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


class NemotronHMLP(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
shaharmor98's avatar
shaharmor98 committed
91
        hidden_size: int,
92
        intermediate_size: int,
93
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
94
        bias: bool = False,
95
96
        reduce_results: bool = True,
        is_sequence_parallel: bool = False,
97
        prefix: str = "",
Luis Vega's avatar
Luis Vega committed
98
99
    ) -> None:
        super().__init__()
100

101
        self.up_proj = ColumnParallelLinear(
shaharmor98's avatar
shaharmor98 committed
102
            input_size=hidden_size,
103
            output_size=intermediate_size,
Luis Vega's avatar
Luis Vega committed
104
105
            bias=bias,
            quant_config=quant_config,
106
            disable_tp=is_sequence_parallel,
107
            prefix=f"{prefix}.up_proj",
Luis Vega's avatar
Luis Vega committed
108
109
        )
        self.down_proj = RowParallelLinear(
110
            input_size=intermediate_size,
shaharmor98's avatar
shaharmor98 committed
111
            output_size=hidden_size,
Luis Vega's avatar
Luis Vega committed
112
113
            bias=bias,
            quant_config=quant_config,
114
115
            reduce_results=reduce_results,
            disable_tp=is_sequence_parallel,
116
            prefix=f"{prefix}.down_proj",
Luis Vega's avatar
Luis Vega committed
117
118
119
120
121
122
123
124
125
126
        )
        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


127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
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
144
145
146
147
        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
        )
148
149
150

        self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe

151
        self.gate = GateLinear(
152
153
            config.hidden_size,
            config.n_routed_experts,
154
155
            out_dtype=torch.float32,
            force_fp32_compute=True,
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
            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
184
                hidden_size=config.hidden_size,
185
186
187
188
189
190
191
                intermediate_size=intermediate_size,
                quant_config=quant_config,
                reduce_results=False,
                is_sequence_parallel=self.is_sequence_parallel,
                prefix=f"{prefix}.shared_experts",
            )

192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
        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

213
        self.experts = SharedFusedMoE(
214
            shared_experts=self.shared_experts,
215
216
            num_experts=config.n_routed_experts,
            top_k=config.num_experts_per_tok,
shaharmor98's avatar
shaharmor98 committed
217
            hidden_size=self.moe_hidden_size,
218
219
220
221
222
223
224
225
226
227
228
229
230
231
            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,
232
            routed_input_transform=self.fc1_latent_proj,
233
234
235
            routed_output_transform=self.fc2_latent_proj,
            routed_scaling_factor=self.routed_scaling_factor,
            apply_routed_scale_to_output=True,
236
            router_logits_dtype=self.gate.out_dtype,
237
238
239
240
241
242
243
244
245
246
        )

    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)
247
        router_logits, _ = self.gate(hidden_states)
248

249
        final_hidden_states = self.experts(
250
251
252
253
254
255
256
257
258
259
260
261
            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
262
263
264
265
266
class NemotronHMLPDecoderLayer(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
267
268
269
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
270
        parallel_config: ParallelConfig | None = None,
Luis Vega's avatar
Luis Vega committed
271
272
273
274
275
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config

276
277
        hybrid_override_pattern = config.hybrid_override_pattern
        mlp_index = hybrid_override_pattern[: layer_idx + 1].count("-") - 1
278
279
280
281
282
        # 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

283
284
285
286
287
288
289
290
        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

291
292
        self.mixer = NemotronHMLP(
            config,
shaharmor98's avatar
shaharmor98 committed
293
            hidden_size=config.hidden_size,
294
            intermediate_size=intermediate_size,
295
296
297
298
            quant_config=quant_config,
            bias=config.mlp_bias,
            prefix=f"{prefix}.mixer",
        )
Luis Vega's avatar
Luis Vega committed
299

300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
        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
332
        # Get per-layer config for heterogeneous models if exists
333
334
335
        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

336
        self.mixer = NemotronHMoE(
337
            layer_config,
338
339
340
341
342
343
            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
344
345
346
347

    def forward(
        self,
        hidden_states: torch.Tensor,
348
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
        **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,
366
367
368
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
369
        parallel_config: ParallelConfig | None = None,
Luis Vega's avatar
Luis Vega committed
370
371
372
373
374
375
376
377
        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,
378
            intermediate_size=config.mamba_num_heads * config.mamba_head_dim,
Luis Vega's avatar
Luis Vega committed
379
380
381
382
383
            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,
384
            rms_norm_eps=config.layer_norm_epsilon,
Luis Vega's avatar
Luis Vega committed
385
            activation=config.mamba_hidden_act,
386
387
            model_config=model_config,
            cache_config=cache_config,
Luis Vega's avatar
Luis Vega committed
388
            quant_config=quant_config,
389
            prefix=f"{prefix}.mixer",
Luis Vega's avatar
Luis Vega committed
390
391
        )

392
        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
Luis Vega's avatar
Luis Vega committed
393
394
395
396

    def forward(
        self,
        hidden_states: torch.Tensor,
397
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
398
399
400
401
402
403
404
405
        **kwargs,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.norm(hidden_states)
        else:
            hidden_states, residual = self.norm(hidden_states, residual)

