vllm.model_executor.models.phimoe

Inference-only PhiMoE model.

PhiMoE

Bases: Module

A tensor-parallel MoE implementation for PhiMoE that shards each expert across all ranks.

Each expert's weights are sharded across all ranks and a fused MoE kernel is used for the forward pass, and finally we reduce the outputs across ranks.

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

class PhiMoE(nn.Module):
    """A tensor-parallel MoE implementation for PhiMoE that shards each expert
    across all ranks.

    Each expert's weights are sharded across all ranks and a fused MoE
    kernel is used for the forward pass, and finally we reduce the outputs
    across ranks.
    """

    def __init__(
        self,
        num_experts: int,
        top_k: int,
        hidden_size: int,
        intermediate_size: int,
        params_dtype: Optional[torch.dtype] = None,
        quant_config: Optional[QuantizationConfig] = None,
        tp_size: Optional[int] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = hidden_size

        # Gate always runs at half / full precision for now.
        self.gate = ReplicatedLinear(
            hidden_size,
            num_experts,
            bias=False,
            params_dtype=params_dtype,
            quant_config=None,
        )

        self.experts = FusedMoE(
            num_experts=num_experts,
            top_k=top_k,
            hidden_size=hidden_size,
            intermediate_size=intermediate_size,
            params_dtype=params_dtype,
            reduce_results=True,
            renormalize=False,
            quant_config=quant_config,
            tp_size=tp_size,
            custom_routing_function=phimoe_routing_function,
            prefix=f"{prefix}.experts")

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # NOTE: hidden_states can have either 1D or 2D shape.
        orig_shape = hidden_states.shape
        hidden_states = hidden_states.view(-1, self.hidden_size)
        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
        final_hidden_states = self.experts(hidden_states, router_logits)
        return final_hidden_states.view(orig_shape)

experts instance-attribute

experts = FusedMoE(
    num_experts=num_experts,
    top_k=top_k,
    hidden_size=hidden_size,
    intermediate_size=intermediate_size,
    params_dtype=params_dtype,
    reduce_results=True,
    renormalize=False,
    quant_config=quant_config,
    tp_size=tp_size,
    custom_routing_function=phimoe_routing_function,
    prefix=f"{prefix}.experts",
)

gate instance-attribute

gate = ReplicatedLinear(
    hidden_size,
    num_experts,
    bias=False,
    params_dtype=params_dtype,
    quant_config=None,
)

hidden_size instance-attribute

hidden_size = hidden_size

__init__

__init__(
    num_experts: int,
    top_k: int,
    hidden_size: int,
    intermediate_size: int,
    params_dtype: Optional[dtype] = None,
    quant_config: Optional[QuantizationConfig] = None,
    tp_size: Optional[int] = None,
    prefix: str = "",
)

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

def __init__(
    self,
    num_experts: int,
    top_k: int,
    hidden_size: int,
    intermediate_size: int,
    params_dtype: Optional[torch.dtype] = None,
    quant_config: Optional[QuantizationConfig] = None,
    tp_size: Optional[int] = None,
    prefix: str = "",
):
    super().__init__()
    self.hidden_size = hidden_size

    # Gate always runs at half / full precision for now.
    self.gate = ReplicatedLinear(
        hidden_size,
        num_experts,
        bias=False,
        params_dtype=params_dtype,
        quant_config=None,
    )

    self.experts = FusedMoE(
        num_experts=num_experts,
        top_k=top_k,
        hidden_size=hidden_size,
        intermediate_size=intermediate_size,
        params_dtype=params_dtype,
        reduce_results=True,
        renormalize=False,
        quant_config=quant_config,
        tp_size=tp_size,
        custom_routing_function=phimoe_routing_function,
        prefix=f"{prefix}.experts")

forward

forward(hidden_states: Tensor) -> Tensor

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

def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    # NOTE: hidden_states can have either 1D or 2D shape.
    orig_shape = hidden_states.shape
    hidden_states = hidden_states.view(-1, self.hidden_size)
    # router_logits: (num_tokens, n_experts)
    router_logits, _ = self.gate(hidden_states)
    final_hidden_states = self.experts(hidden_states, router_logits)
    return final_hidden_states.view(orig_shape)

PhiMoEAttention

Bases: Module

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

class PhiMoEAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        head_dim: Optional[int] = None,
        max_position: int = 4096 * 32,
        rope_theta: float = 10000,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        rope_scaling: Optional[dict] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        if head_dim is None:
            head_dim = hidden_size // num_heads
        self.head_dim = head_dim
        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.rope_scaling = rope_scaling

