vllm.model_executor.models.ernie45_moe

Inference-only ErineMoE model compatible with HuggingFace weights.

logger module-attribute

logger = init_logger(__name__)

Ernie4_5_MoeAttention

Bases: Module

Source code in vllm/model_executor/models/ernie45_moe.py

class Ernie4_5_MoeAttention(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,
        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")

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            is_neox_style=False,
            rope_scaling=rope_scaling,
        )
        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 = get_rope(
    head_dim,
    rotary_dim=head_dim,
    max_position=max_position_embeddings,
    base=rope_theta,
    is_neox_style=False,
    rope_scaling=rope_scaling,
)

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,
    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_moe.py

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,
    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")

    self.rotary_emb = get_rope(
        self.head_dim,
        rotary_dim=self.head_dim,
        max_position=max_position_embeddings,
        base=rope_theta,
        is_neox_style=False,
        rope_scaling=rope_scaling,
    )
    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_moe.py

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_MoeDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/ernie45_moe.py

class Ernie4_5_MoeDecoderLayer(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)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          131072)
        self.self_attn = Ernie4_5_MoeAttention(
            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,
            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_num_experts = getattr(config, "moe_num_experts", 0)
        moe_layer_start_index = getattr(config, "moe_layer_start_index", 0)
        moe_layer_end_index = getattr(config, "moe_layer_end_index",
                                      config.num_hidden_layers - 1)
        moe_layer_interval = getattr(config, "moe_layer_interval", 1)
        use_moe = getattr(config, "use_moe", moe_num_experts > 0)

        if (use_moe and ((layer_idx + 1) % moe_layer_interval == 0)
                and layer_idx >= moe_layer_start_index
                and layer_idx <= moe_layer_end_index):
            self.mlp = Ernie4_5_MoeMoE(config=config,
                                       quant_config=quant_config,
                                       prefix=f"{prefix}.mlp")
        else:
            self.mlp = Ernie4_5_MoeMLP(
                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],
    ) -> 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)

        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_MoeMoE(
    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_MoeAttention(
    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,
    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_moe.py

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)
    max_position_embeddings = getattr(config, "max_position_embeddings",
                                      131072)
    self.self_attn = Ernie4_5_MoeAttention(
        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,
        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_num_experts = getattr(config, "moe_num_experts", 0)
    moe_layer_start_index = getattr(config, "moe_layer_start_index", 0)
    moe_layer_end_index = getattr(config, "moe_layer_end_index",
                                  config.num_hidden_layers - 1)
    moe_layer_interval = getattr(config, "moe_layer_interval", 1)
    use_moe = getattr(config, "use_moe", moe_num_experts > 0)

    if (use_moe and ((layer_idx + 1) % moe_layer_interval == 0)
            and layer_idx >= moe_layer_start_index
            and layer_idx <= moe_layer_end_index):
        self.mlp = Ernie4_5_MoeMoE(config=config,
                                   quant_config=quant_config,
                                   prefix=f"{prefix}.mlp")
    else:
        self.mlp = Ernie4_5_MoeMLP(
            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],
) -> Tensor

Source code in vllm/model_executor/models/ernie45_moe.py

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

    hidden_states = self.mlp(hidden_states)

    return hidden_states, residual

Ernie4_5_MoeForCausalLM

Bases: Module, SupportsPP

Source code in vllm/model_executor/models/ernie45_moe.py

class Ernie4_5_MoeForCausalLM(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_MoeModel(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,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds)
        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=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"):
                continue
            # MTP will be supported soon.
            if "mtp" 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:
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping

                    if weight_name not in name:
                        continue

                    name = name.replace(weight_name, param_name)
                    # 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:
                    # 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_MoeModel(
    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_moe.py

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_MoeModel(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_moe.py

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,
) -> Union[Tensor, IntermediateTensors]

