glm4_moe.py 27.5 KB
Newer Older
Yuxuan Zhang's avatar
Yuxuan Zhang committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

# Copyright 2025 The ZhipuAI Team.
# Copyright 2023 The vLLM team.
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# 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.
24
"""Inference-only GLM-4.5, GLM-4.6 model compatible with HuggingFace weights."""
25

Yuxuan Zhang's avatar
Yuxuan Zhang committed
26
27
import typing
from collections.abc import Callable, Iterable
28
from itertools import islice
Yuxuan Zhang's avatar
Yuxuan Zhang committed
29
30
31
32
from typing import Any, Optional, Union

import torch
from torch import nn
33
from transformers.models.glm4_moe import Glm4MoeConfig
Yuxuan Zhang's avatar
Yuxuan Zhang committed
34
35
36
37

from vllm.attention import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
38
39
40
41
42
from vllm.distributed import (
    get_ep_group,
    get_pp_group,
    get_tensor_model_parallel_world_size,
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
43
44
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
45
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
Yuxuan Zhang's avatar
Yuxuan Zhang committed
46
from vllm.model_executor.layers.layernorm import RMSNorm
47
48
49
50
51
from vllm.model_executor.layers.linear import (
    MergedColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
52
53
54
55
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
56
57
58
    ParallelLMHead,
    VocabParallelEmbedding,
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
59
from vllm.model_executor.model_loader.weight_utils import (
60
61
62
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
63
64
from vllm.sequence import IntermediateTensors

65
from .interfaces import SupportsLoRA, SupportsPP
66
67
68
69
70
71
72
73
from .utils import (
    AutoWeightsLoader,
    PPMissingLayer,
    is_pp_missing_parameter,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89

logger = init_logger(__name__)


class Glm4MoeMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: bool = True,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
90
91
            hidden_size,
            [intermediate_size] * 2,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
92
93
            bias=False,
            quant_config=quant_config,
94
95
96
97
98
99
100
101
102
103
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=f"{prefix}.down_proj",
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
104
        if hidden_act != "silu":
105
106
107
            raise ValueError(
                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
108
109
110
111
112
113
114
115
116
117
118
119
        self.act_fn = SiluAndMul()

    def forward(self, x):
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class Glm4MoE(nn.Module):
    def __init__(
        self,
120
        config: Glm4MoeConfig,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        enable_eplb: bool = False,
    ):
        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

        if config.hidden_act != "silu":
136
137
138
139
            raise ValueError(
                f"Unsupported activation: {config.hidden_act}. "
                "Only silu is supported for now."
            )
140
141
142
143
144
145
146
147
148
        # NOTE In the transformers implementation, the gate isn't an nn.Linear,
        # so we cannot use ReplicatedLinear here.
        # See: https://github.com/huggingface/transformers/blob/v4.55.1/src/transformers/models/glm4_moe/modeling_glm4_moe.py#L260
        self.gate = nn.Linear(
            config.hidden_size,
            config.n_routed_experts,
            bias=False,
            dtype=torch.float32,
        )
149
        self.gate.e_score_correction_bias = nn.Parameter(
150
151
            torch.empty(config.n_routed_experts, dtype=torch.float32)
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
152
153
154

        # Load balancing settings.
        vllm_config = get_current_vllm_config()
155
        eplb_config = vllm_config.parallel_config.eplb_config
Yuxuan Zhang's avatar
Yuxuan Zhang committed
156
157
        self.enable_eplb = enable_eplb

158
        self.n_redundant_experts = eplb_config.num_redundant_experts
Yuxuan Zhang's avatar
Yuxuan Zhang committed
159
        self.n_logical_experts = self.n_routed_experts
160
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
Yuxuan Zhang's avatar
Yuxuan Zhang committed
161
162
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

163
164
165
166
        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
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
167
168

        if config.n_shared_experts is not None:
169
            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
Yuxuan Zhang's avatar
Yuxuan Zhang committed
170
171
172
173
174
            self.shared_experts = Glm4MoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
175
                reduce_results=False,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
176
177
                prefix=f"{prefix}.shared_experts",
            )
178
        else:
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
            self.shared_experts = None

        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",
            # we do scaling outside, set factor to 1.0 to avoid double mul
            routed_scaling_factor=1.0,
            e_score_correction_bias=self.gate.e_score_correction_bias,
            enable_eplb=self.enable_eplb,
            num_redundant_experts=self.n_redundant_experts,
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
201
202
203
204
205

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

206
        # router_logits: (num_tokens, n_experts)
207
        router_logits = self.gate(hidden_states.to(dtype=torch.float32))
208

