nemotron_h.py 33.5 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
23
from itertools import islice
Luis Vega's avatar
Luis Vega committed
24
25
26
27

import torch
from torch import nn

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
from vllm.model_executor.layers.attention import Attention
36
from vllm.model_executor.layers.fused_moe import (
37
    FusedMoE,
38
    GateLinear,
39
    activation_without_mul,
40
    fused_moe_make_expert_params_mapping,
41
)
Luis Vega's avatar
Luis Vega committed
42
from vllm.model_executor.layers.layernorm import RMSNorm
43
44
45
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    QKVParallelLinear,
46
    ReplicatedLinear,
47
48
    RowParallelLinear,
)
Luis Vega's avatar
Luis Vega committed
49
from vllm.model_executor.layers.logits_processor import LogitsProcessor
50
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
51
from vllm.model_executor.layers.mamba.mamba_utils import (
52
53
    MambaStateCopyFunc,
    MambaStateCopyFuncCalculator,
54
55
56
    MambaStateDtypeCalculator,
    MambaStateShapeCalculator,
)
Luis Vega's avatar
Luis Vega committed
57
58
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import (
59
60
61
    ParallelLMHead,
    VocabParallelEmbedding,
)
62
from vllm.model_executor.model_loader.weight_utils import (
63
64
65
66
67
68
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
from vllm.model_executor.models.interfaces import (
    HasInnerState,
    IsHybrid,
69
    MixtureOfExperts,
70
    SupportsLoRA,
71
    SupportsMambaPrefixCaching,
72
73
74
    SupportsPP,
    SupportsQuant,
)
Luis Vega's avatar
Luis Vega committed
75
from vllm.model_executor.models.utils import (
76
77
    AutoWeightsLoader,
    WeightsMapper,
78
    is_pp_missing_parameter,
79
80
81
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
82
    sequence_parallel_chunk,
83
)
Luis Vega's avatar
Luis Vega committed
84
from vllm.sequence import IntermediateTensors
85
from vllm.transformers_utils.configs.nemotron_h import NemotronHConfig
Luis Vega's avatar
Luis Vega committed
86
87
88
89
90
91


class NemotronHMLP(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
shaharmor98's avatar
shaharmor98 committed
92
        hidden_size: int,
93
        intermediate_size: int,
94
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
95
        bias: bool = False,
96
97
        reduce_results: bool = True,
        is_sequence_parallel: bool = False,
98
        prefix: str = "",
Luis Vega's avatar
Luis Vega committed
99
100
    ) -> None:
        super().__init__()
101

102
        self.up_proj = ColumnParallelLinear(
shaharmor98's avatar
shaharmor98 committed
103
            input_size=hidden_size,
104
            output_size=intermediate_size,
Luis Vega's avatar
Luis Vega committed
105
106
            bias=bias,
            quant_config=quant_config,
107
            disable_tp=is_sequence_parallel,
108
            prefix=f"{prefix}.up_proj",
Luis Vega's avatar
Luis Vega committed
109
110
        )
        self.down_proj = RowParallelLinear(
111
            input_size=intermediate_size,
shaharmor98's avatar
shaharmor98 committed
112
            output_size=hidden_size,
Luis Vega's avatar
Luis Vega committed
113
114
            bias=bias,
            quant_config=quant_config,
115
116
            reduce_results=reduce_results,
            disable_tp=is_sequence_parallel,
117
            prefix=f"{prefix}.down_proj",
Luis Vega's avatar
Luis Vega committed
118
119
120
121
122
123
124
125
126
127
        )
        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


128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
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
shaharmor98's avatar
shaharmor98 committed
145
146
147
148
        self.use_latent_moe: bool = getattr(config, "moe_latent_size", None) is not None
        self.moe_hidden_size: int = (
            config.moe_latent_size if self.use_latent_moe else config.hidden_size
        )
149
150
151

        self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe

152
        self.gate = GateLinear(
153
154
            config.hidden_size,
            config.n_routed_experts,
155
156
            out_dtype=torch.float32,
            force_fp32_compute=True,
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
            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,
shaharmor98's avatar
shaharmor98 committed
185
                hidden_size=config.hidden_size,
186
187
188
189
190
191
192
                intermediate_size=intermediate_size,
                quant_config=quant_config,
                reduce_results=False,
                is_sequence_parallel=self.is_sequence_parallel,
                prefix=f"{prefix}.shared_experts",
            )

