Skip to content

vllm.model_executor.models.ernie45_vl_moe

Inference-only Erine VL model compatible with HuggingFace weights.

logger module-attribute 

logger = init_logger(__name__)

Ernie4_5_VLMoeAttention

Bases: Module

Source code in vllm/model_executor/models/ernie45_vl_moe.py 
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
class Ernie4_5_VLMoeAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        head_dim: Optional[int] = None,
        rope_theta: float = 500000,
        rope_scaling: Optional[dict[str, Any]] = None,
        freq_allocation: int = 20,
        max_position_embeddings: int = 131072,
        rms_norm_eps: float = 1e-05,
        qkv_bias: bool = False,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        layer_idx = extract_layer_index(prefix) if len(prefix) > 0 else 0
        self.layer_idx = layer_idx
        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.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")

        self.o_proj = RowParallelLinear(self.total_num_heads * self.head_dim,
                                        hidden_size,
                                        bias=False,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")

        t_rope = freq_allocation
        h_rope = (self.head_dim // 2 - freq_allocation) // 2
        w_rope = (self.head_dim // 2 - freq_allocation) // 2

        self.rotary_emb = Ernie4_5_VLRotaryEmbedding(
            head_size=self.head_dim,
            rotary_dim=self.head_dim,
            max_position_embeddings=max_position_embeddings,
            base=rope_theta,
            is_neox_style=False,
            dtype=torch.get_default_dtype(),
            mrope_section=[h_rope, w_rope, t_rope])

        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")

    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)
        q, k = self.rotary_emb(positions, q, k)

        # Attention
        attn_output = self.attn(q, k, v)
        # Output projection
        output, _ = self.o_proj(attn_output)
        return output

attn instance-attribute 

attn = Attention(
    num_heads,
    head_dim,
    scaling,
    num_kv_heads=num_kv_heads,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.attn",
)

head_dim instance-attribute 

head_dim = head_dim or hidden_size // total_num_heads

hidden_size instance-attribute 

hidden_size = hidden_size

kv_size instance-attribute 

kv_size = num_kv_heads * head_dim

layer_idx instance-attribute 

layer_idx = layer_idx

max_position_embeddings instance-attribute 

max_position_embeddings = max_position_embeddings

num_heads instance-attribute 

num_heads = total_num_heads // tp_size

num_kv_heads instance-attribute 

num_kv_heads = max(1, total_num_kv_heads // tp_size)

o_proj instance-attribute 

o_proj = RowParallelLinear(
    total_num_heads * head_dim,
    hidden_size,
    bias=False,
    quant_config=quant_config,
    prefix=f"{prefix}.o_proj",
)

q_size instance-attribute 

q_size = num_heads * head_dim

qkv_proj instance-attribute 

qkv_proj = QKVParallelLinear(
    hidden_size,
    head_dim,
    total_num_heads,
    total_num_kv_heads,
    bias=qkv_bias,
    quant_config=quant_config,
    prefix=f"{prefix}.qkv_proj",
)

rope_theta instance-attribute 

rope_theta = rope_theta

rotary_emb instance-attribute 

rotary_emb = Ernie4_5_VLRotaryEmbedding(
    head_size=head_dim,
    rotary_dim=head_dim,
    max_position_embeddings=max_position_embeddings,
    base=rope_theta,
    is_neox_style=False,
    dtype=get_default_dtype(),
    mrope_section=[h_rope, w_rope, t_rope],
)

scaling instance-attribute 

scaling = head_dim ** -0.5

total_num_heads instance-attribute 

total_num_heads = num_heads

total_num_kv_heads instance-attribute 

total_num_kv_heads = num_kv_heads

__init__ 

__init__(
    hidden_size: int,
    num_heads: int,
    num_kv_heads: int,
    head_dim: Optional[int] = None,
    rope_theta: float = 500000,
    rope_scaling: Optional[dict[str, Any]] = None,
    freq_allocation: int = 20,
    max_position_embeddings: int = 131072,
    rms_norm_eps: float = 1e-05,
    qkv_bias: bool = False,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
def __init__(
    self,
    hidden_size: int,
    num_heads: int,
    num_kv_heads: int,
    head_dim: Optional[int] = None,
    rope_theta: float = 500000,
    rope_scaling: Optional[dict[str, Any]] = None,
    freq_allocation: int = 20,
    max_position_embeddings: int = 131072,
    rms_norm_eps: float = 1e-05,
    qkv_bias: bool = False,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    layer_idx = extract_layer_index(prefix) if len(prefix) > 0 else 0
    self.layer_idx = layer_idx
    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.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")

    self.o_proj = RowParallelLinear(self.total_num_heads * self.head_dim,
                                    hidden_size,
                                    bias=False,
                                    quant_config=quant_config,
                                    prefix=f"{prefix}.o_proj")

    t_rope = freq_allocation
    h_rope = (self.head_dim // 2 - freq_allocation) // 2
    w_rope = (self.head_dim // 2 - freq_allocation) // 2

    self.rotary_emb = Ernie4_5_VLRotaryEmbedding(
        head_size=self.head_dim,
        rotary_dim=self.head_dim,
        max_position_embeddings=max_position_embeddings,
        base=rope_theta,
        is_neox_style=False,
        dtype=torch.get_default_dtype(),
        mrope_section=[h_rope, w_rope, t_rope])

