glm4_moe.py 31.2 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
25
"""Inference-only GLM-4.5, GLM-4.6, GLM-4.7 model
compatible with HuggingFace weights."""
26

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

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

35
36
from vllm import envs
from vllm.attention.layer import Attention, FusedQkvSplitRmsNormRopeAttention
Yuxuan Zhang's avatar
Yuxuan Zhang committed
37
38
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
39
40
41
42
43
from vllm.distributed import (
    get_ep_group,
    get_pp_group,
    get_tensor_model_parallel_world_size,
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
44
45
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
46
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
47
from vllm.model_executor.layers.layernorm import RMSNorm, FusedRMSNormQuant
48
49
50
51
52
from vllm.model_executor.layers.linear import (
    MergedColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
53
54
55
56
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 (
57
58
59
    ParallelLMHead,
    VocabParallelEmbedding,
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
60
from vllm.model_executor.model_loader.weight_utils import (
61
62
63
    default_weight_loader,
    maybe_remap_kv_scale_name,
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
64
65
from vllm.sequence import IntermediateTensors

66
from .interfaces import MixtureOfExperts, SupportsLoRA, SupportsPP
67
68
69
70
71
72
73
74
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
75
76
77
78
79
80
81
82
83
84

logger = init_logger(__name__)


class Glm4MoeMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
85
        quant_config: QuantizationConfig | None = None,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
86
87
88
89
90
        reduce_results: bool = True,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
91
92
            hidden_size,
            [intermediate_size] * 2,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
93
94
            bias=False,
            quant_config=quant_config,
95
96
97
98
99
100
101
102
103
104
            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
105
        if hidden_act != "silu":
106
107
108
            raise ValueError(
                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
109
110
        self.act_fn = SiluAndMul()

111
112
    def forward(self, x, 
                iqis: tuple[torch.Tensor, torch.Tensor] | None = None):
113
114
        if envs.USE_FUSED_SILU_MUL_QUANT and iqis is not None:
            gate_up, _ = self.gate_up_proj(x, iqis=iqis)
115
116
117
118
            from lmslim.quantize.quant_ops import lm_fuse_silu_mul_quant
            xq, xs = lm_fuse_silu_mul_quant(gate_up)
            x, _ = self.down_proj(gate_up, iqis=(xq, xs))
        else:
119
            gate_up, _ = self.gate_up_proj(x)
120
121
            x = self.act_fn(gate_up)
            x, _ = self.down_proj(x)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
122
123
124
125
126
127
        return x


class Glm4MoE(nn.Module):
    def __init__(
        self,
128
        config: Glm4MoeConfig,
129
        quant_config: QuantizationConfig | None = None,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
130
131
132
133
134
135
136
137
        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
138
        self.ep_rank = get_ep_group().rank_in_group
Yuxuan Zhang's avatar
Yuxuan Zhang committed
139
140
141
142
143
        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":
144
145
146
147
            raise ValueError(
                f"Unsupported activation: {config.hidden_act}. "
                "Only silu is supported for now."
            )
148
149
150
151
152
153
154
155
156
        # 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,
        )
157
        self.gate.e_score_correction_bias = nn.Parameter(
158
159
            torch.empty(config.n_routed_experts, dtype=torch.float32)
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
160
161
162

        # Load balancing settings.
        vllm_config = get_current_vllm_config()
163
        eplb_config = vllm_config.parallel_config.eplb_config
Yuxuan Zhang's avatar
Yuxuan Zhang committed
164
165
        self.enable_eplb = enable_eplb

166
        self.n_redundant_experts = eplb_config.num_redundant_experts
Yuxuan Zhang's avatar
Yuxuan Zhang committed
167
        self.n_logical_experts = self.n_routed_experts
168
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
Yuxuan Zhang's avatar
Yuxuan Zhang committed
169
170
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

171
172
173
174
        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
175
176

        if config.n_shared_experts is not None:
177
            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
Yuxuan Zhang's avatar
Yuxuan Zhang committed
178
179
180
181
182
            self.shared_experts = Glm4MoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
183
                reduce_results=False,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
184
185
                prefix=f"{prefix}.shared_experts",
            )
186
        else:
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
            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,
208
            router_logits_dtype=torch.float32,
209
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
210

