granite.py 18.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

Shawn Tan's avatar
Shawn Tan committed
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 IBM Granite model compatible with HuggingFace weights."""
26

27
from collections.abc import Iterable
28
from itertools import islice
Shawn Tan's avatar
Shawn Tan committed
29
30
31

import torch
from torch import nn
32
from transformers import GraniteConfig
Shawn Tan's avatar
Shawn Tan committed
33

34
from vllm.attention import Attention
35
from vllm.compilation.decorators import support_torch_compile
36
from vllm.config import CacheConfig, VllmConfig
37
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
Shawn Tan's avatar
Shawn Tan committed
38
39
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
40
41
42
43
44
from vllm.model_executor.layers.linear import (
    MergedColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
Shawn Tan's avatar
Shawn Tan committed
45
from vllm.model_executor.layers.logits_processor import LogitsProcessor
46
from vllm.model_executor.layers.quantization import QuantizationConfig
Shawn Tan's avatar
Shawn Tan committed
47
48
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
49
50
51
52
    DEFAULT_VOCAB_PADDING_SIZE,
    ParallelLMHead,
    VocabParallelEmbedding,
)
Shawn Tan's avatar
Shawn Tan committed
53
from vllm.model_executor.model_loader.weight_utils import (
54
55
56
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
Shawn Tan's avatar
Shawn Tan committed
57
58
from vllm.sequence import IntermediateTensors

59
from .interfaces import SupportsLoRA, SupportsPP
60
61
62
63
64
65
66
from .utils import (
    AutoWeightsLoader,
    PPMissingLayer,
    is_pp_missing_parameter,
    make_layers,
    maybe_prefix,
)
Shawn Tan's avatar
Shawn Tan committed
67
68
69
70
71
72
73
74


class GraniteMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
75
        quant_config: QuantizationConfig | None = None,
Shawn Tan's avatar
Shawn Tan committed
76
77
78
79
80
81
82
83
84
        bias: bool = False,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            input_size=hidden_size,
            output_sizes=[intermediate_size] * 2,
            bias=bias,
            quant_config=quant_config,
85
86
87
88
89
90
91
92
93
            prefix=f"{prefix}.gate_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",
        )
Shawn Tan's avatar
Shawn Tan committed
94
        if hidden_act != "silu":
95
96
97
            raise ValueError(
                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
            )
Shawn Tan's avatar
Shawn Tan committed
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
        self.act_fn = SiluAndMul()

    def forward(self, x):
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class GraniteAttention(nn.Module):
    def __init__(
        self,
        config: GraniteConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position_embeddings: int = 8192,
115
        quant_config: QuantizationConfig | None = None,
Shawn Tan's avatar
Shawn Tan committed
116
        bias: bool = False,
117
        cache_config: CacheConfig | None = None,
Shawn Tan's avatar
Shawn Tan committed
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
        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
137
138
139
        self.head_dim = getattr(config, "head_dim", None)
        if self.head_dim is None:
            self.head_dim = self.hidden_size // self.total_num_heads
Shawn Tan's avatar
Shawn Tan committed
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
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = config.attention_multiplier
        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,
166
            rope_parameters=config.rope_parameters,
Shawn Tan's avatar
Shawn Tan committed
167
        )
168
169
170
171
172
173
174
175
176
        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",
        )
Shawn Tan's avatar
Shawn Tan committed
177
178
179
180
181
182
183
184
185

    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)
186
        attn_output = self.attn(q, k, v)
Shawn Tan's avatar
Shawn Tan committed
187
188
189
190
191
192
193
194
        output, _ = self.o_proj(attn_output)
        return output


class GraniteDecoderLayer(nn.Module):
    def __init__(
        self,
        config: GraniteConfig,
195
196
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
Shawn Tan's avatar
Shawn Tan committed
197
198
199
200
201
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        self.residual_multiplier = config.residual_multiplier
202
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
Shawn Tan's avatar
Shawn Tan committed
203
204
205
        # Support abacusai/Smaug-72B-v0.1 with attention_bias
        # Support internlm/internlm-7b with bias
        attention_bias = getattr(config, "attention_bias", False) or getattr(
206
207
            config, "bias", False
        )
Shawn Tan's avatar
Shawn Tan committed
208
209
210
211
        self.self_attn = GraniteAttention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
212
213
214
            num_kv_heads=getattr(
                config, "num_key_value_heads", config.num_attention_heads
            ),
Shawn Tan's avatar
Shawn Tan committed
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            bias=attention_bias,
            cache_config=cache_config,
            prefix=f"{prefix}.self_attn",
        )

        self.mlp = GraniteMLP(
            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",
        )
230
231
232
233
        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )
Shawn Tan's avatar
Shawn Tan committed
234
235
236
237
238

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
239
    ) -> tuple[torch.Tensor, torch.Tensor]:
Shawn Tan's avatar
Shawn Tan committed
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
        # Self Attention
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )
        hidden_states = residual + hidden_states * self.residual_multiplier
        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states * self.residual_multiplier
        return hidden_states


256
@support_torch_compile
Shawn Tan's avatar
Shawn Tan committed
257
class GraniteModel(nn.Module):
258
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
Shawn Tan's avatar
Shawn Tan committed
259
        super().__init__()
260
261
262
263
264
265