    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")

forward 

forward(positions: Tensor, hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
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)
    q, k = self.rotary_emb(positions, q, k)

    # Attention
    attn_output = self.attn(q, k, v)
    # Output projection
    output, _ = self.o_proj(attn_output)
    return output

Ernie4_5_VLMoeDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/ernie45_vl_moe.py 
331
332
333
334
335
336
337
338
339
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
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
class Ernie4_5_VLMoeDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 500000)
        rope_scaling = getattr(config, "rope_scaling", None)
        freq_allocation = getattr(config, "freq_allocation", 20)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          131072)

        self.self_attn = Ernie4_5_VLMoeAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            head_dim=getattr(config, 'head_dim', None),
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            freq_allocation=freq_allocation,
            max_position_embeddings=max_position_embeddings,
            rms_norm_eps=config.rms_norm_eps,
            qkv_bias=getattr(config, 'use_bias', False),
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
        )

        layer_idx = extract_layer_index(prefix)
        self.layer_idx = layer_idx

        # MoE
        moe_layer_start_index = config.moe_layer_start_index
        min_moe_layer_start_index = min(moe_layer_start_index)
        moe_layer_end_index = getattr(
            config, "moe_layer_end_index",
            [config.num_hidden_layers - 1, config.num_hidden_layers - 1])
        max_moe_layer_end_index = max(moe_layer_end_index)
        assert min_moe_layer_start_index <= max_moe_layer_end_index
        moe_num_experts = config.moe_num_experts
        max_moe_num_experts = max(moe_num_experts)
        moe_layer_interval = getattr(config, "moe_layer_interval", 1)
        use_moe = getattr(config, "use_moe", max_moe_num_experts > 0)

        if (use_moe and ((layer_idx + 1) % moe_layer_interval == 0)
                and layer_idx >= min_moe_layer_start_index
                and layer_idx <= max_moe_layer_end_index):
            self.mlp = Ernie4_5_VLMoeMoE(config=config,
                                         quant_config=quant_config,
                                         prefix=f"{prefix}.mlp")
        else:
            self.mlp = Ernie4_5_VLMoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                use_bias=getattr(config, 'use_bias', False),
                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)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        visual_token_mask: Optional[torch.Tensor],
        **kwargs: object,
    ) -> torch.Tensor:

        # Self Attention
        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,
        )

        # Fully Connected
        hidden_states, residual = self.post_attention_layernorm(
            hidden_states, residual)

        if isinstance(self.mlp, Ernie4_5_VLMoeMoE):
            hidden_states = self.mlp(hidden_states, visual_token_mask,
                                     **kwargs)
        else:
            hidden_states = self.mlp(hidden_states)

        return hidden_states, residual

hidden_size instance-attribute 

hidden_size = hidden_size

input_layernorm instance-attribute 

input_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)

layer_idx instance-attribute 

layer_idx = layer_idx

mlp instance-attribute 

mlp = Ernie4_5_VLMoeMoE(
    config=config,
    quant_config=quant_config,
    prefix=f"{prefix}.mlp",
)

post_attention_layernorm instance-attribute 

post_attention_layernorm = RMSNorm(
    hidden_size, eps=rms_norm_eps
)

self_attn instance-attribute 

self_attn = Ernie4_5_VLMoeAttention(
    hidden_size=hidden_size,
    num_heads=num_attention_heads,
    num_kv_heads=num_key_value_heads,
    head_dim=getattr(config, "head_dim", None),
    rope_theta=rope_theta,
    rope_scaling=rope_scaling,
    freq_allocation=freq_allocation,
    max_position_embeddings=max_position_embeddings,
    rms_norm_eps=rms_norm_eps,
    qkv_bias=getattr(config, "use_bias", False),
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.self_attn",
)