193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
        if self.use_latent_moe:
            self.fc1_latent_proj = ReplicatedLinear(
                input_size=config.hidden_size,
                output_size=self.moe_hidden_size,
                bias=config.mlp_bias,
                quant_config=quant_config,
                disable_tp=self.is_sequence_parallel,
                prefix=f"{prefix}.fc1_latent_proj",
            )
            self.fc2_latent_proj = ReplicatedLinear(
                input_size=self.moe_hidden_size,
                output_size=config.hidden_size,
                bias=config.mlp_bias,
                quant_config=quant_config,
                disable_tp=self.is_sequence_parallel,
                prefix=f"{prefix}.fc2_latent_proj",
            )
        else:
            self.fc1_latent_proj = None
            self.fc2_latent_proj = None

214
        self.experts = FusedMoE(
215
            shared_experts=self.shared_experts,
216
217
            num_experts=config.n_routed_experts,
            top_k=config.num_experts_per_tok,
shaharmor98's avatar
shaharmor98 committed
218
            hidden_size=self.moe_hidden_size,
219
220
221
222
223
224
225
226
227
228
229
230
231
232
            intermediate_size=config.moe_intermediate_size,
            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,
233
            routed_input_transform=self.fc1_latent_proj,
234
235
236
            routed_output_transform=self.fc2_latent_proj,
            routed_scaling_factor=self.routed_scaling_factor,
            apply_routed_scale_to_output=True,
237
            router_logits_dtype=self.gate.out_dtype,
238
239
240
241
242
243
244
245
246
247
        )

    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)
248
        router_logits, _ = self.gate(hidden_states)
249

250
        final_hidden_states = self.experts(
251
252
253
254
255
256
257
258
259
260
261
262
            hidden_states=hidden_states, router_logits=router_logits
        )

        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]

        return final_hidden_states.view(num_tokens, hidden_dim)


Luis Vega's avatar
Luis Vega committed
263
264
265
266
267
class NemotronHMLPDecoderLayer(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
268
269
270
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
271
        parallel_config: ParallelConfig | None = None,
Luis Vega's avatar
Luis Vega committed
272
273
274
275
276
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config

277
278
        hybrid_override_pattern = config.hybrid_override_pattern
        mlp_index = hybrid_override_pattern[: layer_idx + 1].count("-") - 1
279
280
281
282
283
        # Get per-layer config for heterogeneous models if exist
        get_layer_config = getattr(config, "get_nemotron_h_config_for_layer", None)
        layer_config = get_layer_config(layer_idx) if get_layer_config else config
        config = layer_config

284
285
286
287
288
289
290
291
        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

292
293
        self.mixer = NemotronHMLP(
            config,
shaharmor98's avatar
shaharmor98 committed
294
            hidden_size=config.hidden_size,
295
            intermediate_size=intermediate_size,
296
297
298
299
            quant_config=quant_config,
            bias=config.mlp_bias,
            prefix=f"{prefix}.mixer",
        )
Luis Vega's avatar
Luis Vega committed
300

301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
        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

Jiayi Yan's avatar
Jiayi Yan committed
333
        # Get per-layer config for heterogeneous models if exists
334
335
336
        get_layer_config = getattr(config, "get_nemotron_h_config_for_layer", None)
        layer_config = get_layer_config(layer_idx) if get_layer_config else config

