File size: 17,478 Bytes
d3ab6f5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Capability-first Vortex 175M decoder.

This file intentionally uses only standard PyTorch CUDA primitives.  The
model is deep-and-thin, uses GQA and SwiGLU, and ties the input/output
embedding.  Those choices are much easier to train and export than a custom
SSM kernel while retaining the main sub-billion-parameter wins reported by
MobileLLM-style studies.
"""

from __future__ import annotations

import math
from dataclasses import asdict, dataclass

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


@dataclass
class VortexConfig:
    vocab_size: int = 8_192
    max_seq_len: int = 4_096
    n_layer: int = 12
    n_embd: int = 1024
    n_head: int = 16
    n_kv_head: int = 4
    head_dim: int = 64
    intermediate_size: int = 3_664
    rope_theta: float = 100_000.0
    norm_eps: float = 1e-5
    logits_chunk_tokens: int = 16_384
    gradient_checkpointing: bool = False
    use_transformer_engine: bool = False
    attn_input_format: str = "bshd"

    def to_dict(self) -> dict:
        return asdict(self)


class RMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float) -> None:
        super().__init__()
        self.weight = nn.Parameter(torch.ones(dim))
        self.eps = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return F.rms_norm(x, (x.shape[-1],), self.weight, self.eps)


def _linear(config: VortexConfig, in_features: int, out_features: int) -> nn.Module:
    """Create a bias-free projection, optionally backed by Transformer Engine."""
    if not config.use_transformer_engine:
        return nn.Linear(in_features, out_features, bias=False)
    try:
        import transformer_engine.pytorch as te
    except ImportError as exc:  # pragma: no cover - exercised only on TE runs
        raise RuntimeError(
            "use_transformer_engine=True requires transformer-engine[pytorch]"
        ) from exc
    return te.Linear(
        in_features,
        out_features,
        bias=False,
        params_dtype=torch.bfloat16,
        device="cuda",
    )


class RotaryEmbedding(nn.Module):
    def __init__(self, dim: int, max_seq_len: int, theta: float) -> None:
        super().__init__()
        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
        positions = torch.arange(max_seq_len, dtype=torch.float32)
        frequencies = torch.outer(positions, inv_freq)
        # NeoX-style rotate-half layout: the first and second halves share
        # the same frequencies, so rotate_half() remains allocation-free in
        # the hot attention path apart from its concatenation.
        angles = torch.cat((frequencies, frequencies), dim=-1)
        self.register_buffer("cos_cached", angles.cos()[None, None], persistent=False)
        self.register_buffer("sin_cached", angles.sin()[None, None], persistent=False)

    @staticmethod
    def rotate_half(x: torch.Tensor) -> torch.Tensor:
        half = x.shape[-1] // 2
        return torch.cat((-x[..., half:], x[..., :half]), dim=-1)

    def forward(self, q: torch.Tensor, k: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        seq_len = q.shape[-2]
        if seq_len > self.cos_cached.shape[-2]:
            raise ValueError(f"sequence length {seq_len} exceeds configured maximum")
        cos = self.cos_cached[:, :, :seq_len].to(dtype=q.dtype)
        sin = self.sin_cached[:, :, :seq_len].to(dtype=q.dtype)
        return (
            q * cos + self.rotate_half(q) * sin,
            k * cos + self.rotate_half(k) * sin,
        )


