"examples/template_vlm2vec.jinja" did not exist on "30a2e8074246e11a1452ab5e84a7be65ecac6119"
nemotron_h.py 31 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

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

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
31
32
from vllm.config.parallel import ParallelConfig
from vllm.distributed import get_ep_group, get_tensor_model_parallel_world_size
from vllm.distributed.communication_op import tensor_model_parallel_all_gather
Luis Vega's avatar
Luis Vega committed
33
34
from vllm.distributed.parallel_state import get_pp_group
from vllm.model_executor.layers.activation import ReLUSquaredActivation
35
36
from vllm.model_executor.layers.fused_moe import FusedMoE, SharedFusedMoE
from vllm.model_executor.layers.fused_moe.utils import activation_without_mul
Luis Vega's avatar
Luis Vega committed
37
from vllm.model_executor.layers.layernorm import RMSNorm
38
39
40
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    QKVParallelLinear,
41
    ReplicatedLinear,
42
43
    RowParallelLinear,
)
Luis Vega's avatar
Luis Vega committed
44
from vllm.model_executor.layers.logits_processor import LogitsProcessor
45
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
46
from vllm.model_executor.layers.mamba.mamba_utils import (
47
48
49
    MambaStateDtypeCalculator,
    MambaStateShapeCalculator,
)
Luis Vega's avatar
Luis Vega committed
50
51
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import (
52
53
54
55
    DEFAULT_VOCAB_PADDING_SIZE,
    ParallelLMHead,
    VocabParallelEmbedding,
)
56
from vllm.model_executor.model_loader.weight_utils import (
57
58
59
60
61
62
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
from vllm.model_executor.models.interfaces import (
    HasInnerState,
    IsHybrid,
63
    MixtureOfExperts,
64
65
66
67
    SupportsLoRA,
    SupportsPP,
    SupportsQuant,
)
Luis Vega's avatar
Luis Vega committed
68
from vllm.model_executor.models.utils import (
69
70
    AutoWeightsLoader,
    WeightsMapper,
71
    is_pp_missing_parameter,
72
73
74
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
75
    sequence_parallel_chunk,
76
)
Luis Vega's avatar
Luis Vega committed
77
78
79
80
81
82
83
84
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs import NemotronHConfig


class NemotronHMLP(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
85
        intermediate_size: int,
86
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
87
        bias: bool = False,
88
89
        reduce_results: bool = True,
        is_sequence_parallel: bool = False,
90
        prefix: str = "",
Luis Vega's avatar
Luis Vega committed
91
92
    ) -> None:
        super().__init__()
93

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
            disable_tp=is_sequence_parallel,
100
            prefix=f"{prefix}.up_proj",
Luis Vega's avatar
Luis Vega committed
101
102
        )
        self.down_proj = RowParallelLinear(
103
            input_size=intermediate_size,
Luis Vega's avatar
Luis Vega committed
104
105
106
            output_size=config.hidden_size,
            bias=bias,
            quant_config=quant_config,
107
108
            reduce_results=reduce_results,
            disable_tp=is_sequence_parallel,
109
            prefix=f"{prefix}.down_proj",
Luis Vega's avatar
Luis Vega committed
110
111
112
113
114
115
116
117
118
119
        )
        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


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
class NemotronHMoE(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        quant_config: QuantizationConfig | None = None,
        parallel_config: ParallelConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.routed_scaling_factor = config.routed_scaling_factor

        self.ep_group = get_ep_group().device_group
        self.ep_rank = self.ep_group.rank()
        self.ep_size = self.ep_group.size()
        self.n_routed_experts: int = config.n_routed_experts
        self.n_shared_experts: int = config.n_shared_experts

        self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe

        self.gate = ReplicatedLinear(
            config.hidden_size,
            config.n_routed_experts,
            bias=False,
            params_dtype=torch.float32,
            quant_config=None,
            prefix=f"{prefix}.gate",
        )