211
212
213
214
    def forward(self, 
                hidden_states: torch.Tensor,
                iqis: tuple[torch.Tensor, torch.Tensor] | None = None
                ) -> torch.Tensor:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
215
216
217
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)

218
        # router_logits: (num_tokens, n_experts)
219
        router_logits = self.gate(hidden_states.to(dtype=torch.float32))
220
221
222
223
224
225
226
        if envs.USE_FUSED_RMS_QUANT and iqis is not None:
            fused_moe_out = self.experts(
                hidden_states=hidden_states, router_logits=router_logits,
                iqis=iqis)
        else:
            fused_moe_out = self.experts(
            hidden_states=hidden_states, router_logits=router_logits)
227
228
229
230

        if self.shared_experts is not None:
            shared_output, final_hidden_states = fused_moe_out
            assert shared_output is not None
231
232
233
            final_hidden_states = (
                final_hidden_states * self.routed_scaling_factor + shared_output
            )
234
235
236
        else:
            final_hidden_states = fused_moe_out * self.routed_scaling_factor

Yuxuan Zhang's avatar
Yuxuan Zhang committed
237
        if self.tp_size > 1:
238
239
240
            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
                final_hidden_states
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
241
242
243
244
245
246
        return final_hidden_states.view(num_tokens, hidden_dim)


class Glm4MoeAttention(nn.Module):
    def __init__(
        self,
247
        config: Glm4MoeConfig,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
248
249
250
251
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position_embeddings: int = 131072,
252
        head_dim: int | None = None,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
253
254
255
        rms_norm_eps: float = 1e-05,
        qkv_bias: bool = False,
        use_qk_norm: bool = False,
256
257
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
        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.max_position_embeddings = max_position_embeddings
        self.use_qk_norm = use_qk_norm

283
284
285
286
287
288
289
290
291
        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
292

293
294
295
296
297
298
299
        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
300

301
        config.rope_parameters.setdefault("partial_rotary_factor", 0.5)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
302
303
304
        self.rotary_emb = get_rope(
            self.head_dim,
            max_position=max_position_embeddings,
305
            rope_parameters=config.rope_parameters,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
306
        )
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327

        if not envs.VLLM_V1_USE_FUSED_QKV_SPLIT_RMS_ROPE_KVSTORE:
            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",
            )
        else:
            self.attn = FusedQkvSplitRmsNormRopeAttention(
                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",
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
328
329
330
331
332
333
334
335
336

        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,
337
        iqis: tuple[torch.Tensor, torch.Tensor] | None = None
Yuxuan Zhang's avatar
Yuxuan Zhang committed
338
    ) -> torch.Tensor:
339
        qkv, _ = self.qkv_proj(hidden_states, iqis=iqis)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
340

341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
        if not envs.VLLM_V1_USE_FUSED_QKV_SPLIT_RMS_ROPE_KVSTORE:
            q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
            if self.use_qk_norm:
                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
                )

            q, k = self.rotary_emb(positions, q, k)
            attn_output = self.attn(q, k, v)
        else:
            cos_sin_cache = self.rotary_emb.cos_sin_cache
            if (cos_sin_cache.device != qkv.device
                    or cos_sin_cache.dtype != qkv.dtype):
                cos_sin_cache = cos_sin_cache.to(qkv.device,
                                                dtype=qkv.dtype,
                                                non_blocking=True)
                # Persist the converted cache so we don't re-copy/re-allocate
                # on every forward when the original buffer starts on CPU.
                self.rotary_emb.cos_sin_cache = cos_sin_cache
            attn_output = self.attn(qkv,
                                    positions,
                                    cos_sin_cache,
                                    self.q_norm.weight,
                                    self.k_norm.weight,
                                    self.q_norm.variance_epsilon,
                                    is_neox=self.rotary_emb.is_neox_style)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
370
371
372
373
374
375
376
        output, _ = self.o_proj(attn_output)
        return output


class Glm4MoeDecoderLayer(nn.Module):
    def __init__(
        self,
377
        config: Glm4MoeConfig,
378
379
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
380
381
382
383
384
        prefix: str = "",
        enable_eplb: bool = False,
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
385
        max_position_embeddings = getattr(config, "max_position_embeddings", 131072)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
386
387
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
388
        layer_idx = int(prefix.split(sep=".")[-1])
Yuxuan Zhang's avatar
Yuxuan Zhang committed
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
        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,
            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,
        )