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=True,
            quant_config=quant_config,
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=True,
            quant_config=quant_config,
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position,
            base=int(self.rope_theta),
            is_neox_style=True,
            rope_scaling=self.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)
        attn_output = self.attn(q, k, v)
        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

hidden_size instance-attribute

hidden_size = hidden_size

kv_size instance-attribute

kv_size = num_kv_heads * head_dim

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=True,
    quant_config=quant_config,
)

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=True,
    quant_config=quant_config,
)

rope_scaling instance-attribute

rope_scaling = rope_scaling

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,
    base=int(rope_theta),
    is_neox_style=True,
    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,
    max_position: int = 4096 * 32,
    rope_theta: float = 10000,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    rope_scaling: Optional[dict] = None,
    prefix: str = "",
) -> None

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

def __init__(
    self,
    hidden_size: int,
    num_heads: int,
    num_kv_heads: int,
    head_dim: Optional[int] = None,
    max_position: int = 4096 * 32,
    rope_theta: float = 10000,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    rope_scaling: Optional[dict] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.hidden_size = hidden_size
    tp_size = get_tensor_model_parallel_world_size()
    self.total_num_heads = num_heads
    assert self.total_num_heads % tp_size == 0
    self.num_heads = self.total_num_heads // tp_size
    self.total_num_kv_heads = num_kv_heads
    if self.total_num_kv_heads >= tp_size:
        # Number of KV heads is greater than TP size, so we partition
        # the KV heads across multiple tensor parallel GPUs.
        assert self.total_num_kv_heads % tp_size == 0
    else:
        # Number of KV heads is less than TP size, so we replicate
        # the KV heads across multiple tensor parallel GPUs.
        assert tp_size % self.total_num_kv_heads == 0
    self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
    if head_dim is None:
        head_dim = hidden_size // num_heads
    self.head_dim = head_dim
    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.rope_scaling = rope_scaling

    self.qkv_proj = QKVParallelLinear(
        hidden_size,
        self.head_dim,
        self.total_num_heads,
        self.total_num_kv_heads,
        bias=True,
        quant_config=quant_config,
    )
    self.o_proj = RowParallelLinear(
        self.total_num_heads * self.head_dim,
        hidden_size,
        bias=True,
        quant_config=quant_config,
    )
    self.rotary_emb = get_rope(
        self.head_dim,
        rotary_dim=self.head_dim,
        max_position=max_position,
        base=int(self.rope_theta),
        is_neox_style=True,
        rope_scaling=self.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/phimoe.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)
    attn_output = self.attn(q, k, v)
    output, _ = self.o_proj(attn_output)
    return output

PhiMoEConfig

Bases: PretrainedConfig

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

class PhiMoEConfig(PretrainedConfig):

    model_type = "phimoe"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=32000,
        hidden_size=4096,
        intermediate_size=14336,
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=8,
        head_dim=None,
        hidden_act="silu",
        max_position_embeddings=4096 * 32,
        initializer_range=0.02,
        rms_norm_eps=1e-5,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=1,
        eos_token_id=2,
        tie_word_embeddings=False,
        rope_theta=1e6,
        sliding_window=None,
        attention_dropout=0.0,
        num_experts_per_tok=2,
        num_local_experts=16,
        output_router_logits=False,
        router_aux_loss_coef=0.001,
        router_jitter_noise=0.0,
        attention_bias=False,
        lm_head_bias=False,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.sliding_window = sliding_window
        self.attention_bias = attention_bias
        self.lm_head_bias = lm_head_bias
        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads
        if head_dim is None:
            head_dim = hidden_size // num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.head_dim = head_dim
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout

        self.num_experts_per_tok = num_experts_per_tok
        self.num_local_experts = num_local_experts
        self.output_router_logits = output_router_logits
        self.router_aux_loss_coef = router_aux_loss_coef
        self.router_jitter_noise = router_jitter_noise
        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )

attention_bias instance-attribute

attention_bias = attention_bias

attention_dropout instance-attribute

attention_dropout = attention_dropout

head_dim instance-attribute

head_dim = head_dim

hidden_act instance-attribute

hidden_act = hidden_act

hidden_size instance-attribute

hidden_size = hidden_size

initializer_range instance-attribute

initializer_range = initializer_range

intermediate_size instance-attribute

intermediate_size = intermediate_size

keys_to_ignore_at_inference class-attribute instance-attribute

keys_to_ignore_at_inference = ['past_key_values']