        self.gate.e_score_correction_bias = nn.Parameter(
            torch.empty(config.n_routed_experts, dtype=torch.float32)
        )
        # Load balancing settings.
        self.enable_eplb = parallel_config.enable_eplb

        self.n_redundant_experts = parallel_config.eplb_config.num_redundant_experts  # noqa: E501
        self.n_logical_experts = self.n_routed_experts
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

        self.physical_expert_start = self.ep_rank * self.n_local_physical_experts
        self.physical_expert_end = (
            self.physical_expert_start + self.n_local_physical_experts
        )

        if config.n_shared_experts is None or config.n_shared_experts == 0:
            self.shared_experts = None
        else:
            intermediate_size = (
                config.moe_shared_expert_intermediate_size * config.n_shared_experts
            )

            self.shared_experts = NemotronHMLP(
                config=config,
                intermediate_size=intermediate_size,
                quant_config=quant_config,
                reduce_results=False,
                is_sequence_parallel=self.is_sequence_parallel,
                prefix=f"{prefix}.shared_experts",
            )

        self.experts = SharedFusedMoE(
            shared_experts=self.shared_experts,
            num_experts=config.n_routed_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            reduce_results=False,
            renormalize=config.norm_topk_prob,
            quant_config=quant_config,
            use_grouped_topk=True,
            num_expert_group=config.n_group,
            topk_group=config.topk_group,
            prefix=f"{prefix}.experts",
            scoring_func="sigmoid",
            e_score_correction_bias=self.gate.e_score_correction_bias,
            activation=activation_without_mul(config.mlp_hidden_act),
            is_act_and_mul=False,  # non-gated MoE
            enable_eplb=self.enable_eplb,
            num_redundant_experts=self.n_redundant_experts,
            is_sequence_parallel=self.is_sequence_parallel,
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)

        if self.is_sequence_parallel:
            hidden_states = sequence_parallel_chunk(hidden_states)

        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states.to(dtype=torch.float32))

        fused_moe_out = self.experts(
            hidden_states=hidden_states, router_logits=router_logits
        )

        shared_output, final_hidden_states = fused_moe_out

        # Fix FP16 overflow
        # See DeepseekV2DecoderLayer for more details.
        if hidden_states.dtype != torch.float16:
            final_hidden_states *= self.routed_scaling_factor
        elif self.shared_experts is not None:
            assert shared_output is not None
            shared_output *= 1.0 / self.routed_scaling_factor

        if self.shared_experts is not None:
            assert shared_output is not None
            final_hidden_states += shared_output

        if self.is_sequence_parallel:
            final_hidden_states = tensor_model_parallel_all_gather(
                final_hidden_states, 0
            )
            final_hidden_states = final_hidden_states[:num_tokens]
        elif self.tp_size > 1:
            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
                final_hidden_states
            )

        return final_hidden_states.view(num_tokens, hidden_dim)


Luis Vega's avatar
Luis Vega committed
244
245
246
247
248
class NemotronHMLPDecoderLayer(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
249
250
251
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
252
        parallel_config: ParallelConfig | None = None,
Luis Vega's avatar
Luis Vega committed
253
254
255
256
257
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config

258
259
260
261
262
263
264
265
266
267
        hybrid_override_pattern = config.hybrid_override_pattern
        mlp_index = hybrid_override_pattern[: layer_idx + 1].count("-") - 1
        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

268
269
        self.mixer = NemotronHMLP(
            config,
270
            intermediate_size=intermediate_size,
271
272
273
274
            quant_config=quant_config,
            bias=config.mlp_bias,
            prefix=f"{prefix}.mixer",
        )
Luis Vega's avatar
Luis Vega committed
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
        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

    def forward(
        self,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
        **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 NemotronHMoEDecoderLayer(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        parallel_config: ParallelConfig | None = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config

        self.mixer = NemotronHMoE(
            config,
            quant_config=quant_config,
            parallel_config=parallel_config,
            prefix=f"{prefix}.mixer",
        )