337
        self.mixer = NemotronHMoE(
338
            layer_config,
339
340
341
342
343
344
            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
345
346
347
348

    def forward(
        self,
        hidden_states: torch.Tensor,
349
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
        **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,
367
368
369
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
370
        parallel_config: ParallelConfig | None = None,
Luis Vega's avatar
Luis Vega committed
371
372
373
374
375
376
377
378
        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,
379
            intermediate_size=config.mamba_num_heads * config.mamba_head_dim,
Luis Vega's avatar
Luis Vega committed
380
381
382
383
384
            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,
385
            rms_norm_eps=config.layer_norm_epsilon,
Luis Vega's avatar
Luis Vega committed
386
            activation=config.mamba_hidden_act,
387
388
            model_config=model_config,
            cache_config=cache_config,
Luis Vega's avatar
Luis Vega committed
389
            quant_config=quant_config,
390
            prefix=f"{prefix}.mixer",
Luis Vega's avatar
Luis Vega committed
391
392
        )

393
        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
Luis Vega's avatar
Luis Vega committed
394
395
396
397

    def forward(
        self,
        hidden_states: torch.Tensor,
398
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
399
400
401
402
403
404
405
406
        **kwargs,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.norm(hidden_states)
        else:
            hidden_states, residual = self.norm(hidden_states, residual)

407
        output = self.mixer(hidden_states)
408
        return output, residual
Luis Vega's avatar
Luis Vega committed
409
410
411
412
413
414
415


class NemotronHAttention(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
416
417
418
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
        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)
437
438
439
440
        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
441
442
443
444
445
446
447
448
449
450
451
        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,
452
            prefix=f"{prefix}.qkv_proj",
Luis Vega's avatar
Luis Vega committed
453
454
455
456
457
458
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            config.hidden_size,
            bias=False,
            quant_config=quant_config,
459
            prefix=f"{prefix}.o_proj",
Luis Vega's avatar
Luis Vega committed
460
461
        )

462
463
464
        # Get per-layer sliding window from config (for heterogeneous models)
        sliding_window = getattr(config, "sliding_window", None)

Luis Vega's avatar
Luis Vega committed
465
466
467
468
469
470
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
471
            quant_config=quant_config,
Luis Vega's avatar
Luis Vega committed
472
            prefix=f"{prefix}.attn",
473
            per_layer_sliding_window=sliding_window,
Luis Vega's avatar
Luis Vega committed
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
        )

    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,
493
494
495
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
496
        parallel_config: ParallelConfig | None = None,
Luis Vega's avatar
Luis Vega committed
497
498
499
500
        prefix: str = "",
    ) -> None:
        super().__init__()

Jiayi Yan's avatar
Jiayi Yan committed
501
        # Get per-layer config for heterogeneous models if exists
502
503
504
        get_layer_config = getattr(config, "get_nemotron_h_config_for_layer", None)
        layer_config = get_layer_config(layer_idx) if get_layer_config else config

Luis Vega's avatar
Luis Vega committed
505
        self.mixer = NemotronHAttention(
506
            layer_config,
Luis Vega's avatar
Luis Vega committed
507
            layer_idx,
508
            model_config,
Luis Vega's avatar
Luis Vega committed
509
510
            cache_config,
            quant_config,
511
            prefix=f"{prefix}.mixer",
Luis Vega's avatar
Luis Vega committed
512
513
        )

514
        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
Luis Vega's avatar
Luis Vega committed
515
516
517
518
519

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
520
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
        **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,
537
    "E": NemotronHMoEDecoderLayer,
Luis Vega's avatar
Luis Vega committed
538
539
540
}


541
@support_torch_compile
Luis Vega's avatar
Luis Vega committed
542
543
544
545
546
class NemotronHModel(nn.Module):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config: NemotronHConfig = vllm_config.model_config.hf_config
547
        model_config = vllm_config.model_config
Luis Vega's avatar
Luis Vega committed
548
549
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
550
        parallel_config = vllm_config.parallel_config
Luis Vega's avatar
Luis Vega committed
551
552

        self.config = config
553
554

        self.vocab_size = config.vocab_size
Luis Vega's avatar
Luis Vega committed
555
556
557
558
559
560

