.head_dim, bias=self.attention_bias)
        self.v_proj = nn.Linear(self.hidden_size, self.n_kv_heads * self.head_dim, bias=self.attention_bias)
        self.w_o = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
        self.rope = RotaryEmbedding(self.head_dim, config.max_position_embeddings)

    def forward(self, x):
        B, T, _ = x.shape
        device = x.device
        dtype = x.dtype

        q = self.q_proj(x).view(B, T, self.n_heads, self.head_dim).transpose(1, 2)
        k = self.k_proj(x).view(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2)
        v = self.v_proj(x).view(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2)

        cos, sin = self.rope(T, device=device, dtype=dtype)
        q = apply_rotary(q, cos, sin)
        k = apply_rotary(k, cos, sin)

        if self.n_kv_groups != 1:
            k = k.repeat_interleave(self.n_kv_groups, dim=1)
            v = v.repeat_interleave(self.n_kv_groups, dim=1)

        scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)

        # causal mask
        mask = torch.triu(torch.full((T, T), float("-inf"), device=device, dtype=dtype), diagonal=1)
        scores = scores + mask.unsqueeze(0).unsqueeze(0)
        attn = torch.softmax(scores, dim=-1)
        attn = self.dropout(attn)

        out = torch.matmul(attn, v)
        out = out.transpose(1, 2).contiguous().view(B, T, self.hidden_size)
        return self.w_o(out)


class SwiGLUFeedForward(nn.Module):
    def __init__(self, config: ModelConfig):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.dropout = nn.Dropout(config.dropout)

        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size)
        self.act = nn.SiLU()

    def forward(self, x):
        x = self.act(self.gate_proj(x)) * self.up_proj(x)
        x = self.down_proj(self.dropout(x))
        return x


class TransformerBlock(nn.Module):
    def __init__(self, config: ModelConfig):
        super().__init__()
        self.attention = GroupedMultiQueryAttention(config)
        self.feed_forward = SwiGLUFeedForward(config)
        self.attn_norm = RMSNorm(config)
        self.ffn_norm = RMSNorm(config)
        self.dropout = nn.Dropout(config.dropout)
        self.pre_norm = config.pre_norm

    def forward(self, x):
        if self.pre_norm:
            x = x + self.dropout(self.attention(self.attn_norm(x)))
            x = x + self.dropout(self.feed_forward(self.ffn_norm(x)))
        else:
            x = self.attn_norm(x + self.dropout(self.attention(x)))
            x = self.ffn_norm(x + self.dropout(self.feed_forward(x)))
        return x


class Transformer(nn.Module):
    def __init__(self, config: ModelConfig):
        super().__init__()
        self.config = config
        self.token_embedding = nn.Embedding(config.vocab_size, config.hidden_size)
        self.blocks = nn.ModuleList([TransformerBlock(config) for _ in range(config.n_layers)])
        self.embedding_dropout = nn.Dropout(config.dropout)
        self.final_norm = RMSNorm(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        if config.tie_weights:
            self.lm_head.weight = self.token_embedding.weight

        self.apply(self._init_weights)

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            nn.init.normal_(module.weight, mean=0.0, std=0.02 / math.sqrt(max(1, self.config.n_layers)))
            if module.bias is not None:
                nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def forward(self, input_ids: torch.Tensor, targets: Optional[torch.Tensor] = None):
        x = self.token_embedding(input_ids) * math.sqrt(self.config.hidden_size)
        x = self.embedding_dropout(x)
        for block in self.blocks:
            x = block(x)
        x = self.final_norm(x)
        logits = self.lm_head(x)
        return logits

def top_k_top_p_filtering(logits: torch.Tensor, top_k: int = 0, top_p: float = 0.0, filter_value: float = -float('Inf')) -> torch.Tensor:
    """
    Filter a distribution of logits using top-k and/or nucleus (top-p) filtering.
    This is taken from common implementations (Hugging Face transformers style).
    Args:
        logits: logits distribution shape (batch, vocab)
        top_k: keep only top k tokens with highest probability (0 = no top-k)
        top_p: keep the top tokens with cumulative probability >= top_p (0.0 = no nucleus)
        filter_value: value to set for filtered logits
    Returns:
        filtered logits with the same shape
    """
    top_k = max(top_k, 0)
    batch_size, vocab_size = logits.size()

    if top_k > 0:
        # Remove all tokens with a probability less than the top-k tokens
        top_k = min(max(top_k, 1), vocab_size)
        values_to_keep, _ = torch.topk(logits, top_k)
        min_values = values_to_keep[:, -1].unsqueeze(1).expand_as(logits)
        logits = torch.where(logits < min_values, torch.full_like(logits, filter_value), logits)

    if top_p > 0.0:
        sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
        sorted_probs = F.softmax(sorted_logits, dim=-1)
        cumulative_probs = torch.cumsum(sorted_probs, dim=-1)