406
        output = self.mixer(hidden_states)
407
        return output, residual
Luis Vega's avatar
Luis Vega committed
408
409
410
411
412
413
414


class NemotronHAttention(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
415
416
417
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
Luis Vega's avatar
Luis Vega committed
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
        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)
436
437
438
439
        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
440
441
442
443
444
445
446
447
448
449
450
        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,
451
            prefix=f"{prefix}.qkv_proj",
Luis Vega's avatar
Luis Vega committed
452
453
454
455
456
457
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            config.hidden_size,
            bias=False,
            quant_config=quant_config,
458
            prefix=f"{prefix}.o_proj",
Luis Vega's avatar
Luis Vega committed
459
460
        )

461
462
463
        # Get per-layer sliding window from config (for heterogeneous models)
        sliding_window = getattr(config, "sliding_window", None)

Luis Vega's avatar
Luis Vega committed
464
465
466
467
468
469
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
470
            quant_config=quant_config,
Luis Vega's avatar
Luis Vega committed
471
            prefix=f"{prefix}.attn",
472
            per_layer_sliding_window=sliding_window,
Luis Vega's avatar
Luis Vega committed
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
        )

    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,
492
493
494
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
495
        parallel_config: ParallelConfig | None = None,
Luis Vega's avatar
Luis Vega committed
496
497
498
499
        prefix: str = "",
    ) -> None:
        super().__init__()

Jiayi Yan's avatar
Jiayi Yan committed
500
        # Get per-layer config for heterogeneous models if exists
501
502
503
        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
504
        self.mixer = NemotronHAttention(
505
            layer_config,
Luis Vega's avatar
Luis Vega committed
506
            layer_idx,
507
            model_config,
Luis Vega's avatar
Luis Vega committed
508
509
            cache_config,
            quant_config,
510
            prefix=f"{prefix}.mixer",
Luis Vega's avatar
Luis Vega committed
511
512
        )

513
        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
Luis Vega's avatar
Luis Vega committed
514
515
516
517
518

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
519
        residual: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
        **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,
536
    "E": NemotronHMoEDecoderLayer,
Luis Vega's avatar
Luis Vega committed
537
538
539
}


540
@support_torch_compile
Luis Vega's avatar
Luis Vega committed
541
542
543
544
545
class NemotronHModel(nn.Module):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config: NemotronHConfig = vllm_config.model_config.hf_config
546
        model_config = vllm_config.model_config
Luis Vega's avatar
Luis Vega committed
547
548
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
549
        parallel_config = vllm_config.parallel_config
Luis Vega's avatar
Luis Vega committed
550
551

        self.config = config
552
553

        self.vocab_size = config.vocab_size
Luis Vega's avatar
Luis Vega committed
554
555
556
557
558
559

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

560
561
        self.has_moe = "E" in config.hybrid_override_pattern

Luis Vega's avatar
Luis Vega committed
562
563
564
        def get_layer(prefix: str):
            layer_idx = int(prefix.rsplit(".", 1)[1])
            layer_class = ALL_DECODER_LAYER_TYPES[
565
566
                config.hybrid_override_pattern[layer_idx]
            ]
Luis Vega's avatar
Luis Vega committed
567
            return layer_class(
568
569
570
571
                config=config,
                layer_idx=layer_idx,
                model_config=model_config,
                cache_config=cache_config,
Luis Vega's avatar
Luis Vega committed
572
                quant_config=quant_config,
573
                parallel_config=parallel_config,
Luis Vega's avatar
Luis Vega committed
574
575
576
577
                prefix=prefix,
            )

        self.start_layer, self.end_layer, self.layers = make_layers(
578
579
            len(config.hybrid_override_pattern), get_layer, prefix=f"{prefix}.layers"
        )
580
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
581
582
            ["hidden_states", "residual"], config.hidden_size
        )
Luis Vega's avatar
Luis Vega committed
583

584
        self.norm_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
Luis Vega's avatar
Luis Vega committed
585

586
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
Luis Vega's avatar
Luis Vega committed
587
588
589
590
        return self.embed_tokens(input_ids)

    def forward(
        self,
591
        input_ids: torch.Tensor | None,
Luis Vega's avatar
Luis Vega committed
592
        positions: torch.Tensor,
593
594
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
595
    ) -> torch.Tensor | IntermediateTensors:
Luis Vega's avatar
Luis Vega committed
596
597
598
599
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
600
                hidden_states = self.embed_input_ids(input_ids)
Luis Vega's avatar
Luis Vega committed
601
602
603
604
605
606
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

607
        for layer in islice(self.layers, self.start_layer, self.end_layer):
Luis Vega's avatar
Luis Vega committed
608
609
610
611
612
613
614
            hidden_states, residual = layer(
                positions=positions,
                hidden_states=hidden_states,
                residual=residual,
            )

        if not get_pp_group().is_last_rank:
615
616
617
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
Luis Vega's avatar
Luis Vega committed
618
619
620
        hidden_states, _ = self.norm_f(hidden_states, residual)
        return hidden_states

621
622
623
    def is_spec_layer(self, config: NemotronHConfig, weight_name: str) -> bool:
        return weight_name.startswith("mtp.")