__init__ 

__init__(
    config: PretrainedConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
333
334
335
336
337
338
339
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
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
def __init__(
    self,
    config: PretrainedConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.hidden_size = config.hidden_size
    rope_theta = getattr(config, "rope_theta", 500000)
    rope_scaling = getattr(config, "rope_scaling", None)
    freq_allocation = getattr(config, "freq_allocation", 20)
    max_position_embeddings = getattr(config, "max_position_embeddings",
                                      131072)

    self.self_attn = Ernie4_5_VLMoeAttention(
        hidden_size=self.hidden_size,
        num_heads=config.num_attention_heads,
        num_kv_heads=config.num_key_value_heads,
        head_dim=getattr(config, 'head_dim', None),
        rope_theta=rope_theta,
        rope_scaling=rope_scaling,
        freq_allocation=freq_allocation,
        max_position_embeddings=max_position_embeddings,
        rms_norm_eps=config.rms_norm_eps,
        qkv_bias=getattr(config, 'use_bias', False),
        cache_config=cache_config,
        quant_config=quant_config,
        prefix=f"{prefix}.self_attn",
    )

    layer_idx = extract_layer_index(prefix)
    self.layer_idx = layer_idx

    # MoE
    moe_layer_start_index = config.moe_layer_start_index
    min_moe_layer_start_index = min(moe_layer_start_index)
    moe_layer_end_index = getattr(
        config, "moe_layer_end_index",
        [config.num_hidden_layers - 1, config.num_hidden_layers - 1])
    max_moe_layer_end_index = max(moe_layer_end_index)
    assert min_moe_layer_start_index <= max_moe_layer_end_index
    moe_num_experts = config.moe_num_experts
    max_moe_num_experts = max(moe_num_experts)
    moe_layer_interval = getattr(config, "moe_layer_interval", 1)
    use_moe = getattr(config, "use_moe", max_moe_num_experts > 0)

    if (use_moe and ((layer_idx + 1) % moe_layer_interval == 0)
            and layer_idx >= min_moe_layer_start_index
            and layer_idx <= max_moe_layer_end_index):
        self.mlp = Ernie4_5_VLMoeMoE(config=config,
                                     quant_config=quant_config,
                                     prefix=f"{prefix}.mlp")
    else:
        self.mlp = Ernie4_5_VLMoeMLP(
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            use_bias=getattr(config, 'use_bias', False),
            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)

forward 

forward(
    positions: Tensor,
    hidden_states: Tensor,
    residual: Optional[Tensor],
    visual_token_mask: Optional[Tensor],
    **kwargs: object,
) -> Tensor
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
    visual_token_mask: Optional[torch.Tensor],
    **kwargs: object,
) -> torch.Tensor:

    # Self Attention
    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,
    )

    # Fully Connected
    hidden_states, residual = self.post_attention_layernorm(
        hidden_states, residual)

    if isinstance(self.mlp, Ernie4_5_VLMoeMoE):
        hidden_states = self.mlp(hidden_states, visual_token_mask,
                                 **kwargs)
    else:
        hidden_states = self.mlp(hidden_states)

    return hidden_states, residual

Ernie4_5_VLMoeForCausalLM

Bases: Module, SupportsPP

Source code in vllm/model_executor/models/ernie45_vl_moe.py 
528
529
530
531
532
533
534
535
536
537
538
539
540
541
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
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
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
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
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
710
711
712
713
714
715
716
717
718
719
720
721
722
723
class Ernie4_5_VLMoeForCausalLM(nn.Module, SupportsPP):
    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
        self.model = Ernie4_5_VLMoeModel(vllm_config=vllm_config,
                                         prefix=maybe_prefix(prefix, "model"))

        if get_pp_group().is_last_rank:
            self.lm_head = ParallelLMHead(config.vocab_size,
                                          config.hidden_size,
                                          quant_config=quant_config)
        else:
            self.lm_head = PPMissingLayer()

        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

    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,
        **kwargs: object,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds, **kwargs)
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states,
                                       sampling_metadata)
        return logits

    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"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]

        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        expert_params_mapping = FusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
            num_experts=max(self.config.moe_num_experts))

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if self.config.tie_word_embeddings and name.endswith(
                    "lm_head.weight"):
                loaded_params.add("lm_head.weight")
                continue
            # MTP will be supported soon.
            if "mtp" in name or \
               "vision_model" in name or \
               "resampler_model" in name:
                continue

            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                # Skip non-stacked layers and experts (experts handled below).
                if weight_name not in name:
                    continue

                if (("mlp.experts." in name) and name not in params_dict):
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if ((name.endswith(".bias") or name.endswith("_bias"))
                        and name not in params_dict):
                    continue
                # Skip layers on other devices.
                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:
                # Distinguish between vision experts and text experts
                if "mlp.experts" in name:
                    moe_offset = int(name.split(".")[-3])
                    vision_expert_start_idx = self.config.moe_num_experts[0]
                    is_text_expert = \
                        moe_offset <= vision_expert_start_idx - 1
                    if is_text_expert:
                        name = name.replace(".experts.", ".text_experts.")
                    else:
                        name = name.replace(
                            f".experts.{moe_offset}",
                            f".vision_experts.{moe_offset-vision_expert_start_idx}"
                        )

                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping

                    if weight_name not in name:
                        continue

                    # Distinguish between vision experts and text experts
                    moe_offset = int(name.split(".")[-3])
                    is_text_expert = \
                        moe_offset <= self.config.moe_num_experts[0] - 1

                    name = name.replace(weight_name, param_name)
                    if is_text_expert:
                        name = name.replace(".experts.", ".text_experts.")
                    else:
                        name = name.replace(".experts.", ".vision_experts.")