lm_head_bias instance-attribute

lm_head_bias = lm_head_bias

max_position_embeddings instance-attribute

max_position_embeddings = max_position_embeddings

model_type class-attribute instance-attribute

model_type = 'phimoe'

num_attention_heads instance-attribute

num_attention_heads = num_attention_heads

num_experts_per_tok instance-attribute

num_experts_per_tok = num_experts_per_tok

num_hidden_layers instance-attribute

num_hidden_layers = num_hidden_layers

num_key_value_heads instance-attribute

num_key_value_heads = num_key_value_heads

num_local_experts instance-attribute

num_local_experts = num_local_experts

output_router_logits instance-attribute

output_router_logits = output_router_logits

rms_norm_eps instance-attribute

rms_norm_eps = rms_norm_eps

rope_theta instance-attribute

rope_theta = rope_theta

router_aux_loss_coef instance-attribute

router_aux_loss_coef = router_aux_loss_coef

router_jitter_noise instance-attribute

router_jitter_noise = router_jitter_noise

sliding_window instance-attribute

sliding_window = sliding_window

use_cache instance-attribute

use_cache = use_cache

vocab_size instance-attribute

vocab_size = vocab_size

__init__

__init__(
    vocab_size=32000,
    hidden_size=4096,
    intermediate_size=14336,
    num_hidden_layers=32,
    num_attention_heads=32,
    num_key_value_heads=8,
    head_dim=None,
    hidden_act="silu",
    max_position_embeddings=4096 * 32,
    initializer_range=0.02,
    rms_norm_eps=1e-05,
    use_cache=True,
    pad_token_id=None,
    bos_token_id=1,
    eos_token_id=2,
    tie_word_embeddings=False,
    rope_theta=1000000.0,
    sliding_window=None,
    attention_dropout=0.0,
    num_experts_per_tok=2,
    num_local_experts=16,
    output_router_logits=False,
    router_aux_loss_coef=0.001,
    router_jitter_noise=0.0,
    attention_bias=False,
    lm_head_bias=False,
    **kwargs,
)

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

def __init__(
    self,
    vocab_size=32000,
    hidden_size=4096,
    intermediate_size=14336,
    num_hidden_layers=32,
    num_attention_heads=32,
    num_key_value_heads=8,
    head_dim=None,
    hidden_act="silu",
    max_position_embeddings=4096 * 32,
    initializer_range=0.02,
    rms_norm_eps=1e-5,
    use_cache=True,
    pad_token_id=None,
    bos_token_id=1,
    eos_token_id=2,
    tie_word_embeddings=False,
    rope_theta=1e6,
    sliding_window=None,
    attention_dropout=0.0,
    num_experts_per_tok=2,
    num_local_experts=16,
    output_router_logits=False,
    router_aux_loss_coef=0.001,
    router_jitter_noise=0.0,
    attention_bias=False,
    lm_head_bias=False,
    **kwargs,
):
    self.vocab_size = vocab_size
    self.max_position_embeddings = max_position_embeddings
    self.hidden_size = hidden_size
    self.intermediate_size = intermediate_size
    self.num_hidden_layers = num_hidden_layers
    self.num_attention_heads = num_attention_heads
    self.sliding_window = sliding_window
    self.attention_bias = attention_bias
    self.lm_head_bias = lm_head_bias
    # for backward compatibility
    if num_key_value_heads is None:
        num_key_value_heads = num_attention_heads
    if head_dim is None:
        head_dim = hidden_size // num_attention_heads

    self.num_key_value_heads = num_key_value_heads
    self.head_dim = head_dim
    self.hidden_act = hidden_act
    self.initializer_range = initializer_range
    self.rms_norm_eps = rms_norm_eps
    self.use_cache = use_cache
    self.rope_theta = rope_theta
    self.attention_dropout = attention_dropout

    self.num_experts_per_tok = num_experts_per_tok
    self.num_local_experts = num_local_experts
    self.output_router_logits = output_router_logits
    self.router_aux_loss_coef = router_aux_loss_coef
    self.router_jitter_noise = router_jitter_noise
    super().__init__(
        pad_token_id=pad_token_id,
        bos_token_id=bos_token_id,
        eos_token_id=eos_token_id,
        tie_word_embeddings=tie_word_embeddings,
        **kwargs,
    )