class GQAAttention(nn.Module):
    def __init__(self, config: VortexConfig) -> None:
        super().__init__()
        if config.n_head % config.n_kv_head:
            raise ValueError("n_head must be divisible by n_kv_head")
        if config.n_head * config.head_dim != config.n_embd:
            raise ValueError("n_head * head_dim must equal n_embd")
        self.n_head = config.n_head
        self.n_kv_head = config.n_kv_head
        self.head_dim = config.head_dim
        kv_dim = config.n_kv_head * config.head_dim
        self.q_proj = _linear(config, config.n_embd, config.n_embd)
        self.k_proj = _linear(config, config.n_embd, kv_dim)
        self.v_proj = _linear(config, config.n_embd, kv_dim)
        self.o_proj = _linear(config, config.n_embd, config.n_embd)
        # QK-Norm keeps attention logits well-conditioned at the deliberately
        # high pretraining learning rate.  These are per-head, parameter-light
        # norms, not full hidden-size projections.
        self.q_norm = RMSNorm(config.head_dim, config.norm_eps)
        self.k_norm = RMSNorm(config.head_dim, config.norm_eps)
        self.rope = RotaryEmbedding(config.head_dim, config.max_seq_len, config.rope_theta)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        batch, seq_len, _ = x.shape
        q = self.q_proj(x).view(batch, seq_len, self.n_head, self.head_dim).transpose(1, 2)
        k = self.k_proj(x).view(batch, seq_len, self.n_kv_head, self.head_dim).transpose(1, 2)
        v = self.v_proj(x).view(batch, seq_len, self.n_kv_head, self.head_dim).transpose(1, 2)
        q, k = self.rope(self.q_norm(q), self.k_norm(k))
        # PyTorch dispatches this to the fused flash/efficient causal kernel
        # when the local CUDA build supports it.  enable_gqa avoids material-
        # izing repeated K/V heads.
        y = F.scaled_dot_product_attention(
            q, k, v, is_causal=True, enable_gqa=True
        )
        y = y.transpose(1, 2).contiguous().view(batch, seq_len, -1)
        return self.o_proj(y)


class SwiGLU(nn.Module):
    def __init__(self, config: VortexConfig) -> None:
        super().__init__()
        self.gate_proj = _linear(config, config.n_embd, config.intermediate_size)
        self.up_proj = _linear(config, config.n_embd, config.intermediate_size)
        self.down_proj = _linear(config, config.intermediate_size, config.n_embd)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))


class VortexBlock(nn.Module):
    def __init__(self, config: VortexConfig, layer_number: int | None = None) -> None:
        super().__init__()
        self.use_transformer_engine = config.use_transformer_engine
        self.attn_input_format = config.attn_input_format
        if self.use_transformer_engine:
            try:
                import transformer_engine.pytorch as te
            except ImportError as exc:  # pragma: no cover - TE-only path
                raise RuntimeError(
                    "use_transformer_engine=True requires transformer-engine[pytorch]"
                ) from exc
            # This fused layer has the same parameter shapes as the explicit
            # reference block: fused GQA QKV, RMSNorm, QK-Norm, SwiGLU, and
            # the causal attention kernel.  It is used only for the TE backend;
            # the reference PyTorch path remains readable and exportable.
            self.te_layer = te.TransformerLayer(
                hidden_size=config.n_embd,
                ffn_hidden_size=config.intermediate_size,
                num_attention_heads=config.n_head,
                num_gqa_groups=config.n_kv_head,
                layernorm_epsilon=config.norm_eps,
                hidden_dropout=0.0,
                attention_dropout=0.0,
                kv_channels=config.head_dim,
                layer_number=layer_number,
                bias=False,
                activation="swiglu",
                normalization="RMSNorm",
                qk_norm_type="RMSNorm",
                qk_norm_before_rope=True,
                fuse_qkv_params=True,
                self_attn_mask_type="causal",
                attn_input_format=config.attn_input_format,
                params_dtype=torch.bfloat16,
                device="cuda",
            )
            return
        self.norm1 = RMSNorm(config.n_embd, config.norm_eps)
        self.attn = GQAAttention(config)
        self.norm2 = RMSNorm(config.n_embd, config.norm_eps)
        self.ffn = SwiGLU(config)