                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue

                    # Skip loading extra bias for GPTQ models.
                    if ((name.endswith(".bias") or name.endswith("_bias"))
                            and name not in params_dict):
                        continue
                    param = params_dict[name]

                    weight_loader = param.weight_loader
                    weight_loader(param,
                                  loaded_weight,
                                  name,
                                  shard_id=shard_id,
                                  expert_id=expert_id)
                    break
                else:
                    # Distinguish between vision expert gate
                    # and text expert gate
                    if name.endswith("mlp.gate.weight"):
                        name = name.replace("gate.weight",
                                            "text_experts_gate.weight")
                        loaded_weight = loaded_weight.T
                    elif name.endswith("mlp.gate.weight_1"):
                        name = name.replace("gate.weight_1",
                                            "vision_experts_gate.weight")
                        loaded_weight = loaded_weight.T

                    if "e_score_correction_bias" in name:
                        name = name.replace(".moe_statics.", ".")

                    # Skip loading extra bias for GPTQ models.
                    if ((name.endswith(".bias") or name.endswith("_bias"))
                            and name not in params_dict):
                        continue
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

                    param = params_dict[name]

                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

config instance-attribute 

config = config

fall_back_to_pt_during_load class-attribute instance-attribute 

fall_back_to_pt_during_load = False

lm_head instance-attribute 

lm_head = ParallelLMHead(
    vocab_size, hidden_size, quant_config=quant_config
)

logits_processor instance-attribute 

logits_processor = LogitsProcessor(vocab_size)

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

model instance-attribute 

model = Ernie4_5_VLMoeModel(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "model"),
)

packed_modules_mapping class-attribute instance-attribute 

packed_modules_mapping = {
    "qkv_proj": ["q_proj", "k_proj", "v_proj"],
    "gate_up_proj": ["gate_proj", "up_proj"],
}

quant_config instance-attribute 

quant_config = quant_config

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
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
    self.model = Ernie4_5_VLMoeModel(vllm_config=vllm_config,
                                     prefix=maybe_prefix(prefix, "model"))

    if get_pp_group().is_last_rank:
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
                                      quant_config=quant_config)
    else:
        self.lm_head = PPMissingLayer()

    if self.config.tie_word_embeddings:
        self.lm_head.weight = self.model.embed_tokens.weight
    self.logits_processor = LogitsProcessor(config.vocab_size)
    self.make_empty_intermediate_tensors = (
        self.model.make_empty_intermediate_tensors)

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
580
581
582
583
584
585
586
587
def compute_logits(
    self,
    hidden_states: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[torch.Tensor]:
    logits = self.logits_processor(self.lm_head, hidden_states,
                                   sampling_metadata)
    return logits

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    **kwargs: object,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
568
569
570
571
572
573
574
575
576
577
578
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
    **kwargs: object,
) -> Union[torch.Tensor, IntermediateTensors]:
    hidden_states = self.model(input_ids, positions, intermediate_tensors,
                               inputs_embeds, **kwargs)
    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
565
566
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.model.get_input_embeddings(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
589
590
591
592
593
594
595
596
597
598
599
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
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
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
710
711
712
713
714
715
716
717
718
719
720
721
722
723
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"),
        ("gate_up_proj", "gate_proj", 0),
        ("gate_up_proj", "up_proj", 1),
    ]

    # Params for weights, fp8 weight scales, fp8 activation scales
    # (param_name, weight_name, expert_id, shard_id)
    expert_params_mapping = FusedMoE.make_expert_params_mapping(
        ckpt_gate_proj_name="gate_proj",
        ckpt_down_proj_name="down_proj",
        ckpt_up_proj_name="up_proj",
        num_experts=max(self.config.moe_num_experts))