Source code in vllm/model_executor/models/ernie45_moe.py

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

get_input_embeddings

get_input_embeddings(input_ids: Tensor) -> Tensor

Source code in vllm/model_executor/models/ernie45_moe.py

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_moe.py

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=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"):
            continue
        # MTP will be supported soon.
        if "mtp" 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:
            for mapping in expert_params_mapping:
                param_name, weight_name, expert_id, shard_id = mapping

                if weight_name not in name:
                    continue

                name = name.replace(weight_name, param_name)
                # 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:
                # 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_MoeMLP

Bases: Module

Source code in vllm/model_executor/models/ernie45_moe.py

class Ernie4_5_MoeMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        use_bias: bool = False,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: bool = True,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=use_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj")
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=use_bias,
                                           quant_config=quant_config,
                                           reduce_results=reduce_results,
                                           prefix=f"{prefix}.down_proj")
        if hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {hidden_act}. "
                             "Only silu is supported for now.")
        self.act_fn = SiluAndMul()

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

act_fn instance-attribute

act_fn = SiluAndMul()

down_proj instance-attribute

down_proj = RowParallelLinear(
    intermediate_size,
    hidden_size,
    bias=use_bias,
    quant_config=quant_config,
    reduce_results=reduce_results,
    prefix=f"{prefix}.down_proj",
)

gate_up_proj instance-attribute

gate_up_proj = MergedColumnParallelLinear(
    hidden_size,
    [intermediate_size] * 2,
    bias=use_bias,
    quant_config=quant_config,
    prefix=f"{prefix}.gate_up_proj",
)

__init__

__init__(
    hidden_size: int,
    intermediate_size: int,
    hidden_act: str,
    use_bias: bool = False,
    quant_config: Optional[QuantizationConfig] = None,
    reduce_results: bool = True,
    prefix: str = "",
) -> None

Source code in vllm/model_executor/models/ernie45_moe.py

def __init__(
    self,
    hidden_size: int,
    intermediate_size: int,
    hidden_act: str,
    use_bias: bool = False,
    quant_config: Optional[QuantizationConfig] = None,
    reduce_results: bool = True,
    prefix: str = "",
) -> None:
    super().__init__()
    self.gate_up_proj = MergedColumnParallelLinear(
        hidden_size, [intermediate_size] * 2,
        bias=use_bias,
        quant_config=quant_config,
        prefix=f"{prefix}.gate_up_proj")
    self.down_proj = RowParallelLinear(intermediate_size,
                                       hidden_size,
                                       bias=use_bias,
                                       quant_config=quant_config,
                                       reduce_results=reduce_results,
                                       prefix=f"{prefix}.down_proj")
    if hidden_act != "silu":
        raise ValueError(f"Unsupported activation: {hidden_act}. "
                         "Only silu is supported for now.")
    self.act_fn = SiluAndMul()

forward

forward(x)

Source code in vllm/model_executor/models/ernie45_moe.py

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

Ernie4_5_MoeMoE

Bases: Module

Source code in vllm/model_executor/models/ernie45_moe.py

class Ernie4_5_MoeMoE(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.moe_num_shared_experts = getattr(config, "moe_num_shared_experts",
                                              None)

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

        self.gate = ReplicatedLinear(config.hidden_size,
                                     config.moe_num_experts,
                                     bias=False,
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")

        self.experts = FusedMoE(num_experts=config.moe_num_experts,
                                top_k=config.moe_k,
                                hidden_size=config.hidden_size,
                                intermediate_size=config.moe_intermediate_size,
                                reduce_results=False,
                                renormalize=True,
                                quant_config=quant_config,
                                prefix=f"{prefix}.experts")

        if self.moe_num_shared_experts is not None:
            intermediate_size = (config.moe_intermediate_size *
                                 config.moe_num_shared_experts)
            self.shared_experts = Ernie4_5_MoeMLP(
                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.experts.must_reduce_shared_expert_outputs(
                ))