        # Remove tokens with cumulative probability above the threshold
        sorted_mask = cumulative_probs > top_p

        # Shift the mask right to keep at least one token
        sorted_mask[..., 1:] = sorted_mask[..., :-1].clone()
        sorted_mask[..., 0] = False

        indices_to_remove = sorted_mask.scatter(1, sorted_indices, sorted_mask)
        logits = logits.masked_fill(indices_to_remove, filter_value)

    return logits


@torch.no_grad()
def generate(
    model: Transformer,
    input_ids: torch.LongTensor,
    max_new_tokens: int = 50,
    temperature: float = 1.0,
    top_k: int = 0,
    top_p: float = 0.0,
    do_sample: bool = True,
    eos_token_id: Optional[int] = None,
    pad_token_id: Optional[int] = None,
    device: Optional[torch.device] = None,
):
    """
    Autoregressive generation helper for the model. This implementation does NOT use KV cache
    (the model defined in this file does not implement a cache), so generation is performed
    by repeatedly calling the model on the growing sequence. It supports temperature,
    top-k and nucleus (top-p) sampling, greedy decoding, and optional early stopping
    on an `eos_token_id`.

    Args:
        model: the Transformer instance
        input_ids: (batch, seq_len) input token ids
        max_new_tokens: number of tokens to generate
        temperature: sampling temperature (<=0 or do_sample=False => greedy)
        top_k: top-k filtering (0 disables)
        top_p: nucleus/top-p filtering (0.0 disables)
        do_sample: whether to sample (True) or do greedy decoding (False)
        eos_token_id: optional EOS id to stop generation for individual sequences
        pad_token_id: optional pad id to use for finished sequences
        device: optional torch.device to run on; if None uses model's device

    Returns:
        tensor of shape (batch, seq_len + generated) with generated tokens appended
    """
    model.eval()
    if device is None:
        # try to infer device
        try:
            device = next(model.parameters()).device
        except StopIteration:
            device = torch.device('cpu')

    input_ids = input_ids.to(device)

    batch_size, seq_len = input_ids.shape
    generated = 0
    unfinished = torch.ones(batch_size, dtype=torch.bool, device=device)

    for _ in range(max_new_tokens):
        logits = model(input_ids)
        # logits shape: (batch, seq_len_total, vocab)
        next_token_logits = logits[:, -1, :]

        if temperature <= 0 or not do_sample:
            # Greedy
            next_tokens = torch.argmax(next_token_logits, dim=-1)
        else:
            logits_proc = next_token_logits / max(temperature, 1e-8)
            logits_proc = top_k_top_p_filtering(logits_proc, top_k=top_k, top_p=top_p)
            probs = F.softmax(logits_proc, dim=-1)
            next_tokens = torch.multinomial(probs, num_samples=1).squeeze(-1)

        # If EOS is provided, update finished sequences and pad further tokens
        if eos_token_id is not None:
            is_eos = next_tokens.eq(eos_token_id)
            # sequences that have just finished
            just_finished = unfinished & is_eos
            unfinished = unfinished & (~is_eos)

        # For sequences already finished, append pad_token_id (if provided), otherwise keep EOS or sampled token
        if pad_token_id is not None and not unfinished.all():
            finished_mask = ~unfinished
            if finished_mask.any():
                next_tokens = next_tokens.masked_fill(finished_mask, pad_token_id)

        # append
        input_ids = torch.cat([input_ids, next_tokens.unsqueeze(-1)], dim=1)
        generated += 1

        if eos_token_id is not None and not unfinished.any():
            break

    return input_ids


def _smoke_test():
    config = ModelConfig(
        vocab_size=128,
        hidden_size=64,
        n_heads=4,
        n_kv_heads=4,
        n_kv_groups=None,
        head_dim=16,
        n_layers=2,
        attention_bias=False,
        intermediate_size=256,
        mlp_bias=False,
        eps=1e-5,
    )
    model = Transformer(config)
    model.eval()

    batch, seq_len = 2, 8
    input_ids = torch.randint(0, config.vocab_size, (batch, seq_len))
    logits, loss = model(input_ids, targets=input_ids)

    assert logits.shape == (batch, seq_len, config.vocab_size)
    assert loss.dim() == 0
    print("Smoke test passed: logits shape", logits.shape, "loss", loss.detach().item())


if __name__ == "__main__":
    _smoke_test()