        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
Luis Vega's avatar
Luis Vega committed
316
317
318
319

    def forward(
        self,
        hidden_states: torch.Tensor,
320
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
        **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,
338
339
340
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
341
        parallel_config: ParallelConfig | None = None,
Luis Vega's avatar
Luis Vega committed
342
343
344
345
346
347
348
349
        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,
350
            intermediate_size=config.mamba_num_heads * config.mamba_head_dim,
Luis Vega's avatar
Luis Vega committed
351
352
353
354
355
            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,
356
            rms_norm_eps=config.layer_norm_epsilon,
Luis Vega's avatar
Luis Vega committed
357
            activation=config.mamba_hidden_act,
358
359
            model_config=model_config,
            cache_config=cache_config,
Luis Vega's avatar
Luis Vega committed
360
            quant_config=quant_config,
361
            prefix=f"{prefix}.mixer",
Luis Vega's avatar
Luis Vega committed
362
363
        )

364
        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
Luis Vega's avatar
Luis Vega committed
365
366
367
368

    def forward(
        self,
        hidden_states: torch.Tensor,
369
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
370
371
372
373
374
375
376
377
        **kwargs,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.norm(hidden_states)
        else:
            hidden_states, residual = self.norm(hidden_states, residual)

378
        output = torch.empty_like(hidden_states)
379
        self.mixer(hidden_states, output)
380
        return output, residual
Luis Vega's avatar
Luis Vega committed
381
382
383
384
385
386
387


class NemotronHAttention(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
388
389
390
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
        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)
409
410
411
412
        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
413
414
415
416
417
418
419
420
421
422
423
        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,
424
            prefix=f"{prefix}.qkv_proj",
Luis Vega's avatar
Luis Vega committed
425
426
427
428
429
430
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            config.hidden_size,
            bias=False,
            quant_config=quant_config,
431
            prefix=f"{prefix}.o_proj",
Luis Vega's avatar
Luis Vega committed
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
        )

        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,
460
461
462
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
463
        parallel_config: ParallelConfig | None = None,
Luis Vega's avatar
Luis Vega committed
464
465
466
467
468
469
470
        prefix: str = "",
    ) -> None:
        super().__init__()

        self.mixer = NemotronHAttention(
            config,
            layer_idx,
471
            model_config,
Luis Vega's avatar
Luis Vega committed
472
473
            cache_config,
            quant_config,
474
            prefix=f"{prefix}.mixer",
Luis Vega's avatar
Luis Vega committed
475
476
        )

477
        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
Luis Vega's avatar
Luis Vega committed
478
479
480
481
482

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
483
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
        **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,
500
    "E": NemotronHMoEDecoderLayer,
Luis Vega's avatar
Luis Vega committed
501
502
503
}


504
@support_torch_compile
Luis Vega's avatar
Luis Vega committed
505
506
507
508
509
class NemotronHModel(nn.Module):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config: NemotronHConfig = vllm_config.model_config.hf_config
510
        model_config = vllm_config.model_config
Luis Vega's avatar
Luis Vega committed
511
512
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
513
        parallel_config = vllm_config.parallel_config
Luis Vega's avatar
Luis Vega committed
514
515
516
        lora_config = vllm_config.lora_config

        self.config = config
517
518
519
520
521
        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
522
523
524
525
526
527
528
529
530
        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,
        )

531
532
        self.has_moe = "E" in config.hybrid_override_pattern

Luis Vega's avatar
Luis Vega committed
533
534
535
        def get_layer(prefix: str):
            layer_idx = int(prefix.rsplit(".", 1)[1])
            layer_class = ALL_DECODER_LAYER_TYPES[
536
537
                config.hybrid_override_pattern[layer_idx]
            ]
Luis Vega's avatar
Luis Vega committed
538
            return layer_class(
539
540
541
542
                config=config,
                layer_idx=layer_idx,
                model_config=model_config,
                cache_config=cache_config,
Luis Vega's avatar
Luis Vega committed
543
                quant_config=quant_config,
544
                parallel_config=parallel_config,
Luis Vega's avatar
Luis Vega committed
545
546
547
548
                prefix=prefix,
            )

        self.start_layer, self.end_layer, self.layers = make_layers(
549
550
            len(config.hybrid_override_pattern), get_layer, prefix=f"{prefix}.layers"
        )
Luis Vega's avatar
Luis Vega committed
551
        self.make_empty_intmd_tensors = make_empty_intermediate_tensors_factory(
552
553
            ["hidden_states", "residual"], config.hidden_size
        )
Luis Vega's avatar
Luis Vega committed
554

