nemotron_h.py 19.2 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
from collections.abc import Iterable

import torch
from torch import nn

from vllm.attention.layer import Attention
27
from vllm.compilation.decorators import support_torch_compile
28
from vllm.config import CacheConfig, ModelConfig, VllmConfig
29
from vllm.distributed import get_tensor_model_parallel_world_size
Luis Vega's avatar
Luis Vega committed
30
31
32
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
33
34
35
36
37
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
Luis Vega's avatar
Luis Vega committed
38
from vllm.model_executor.layers.logits_processor import LogitsProcessor
39
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
40
from vllm.model_executor.layers.mamba.mamba_utils import (
41
42
43
    MambaStateDtypeCalculator,
    MambaStateShapeCalculator,
)
Luis Vega's avatar
Luis Vega committed
44
45
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import (
46
47
48
49
    DEFAULT_VOCAB_PADDING_SIZE,
    ParallelLMHead,
    VocabParallelEmbedding,
)
50
from vllm.model_executor.model_loader.weight_utils import (
51
52
53
54
55
56
57
58
59
60
    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
61
from vllm.model_executor.models.utils import (
62
63
64
65
66
67
    AutoWeightsLoader,
    WeightsMapper,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
Luis Vega's avatar
Luis Vega committed
68
69
70
71
72
73
74
75
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs import NemotronHConfig


class NemotronHMLP(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
76
        layer_idx: int,
77
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
78
        bias: bool = False,
79
        prefix: str = "",
Luis Vega's avatar
Luis Vega committed
80
81
    ) -> None:
        super().__init__()
82
83

        hybrid_override_pattern = config.hybrid_override_pattern
84
        mlp_index = hybrid_override_pattern[: layer_idx + 1].count("-") - 1
85
86
87
88
89
90
91
92
        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

93
        self.up_proj = ColumnParallelLinear(
Luis Vega's avatar
Luis Vega committed
94
            input_size=config.hidden_size,
95
            output_size=intermediate_size,
Luis Vega's avatar
Luis Vega committed
96
97
            bias=bias,
            quant_config=quant_config,
98
            prefix=f"{prefix}.up_proj",
Luis Vega's avatar
Luis Vega committed
99
100
        )
        self.down_proj = RowParallelLinear(
101
            input_size=intermediate_size,
Luis Vega's avatar
Luis Vega committed
102
103
104
            output_size=config.hidden_size,
            bias=bias,
            quant_config=quant_config,
105
            prefix=f"{prefix}.down_proj",
Luis Vega's avatar
Luis Vega committed
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
        )
        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,
121
122
123
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
124
125
126
127
128
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config

129
130
131
132
133
        self.mixer = NemotronHMLP(
            config,
            quant_config=quant_config,
            bias=config.mlp_bias,
            prefix=f"{prefix}.mixer",
134
            layer_idx=layer_idx,
135
        )
Luis Vega's avatar
Luis Vega committed
136
137
138
139
140
141

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

    def forward(
        self,
        hidden_states: torch.Tensor,
142
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
        **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,
160
161
162
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
163
164
165
166
167
168
169
170
        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,
171
            intermediate_size=config.mamba_num_heads * config.mamba_head_dim,
Luis Vega's avatar
Luis Vega committed
172
173
174
175
176
177
178
            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,
179
180
            model_config=model_config,
            cache_config=cache_config,
Luis Vega's avatar
Luis Vega committed
181
            quant_config=quant_config,
182
            prefix=f"{prefix}.mixer",
Luis Vega's avatar
Luis Vega committed
183
184
185
186
187
188
189
        )

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

    def forward(
        self,
        hidden_states: torch.Tensor,
190
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
191
192
193
194
195
196
197
198
        **kwargs,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.norm(hidden_states)
        else:
            hidden_states, residual = self.norm(hidden_states, residual)

199
        output = torch.empty_like(hidden_states)
200
        self.mixer(hidden_states, output)
201
        return output, residual
Luis Vega's avatar
Luis Vega committed
202
203
204
205
206
207
208


class NemotronHAttention(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
209
210
211
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
        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)
230
231
232
233
        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
234
235
236
237
238
239
240
241
242
243
244
        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,
245
            prefix=f"{prefix}.qkv_proj",
Luis Vega's avatar
Luis Vega committed
246
247
248
249
250
251
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            config.hidden_size,
            bias=False,
            quant_config=quant_config,
252
            prefix=f"{prefix}.o_proj",
Luis Vega's avatar
Luis Vega committed
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
        )

        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,
281
282
283
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
284
285
286
287
288
289
290
        prefix: str = "",
    ) -> None:
        super().__init__()

        self.mixer = NemotronHAttention(
            config,
            layer_idx,
291
            model_config,
Luis Vega's avatar
Luis Vega committed
292
293
            cache_config,
            quant_config,
294
            prefix=f"{prefix}.mixer",
Luis Vega's avatar
Luis Vega committed
295
296
297
298
299
300
301
302
        )