    def forward(
        self,
        x: torch.Tensor,
        rotary_pos_emb: torch.Tensor | None = None,
        is_first_microbatch: bool | None = None,
        inference_params=None,
        attention_mask: torch.Tensor | None = None,
        inference_decode_bshd: bool = False,
    ) -> torch.Tensor:
        if self.use_transformer_engine:
            if inference_params is not None:
                # A packed THD prompt handles variable-length prefill
                # efficiently.  Subsequent one-token decode is cheaper and
                # more numerically stable through the regular BSHD cache
                # path; switch the TE attention format explicitly between the
                # two phases.
                effective_format = "bshd" if inference_decode_bshd else self.attn_input_format
                self.te_layer.self_attention.qkv_format = effective_format
            te_kwargs = {
                "attention_mask": attention_mask,
                "self_attn_mask_type": (
                    "padding_causal" if inference_params is not None else None
                ),
                "rotary_pos_emb": rotary_pos_emb,
                "is_first_microbatch": is_first_microbatch,
                "inference_params": inference_params,
            }
            if inference_params is not None and not inference_decode_bshd and self.attn_input_format == "thd":
                batch_size = len(inference_params.sequences)
                cu_seqlens = inference_params.cu_seqlens_q[: batch_size + 1]
                sequence_lengths = cu_seqlens[1:] - cu_seqlens[:-1]
                te_kwargs.update(
                    {
                        "cu_seqlens_q": cu_seqlens,
                        "cu_seqlens_q_padded": cu_seqlens,
                        "max_seqlen_q": int(sequence_lengths.max().item()),
                        "max_seqlen_kv": int(sequence_lengths.max().item()),
                    }
                )
            return self.te_layer(
                x,
                **te_kwargs,
            )
        x = x + self.attn(self.norm1(x))
        x = x + self.ffn(self.norm2(x))
        return x


class VortexForCausalLM(nn.Module):
    def __init__(self, config: VortexConfig | None = None) -> None:
        super().__init__()
        self.config = config or VortexConfig()
        self.embed_tokens = nn.Embedding(self.config.vocab_size, self.config.n_embd)
        self.layers = nn.ModuleList(
            VortexBlock(self.config, layer_number=index + 1)
            for index in range(self.config.n_layer)
        )
        self.norm = RMSNorm(self.config.n_embd, self.config.norm_eps)
        if self.config.use_transformer_engine:
            import transformer_engine.pytorch as te

            self.rotary = te.RotaryPositionEmbedding(
                self.config.head_dim,
                rotary_base=self.config.rope_theta,
                interleaved=False,
            )
        self._initialize_weights()

    def _initialize_weights(self) -> None:
        # Scale residual outputs down with depth; this gives a forgiving high-
        # LR start without adding trainable parameters.
        output_std = 0.02 / math.sqrt(2.0 * self.config.n_layer)
        nn.init.normal_(self.embed_tokens.weight, mean=0.0, std=0.02)
        for block in self.layers:
            if self.config.use_transformer_engine:
                for name, parameter in block.named_parameters():
                    if parameter.ndim == 1:
                        nn.init.ones_(parameter)
                    else:
                        is_output = (
                            name.endswith("self_attention.proj.weight")
                            or name.endswith("layernorm_mlp.fc2_weight")
                        )
                        nn.init.normal_(
                            parameter,
                            mean=0.0,
                            std=output_std if is_output else 0.02,
                        )
            else:
                for child in block.modules():
                    if isinstance(child, nn.Linear):
                        is_output = child is block.attn.o_proj or child is block.ffn.down_proj
                        nn.init.normal_(
                            child.weight,
                            mean=0.0,
                            std=output_std if is_output else 0.02,
                        )

    def parameter_count(self) -> int:
        return sum(parameter.numel() for parameter in self.parameters())