PhiMoEDecoderLayer

Bases: Module

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

class PhiMoEDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PhiMoEConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        # Requires transformers > 4.32.0
        rope_theta = getattr(config, "rope_theta", 10000)
        self.self_attn = PhiMoEAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            max_position=config.max_position_embeddings,
            num_kv_heads=config.num_key_value_heads,
            head_dim=getattr(config, "head_dim",
                             self.hidden_size // config.num_attention_heads),
            rope_theta=rope_theta,
            cache_config=cache_config,
            quant_config=quant_config,
            rope_scaling=config.rope_scaling,
            prefix=f"{prefix}.self_attn",
        )
        self.block_sparse_moe = PhiMoE(
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            quant_config=quant_config,
            prefix=f"{prefix}.block_sparse_moe",
        )
        self.input_layernorm = nn.LayerNorm(config.hidden_size,
                                            eps=config.rms_norm_eps,
                                            elementwise_affine=True)
        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
                                                     eps=config.rms_norm_eps,
                                                     elementwise_affine=True)

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

        # Self Attention
        hidden_states = self.input_layernorm(hidden_states)

        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )
        hidden_states = hidden_states + residual

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

        hidden_states = hidden_states + residual
        return hidden_states, residual

block_sparse_moe instance-attribute

block_sparse_moe = PhiMoE(
    num_experts=num_local_experts,
    top_k=num_experts_per_tok,
    hidden_size=hidden_size,
    intermediate_size=intermediate_size,
    quant_config=quant_config,
    prefix=f"{prefix}.block_sparse_moe",
)

hidden_size instance-attribute

hidden_size = hidden_size

input_layernorm instance-attribute

input_layernorm = LayerNorm(
    hidden_size, eps=rms_norm_eps, elementwise_affine=True
)

post_attention_layernorm instance-attribute

post_attention_layernorm = LayerNorm(
    hidden_size, eps=rms_norm_eps, elementwise_affine=True
)

self_attn instance-attribute

self_attn = PhiMoEAttention(
    hidden_size=hidden_size,
    num_heads=num_attention_heads,
    max_position=max_position_embeddings,
    num_kv_heads=num_key_value_heads,
    head_dim=getattr(
        config,
        "head_dim",
        hidden_size // num_attention_heads,
    ),
    rope_theta=rope_theta,
    cache_config=cache_config,
    quant_config=quant_config,
    rope_scaling=rope_scaling,
    prefix=f"{prefix}.self_attn",
)

__init__

__init__(
    config: PhiMoEConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None

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

def __init__(
    self,
    config: PhiMoEConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.hidden_size = config.hidden_size
    # Requires transformers > 4.32.0
    rope_theta = getattr(config, "rope_theta", 10000)
    self.self_attn = PhiMoEAttention(
        hidden_size=self.hidden_size,
        num_heads=config.num_attention_heads,
        max_position=config.max_position_embeddings,
        num_kv_heads=config.num_key_value_heads,
        head_dim=getattr(config, "head_dim",
                         self.hidden_size // config.num_attention_heads),
        rope_theta=rope_theta,
        cache_config=cache_config,
        quant_config=quant_config,
        rope_scaling=config.rope_scaling,
        prefix=f"{prefix}.self_attn",
    )
    self.block_sparse_moe = PhiMoE(
        num_experts=config.num_local_experts,
        top_k=config.num_experts_per_tok,
        hidden_size=config.hidden_size,
        intermediate_size=config.intermediate_size,
        quant_config=quant_config,
        prefix=f"{prefix}.block_sparse_moe",
    )
    self.input_layernorm = nn.LayerNorm(config.hidden_size,
                                        eps=config.rms_norm_eps,
                                        elementwise_affine=True)
    self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
                                                 eps=config.rms_norm_eps,
                                                 elementwise_affine=True)

forward

forward(
    positions: Tensor,
    hidden_states: Tensor,
    residual: Optional[Tensor],
) -> Tensor

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

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

    # Self Attention
    hidden_states = self.input_layernorm(hidden_states)

    hidden_states = self.self_attn(
        positions=positions,
        hidden_states=hidden_states,
    )
    hidden_states = hidden_states + residual