    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if self.config.tie_word_embeddings and name.endswith(
                "lm_head.weight"):
            loaded_params.add("lm_head.weight")
            continue
        # MTP will be supported soon.
        if "mtp" in name or \
           "vision_model" in name or \
           "resampler_model" in name:
            continue

        for (param_name, weight_name, shard_id) in stacked_params_mapping:
            # Skip non-stacked layers and experts (experts handled below).
            if weight_name not in name:
                continue

            if (("mlp.experts." in name) and name not in params_dict):
                continue
            name = name.replace(weight_name, param_name)
            # Skip loading extra bias for GPTQ models.
            if ((name.endswith(".bias") or name.endswith("_bias"))
                    and name not in params_dict):
                continue
            # Skip layers on other devices.
            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:
            # Distinguish between vision experts and text experts
            if "mlp.experts" in name:
                moe_offset = int(name.split(".")[-3])
                vision_expert_start_idx = self.config.moe_num_experts[0]
                is_text_expert = \
                    moe_offset <= vision_expert_start_idx - 1
                if is_text_expert:
                    name = name.replace(".experts.", ".text_experts.")
                else:
                    name = name.replace(
                        f".experts.{moe_offset}",
                        f".vision_experts.{moe_offset-vision_expert_start_idx}"
                    )

            for mapping in expert_params_mapping:
                param_name, weight_name, expert_id, shard_id = mapping

                if weight_name not in name:
                    continue

                # Distinguish between vision experts and text experts
                moe_offset = int(name.split(".")[-3])
                is_text_expert = \
                    moe_offset <= self.config.moe_num_experts[0] - 1

                name = name.replace(weight_name, param_name)
                if is_text_expert:
                    name = name.replace(".experts.", ".text_experts.")
                else:
                    name = name.replace(".experts.", ".vision_experts.")

                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    continue

                # Skip loading extra bias for GPTQ models.
                if ((name.endswith(".bias") or name.endswith("_bias"))
                        and name not in params_dict):
                    continue
                param = params_dict[name]

                weight_loader = param.weight_loader
                weight_loader(param,
                              loaded_weight,
                              name,
                              shard_id=shard_id,
                              expert_id=expert_id)
                break
            else:
                # Distinguish between vision expert gate
                # and text expert gate
                if name.endswith("mlp.gate.weight"):
                    name = name.replace("gate.weight",
                                        "text_experts_gate.weight")
                    loaded_weight = loaded_weight.T
                elif name.endswith("mlp.gate.weight_1"):
                    name = name.replace("gate.weight_1",
                                        "vision_experts_gate.weight")
                    loaded_weight = loaded_weight.T

                if "e_score_correction_bias" in name:
                    name = name.replace(".moe_statics.", ".")

                # Skip loading extra bias for GPTQ models.
                if ((name.endswith(".bias") or name.endswith("_bias"))
                        and name not in params_dict):
                    continue
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    continue
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue

                param = params_dict[name]

                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
        loaded_params.add(name)
    return loaded_params

Ernie4_5_VLMoeMLP

Bases: Ernie4_5_MoeMLP

Source code in vllm/model_executor/models/ernie45_vl_moe.py 
64
65
class Ernie4_5_VLMoeMLP(Ernie4_5_MoeMLP):
    pass

Ernie4_5_VLMoeMoE

Bases: Module

Source code in vllm/model_executor/models/ernie45_vl_moe.py 
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
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
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
class Ernie4_5_VLMoeMoE(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()

        layer_idx = extract_layer_index(prefix)
        self.layer_idx = layer_idx
        self.tp_size = get_tensor_model_parallel_world_size()
        self.has_shared_experts = (getattr(config, "moe_num_shared_experts", 0)
                                   > 0)
        self.hidden_size = config.hidden_size

        moe_num_experts = config.moe_num_experts
        max_moe_num_experts = max(moe_num_experts)

        if self.tp_size > max_moe_num_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {moe_num_experts}.")

        moe_layer_start_index = config.moe_layer_start_index
        text_moe_layer_start_index = moe_layer_start_index[0]
        vision_moe_layer_start_index = moe_layer_start_index[1]
        moe_layer_end_index = config.moe_layer_end_index
        moe_layer_end_index = getattr(
            config, "moe_layer_end_index",
            [config.num_hidden_layers - 1, config.num_hidden_layers - 1])
        text_moe_layer_end_index = moe_layer_end_index[0]
        vision_moe_layer_end_index = moe_layer_end_index[1]

        assert config.moe_num_experts[0] == config.moe_num_experts[1]
        self.e_score_correction_bias = nn.Parameter(
            torch.empty(2, config.moe_num_experts[0]))

        assert text_moe_layer_start_index <= text_moe_layer_end_index

        if layer_idx >= text_moe_layer_start_index and \
            layer_idx <= text_moe_layer_end_index:
            self.text_experts_gate = ReplicatedLinear(
                config.hidden_size,
                config.moe_num_experts[0],
                bias=False,
                quant_config=quant_config,
                prefix=f"{prefix}.text_experts_gate")