406
407
408
409
        if (
            config.n_routed_experts is not None
            and layer_idx >= config.first_k_dense_replace
        ):
Yuxuan Zhang's avatar
Yuxuan Zhang committed
410
411
412
413
414
415
416
            self.mlp = Glm4MoE(
                config=config,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
                enable_eplb=enable_eplb,
            )
        else:
417
418
419
420
421
422
423
424
            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",
            )

425
426
427
428
429
        if envs.USE_FUSED_RMS_QUANT:
            self.input_layernorm = FusedRMSNormQuant(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

430
        if envs.USE_FUSED_RMS_QUANT:
431
432
433
434
435
436
437
            self.post_attention_layernorm = FusedRMSNormQuant(
                config.hidden_size, eps=config.rms_norm_eps
            )
        else:
            self.post_attention_layernorm = RMSNorm(
                config.hidden_size, eps=config.rms_norm_eps
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
438
439
440
441
442
443
        self.routed_scaling_factor = config.routed_scaling_factor

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
444
        residual: torch.Tensor | None,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
445
    ) -> tuple[torch.Tensor, torch.Tensor]:
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
        if not envs.USE_FUSED_RMS_QUANT:
            if residual is None:
                residual = hidden_states
                hidden_states = self.input_layernorm(hidden_states)
            else:
                hidden_states, residual = self.input_layernorm(hidden_states, residual)
            hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states)
        else:
            if residual is None:
                residual = hidden_states.clone()
                i_q, i_s, _ = self.input_layernorm(x=hidden_states, 
                                                residual=None, 
                                                quant_dtype=torch.int8,
                                                update_input=False
                                                )
            else:
                i_q, i_s, residual = self.input_layernorm(x=hidden_states, 
                                                      residual=residual, 
                                                      quant_dtype=torch.int8,
                                                      update_input=False
                                                      )
            hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states, iqis=(i_q, i_s))
        
469
470
471
472
473
474
        if envs.USE_FUSED_RMS_QUANT:
            update_hs = True if isinstance(self.mlp, Glm4MoE) else False
            i_q, i_s, residual = self.post_attention_layernorm(x=hidden_states,
                                                               residual=residual,
                                                               quant_dtype=torch.int8,
                                                               update_input=update_hs)
475
            hidden_states = self.mlp(hidden_states, iqis=(i_q, i_s))
Yuxuan Zhang's avatar
Yuxuan Zhang committed
476
        else:
477
478
            hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
            hidden_states = self.mlp(hidden_states)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
479
480
481
        return hidden_states, residual


482
483
484
485
486
487
@support_torch_compile(
    dynamic_arg_dims={
        "input_ids": 0,
        "positions": -1,
        "intermediate_tensors": 0,
        "inputs_embeds": 0,
488
489
    }
)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
490
491
492
493
494
495
496
497
498
499
500
501
502
503
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(
504
505
                config.vocab_size, config.hidden_size, prefix=f"{prefix}.embed_tokens"
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
506
507
508
509
510
511
512
513
514
515
516
517
        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,
            ),
518
519
            prefix=f"{prefix}.layers",
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
520
521
522
523
524

        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
525
526
527
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
528

529
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
530
531
532
533
        return self.embed_tokens(input_ids)

    def forward(
        self,
zhuwenwen's avatar
zhuwenwen committed
534
        input_ids: torch.Tensor,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
535
        positions: torch.Tensor,
536
537
538
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
539
540
541
542
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
543
                hidden_states = self.embed_input_ids(input_ids)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
544
545
546
547
548
549
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

550
        for layer in islice(self.layers, self.start_layer, self.end_layer):
Yuxuan Zhang's avatar
Yuxuan Zhang committed
551
552
553
            hidden_states, residual = layer(positions, hidden_states, residual)

        if not get_pp_group().is_last_rank:
554
555
556
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
557
558
559
560