209
210
211
        fused_moe_out = self.experts(
            hidden_states=hidden_states, router_logits=router_logits
        )
212
213
214
215

        if self.shared_experts is not None:
            shared_output, final_hidden_states = fused_moe_out
            assert shared_output is not None
216
217
218
            final_hidden_states = (
                final_hidden_states * self.routed_scaling_factor + shared_output
            )
219
220
221
        else:
            final_hidden_states = fused_moe_out * self.routed_scaling_factor

Yuxuan Zhang's avatar
Yuxuan Zhang committed
222
        if self.tp_size > 1:
223
224
225
            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
                final_hidden_states
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
226
227
228
229
230
231
        return final_hidden_states.view(num_tokens, hidden_dim)


class Glm4MoeAttention(nn.Module):
    def __init__(
        self,
232
        config: Glm4MoeConfig,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        max_position_embeddings: int = 131072,
        head_dim: Optional[int] = None,
        rms_norm_eps: float = 1e-05,
        qkv_bias: bool = False,
        use_qk_norm: bool = False,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_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)
        self.head_dim = head_dim or (hidden_size // self.total_num_heads)
        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.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings
        self.use_qk_norm = use_qk_norm

271
272
273
274
275
276
277
278
279
        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=qkv_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
280

281
282
283
284
285
286
287
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
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
316
317
318
319

        partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5)
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
            partial_rotary_factor=partial_rotary_factor,
        )
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )

        if self.use_qk_norm:
            self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
            self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        if self.use_qk_norm:
320
321
322
323
324
325
            q = self.q_norm(q.reshape(-1, self.num_heads, self.head_dim)).reshape(
                q.shape
            )
            k = self.k_norm(k.reshape(-1, self.num_kv_heads, self.head_dim)).reshape(
                k.shape
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
326
327
328
329
330
331
332
333
334
335

        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output


class Glm4MoeDecoderLayer(nn.Module):
    def __init__(
        self,
336
        config: Glm4MoeConfig,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
337
338
339
340
341
342
343
344
345
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        enable_eplb: bool = False,
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
346
        max_position_embeddings = getattr(config, "max_position_embeddings", 131072)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
347
348
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
349
        layer_idx = int(prefix.split(sep=".")[-1])
Yuxuan Zhang's avatar
Yuxuan Zhang committed
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
        self.layer_idx = layer_idx

        self.self_attn = Glm4MoeAttention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            head_dim=config.head_dim,
            rms_norm_eps=config.rms_norm_eps,
            qkv_bias=config.attention_bias,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
            use_qk_norm=config.use_qk_norm,
        )

369
370
371
372
        if (
            config.n_routed_experts is not None
            and layer_idx >= config.first_k_dense_replace
        ):
Yuxuan Zhang's avatar
Yuxuan Zhang committed
373
374
375
376
377
378
379
            self.mlp = Glm4MoE(
                config=config,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
                enable_eplb=enable_eplb,
            )
        else:
380
381
382
383
384
385
386
387
388
389
390
391
            self.mlp = Glm4MoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )

        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
392
393
394
395
396
397
398
399
400
401
402
403
        self.routed_scaling_factor = config.routed_scaling_factor

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
404
405
406
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
        hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states)
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
407
408
409
410
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual


411
412
413
414
415
416
@support_torch_compile(
    dynamic_arg_dims={
        "input_ids": 0,
        "positions": -1,
        "intermediate_tensors": 0,
        "inputs_embeds": 0,
417
418
    }
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
419
420
421
422
423
424
425
426
427
428
429
430
431
432
class Glm4MoeModel(nn.Module):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        enable_eplb = vllm_config.parallel_config.enable_eplb
        self.config = config

        self.vocab_size = config.vocab_size

        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
433
434
                config.vocab_size, config.hidden_size, prefix=f"{prefix}.embed_tokens"
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
435
436
437
438
439
440
441
442
443
444
445
446
        else:
            self.embed_tokens = PPMissingLayer()

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: Glm4MoeDecoderLayer(
                config=config,
                cache_config=cache_config,
                quant_config=quant_config,
                prefix=prefix,
                enable_eplb=enable_eplb,
            ),
447
448
            prefix=f"{prefix}.layers",
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
449
450
451
452
453

        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
454
455
456
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        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"]

479
        for layer in islice(self.layers, self.start_layer, self.end_layer):
Yuxuan Zhang's avatar
Yuxuan Zhang committed
480
481
482
            hidden_states, residual = layer(positions, hidden_states, residual)

        if not get_pp_group().is_last_rank:
483
484
485
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
486
487
488
489
490