    def parameter_breakdown(self) -> dict[str, int]:
        c = self.config
        embedding = c.vocab_size * c.n_embd
        q = c.n_layer * c.n_embd * c.n_embd
        k = c.n_layer * c.n_embd * (c.n_kv_head * c.head_dim)
        v = k
        o = q
        # One learned head-dimension scale is shared across all query heads,
        # and another across all KV heads, matching the module definitions
        # above (not one scale vector per physical head).
        qk_norm = c.n_layer * 2 * c.head_dim
        ffn = c.n_layer * 3 * c.n_embd * c.intermediate_size
        block_norm = c.n_layer * 2 * c.n_embd
        final_norm = c.n_embd
        return {
            "input_embedding_and_tied_output": embedding,
            "attention_q_projection": q,
            "attention_k_projection": k,
            "attention_v_projection": v,
            "attention_o_projection": o,
            "attention_qk_norm": qk_norm,
            "ffn_swiglu": ffn,
            "block_rmsnorm": block_norm,
            "final_rmsnorm": final_norm,
            "total": self.parameter_count(),
        }

    def _chunked_tied_loss(self, hidden: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
        chunk = self.config.logits_chunk_tokens
        total = hidden.new_zeros((), dtype=torch.float32)
        for start in range(0, hidden.shape[0], chunk):
            end = min(hidden.shape[0], start + chunk)
            logits = F.linear(hidden[start:end], self.embed_tokens.weight)
            total = total + F.cross_entropy(logits, targets[start:end]).float() * (end - start)
        return total / hidden.shape[0]

    def forward(
        self,
        input_ids: torch.Tensor,
        labels: torch.Tensor | None = None,
        is_first_microbatch: bool | None = None,
        inference_params=None,
        inference_attention_mask=None,
        inference_decode_bshd: bool = False,
    ) -> tuple[torch.Tensor | None, torch.Tensor | None]:
        x = self.embed_tokens(input_ids)
        rotary_pos_emb = None
        attention_mask = None
        if self.config.use_transformer_engine:
            if inference_params is not None:
                # TE's cached-attention path applies the correct absolute
                # offset from inference_params.  Supplying the full table is
                # necessary when the current query is only one token long but
                # starts after a long cached prefix.
                rotary_pos_emb = self.rotary(self.config.max_seq_len)
                if self.config.attn_input_format == "thd" and not inference_decode_bshd:
                    attention_mask = None
                elif inference_attention_mask is None:
                    query_padding_mask = torch.zeros(
                        (x.shape[0], 1, 1, x.shape[1]),
                        dtype=torch.bool,
                        device=x.device,
                    )
                    key_padding_mask = torch.ones(
                        (x.shape[0], 1, 1, inference_params.max_sequence_length),
                        dtype=torch.bool,
                        device=x.device,
                    )
                    for batch_index, sequence_length in enumerate(
                        inference_params.sequences.values()
                    ):
                        key_padding_mask[batch_index, :, :, :sequence_length] = False
                    # TE switches the cached self-attention implementation to
                    # its cross-attention backend internally; that backend
                    # expects the query and key padding masks as a pair.
                    attention_mask = (query_padding_mask, key_padding_mask)
                else:
                    attention_mask = inference_attention_mask
            else:
                rotary_pos_emb = self.rotary(x.shape[1])
        for block in self.layers:
            if self.training and self.config.gradient_checkpointing:
                x = checkpoint(
                    block,
                    x,
                    rotary_pos_emb,
                    is_first_microbatch,
                    use_reentrant=False,
                )
            else:
                x = block(
                    x,
                    rotary_pos_emb,
                    is_first_microbatch,
                    inference_params,
                    attention_mask,
                    inference_decode_bshd,
                )
        x = self.norm(x)
        if labels is None:
            return F.linear(x, self.embed_tokens.weight), None
        hidden = x[:, :-1].reshape(-1, x.shape[-1])
        targets = labels[:, 1:].reshape(-1)
        return None, self._chunked_tied_loss(hidden, targets)


if __name__ == "__main__":
    config = VortexConfig()
    model = VortexForCausalLM(config)
    print(config.to_dict())
    print(model.parameter_breakdown())