            self.text_experts = FusedMoE(
                num_experts=config.moe_num_experts[0],
                top_k=config.moe_k,
                hidden_size=config.hidden_size,
                intermediate_size=config.moe_intermediate_size[0],
                reduce_results=False,
                renormalize=True,
                quant_config=quant_config,
                e_score_correction_bias=self.e_score_correction_bias[0],
                prefix=f"{prefix}.text_experts")
        else:
            self.text_experts = Ernie4_5_VLMoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                use_bias=getattr(config, 'use_bias', False),
                quant_config=quant_config,
                prefix=f"{prefix}.mlp")

        assert vision_moe_layer_start_index <= vision_moe_layer_end_index
        if layer_idx >= vision_moe_layer_start_index and \
            layer_idx <= vision_moe_layer_end_index:
            self.vision_experts_gate = ReplicatedLinear(
                config.hidden_size,
                config.moe_num_experts[1],
                bias=False,
                quant_config=quant_config,
                prefix=f"{prefix}.vision_experts_gate")

            self.vision_experts = FusedMoE(
                num_experts=config.moe_num_experts[1],
                top_k=config.moe_k,
                hidden_size=config.hidden_size,
                intermediate_size=config.moe_intermediate_size[1],
                reduce_results=False,
                renormalize=True,
                quant_config=quant_config,
                e_score_correction_bias=self.e_score_correction_bias[1],
                prefix=f"{prefix}.vision_experts")
        else:
            self.vision_experts = Ernie4_5_VLMoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                use_bias=getattr(config, 'use_bias', False),
                quant_config=quant_config,
                prefix=f"{prefix}.mlp")

        if self.has_shared_experts:
            intermediate_size = (config.moe_intermediate_size[0] *
                                 config.moe_num_shared_experts)
            self.shared_experts = Ernie4_5_VLMoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                prefix=f"{prefix}.shared_experts",
                reduce_results=self.text_experts.
                must_reduce_shared_expert_outputs())

    def forward(
        self,
        hidden_states: torch.Tensor,
        visual_token_mask: torch.Tensor,
        **kwargs: object,
    ) -> torch.Tensor:

        orig_shape = hidden_states.shape
        hidden_dim = hidden_states.shape[-1]
        hidden_states = hidden_states.view(-1, hidden_dim)

        if self.has_shared_experts:
            shared_output = self.shared_experts(hidden_states)

        if visual_token_mask is not None and visual_token_mask.any():
            # assert visual_token_mask.shape[0] != hidden_states.shape[0]
            visual_token_mask = visual_token_mask.repeat(
                1, self.hidden_size).bool()
            text_token_mask = ~visual_token_mask
            final_hidden_states = torch.zeros_like(hidden_states)

            text_hidden_states = hidden_states[text_token_mask].reshape(
                -1, self.hidden_size)
            vision_hidden_states = hidden_states[visual_token_mask].reshape(
                -1, self.hidden_size)

            text_router_logits, _ = self.text_experts_gate(text_hidden_states)
            final_hidden_states[text_token_mask] = self.text_experts(
                hidden_states=text_hidden_states,
                router_logits=text_router_logits).flatten()

            vision_router_logits, _ = self.vision_experts_gate(
                vision_hidden_states)
            final_hidden_states[visual_token_mask] = self.vision_experts(
                hidden_states=vision_hidden_states,
                router_logits=vision_router_logits).flatten()
        else:
            # text modal input processing directly
            text_router_logits, _ = self.text_experts_gate(hidden_states)

            final_hidden_states = self.text_experts(
                hidden_states=hidden_states, router_logits=text_router_logits)

        if self.has_shared_experts and \
              shared_output is not None:
            final_hidden_states = final_hidden_states + shared_output

        if self.tp_size > 1:
            final_hidden_states = (
                self.text_experts.maybe_all_reduce_tensor_model_parallel(
                    final_hidden_states))

        return final_hidden_states.view(orig_shape)

e_score_correction_bias instance-attribute 

e_score_correction_bias = Parameter(
    empty(2, moe_num_experts[0])
)

has_shared_experts instance-attribute 

has_shared_experts = (
    getattr(config, "moe_num_shared_experts", 0) > 0
)

hidden_size instance-attribute 

hidden_size = hidden_size

layer_idx instance-attribute 

layer_idx = layer_idx

shared_experts instance-attribute 

shared_experts = Ernie4_5_VLMoeMLP(
    hidden_size=hidden_size,
    intermediate_size=intermediate_size,
    hidden_act=hidden_act,
    quant_config=quant_config,
    prefix=f"{prefix}.shared_experts",
    reduce_results=must_reduce_shared_expert_outputs(),
)

text_experts instance-attribute 

text_experts = FusedMoE(
    num_experts=moe_num_experts[0],
    top_k=moe_k,
    hidden_size=hidden_size,
    intermediate_size=moe_intermediate_size[0],
    reduce_results=False,
    renormalize=True,
    quant_config=quant_config,
    e_score_correction_bias=e_score_correction_bias[0],
    prefix=f"{prefix}.text_experts",
)