        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states

    def make_empty_intermediate_tensors(
491
492
493
494
495
496
497
498
499
500
501
502
        self, batch_size: int, dtype: torch.dtype, device: torch.device
    ) -> IntermediateTensors:
        return IntermediateTensors(
            {
                "hidden_states": torch.zeros(
                    (batch_size, self.config.hidden_size), dtype=dtype, device=device
                ),
                "residual": torch.zeros(
                    (batch_size, self.config.hidden_size), dtype=dtype, device=device
                ),
            }
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
503

504
505
506
    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
507
        return SharedFusedMoE.make_expert_params_mapping(
508
509
510
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
511
512
            num_experts=self.config.n_routed_experts,
        )
513

514
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
515
516
517
518
519
520
521
522
523
524
525
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
526
        expert_params_mapping = self.get_expert_mapping()
Yuxuan Zhang's avatar
Yuxuan Zhang committed
527
528
529
530
        for name, loaded_weight in weights:
            spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
            if spec_layer is not None:
                continue
531
            for param_name, weight_name, shard_id in stacked_params_mapping:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
532
533
534
535
536
537
538
539
540
                # Skip non-stacked layers and experts (experts handled below).
                if weight_name not in name:
                    continue
                # We have mlp.experts[0].gate_proj in the checkpoint.
                # Since we handle the experts below in expert_params_mapping,
                # we need to skip here BEFORE we update the name, otherwise
                # name will be updated to mlp.experts[0].gate_up_proj, which
                # will then be updated below in expert_params_mapping
                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
541
                if ("mlp.experts." in name) and name not in params_dict:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
                    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
                if is_pp_missing_parameter(name, self):
                    continue

                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                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.
576
577
578
579
580
581
582
583
584
585
586
                    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,
                    )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
                    if success:
                        name = name_mapped
                        break
                else:
                    if is_expert_weight:
                        # We've checked that this is an expert weight
                        # However it's not mapped locally to this rank
                        # So we simply skip it
                        continue

                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue

                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

                    if is_pp_missing_parameter(name, self):
                        continue

                    param = params_dict[name]
610
611
612
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
613
614
615
616
617
618
                    weight_loader(param, loaded_weight)
            loaded_params.add(name)

        return loaded_params


619
class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
Yuxuan Zhang's avatar
Yuxuan Zhang committed
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    fall_back_to_pt_during_load = False

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        self.config = config
        self.quant_config = quant_config
640
641
642
        self.model = Glm4MoeModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
643
        if get_pp_group().is_last_rank:
644
645
646
647
648
649
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
650
651
652
653
        else:
            self.lm_head = PPMissingLayer()
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
654
655
            self.model.make_empty_intermediate_tensors
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
656
657
658
        self.expert_weights = []

        # Set MoE hyperparameters
659
        self.num_moe_layers = config.num_hidden_layers - config.first_k_dense_replace
Yuxuan Zhang's avatar
Yuxuan Zhang committed
660
661
        self.num_expert_groups = config.n_group

662
        self.moe_layers: list[SharedFusedMoE] = []
663
        example_moe = None
Yuxuan Zhang's avatar
Yuxuan Zhang committed
664
        for layer in self.model.layers:
665
666
667
            if isinstance(layer, PPMissingLayer):
                continue

Yuxuan Zhang's avatar
Yuxuan Zhang committed
668
669
            assert isinstance(layer, Glm4MoeDecoderLayer)
            if isinstance(layer.mlp, Glm4MoE):
670
671
                # Pick last one layer since the first ones may be dense layers.
                example_moe = layer.mlp
Yuxuan Zhang's avatar
Yuxuan Zhang committed
672
673
                self.moe_layers.append(layer.mlp.experts)

674
675
676
        if example_moe is None:
            raise RuntimeError("No Glm4MoE layer found in model.layers.")

Yuxuan Zhang's avatar
Yuxuan Zhang committed
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
        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
        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 get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
710
711
712
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
713
714
715
716
717
718
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> Optional[torch.Tensor]:
719
        logits = self.logits_processor(self.lm_head, hidden_states)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
720
721
        return logits

722
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
723
724
725
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)

726
727
728
    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        return self.model.get_expert_mapping()

Yuxuan Zhang's avatar
Yuxuan Zhang committed
729

730
731
732
733
734
735
def get_spec_layer_idx_from_weight_name(
    config: Glm4MoeConfig, weight_name: str
) -> Optional[int]:
    if hasattr(config, "num_nextn_predict_layers") and (
        config.num_nextn_predict_layers > 0
    ):
Yuxuan Zhang's avatar
Yuxuan Zhang committed
736
737
        layer_idx = config.num_hidden_layers
        for i in range(config.num_nextn_predict_layers):
738
            if f"layers.{layer_idx + i}." in weight_name:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
739
740
                return layer_idx + i
    return None