555
        self.norm_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
Luis Vega's avatar
Luis Vega committed
556
557
558
559
560
561
562
563

    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,
564
565
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
Luis Vega's avatar
Luis Vega committed
566
567
568
569
570
571
572
573
574
575
576
577
578
    ) -> 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
579
        for i, layer in enumerate(self.layers):
Luis Vega's avatar
Luis Vega committed
580
581
582
583
584
585
586
            hidden_states, residual = layer(
                positions=positions,
                hidden_states=hidden_states,
                residual=residual,
            )

        if not get_pp_group().is_last_rank:
587
588
589
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
Luis Vega's avatar
Luis Vega committed
590
591
592
        hidden_states, _ = self.norm_f(hidden_states, residual)
        return hidden_states

593
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
594
595
596
597
598
599
        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
600

601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
        if self.has_moe:
            # (param_name, weight_name, expert_id, shard_id)
            expert_params_mapping = FusedMoE.make_expert_params_mapping(
                # - FusedMoe.w1 (aka gate_proj) should be up_proj since that's
                #   what the activation is applied to
                # - FusedMoe.w3 (aka up_proj) should be ignored since we're
                #   using non-gated MoE
                ckpt_gate_proj_name="up_proj",
                ckpt_down_proj_name="down_proj",
                ckpt_up_proj_name="",
                num_experts=self.config.n_routed_experts,
                num_redundant_experts=getattr(self, "num_redundant_experts", 0),
            )
        else:
            expert_params_mapping = []

Luis Vega's avatar
Luis Vega committed
617
618
619
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
620
621
622
623
624
625
626
627
628
629
630
631
632
633
            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
634
635
636

                param = params_dict[name]
                weight_loader = param.weight_loader
637
638
639
                weight_loader(param, loaded_weight, shard_id)
                break

Luis Vega's avatar
Luis Vega committed
640
641
            # load other params
            else:
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
                is_expert_weight = False
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue

                    # Anyway, this is an expert weight and should not be
                    # attempted to load as other weights later
                    is_expert_weight = True

                    # Do not modify `name` since the loop may continue here
                    # Instead, create a new variable
                    name_mapped = name.replace(weight_name, param_name)

                    if is_pp_missing_parameter(name_mapped, self):
                        continue
                    param = params_dict[name_mapped]
                    # We should ask the weight loader to return success or not
                    # here since otherwise we may skip experts with other
                    # available replicas.
                    weight_loader = typing.cast(
                        Callable[..., bool], param.weight_loader
                    )
                    success = weight_loader(
                        param,
                        loaded_weight,
                        name_mapped,
                        shard_id=shard_id,
                        expert_id=expert_id,
                        return_success=True,
                    )
                    if success:
                        name = name_mapped
                        break
                else:
                    if is_expert_weight:
                        continue

                    param = params_dict[name]
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
                    weight_loader(param, loaded_weight)
Luis Vega's avatar
Luis Vega committed
685
686
687
688
689

            loaded_params.add(name)
        return loaded_params


690
class NemotronHForCausalLM(
691
692
693
694
695
696
697
    nn.Module,
    HasInnerState,
    SupportsLoRA,
    SupportsPP,
    IsHybrid,
    SupportsQuant,
    MixtureOfExperts,
698
):
699
700
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={"backbone": "model"},
701
        orig_to_new_substr={"A_log": "A", "embeddings": "embed_tokens"},
702
703
    )