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

561
562
563
    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)
564
        return SharedFusedMoE.make_expert_params_mapping(
565
            self,
566
567
568
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
569
570
            num_experts=self.config.n_routed_experts,
        )
571

572
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
573
574
575
576
577
578
579
580
581
582
583
        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()
584
        expert_params_mapping = self.get_expert_mapping()
Yuxuan Zhang's avatar
Yuxuan Zhang committed
585
586
587
588
        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
589
            for param_name, weight_name, shard_id in stacked_params_mapping:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
590
591
592
593
594
595
596
597
598
                # 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.
599
                if ("mlp.experts." in name) and name not in params_dict:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
                    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.
634
635
636
637
638
639
640
641
642
643
644
                    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
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
                    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]
668
669
670
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
671
672
673
674
675
676
                    weight_loader(param, loaded_weight)
            loaded_params.add(name)

        return loaded_params


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
class Glm4MixtureOfExperts(MixtureOfExperts):
    def extract_moe_parameters(self, example_moe: Glm4MoE | None) -> None:
        if example_moe is None:
            raise RuntimeError("No Glm4MoE layer found in model.layers.")
        else:
            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 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 moe in self.moe_mlp_layers:
            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()


class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, Glm4MixtureOfExperts):
Yuxuan Zhang's avatar
Yuxuan Zhang committed
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
    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
726
727
728
        self.model = Glm4MoeModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
729
        if get_pp_group().is_last_rank:
730
731
732
733
734
735
            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
736
737
738
739
        else:
            self.lm_head = PPMissingLayer()
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
740
741
            self.model.make_empty_intermediate_tensors
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
742
743
744
        self.expert_weights = []

        # Set MoE hyperparameters
745
        self.num_moe_layers = config.num_hidden_layers - config.first_k_dense_replace
Yuxuan Zhang's avatar
Yuxuan Zhang committed
746
747
        self.num_expert_groups = config.n_group

748
749
750
        self.moe_layers = []
        self.moe_mlp_layers: list[Glm4MoE] = []

751
        example_moe = None
Yuxuan Zhang's avatar
Yuxuan Zhang committed
752
        for layer in self.model.layers:
753
754
755
            if isinstance(layer, PPMissingLayer):
                continue

Yuxuan Zhang's avatar
Yuxuan Zhang committed
756
757
            assert isinstance(layer, Glm4MoeDecoderLayer)
            if isinstance(layer.mlp, Glm4MoE):
758
759
                # Pick last one layer since the first ones may be dense layers.
                example_moe = layer.mlp
760
                self.moe_mlp_layers.append(layer.mlp)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
761
762
                self.moe_layers.append(layer.mlp.experts)

763
        self.extract_moe_parameters(example_moe)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
764

765
766
    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
767
768
769

    def forward(
        self,
zhuwenwen's avatar
zhuwenwen committed
770
        input_ids: torch.Tensor,
Yuxuan Zhang's avatar
Yuxuan Zhang committed
771
        positions: torch.Tensor,
772
773
774
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
775
776
777
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
Yuxuan Zhang's avatar
Yuxuan Zhang committed
778
779
780
781
782
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
783
    ) -> torch.Tensor | None:
784
        logits = self.logits_processor(self.lm_head, hidden_states)
Yuxuan Zhang's avatar
Yuxuan Zhang committed
785
786
        return logits

787
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
788
789
790
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)

791
792
793
    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        return self.model.get_expert_mapping()

Yuxuan Zhang's avatar
Yuxuan Zhang committed
794

795
796
def get_spec_layer_idx_from_weight_name(
    config: Glm4MoeConfig, weight_name: str
797
) -> int | None:
798
799
800
    if hasattr(config, "num_nextn_predict_layers") and (
        config.num_nextn_predict_layers > 0
    ):
Yuxuan Zhang's avatar
Yuxuan Zhang committed
801
802
        layer_idx = config.num_hidden_layers
        for i in range(config.num_nextn_predict_layers):
803
            if f"layers.{layer_idx + i}." in weight_name:
Yuxuan Zhang's avatar
Yuxuan Zhang committed
804
                return layer_idx + i
zhuwenwen's avatar
zhuwenwen committed
805
    return None