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

Shawn Tan's avatar
Shawn Tan committed
266
        self.config = config
267
        self.quant_config = quant_config
268
269
270
271
272
        lora_vocab = (
            (lora_config.lora_extra_vocab_size * (lora_config.max_loras or 1))
            if lora_config
            else 0
        )
Shawn Tan's avatar
Shawn Tan committed
273
274
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size
275
276
277
        if get_pp_group().is_first_rank or (
            config.tie_word_embeddings and get_pp_group().is_last_rank
        ):
Shawn Tan's avatar
Shawn Tan committed
278
279
280
281
            self.embed_tokens = VocabParallelEmbedding(
                self.vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
282
283
284
                padding_size=DEFAULT_VOCAB_PADDING_SIZE
                # We need bigger padding if using lora for kernel
                # compatibility
285
286
                if not lora_config
                else lora_config.lora_vocab_padding_size,
Shawn Tan's avatar
Shawn Tan committed
287
288
289
290
291
292
                quant_config=quant_config,
            )
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
293
294
295
296
297
298
299
300
            lambda prefix: GraniteDecoderLayer(
                config=config,
                cache_config=cache_config,
                quant_config=quant_config,
                prefix=prefix,
            ),
            prefix=f"{prefix}.layers",
        )
Shawn Tan's avatar
Shawn Tan committed
301
302
303
304
305
        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()

306
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
Shawn Tan's avatar
Shawn Tan committed
307
308
309
310
        return self.embed_tokens(input_ids)

    def forward(
        self,
311
        input_ids: torch.Tensor | None,
Shawn Tan's avatar
Shawn Tan committed
312
        positions: torch.Tensor,
313
314
315
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
Shawn Tan's avatar
Shawn Tan committed
316
317
318
319
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
320
                hidden_states = self.embed_input_ids(input_ids)
321
322

            hidden_states *= self.config.embedding_multiplier
Shawn Tan's avatar
Shawn Tan committed
323
324
325
326
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]

327
        for layer in islice(self.layers, self.start_layer, self.end_layer):
328
            hidden_states = layer(positions, hidden_states)
Shawn Tan's avatar
Shawn Tan committed
329
330

        if not get_pp_group().is_last_rank:
331
332
333
334
335
            return IntermediateTensors(
                {
                    "hidden_states": hidden_states,
                }
            )
Shawn Tan's avatar
Shawn Tan committed
336
337
338
339

        hidden_states = self.norm(hidden_states)
        return hidden_states

340
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
341
342
343
344
345
346
347
348
349
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]
        params_dict = dict(self.named_parameters())
350
        loaded_params: set[str] = set()
351
        for name, loaded_weight in weights:
352
353
354
            if self.quant_config is not None and (
                scale_name := self.quant_config.get_cache_scale(name)
            ):
355
356
                # Loading kv cache quantization scales
                param = params_dict[scale_name]
357
358
359
360
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                loaded_weight = (
                    loaded_weight if loaded_weight.dim() == 0 else loaded_weight[0]
                )
361
362
363
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue
364
            for param_name, weight_name, shard_id in stacked_params_mapping:
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
                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]
393
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
394
395
396
397
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

Shawn Tan's avatar
Shawn Tan committed
398

399
class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
Shawn Tan's avatar
Shawn Tan committed
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

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

419
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
Shawn Tan's avatar
Shawn Tan committed
420
        super().__init__()
421
422
423
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
Shawn Tan's avatar
Shawn Tan committed
424
425
426

        self.config = config
        self.lora_config = lora_config
427
        self.quant_config = quant_config
Shawn Tan's avatar
Shawn Tan committed
428

429
430
431
        self.model = GraniteModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
Shawn Tan's avatar
Shawn Tan committed
432
433
434
435
436
437
438
439
440
441
442
        if get_pp_group().is_last_rank:
            self.unpadded_vocab_size = config.vocab_size
            if lora_config:
                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
            self.lm_head = ParallelLMHead(
                self.unpadded_vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
                padding_size=DEFAULT_VOCAB_PADDING_SIZE
                # We need bigger padding if using lora for kernel
                # compatibility
443
444
                if not lora_config
                else lora_config.lora_vocab_padding_size,
Shawn Tan's avatar
Shawn Tan committed
445
                quant_config=quant_config,
446
                prefix=maybe_prefix(prefix, "lm_head"),
Shawn Tan's avatar
Shawn Tan committed
447
448
449
450
451
            )
            if config.tie_word_embeddings:
                self.lm_head.weight = self.model.embed_tokens.weight

            logit_scale = getattr(config, "logit_scale", 1.0)
452
453
            if hasattr(config, "logits_scaling"):
                logit_scale /= config.logits_scaling
454

455
456
457
            self.logits_processor = LogitsProcessor(
                self.unpadded_vocab_size, config.vocab_size, scale=logit_scale
            )
Shawn Tan's avatar
Shawn Tan committed
458
459
460
        else:
            self.lm_head = PPMissingLayer()

461
462
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
463

Shawn Tan's avatar
Shawn Tan committed
464
465
466
467
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
468
469
470
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
471
472
473
        model_output = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
Shawn Tan's avatar
Shawn Tan committed
474
475
        return model_output

476
    def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor | None:
477
        logits = self.logits_processor(self.lm_head, hidden_states)
Shawn Tan's avatar
Shawn Tan committed
478
479
480
        return logits

    def make_empty_intermediate_tensors(
481
482
483
484
485
486
487
488
489
490
491
        self, batch_size: int, dtype: torch.dtype, device: torch.device
    ) -> IntermediateTensors:
        return IntermediateTensors(
            {
                "hidden_states": torch.zeros(
                    (batch_size, self.config.hidden_size), dtype=dtype, device=device
                ),
            }
        )

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
492
493
494
        # With tie_word_embeddings, we can skip lm_head.weight
        # The weight might appear unnecessarily in the files if the model is
        # processed with quantization, LoRA, fine-tuning, etc.
495
        skip_prefixes = ["lm_head."] if self.config.tie_word_embeddings else None
496

497
498
499
500
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=skip_prefixes,
        )
501
        return loader.load_weights(weights)