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

561
562
        self.has_moe = "E" in config.hybrid_override_pattern

Luis Vega's avatar
Luis Vega committed
563
564
565
        def get_layer(prefix: str):
            layer_idx = int(prefix.rsplit(".", 1)[1])
            layer_class = ALL_DECODER_LAYER_TYPES[
566
567
                config.hybrid_override_pattern[layer_idx]
            ]
Luis Vega's avatar
Luis Vega committed
568
            return layer_class(
569
570
571
572
                config=config,
                layer_idx=layer_idx,
                model_config=model_config,
                cache_config=cache_config,
Luis Vega's avatar
Luis Vega committed
573
                quant_config=quant_config,
574
                parallel_config=parallel_config,
Luis Vega's avatar
Luis Vega committed
575
576
577
578
                prefix=prefix,
            )

        self.start_layer, self.end_layer, self.layers = make_layers(
579
580
            len(config.hybrid_override_pattern), get_layer, prefix=f"{prefix}.layers"
        )
581
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
582
583
            ["hidden_states", "residual"], config.hidden_size
        )
Luis Vega's avatar
Luis Vega committed
584

585
        self.norm_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
Luis Vega's avatar
Luis Vega committed
586

587
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
Luis Vega's avatar
Luis Vega committed
588
589
590
591
        return self.embed_tokens(input_ids)

    def forward(
        self,
592
        input_ids: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
593
        positions: torch.Tensor,
594
595
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
596
    ) -> torch.Tensor | IntermediateTensors:
Luis Vega's avatar
Luis Vega committed
597
598
599
600
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
601
                hidden_states = self.embed_input_ids(input_ids)
Luis Vega's avatar
Luis Vega committed
602
603
604
605
606
607
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

608
        for layer in islice(self.layers, self.start_layer, self.end_layer):
Luis Vega's avatar
Luis Vega committed
609
610
611
612
613
614
615
            hidden_states, residual = layer(
                positions=positions,
                hidden_states=hidden_states,
                residual=residual,
            )

        if not get_pp_group().is_last_rank:
616
617
618
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
Luis Vega's avatar
Luis Vega committed
619
620
621
        hidden_states, _ = self.norm_f(hidden_states, residual)
        return hidden_states

622
623
624
    def is_spec_layer(self, config: NemotronHConfig, weight_name: str) -> bool:
        return weight_name.startswith("mtp.")

625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
    def _get_max_n_routed_experts(self) -> int:
        """Get max n_routed_experts from config or block_configs for puzzle models.

        For heterogeneous models with varying expert counts per layer,
        returns the MAX to ensure all expert weights can be loaded.
        """
        # First try top-level attribute
        n_routed_experts = getattr(self.config, "n_routed_experts", None)
        if n_routed_experts is not None:
            return n_routed_experts

        # For puzzle models, get MAX from all MoE blocks in block_configs
        # (different layers may have different expert counts)
        max_experts = 0
        block_configs = getattr(self.config, "block_configs", None)
        if block_configs:
            for block in block_configs:
                if isinstance(block, dict):
                    if block.get("block_type") == "moe":
                        max_experts = max(max_experts, block.get("n_routed_experts", 0))
                else:
                    # HF converts dicts to objects with attributes
                    if getattr(block, "block_type", "") == "moe":
                        max_experts = max(
                            max_experts, getattr(block, "n_routed_experts", 0)
                        )
        return max_experts

653
    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
654
655
        if self.has_moe:
            # (param_name, weight_name, expert_id, shard_id)
656
            expert_params_mapping = fused_moe_make_expert_params_mapping(
657
658
659
660
                # - 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
661
                self,
662
663
664
                ckpt_gate_proj_name="up_proj",
                ckpt_down_proj_name="down_proj",
                ckpt_up_proj_name="",
665
                num_experts=self._get_max_n_routed_experts(),
666
667
                num_redundant_experts=getattr(self, "num_redundant_experts", 0),
            )
668
669
670
671
672
673
674
675
676
677
678
679
680
            return expert_params_mapping

        return []

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]

        expert_params_mapping = self.get_expert_mapping()
681

Luis Vega's avatar
Luis Vega committed
682
683
684
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
685
            if "scale" in name or "zero_point" in name:
686
687
688
689
690
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue

