nemotron_h.py 19.3 KB
Newer Older
Luis Vega's avatar
Luis Vega committed
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
Luis Vega's avatar
Luis Vega committed
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

# Adapted from https://github.com/vllm-project/vllm/blob/94d8ec8d2bcb4ec55e33022b313c7e978edf05e1/vllm/model_executor/models/bamba.py
# Copyright 2024 HuggingFace Inc. team. All rights reserved.
# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
#
# 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 NemotronH model."""
20

Luis Vega's avatar
Luis Vega committed
21
22
23
24
25
26
27
from collections.abc import Iterable
from typing import Optional

import torch
from torch import nn

from vllm.attention.layer import Attention
28
from vllm.compilation.decorators import support_torch_compile
29
from vllm.config import CacheConfig, ModelConfig, VllmConfig
30
from vllm.distributed import get_tensor_model_parallel_world_size
Luis Vega's avatar
Luis Vega committed
31
32
33
from vllm.distributed.parallel_state import get_pp_group
from vllm.model_executor.layers.activation import ReLUSquaredActivation
from vllm.model_executor.layers.layernorm import RMSNorm
34
35
36
37
38
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
Luis Vega's avatar
Luis Vega committed
39
from vllm.model_executor.layers.logits_processor import LogitsProcessor
40
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
41
from vllm.model_executor.layers.mamba.mamba_utils import (
42
43
44
    MambaStateDtypeCalculator,
    MambaStateShapeCalculator,
)
Luis Vega's avatar
Luis Vega committed
45
46
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import (
47
48
49
50
    DEFAULT_VOCAB_PADDING_SIZE,
    ParallelLMHead,
    VocabParallelEmbedding,
)
51
from vllm.model_executor.model_loader.weight_utils import (
52
53
54
55
56
57
58
59
60
61
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
from vllm.model_executor.models.interfaces import (
    HasInnerState,
    IsHybrid,
    SupportsLoRA,
    SupportsPP,
    SupportsQuant,
)
Luis Vega's avatar
Luis Vega committed
62
from vllm.model_executor.models.utils import (
63
64
65
66
67
68
    AutoWeightsLoader,
    WeightsMapper,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
Luis Vega's avatar
Luis Vega committed
69
70
71
72
73
74
75
76
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs import NemotronHConfig


class NemotronHMLP(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
77
        layer_idx: int,
Luis Vega's avatar
Luis Vega committed
78
79
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
80
        prefix: str = "",
Luis Vega's avatar
Luis Vega committed
81
82
    ) -> None:
        super().__init__()
83
84

        hybrid_override_pattern = config.hybrid_override_pattern
85
        mlp_index = hybrid_override_pattern[: layer_idx + 1].count("-") - 1
86
87
88
89
90
91
92
93
        if isinstance(config.intermediate_size, list):
            if len(config.intermediate_size) == 1:
                intermediate_size = config.intermediate_size[0]
            else:
                intermediate_size = config.intermediate_size[mlp_index]
        else:
            intermediate_size = config.intermediate_size

94
        self.up_proj = ColumnParallelLinear(
Luis Vega's avatar
Luis Vega committed
95
            input_size=config.hidden_size,
96
            output_size=intermediate_size,
Luis Vega's avatar
Luis Vega committed
97
98
            bias=bias,
            quant_config=quant_config,
99
            prefix=f"{prefix}.up_proj",
Luis Vega's avatar
Luis Vega committed
100
101
        )
        self.down_proj = RowParallelLinear(
102
            input_size=intermediate_size,
Luis Vega's avatar
Luis Vega committed
103
104
105
            output_size=config.hidden_size,
            bias=bias,
            quant_config=quant_config,
106
            prefix=f"{prefix}.down_proj",
Luis Vega's avatar
Luis Vega committed
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
        )
        self.act_fn = ReLUSquaredActivation()

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


class NemotronHMLPDecoderLayer(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
122
        model_config: Optional[ModelConfig] = None,
Luis Vega's avatar
Luis Vega committed
123
124
125
126
127
128
129
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config

130
131
132
133
134
        self.mixer = NemotronHMLP(
            config,
            quant_config=quant_config,
            bias=config.mlp_bias,
            prefix=f"{prefix}.mixer",
135
            layer_idx=layer_idx,
136
        )
Luis Vega's avatar
Luis Vega committed
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160