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        orig_shape = hidden_states.shape
        hidden_dim = hidden_states.shape[-1]
        hidden_states = hidden_states.view(-1, hidden_dim)
        if self.moe_num_shared_experts is not None:
            shared_output = self.shared_experts(hidden_states)

        router_logits, _ = self.gate(hidden_states)

        final_hidden_states = self.experts(hidden_states=hidden_states,
                                           router_logits=router_logits)

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

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

        return final_hidden_states.view(orig_shape)

experts instance-attribute

experts = FusedMoE(
    num_experts=moe_num_experts,
    top_k=moe_k,
    hidden_size=hidden_size,
    intermediate_size=moe_intermediate_size,
    reduce_results=False,
    renormalize=True,
    quant_config=quant_config,
    prefix=f"{prefix}.experts",
)

gate instance-attribute

gate = ReplicatedLinear(
    hidden_size,
    moe_num_experts,
    bias=False,
    quant_config=None,
    prefix=f"{prefix}.gate",
)

layer_idx instance-attribute

layer_idx = layer_idx

moe_num_shared_experts instance-attribute

moe_num_shared_experts = getattr(
    config, "moe_num_shared_experts", None
)

shared_experts instance-attribute

shared_experts = Ernie4_5_MoeMLP(
    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(),
)

tp_size instance-attribute

tp_size = get_tensor_model_parallel_world_size()

__init__

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

Source code in vllm/model_executor/models/ernie45_moe.py

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.moe_num_shared_experts = getattr(config, "moe_num_shared_experts",
                                          None)

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

    self.gate = ReplicatedLinear(config.hidden_size,
                                 config.moe_num_experts,
                                 bias=False,
                                 quant_config=None,
                                 prefix=f"{prefix}.gate")

    self.experts = FusedMoE(num_experts=config.moe_num_experts,
                            top_k=config.moe_k,
                            hidden_size=config.hidden_size,
                            intermediate_size=config.moe_intermediate_size,
                            reduce_results=False,
                            renormalize=True,
                            quant_config=quant_config,
                            prefix=f"{prefix}.experts")

    if self.moe_num_shared_experts is not None:
        intermediate_size = (config.moe_intermediate_size *
                             config.moe_num_shared_experts)
        self.shared_experts = Ernie4_5_MoeMLP(
            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.experts.must_reduce_shared_expert_outputs(
            ))

forward

forward(hidden_states: Tensor) -> Tensor

Source code in vllm/model_executor/models/ernie45_moe.py

def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    orig_shape = hidden_states.shape
    hidden_dim = hidden_states.shape[-1]
    hidden_states = hidden_states.view(-1, hidden_dim)
    if self.moe_num_shared_experts is not None:
        shared_output = self.shared_experts(hidden_states)

    router_logits, _ = self.gate(hidden_states)

    final_hidden_states = self.experts(hidden_states=hidden_states,
                                       router_logits=router_logits)

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

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

    return final_hidden_states.view(orig_shape)

Ernie4_5_MoeModel

Bases: Module

Source code in vllm/model_executor/models/ernie45_moe.py

@support_torch_compile
class Ernie4_5_MoeModel(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

        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_MoeDecoderLayer(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,
    ) -> 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 i in range(self.start_layer, self.end_layer):
            layer = self.layers[i]
            hidden_states, residual = layer(positions, hidden_states, residual)

        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",
)

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_moe.py

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

    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_MoeDecoderLayer(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,
) -> Union[Tensor, IntermediateTensors]

Source code in vllm/model_executor/models/ernie45_moe.py

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

    if get_pp_group().is_first_rank:
        if inputs_embeds is not None:
            hidden_states = inputs_embeds
        else:
            hidden_states = self.get_input_embeddings(input_ids)
        residual = None
    else:
        assert intermediate_tensors is not None
        hidden_states = intermediate_tensors["hidden_states"]
        residual = intermediate_tensors["residual"]

    for i in range(self.start_layer, self.end_layer):
        layer = self.layers[i]
        hidden_states, residual = layer(positions, hidden_states, residual)

    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_moe.py

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