text_experts_gate instance-attribute 

text_experts_gate = ReplicatedLinear(
    hidden_size,
    moe_num_experts[0],
    bias=False,
    quant_config=quant_config,
    prefix=f"{prefix}.text_experts_gate",
)

tp_size instance-attribute 

tp_size = get_tensor_model_parallel_world_size()

vision_experts instance-attribute 

vision_experts = FusedMoE(
    num_experts=moe_num_experts[1],
    top_k=moe_k,
    hidden_size=hidden_size,
    intermediate_size=moe_intermediate_size[1],
    reduce_results=False,
    renormalize=True,
    quant_config=quant_config,
    e_score_correction_bias=e_score_correction_bias[1],
    prefix=f"{prefix}.vision_experts",
)

vision_experts_gate instance-attribute 

vision_experts_gate = ReplicatedLinear(
    hidden_size,
    moe_num_experts[1],
    bias=False,
    quant_config=quant_config,
    prefix=f"{prefix}.vision_experts_gate",
)

__init__ 

__init__(
    config: PretrainedConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
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
271
272
273
274
def __init__(
    self,
    config: PretrainedConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()

    layer_idx = extract_layer_index(prefix)
    self.layer_idx = layer_idx
    self.tp_size = get_tensor_model_parallel_world_size()
    self.has_shared_experts = (getattr(config, "moe_num_shared_experts", 0)
                               > 0)
    self.hidden_size = config.hidden_size

    moe_num_experts = config.moe_num_experts
    max_moe_num_experts = max(moe_num_experts)

    if self.tp_size > max_moe_num_experts:
        raise ValueError(
            f"Tensor parallel size {self.tp_size} is greater than "
            f"the number of experts {moe_num_experts}.")

    moe_layer_start_index = config.moe_layer_start_index
    text_moe_layer_start_index = moe_layer_start_index[0]
    vision_moe_layer_start_index = moe_layer_start_index[1]
    moe_layer_end_index = config.moe_layer_end_index
    moe_layer_end_index = getattr(
        config, "moe_layer_end_index",
        [config.num_hidden_layers - 1, config.num_hidden_layers - 1])
    text_moe_layer_end_index = moe_layer_end_index[0]
    vision_moe_layer_end_index = moe_layer_end_index[1]

    assert config.moe_num_experts[0] == config.moe_num_experts[1]
    self.e_score_correction_bias = nn.Parameter(
        torch.empty(2, config.moe_num_experts[0]))

    assert text_moe_layer_start_index <= text_moe_layer_end_index

    if layer_idx >= text_moe_layer_start_index and \
        layer_idx <= text_moe_layer_end_index:
        self.text_experts_gate = ReplicatedLinear(
            config.hidden_size,
            config.moe_num_experts[0],
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.text_experts_gate")

        self.text_experts = FusedMoE(
            num_experts=config.moe_num_experts[0],
            top_k=config.moe_k,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size[0],
            reduce_results=False,
            renormalize=True,
            quant_config=quant_config,
            e_score_correction_bias=self.e_score_correction_bias[0],
            prefix=f"{prefix}.text_experts")
    else:
        self.text_experts = Ernie4_5_VLMoeMLP(
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            use_bias=getattr(config, 'use_bias', False),
            quant_config=quant_config,
            prefix=f"{prefix}.mlp")

    assert vision_moe_layer_start_index <= vision_moe_layer_end_index
    if layer_idx >= vision_moe_layer_start_index and \
        layer_idx <= vision_moe_layer_end_index:
        self.vision_experts_gate = ReplicatedLinear(
            config.hidden_size,
            config.moe_num_experts[1],
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.vision_experts_gate")

        self.vision_experts = FusedMoE(
            num_experts=config.moe_num_experts[1],
            top_k=config.moe_k,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size[1],
            reduce_results=False,
            renormalize=True,
            quant_config=quant_config,
            e_score_correction_bias=self.e_score_correction_bias[1],
            prefix=f"{prefix}.vision_experts")
    else:
        self.vision_experts = Ernie4_5_VLMoeMLP(
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            use_bias=getattr(config, 'use_bias', False),
            quant_config=quant_config,
            prefix=f"{prefix}.mlp")

    if self.has_shared_experts:
        intermediate_size = (config.moe_intermediate_size[0] *
                             config.moe_num_shared_experts)
        self.shared_experts = Ernie4_5_VLMoeMLP(
            hidden_size=config.hidden_size,
            intermediate_size=intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            prefix=f"{prefix}.shared_experts",
            reduce_results=self.text_experts.
            must_reduce_shared_expert_outputs())

forward 

forward(
    hidden_states: Tensor,
    visual_token_mask: Tensor,
    **kwargs: object,
) -> Tensor
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
def forward(
    self,
    hidden_states: torch.Tensor,
    visual_token_mask: torch.Tensor,
    **kwargs: object,
) -> torch.Tensor:

