nemotron.py 17.7 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# Copyright 2023 The vLLM team.
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only Nemotron model compatible with HuggingFace weights."""
26

27
from collections.abc import Iterable
28
from itertools import islice
29
30
31
32

import torch
from torch import nn

33
from vllm.attention.layer import Attention
34
from vllm.compilation.decorators import support_torch_compile
35
from vllm.config import CacheConfig, VllmConfig
36
37
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
38
39
40
41
42
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
43
from vllm.model_executor.layers.logits_processor import LogitsProcessor
44
from vllm.model_executor.layers.quantization import QuantizationConfig
45
46
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
47
48
49
    ParallelLMHead,
    VocabParallelEmbedding,
)
50
from vllm.model_executor.model_loader.weight_utils import (
51
52
53
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
54
from vllm.sequence import IntermediateTensors
55
56
from vllm.transformers_utils.configs import NemotronConfig

57
from .interfaces import SupportsLoRA, SupportsPP
58
59
60
61
62
63
64
65
from .utils import (
    AutoWeightsLoader,
    PPMissingLayer,
    is_pp_missing_parameter,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
66
67
68
69
70

# The architecture is pretty similar to Llama, with these changes:
# - There is no gate_proj, just up_proj
# - Normal LayerNorm (with a +1 to the weights) instead of RMSNorm
# - Squared ReLU instead of SwiGLU
71
# - Adds a partial_rotary_factor to RoPE
72
73
74
75
76
77


def _cast_if_autocast_enabled(*args):
    if not torch.is_autocast_enabled():
        return args
    else:
cyyever's avatar
cyyever committed
78
        return torch.amp.autocast_mode._cast(
79
80
            args, device_type="cuda", dtype=torch.get_autocast_gpu_dtype()
        )
81
82
83


class NemotronLayerNorm1P(nn.LayerNorm):
84
85
    def __init__(
        self,
86
        normalized_shape: int | list[int] | torch.Size,
87
88
89
90
91
92
93
        eps: float = 1e-5,
        elementwise_affine: bool = True,
        bias: bool = True,
        device=None,
        dtype=None,
    ):
        super().__init__(normalized_shape, eps, elementwise_affine, bias, device, dtype)
94
95
96
97

    def forward(
        self,
        x: torch.Tensor,
98
        residual: torch.Tensor | None = None,
99
100
101
102
    ) -> torch.Tensor:
        if residual is not None:
            x = x + residual
            residual = x
103
104
105
        args = _cast_if_autocast_enabled(
            x, self.normalized_shape, self.weight + 1, self.bias, self.eps
        )
cyyever's avatar
cyyever committed
106
        with torch.amp.autocast("cuda", enabled=False):
107
108
109
110
111
112
113
114
115
116
            x = torch.nn.functional.layer_norm(*args)
            return x if residual is None else (x, residual)


class NemotronMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
117
        quant_config: QuantizationConfig | None = None,
118
119
120
121
        bias: bool = False,
        prefix: str = "",
    ) -> None:
        super().__init__()
122
123
124
125
126
127
128
129
130
131
132
133
134
135
        self.up_proj = ColumnParallelLinear(
            input_size=hidden_size,
            output_size=intermediate_size,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.up_proj",
        )
        self.down_proj = RowParallelLinear(
            input_size=intermediate_size,
            output_size=hidden_size,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.down_proj",
        )
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
        self.act_fn = get_act_fn(hidden_act)

    def forward(self, x):
        up, _ = self.up_proj(x)
        x = self.act_fn(up)
        x, _ = self.down_proj(x)
        return x


class NemotronAttention(nn.Module):
    def __init__(
        self,
        config: NemotronConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position_embeddings: int = 8192,
153
        quant_config: QuantizationConfig | None = None,
154
        bias: bool = False,
155
        cache_config: CacheConfig | None = None,
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
175
176
177
        self.head_dim = getattr(config, "head_dim", None)
        if self.head_dim is None:
            self.head_dim = self.hidden_size // self.total_num_heads
178
179
180
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
181
        self.partial_rotary_factor = config.partial_rotary_factor
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
        self.max_position_embeddings = max_position_embeddings

        self.qkv_proj = QKVParallelLinear(
            hidden_size=hidden_size,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
        self.o_proj = RowParallelLinear(
            input_size=self.total_num_heads * self.head_dim,
            output_size=hidden_size,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
205
            rope_parameters=config.rope_parameters,
206
            partial_rotary_factor=self.partial_rotary_factor,
207
        )
208
209
210
211
212
213
214
215
216
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )
217
218
219
220
221
222
223
224
225