691
692
693
694
            # Skip MTP/spec decode layers early (before stacked params mapping)
            if name.startswith("mtp."):
                continue

695
696
697
698
699
700
701
702
            # 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
703

704
705
706
                if is_pp_missing_parameter(name, self):
                    continue

Luis Vega's avatar
Luis Vega committed
707
708
                param = params_dict[name]
                weight_loader = param.weight_loader
709
710
711
                weight_loader(param, loaded_weight, shard_id)
                break

Luis Vega's avatar
Luis Vega committed
712
713
            # load other params
            else:
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
                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

752
753
754
                    if is_pp_missing_parameter(name, self):
                        continue

755
756
757
758
759
                    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
760
761
762
763
764

            loaded_params.add(name)
        return loaded_params


765
class NemotronHForCausalLM(
766
767
768
769
770
771
772
    nn.Module,
    HasInnerState,
    SupportsLoRA,
    SupportsPP,
    IsHybrid,
    SupportsQuant,
    MixtureOfExperts,
773
    SupportsMambaPrefixCaching,
774
):
775
776
777
    # Relevant only if self.has_moe is True
    is_non_gated_moe: bool = True

778
779
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={"backbone": "model"},
780
        orig_to_new_substr={"A_log": "A", "embeddings": "embed_tokens"},
781
782
    )

Luis Vega's avatar
Luis Vega committed
783
784
785
786
787
788
789
790
791
792
793
794
795
796
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }

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

797
798
799
    # Skip MTP (Multi-Token Prediction) layers during LoRA loading
    lora_skip_prefixes = ["mtp."]

800
801
802
803
804
805
806
807
808
809
810
    @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,
        )

811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
    @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
828
        intermediate_size = hf_config.mamba_num_heads * hf_config.mamba_head_dim
829

830
        return MambaStateShapeCalculator.mamba2_state_shape(
831
832
833
834
835
836
837
            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,
838
            num_spec=vllm_config.num_speculative_tokens,
839
840
        )

841
842
843
844
    @classmethod
    def get_mamba_state_copy_func(cls) -> tuple[MambaStateCopyFunc, MambaStateCopyFunc]:
        return MambaStateCopyFuncCalculator.mamba2_state_copy_func()

Luis Vega's avatar
Luis Vega committed
845
846
847
848
    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
849

Luis Vega's avatar
Luis Vega committed
850
851
852
853
854
855
856
        scheduler_config = vllm_config.scheduler_config

        self.quant_config = vllm_config.quant_config

        super().__init__()
        self.config = config
        self.scheduler_config = scheduler_config
857
858
859
        self.model = NemotronHModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
860

Luis Vega's avatar
Luis Vega committed
861
        self.lm_head = ParallelLMHead(
862
            config.vocab_size,
Luis Vega's avatar
Luis Vega committed
863
            config.hidden_size,
864
            prefix=maybe_prefix(prefix, "lm_head"),
Luis Vega's avatar
Luis Vega committed
865
866
        )

867
        self.logits_processor = LogitsProcessor(config.vocab_size)
Luis Vega's avatar
Luis Vega committed
868

869
870
871
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )
Luis Vega's avatar
Luis Vega committed
872

873
874
875
876
877
        # Set MoE hyperparameters
        if self.model.has_moe:
            self.expert_weights = []
            self.num_expert_groups = config.n_group

878
            self.moe_layers = []
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
            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 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()

912
913
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
Luis Vega's avatar
Luis Vega committed
914

915
916
    def forward(
        self,
917
        input_ids: torch.Tensor | None,
918
        positions: torch.Tensor,
919
920
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
921
922
923
924
925
        **kwargs,
    ):
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
Luis Vega's avatar
Luis Vega committed
926
927
928
929
930
931

        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
932
    ) -> torch.Tensor | None:
933
        logits = self.logits_processor(self.lm_head, hidden_states)
Luis Vega's avatar
Luis Vega committed
934
935
        return logits

936
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
shaharmor98's avatar
shaharmor98 committed
937
        loader = AutoWeightsLoader(self, skip_prefixes=["mtp"])
938
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)