    orig_shape = hidden_states.shape
    hidden_dim = hidden_states.shape[-1]
    hidden_states = hidden_states.view(-1, hidden_dim)

    if self.has_shared_experts:
        shared_output = self.shared_experts(hidden_states)

    if visual_token_mask is not None and visual_token_mask.any():
        # assert visual_token_mask.shape[0] != hidden_states.shape[0]
        visual_token_mask = visual_token_mask.repeat(
            1, self.hidden_size).bool()
        text_token_mask = ~visual_token_mask
        final_hidden_states = torch.zeros_like(hidden_states)

        text_hidden_states = hidden_states[text_token_mask].reshape(
            -1, self.hidden_size)
        vision_hidden_states = hidden_states[visual_token_mask].reshape(
            -1, self.hidden_size)

        text_router_logits, _ = self.text_experts_gate(text_hidden_states)
        final_hidden_states[text_token_mask] = self.text_experts(
            hidden_states=text_hidden_states,
            router_logits=text_router_logits).flatten()

        vision_router_logits, _ = self.vision_experts_gate(
            vision_hidden_states)
        final_hidden_states[visual_token_mask] = self.vision_experts(
            hidden_states=vision_hidden_states,
            router_logits=vision_router_logits).flatten()
    else:
        # text modal input processing directly
        text_router_logits, _ = self.text_experts_gate(hidden_states)

        final_hidden_states = self.text_experts(
            hidden_states=hidden_states, router_logits=text_router_logits)

    if self.has_shared_experts and \
          shared_output is not None:
        final_hidden_states = final_hidden_states + shared_output

    if self.tp_size > 1:
        final_hidden_states = (
            self.text_experts.maybe_all_reduce_tensor_model_parallel(
                final_hidden_states))

    return final_hidden_states.view(orig_shape)

Ernie4_5_VLMoeModel

Bases: Module

Source code in vllm/model_executor/models/ernie45_vl_moe.py 
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
class Ernie4_5_VLMoeModel(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

        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size
        self.config = config

        self.im_patch_id = config.im_patch_id

        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=f"{prefix}.embed_tokens")
        else:
            self.embed_tokens = PPMissingLayer()

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: Ernie4_5_VLMoeDecoderLayer(
                config=config,
                cache_config=cache_config,
                quant_config=quant_config,
                prefix=prefix),
            prefix=f"{prefix}.layers",
        )

        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()

        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

    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,
        visual_token_mask: Optional[torch.Tensor] = None,
        **kwargs: object,
    ) -> 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"]

        for layer in islice(self.layers, self.start_layer, self.end_layer):
            hidden_states, residual = layer(positions, hidden_states, residual,
                                            visual_token_mask, **kwargs)

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })

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

        return hidden_states

config instance-attribute 

config = config

embed_tokens instance-attribute 

embed_tokens = VocabParallelEmbedding(
    vocab_size,
    hidden_size,
    quant_config=quant_config,
    prefix=f"{prefix}.embed_tokens",
)

im_patch_id instance-attribute 

im_patch_id = im_patch_id

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors_factory(
        ["hidden_states", "residual"], hidden_size
    )
)

norm instance-attribute 

norm = RMSNorm(hidden_size, eps=rms_norm_eps)

padding_idx instance-attribute 

padding_idx = pad_token_id

vocab_size instance-attribute 

vocab_size = vocab_size

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
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

    self.padding_idx = config.pad_token_id
    self.vocab_size = config.vocab_size
    self.config = config

    self.im_patch_id = config.im_patch_id

    if get_pp_group().is_first_rank:
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=f"{prefix}.embed_tokens")
    else:
        self.embed_tokens = PPMissingLayer()

    self.start_layer, self.end_layer, self.layers = make_layers(
        config.num_hidden_layers,
        lambda prefix: Ernie4_5_VLMoeDecoderLayer(
            config=config,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=prefix),
        prefix=f"{prefix}.layers",
    )

    if get_pp_group().is_last_rank:
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
    else:
        self.norm = PPMissingLayer()

    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size))

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    visual_token_mask: Optional[Tensor] = None,
    **kwargs: object,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
    visual_token_mask: Optional[torch.Tensor] = None,
    **kwargs: object,
) -> 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"]

    for layer in islice(self.layers, self.start_layer, self.end_layer):
        hidden_states, residual = layer(positions, hidden_states, residual,
                                        visual_token_mask, **kwargs)

    if not get_pp_group().is_last_rank:
        return IntermediateTensors({
            "hidden_states": hidden_states,
            "residual": residual
        })

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

    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/ernie45_vl_moe.py 
488
489
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.embed_tokens(input_ids)