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

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
303
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
        **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,
}


323
@support_torch_compile
Luis Vega's avatar
Luis Vega committed
324
325
326
327
328
class NemotronHModel(nn.Module):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config: NemotronHConfig = vllm_config.model_config.hf_config
329
        model_config = vllm_config.model_config
Luis Vega's avatar
Luis Vega committed
330
331
332
333
334
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
335
336
337
338
339
        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
340
341
342
343
344
345
346
347
348
349
350
351
        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[
352
353
                config.hybrid_override_pattern[layer_idx]
            ]
Luis Vega's avatar
Luis Vega committed
354
355
356
            return layer_class(
                config,
                layer_idx,
357
                model_config,
Luis Vega's avatar
Luis Vega committed
358
359
360
361
362
363
                cache_config,
                quant_config=quant_config,
                prefix=prefix,
            )

        self.start_layer, self.end_layer, self.layers = make_layers(
364
365
            len(config.hybrid_override_pattern), get_layer, prefix=f"{prefix}.layers"
        )
Luis Vega's avatar
Luis Vega committed
366
        self.make_empty_intmd_tensors = make_empty_intermediate_tensors_factory(
367
368
            ["hidden_states", "residual"], config.hidden_size
        )
Luis Vega's avatar
Luis Vega committed
369
370
371
372
373
374
375
376
377
378

        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,
379
380
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
Luis Vega's avatar
Luis Vega committed
381
382
383
384
385
386
387
388
389
390
391
392
393
    ) -> 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
394
        for i, layer in enumerate(self.layers):
Luis Vega's avatar
Luis Vega committed
395
396
397
398
399
400
401
            hidden_states, residual = layer(
                positions=positions,
                hidden_states=hidden_states,
                residual=residual,
            )

        if not get_pp_group().is_last_rank:
402
403
404
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
Luis Vega's avatar
Luis Vega committed
405
406
407
        hidden_states, _ = self.norm_f(hidden_states, residual)
        return hidden_states

408
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
409
410
411
412
413
414
        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
415
416
417
418

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
419
420
421
422
423
424
425
426
427
428
429
430
431
432
            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
433
434
435

                param = params_dict[name]
                weight_loader = param.weight_loader
436
437
438
                weight_loader(param, loaded_weight, shard_id)
                break

Luis Vega's avatar
Luis Vega committed
439
440
441
            # load other params
            else:
                param = params_dict[name]
442
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
Luis Vega's avatar
Luis Vega committed
443
444
445
446
447
448
                weight_loader(param, loaded_weight)

            loaded_params.add(name)
        return loaded_params


449
450
451
class NemotronHForCausalLM(
    nn.Module, HasInnerState, SupportsLoRA, SupportsPP, IsHybrid, SupportsQuant
):
452
453
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={"backbone": "model"},
454
        orig_to_new_substr={"A_log": "A", "embeddings": "embed_tokens"},
455
456
    )

Luis Vega's avatar
Luis Vega committed
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
    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"]

472
473
474
475
476
477
478
479
480
481
482
    @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,
        )

483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
    @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
500
        intermediate_size = hf_config.mamba_num_heads * hf_config.mamba_head_dim
501

502
        return MambaStateShapeCalculator.mamba2_state_shape(
503
504
505
506
507
508
509
510
511
            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
512
513
514
515
516
517
518
519
520
521
522
523
    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
524
525
526
        self.model = NemotronHModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
Luis Vega's avatar
Luis Vega committed
527
528
529
530
531
532
533
534
535
536
        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
537
538
            if not lora_config
            else lora_config.lora_vocab_padding_size,
539
            prefix=maybe_prefix(prefix, "lm_head"),
Luis Vega's avatar
Luis Vega committed
540
541
        )

542
543
544
        self.logits_processor = LogitsProcessor(
            self.unpadded_vocab_size, config.vocab_size
        )
Luis Vega's avatar
Luis Vega committed
545

546
        self.make_empty_intmd_tensors = self.model.make_empty_intmd_tensors
Luis Vega's avatar
Luis Vega committed
547
548
549
550

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

551
552
553
554
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
555
556
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
557
558
559
560
561
        **kwargs,
    ):
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
Luis Vega's avatar
Luis Vega committed
562
563
564
565
566
567

        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
568
    ) -> torch.Tensor | None:
569
        logits = self.logits_processor(self.lm_head, hidden_states)
Luis Vega's avatar
Luis Vega committed
570
571
        return logits

572
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
Luis Vega's avatar
Luis Vega committed
573
        loader = AutoWeightsLoader(self)
574
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)