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)
226
        attn_output = self.attn(q, k, v)
227
228
229
230
231
232
233
234
        output, _ = self.o_proj(attn_output)
        return output


class NemotronDecoderLayer(nn.Module):
    def __init__(
        self,
        config: NemotronConfig,
235
236
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
237
238
239
240
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
241
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
242
243
244
        # Support abacusai/Smaug-72B-v0.1 with attention_bias
        # Support internlm/internlm-7b with bias
        attention_bias = getattr(config, "attention_bias", False) or getattr(
245
246
            config, "bias", False
        )
247
248
249
250
        self.self_attn = NemotronAttention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
251
252
253
            num_kv_heads=getattr(
                config, "num_key_value_heads", config.num_attention_heads
            ),
254
255
256
257
258
259
260
261
262
263
264
265
266
267
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            bias=attention_bias,
            cache_config=cache_config,
            prefix=f"{prefix}.self_attn",
        )
        self.mlp = NemotronMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            bias=getattr(config, "mlp_bias", False),
            prefix=f"{prefix}.mlp",
        )
268
269
270
        self.input_layernorm = NemotronLayerNorm1P(
            config.hidden_size, eps=config.norm_eps
        )
271
        self.post_attention_layernorm = NemotronLayerNorm1P(
272
273
            config.hidden_size, eps=config.norm_eps
        )
274
275
276
277
278

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
279
        residual: torch.Tensor | None,
280
    ) -> tuple[torch.Tensor, torch.Tensor]:
281
282
283
284
285
        # Self Attention
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
286
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
287
288
289
290
291
292
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )

        # Fully Connected
293
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
294
295
296
297
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual


298
@support_torch_compile
299
class NemotronModel(nn.Module):
300
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
301
        super().__init__()
302
303
304
305
306

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

307
        self.config = config
308
        self.quant_config = quant_config
309
310
311

        self.vocab_size = config.vocab_size

312
313
314
        if get_pp_group().is_first_rank or (
            config.tie_word_embeddings and get_pp_group().is_last_rank
        ):
315
316
317
318
319
320
321
322
            self.embed_tokens = VocabParallelEmbedding(
                self.vocab_size,
                config.hidden_size,
            )
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
323
324
325
326
327
328
329
330
            lambda prefix: NemotronDecoderLayer(
                config=config,
                cache_config=cache_config,
                quant_config=quant_config,
                prefix=prefix,
            ),
            prefix=f"{prefix}.layers",
        )
331
        if get_pp_group().is_last_rank:
332
            self.norm = NemotronLayerNorm1P(config.hidden_size, eps=config.norm_eps)
333
334
        else:
            self.norm = PPMissingLayer()
335
336
337
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
338

339
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
340
341
342
343
        return self.embed_tokens(input_ids)

    def forward(
        self,
344
        input_ids: torch.Tensor | None,
345
        positions: torch.Tensor,
346
347
348
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
349
350
351
352
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
353
                hidden_states = self.embed_input_ids(input_ids)
354
355
356
357
358
359
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

360
        for layer in islice(self.layers, self.start_layer, self.end_layer):
361
            hidden_states, residual = layer(positions, hidden_states, residual)
362
363

        if not get_pp_group().is_last_rank:
364
365
366
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
367
368
369
370

        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states

371
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
372
373
374
375
376
377
378
379
380
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
381
382
383
            if self.quant_config is not None and (
                scale_name := self.quant_config.get_cache_scale(name)
            ):
384
385
                # Loading kv cache quantization scales
                param = params_dict[scale_name]
386
387
388
389
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                loaded_weight = (
                    loaded_weight if loaded_weight.dim() == 0 else loaded_weight[0]
                )
390
391
392
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue
393
            for param_name, weight_name, shard_id in stacked_params_mapping:
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
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue

                if is_pp_missing_parameter(name, self):
                    continue

                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)

                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue

                if is_pp_missing_parameter(name, self):
                    continue

                param = params_dict[name]
422
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
423
424
425
426
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

427

428
class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
429
430
431
432
433
434
435
436
437
438
439
440
441
442
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }

    # LoRA specific attributes
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }

443
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
444
        super().__init__()
445
446
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
447

448
449
450
        assert isinstance(config, NemotronConfig)

        self.config = config
451

452
        self.quant_config = quant_config
453

454
455
456
        self.model = NemotronModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
457
458
        if get_pp_group().is_last_rank:
            self.lm_head = ParallelLMHead(
459
                config.vocab_size,
460
461
                config.hidden_size,
                quant_config=quant_config,
462
                prefix=maybe_prefix(prefix, "lm_head"),
463
464
465
466
467
            )
            if config.tie_word_embeddings:
                self.lm_head.weight = self.model.embed_tokens.weight

            logit_scale = getattr(config, "logit_scale", 1.0)
468
            self.logits_processor = LogitsProcessor(
469
                config.vocab_size, scale=logit_scale
470
            )
471
472
        else:
            self.lm_head = PPMissingLayer()
473

474
        self.make_empty_intermediate_tensors = (
475
476
            self.model.make_empty_intermediate_tensors
        )
477

478
479
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
480

481
482
483
484
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
485
486
487
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
488
489
490
        model_output = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
491
492
        return model_output

493
494
495
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
496
    ) -> torch.Tensor | None:
497
        logits = self.logits_processor(self.lm_head, hidden_states)
498
499
        return logits

500
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
501
        loader = AutoWeightsLoader(self)
502
        return loader.load_weights(weights)