        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        **kwargs,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.norm(hidden_states)
        else:
            hidden_states, residual = self.norm(hidden_states, residual)

        hidden_states = self.mixer(hidden_states)
        return hidden_states, residual


class NemotronHMambaDecoderLayer(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
161
        model_config: Optional[ModelConfig] = None,
Luis Vega's avatar
Luis Vega committed
162
163
164
165
166
167
168
169
170
171
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.mixer = MambaMixer2(
            hidden_size=config.hidden_size,
            ssm_state_size=config.ssm_state_size,
            conv_kernel_size=config.conv_kernel,
172
            intermediate_size=config.mamba_num_heads * config.mamba_head_dim,
Luis Vega's avatar
Luis Vega committed
173
174
175
176
177
178
179
            use_conv_bias=config.use_conv_bias,
            use_bias=config.use_bias,
            n_groups=config.n_groups,
            num_heads=config.mamba_num_heads,
            head_dim=config.mamba_head_dim,
            rms_norm_eps=config.rms_norm_eps,
            activation=config.mamba_hidden_act,
180
181
            model_config=model_config,
            cache_config=cache_config,
Luis Vega's avatar
Luis Vega committed
182
            quant_config=quant_config,
183
            prefix=f"{prefix}.mixer",
Luis Vega's avatar
Luis Vega committed
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
        )

        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        **kwargs,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.norm(hidden_states)
        else:
            hidden_states, residual = self.norm(hidden_states, residual)

200
        output = torch.empty_like(hidden_states)
201
        self.mixer(hidden_states, output)
202
        return output, residual
Luis Vega's avatar
Luis Vega committed
203
204
205
206
207
208
209


class NemotronHAttention(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
210
        model_config: Optional[ModelConfig] = None,
Luis Vega's avatar
Luis Vega committed
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = config.num_attention_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = config.num_key_value_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)
231
232
233
234
        if hasattr(config, "head_dim") and config.head_dim is not None:
            self.head_dim = config.head_dim
        else:
            self.head_dim = config.hidden_size // self.total_num_heads
Luis Vega's avatar
Luis Vega committed
235
236
237
238
239
240
241
242
243
244
245
        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

        self.qkv_proj = QKVParallelLinear(
            config.hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
246
            prefix=f"{prefix}.qkv_proj",
Luis Vega's avatar
Luis Vega committed
247
248
249
250
251
252
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            config.hidden_size,
            bias=False,
            quant_config=quant_config,
253
            prefix=f"{prefix}.o_proj",
Luis Vega's avatar
Luis Vega committed
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
        )

        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            prefix=f"{prefix}.attn",
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        **kwargs,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output


class NemotronHAttentionDecoderLayer(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
282
        model_config: Optional[ModelConfig] = None,
Luis Vega's avatar
Luis Vega committed
283
284
285
286
287
288
289
290
291
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()

        self.mixer = NemotronHAttention(
            config,
            layer_idx,
292
            model_config,
Luis Vega's avatar
Luis Vega committed
293
294
            cache_config,
            quant_config,
295
            prefix=f"{prefix}.mixer",
Luis Vega's avatar
Luis Vega committed
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
        )

        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        **kwargs,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.norm(hidden_states)
        else:
            hidden_states, residual = self.norm(hidden_states, residual)

        hidden_states = self.mixer(hidden_states=hidden_states)
        return hidden_states, residual


ALL_DECODER_LAYER_TYPES = {
    "M": NemotronHMambaDecoderLayer,
    "-": NemotronHMLPDecoderLayer,
    "*": NemotronHAttentionDecoderLayer,
}


324
@support_torch_compile
Luis Vega's avatar
Luis Vega committed
325
326
327
328
329
class NemotronHModel(nn.Module):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config: NemotronHConfig = vllm_config.model_config.hf_config
330
        model_config = vllm_config.model_config
Luis Vega's avatar
Luis Vega committed
331
332
333
334
335
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
336
337
338
339
340
        lora_vocab = (
            (lora_config.lora_extra_vocab_size * (lora_config.max_loras or 1))
            if lora_config
            else 0
        )
Luis Vega's avatar
Luis Vega committed
341
342
343
344
345
346
347
348
349
350
351
352
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            self.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
        )

        def get_layer(prefix: str):
            layer_idx = int(prefix.rsplit(".", 1)[1])
            layer_class = ALL_DECODER_LAYER_TYPES[
353
354
                config.hybrid_override_pattern[layer_idx]
            ]
Luis Vega's avatar
Luis Vega committed
355
356
357
            return layer_class(
                config,
                layer_idx,
358
                model_config,
Luis Vega's avatar
Luis Vega committed
359
360
361
362
363
364
                cache_config,
                quant_config=quant_config,
                prefix=prefix,
            )

        self.start_layer, self.end_layer, self.layers = make_layers(
365
366
            len(config.hybrid_override_pattern), get_layer, prefix=f"{prefix}.layers"
        )
Luis Vega's avatar
Luis Vega committed
367
        self.make_empty_intmd_tensors = make_empty_intermediate_tensors_factory(
368
369
            ["hidden_states", "residual"], config.hidden_size
        )
Luis Vega's avatar
Luis Vega committed
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

        self.norm_f = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        residual = None
395
        for i, layer in enumerate(self.layers):
Luis Vega's avatar
Luis Vega committed
396
397
398
399
400
401
402
            hidden_states, residual = layer(
                positions=positions,
                hidden_states=hidden_states,
                residual=residual,
            )

        if not get_pp_group().is_last_rank:
403
404
405
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
Luis Vega's avatar
Luis Vega committed
406
407
408
        hidden_states, _ = self.norm_f(hidden_states, residual)
        return hidden_states