624
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
    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

652
    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
653
654
        if self.has_moe:
            # (param_name, weight_name, expert_id, shard_id)
655
            expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
656
657
658
659
                # - 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
660
                self,
661
662
663
                ckpt_gate_proj_name="up_proj",
                ckpt_down_proj_name="down_proj",
                ckpt_up_proj_name="",
664
                num_experts=self._get_max_n_routed_experts(),
665
666
                num_redundant_experts=getattr(self, "num_redundant_experts", 0),
            )
667
668
669
670
671
672
673
674
675
676
677
678
679
            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()
680

Luis Vega's avatar
Luis Vega committed
681
682
683
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
684
            if "scale" in name or "zero_point" in name:
685
686
687
688
689
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue

690
691
692
693
            # Skip MTP/spec decode layers early (before stacked params mapping)
            if name.startswith("mtp."):
                continue

694
695
696
697
698
699
700
701
            # 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
702

703
704
705
                if is_pp_missing_parameter(name, self):
                    continue

Luis Vega's avatar
Luis Vega committed
706
707
                param = params_dict[name]
                weight_loader = param.weight_loader
708
709
710
                weight_loader(param, loaded_weight, shard_id)
                break

Luis Vega's avatar
Luis Vega committed
711
712
            # load other params
            else:
713
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
                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

751
752
753
                    if is_pp_missing_parameter(name, self):
                        continue

754
755
756
757
758
                    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
759
760
761
762
763

            loaded_params.add(name)
        return loaded_params


764
class NemotronHForCausalLM(
765
766
767
768
769
770
771
    nn.Module,
    HasInnerState,
    SupportsLoRA,
    SupportsPP,
    IsHybrid,
    SupportsQuant,
    MixtureOfExperts,
772
    SupportsMambaPrefixCaching,
773
):
774
775
776
    # Relevant only if self.has_moe is True
    is_non_gated_moe: bool = True

777
778
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={"backbone": "model"},
779
        orig_to_new_substr={"A_log": "A", "embeddings": "embed_tokens"},
780
781
    )

Luis Vega's avatar
Luis Vega committed
782
783
784
785
786
787
788
789
790
791
792
793
794
795
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }

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

796
797
798
    # Skip MTP (Multi-Token Prediction) layers during LoRA loading
    lora_skip_prefixes = ["mtp."]

799
800
801
802
803
804
805
806
807
808
809
    @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,
        )

810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
    @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
827
        intermediate_size = hf_config.mamba_num_heads * hf_config.mamba_head_dim
828

829
        return MambaStateShapeCalculator.mamba2_state_shape(
830
831
832
833
834
835
836
            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,
837
            num_spec=vllm_config.num_speculative_tokens,
838
839
        )

840
841
842
843
    @classmethod
    def get_mamba_state_copy_func(cls) -> tuple[MambaStateCopyFunc, MambaStateCopyFunc]:
        return MambaStateCopyFuncCalculator.mamba2_state_copy_func()

Luis Vega's avatar
Luis Vega committed
844
845
846
847
    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
848

Luis Vega's avatar
Luis Vega committed
849
850
851
852
853
854
855
        scheduler_config = vllm_config.scheduler_config

        self.quant_config = vllm_config.quant_config

        super().__init__()
        self.config = config
        self.scheduler_config = scheduler_config
856
857
858
        self.model = NemotronHModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
859

Luis Vega's avatar
Luis Vega committed
860
        self.lm_head = ParallelLMHead(
861
            config.vocab_size,
Luis Vega's avatar
Luis Vega committed
862
            config.hidden_size,
863
            prefix=maybe_prefix(prefix, "lm_head"),
Luis Vega's avatar
Luis Vega committed
864
865
        )

866
        self.logits_processor = LogitsProcessor(config.vocab_size)
Luis Vega's avatar
Luis Vega committed
867

868
869
870
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )
Luis Vega's avatar
Luis Vega committed
871

872
873
874
875
876
        # Set MoE hyperparameters
        if self.model.has_moe:
            self.expert_weights = []
            self.num_expert_groups = config.n_group

877
            self.moe_layers = []
878
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
            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()

911
912
    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
913

914
915
    def forward(
        self,
916
        input_ids: torch.Tensor | None,
917
        positions: torch.Tensor,
918
919
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
920
921
922
923
924
        **kwargs,
    ):
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
Luis Vega's avatar
Luis Vega committed
925
926
927
928
929
930

        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
931
    ) -> torch.Tensor | None:
932
        logits = self.logits_processor(self.lm_head, hidden_states)
Luis Vega's avatar
Luis Vega committed
933
934
        return logits

935
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
shaharmor98's avatar
shaharmor98 committed
936
        loader = AutoWeightsLoader(self, skip_prefixes=["mtp"])
937
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)