Luis Vega's avatar
Luis Vega committed
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
    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"]

719
720
721
722
723
724
725
726
727
728
729
    @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,
        )

730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
    @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
747
        intermediate_size = hf_config.mamba_num_heads * hf_config.mamba_head_dim
748

749
        return MambaStateShapeCalculator.mamba2_state_shape(
750
751
752
753
754
755
756
757
758
            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
759
760
761
762
763
764
765
766
767
768
769
770
    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
771
772
773
        self.model = NemotronHModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
Luis Vega's avatar
Luis Vega committed
774
775
776
777
778
779
780
781
782
783
        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
784
785
            if not lora_config
            else lora_config.lora_vocab_padding_size,
786
            prefix=maybe_prefix(prefix, "lm_head"),
Luis Vega's avatar
Luis Vega committed
787
788
        )

789
790
791
        self.logits_processor = LogitsProcessor(
            self.unpadded_vocab_size, config.vocab_size
        )
Luis Vega's avatar
Luis Vega committed
792

793
        self.make_empty_intmd_tensors = self.model.make_empty_intmd_tensors
Luis Vega's avatar
Luis Vega committed
794

795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
        # Set MoE hyperparameters
        if self.model.has_moe:
            self.expert_weights = []
            self.num_expert_groups = config.n_group

            self.moe_layers: list[SharedFusedMoE] = []
            example_moe = None
            for layer in self.model.layers:
                if isinstance(layer, NemotronHMoEDecoderLayer):
                    # Pick last one layer since the first ones
                    # may be dense layers.
                    example_moe = layer.mixer
                    self.moe_layers.append(layer.mixer.experts)

            self.num_moe_layers = len(self.moe_layers)
            self.num_logical_experts = example_moe.n_logical_experts
            self.num_physical_experts = example_moe.n_physical_experts
            self.num_local_physical_experts = example_moe.n_local_physical_experts  # noqa: E501
            self.num_routed_experts = example_moe.n_routed_experts
            self.num_shared_experts = example_moe.n_shared_experts
            self.num_redundant_experts = example_moe.n_redundant_experts

    def set_eplb_state(
        self,
        expert_load_view: torch.Tensor,
        logical_to_physical_map: torch.Tensor,
        logical_replica_count: torch.Tensor,
    ) -> None:
        for layer_idx, layer in enumerate(self.moe_layers):
            # Register the expert weights.
            self.expert_weights.append(layer.get_expert_weights())
            layer.set_eplb_state(
                moe_layer_idx=layer_idx,
                expert_load_view=expert_load_view,
                logical_to_physical_map=logical_to_physical_map,
                logical_replica_count=logical_replica_count,
            )

    def update_physical_experts_metadata(
        self,
        num_physical_experts: int,
        num_local_physical_experts: int,
    ) -> None:
        assert self.num_local_physical_experts == num_local_physical_experts
        self.num_physical_experts = num_physical_experts
        self.num_local_physical_experts = num_local_physical_experts
        self.num_redundant_experts = num_physical_experts - self.num_logical_experts
        for layer in self.model.layers:
            if isinstance(layer, NemotronHMoEDecoderLayer):
                moe = layer.mixer
                moe.n_local_physical_experts = num_local_physical_experts
                moe.n_physical_experts = num_physical_experts
                moe.n_redundant_experts = self.num_redundant_experts
                moe.experts.update_expert_map()

Luis Vega's avatar
Luis Vega committed
850
851
852
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

853
854
855
856
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
857
858
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
859
860
861
862
863
        **kwargs,
    ):
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
Luis Vega's avatar
Luis Vega committed
864
865
866
867
868
869

        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
870
    ) -> torch.Tensor | None:
871
        logits = self.logits_processor(self.lm_head, hidden_states)
Luis Vega's avatar
Luis Vega committed
872
873
        return logits

874
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
Luis Vega's avatar
Luis Vega committed
875
        loader = AutoWeightsLoader(self)
876
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)