409
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
410
411
412
413
414
415
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]
Luis Vega's avatar
Luis Vega committed
416
417
418
419

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
420
421
422
423
424
425
426
427
428
429
430
431
432
433
            if "scale" in name:
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue

            # load stacked params
            for param_name, weight_name, shard_id in stacked_params_mapping:
                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
Luis Vega's avatar
Luis Vega committed
434
435
436

                param = params_dict[name]
                weight_loader = param.weight_loader
437
438
439
                weight_loader(param, loaded_weight, shard_id)
                break

Luis Vega's avatar
Luis Vega committed
440
441
442
            # load other params
            else:
                param = params_dict[name]
443
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
Luis Vega's avatar
Luis Vega committed
444
445
446
447
448
449
                weight_loader(param, loaded_weight)

            loaded_params.add(name)
        return loaded_params


450
451
452
class NemotronHForCausalLM(
    nn.Module, HasInnerState, SupportsLoRA, SupportsPP, IsHybrid, SupportsQuant
):
453
454
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={"backbone": "model"},
455
        orig_to_new_substr={"A_log": "A", "embeddings": "embed_tokens"},
456
457
    )

Luis Vega's avatar
Luis Vega committed
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }

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

473
474
475
476
477
478
479
480
481
482
483
    @classmethod
    def get_mamba_state_dtype_from_config(
        cls,
        vllm_config: "VllmConfig",
    ) -> tuple[torch.dtype, torch.dtype]:
        return MambaStateDtypeCalculator.mamba2_state_dtype(
            vllm_config.model_config.dtype,
            vllm_config.cache_config.mamba_cache_dtype,
            vllm_config.cache_config.mamba_ssm_cache_dtype,
        )

484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
    @classmethod
    def get_mamba_state_shape_from_config(
        cls,
        vllm_config: "VllmConfig",
    ) -> tuple[tuple[int, int], tuple[int, int, int]]:
        """Calculate shapes for Mamba's convolutional and state caches.

        Args:
            vllm_config: vLLM config

        Returns:
            Tuple containing:
            - conv_state_shape: Shape for convolutional state cache
            - temporal_state_shape: Shape for state space model cache
        """
        parallel_config = vllm_config.parallel_config
        hf_config = vllm_config.model_config.hf_config
501
        intermediate_size = hf_config.mamba_num_heads * hf_config.mamba_head_dim
502

503
        return MambaStateShapeCalculator.mamba2_state_shape(
504
505
506
507
508
509
510
511
512
            intermediate_size=intermediate_size,
            tp_world_size=parallel_config.tensor_parallel_size,
            n_groups=hf_config.n_groups,
            num_heads=hf_config.mamba_num_heads,
            head_dim=hf_config.mamba_head_dim,
            state_size=hf_config.ssm_state_size,
            conv_kernel=hf_config.conv_kernel,
        )

Luis Vega's avatar
Luis Vega committed
513
514
515
516
517
518
519
520
521
522
523
524
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        config = vllm_config.model_config.hf_config
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        lora_config = vllm_config.lora_config
        scheduler_config = vllm_config.scheduler_config

        self.quant_config = vllm_config.quant_config

        super().__init__()
        self.config = config
        self.scheduler_config = scheduler_config
525
526
527
        self.model = NemotronHModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
Luis Vega's avatar
Luis Vega committed
528
529
530
531
532
533
534
535
536
537
        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
538
539
            if not lora_config
            else lora_config.lora_vocab_padding_size,
540
            prefix=maybe_prefix(prefix, "lm_head"),
Luis Vega's avatar
Luis Vega committed
541
542
        )

543
544
545
        self.logits_processor = LogitsProcessor(
            self.unpadded_vocab_size, config.vocab_size
        )
Luis Vega's avatar
Luis Vega committed
546

547
        self.make_empty_intmd_tensors = self.model.make_empty_intmd_tensors
Luis Vega's avatar
Luis Vega committed
548
549
550
551

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

552
553
554
555
556
557
558
559
560
561
562
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        **kwargs,
    ):
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
Luis Vega's avatar
Luis Vega committed
563
564
565
566
567
568
569

        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> Optional[torch.Tensor]:
570
        logits = self.logits_processor(self.lm_head, hidden_states)
Luis Vega's avatar
Luis Vega committed
571
572
        return logits

573
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
Luis Vega's avatar
Luis Vega committed
574
        loader = AutoWeightsLoader(self)
575
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)