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

    hidden_states = hidden_states + residual
    return hidden_states, residual

PhiMoEForCausalLM

Bases: Module, SupportsLoRA, SupportsPP

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

class PhiMoEForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
    fall_back_to_pt_during_load = False

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

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

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        lora_config = vllm_config.lora_config
        self.config = config
        self.lora_config = lora_config
        self.quant_config = vllm_config.quant_config

        self.model = PhiMoEModel(vllm_config=vllm_config,
                                 prefix=maybe_prefix(prefix, "model"))
        self.unpadded_vocab_size = config.vocab_size
        if lora_config:
            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
        self.lm_head = ParallelLMHead(
            self.unpadded_vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            padding_size=(
                DEFAULT_VOCAB_PADDING_SIZE
                # We need bigger padding if using lora for kernel
                # compatibility
                if not lora_config else lora_config.lora_vocab_padding_size),
            quant_config=None,
            bias=True,
        )
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                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) -> 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]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)

config instance-attribute

config = config

embedding_modules class-attribute instance-attribute

embedding_modules = {
    "embed_tokens": "input_embeddings",
    "lm_head": "output_embeddings",
}

embedding_padding_modules class-attribute instance-attribute

embedding_padding_modules = ['lm_head']

fall_back_to_pt_during_load class-attribute instance-attribute

fall_back_to_pt_during_load = False

lm_head instance-attribute

lm_head = ParallelLMHead(
    unpadded_vocab_size,
    hidden_size,
    org_num_embeddings=vocab_size,
    padding_size=DEFAULT_VOCAB_PADDING_SIZE
    if not lora_config
    else lora_vocab_padding_size,
    quant_config=None,
    bias=True,
)

logits_processor instance-attribute

logits_processor = LogitsProcessor(
    unpadded_vocab_size, vocab_size
)

lora_config instance-attribute

lora_config = lora_config

make_empty_intermediate_tensors instance-attribute

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

model instance-attribute

model = PhiMoEModel(
    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"]
}

quant_config instance-attribute

quant_config = quant_config

unpadded_vocab_size instance-attribute

unpadded_vocab_size = vocab_size

__init__

__init__(*, vllm_config: VllmConfig, prefix: str = '')

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

def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    lora_config = vllm_config.lora_config
    self.config = config
    self.lora_config = lora_config
    self.quant_config = vllm_config.quant_config

    self.model = PhiMoEModel(vllm_config=vllm_config,
                             prefix=maybe_prefix(prefix, "model"))
    self.unpadded_vocab_size = config.vocab_size
    if lora_config:
        self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
    self.lm_head = ParallelLMHead(
        self.unpadded_vocab_size,
        config.hidden_size,
        org_num_embeddings=config.vocab_size,
        padding_size=(
            DEFAULT_VOCAB_PADDING_SIZE
            # We need bigger padding if using lora for kernel
            # compatibility
            if not lora_config else lora_config.lora_vocab_padding_size),
        quant_config=None,
        bias=True,
    )
    self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                            config.vocab_size)

    self.make_empty_intermediate_tensors = (
        self.model.make_empty_intermediate_tensors)

compute_logits

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Tensor

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

def compute_logits(self, hidden_states: torch.Tensor,
                   sampling_metadata: SamplingMetadata) -> 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/phimoe.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/phimoe.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/phimoe.py

def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(self)
    return loader.load_weights(weights)

PhiMoEModel

Bases: Module

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

@support_torch_compile
class PhiMoEModel(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
        lora_config = vllm_config.lora_config

        lora_vocab = ((lora_config.lora_extra_vocab_size *
                       (lora_config.max_loras or 1)) if lora_config else 0)
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size
        self.config = config
        self.quant_config = quant_config

        self.embed_tokens = VocabParallelEmbedding(
            self.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
        )
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: PhiMoEDecoderLayer(
                config, cache_config, quant_config, prefix=prefix),
            prefix=f"{prefix}.layers")
        self.norm = nn.LayerNorm(config.hidden_size,
                                 eps=config.rms_norm_eps,
                                 elementwise_affine=True)

        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],
        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 layer in self.layers[self.start_layer:self.end_layer]:
            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)
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]

        expert_params_mapping = FusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="w1",
            ckpt_down_proj_name="w2",
            ckpt_up_proj_name="w3",
            num_experts=self.config.num_local_experts)

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if (self.quant_config is not None and
                (scale_name := self.quant_config.get_cache_scale(name))):
                # Loading kv cache quantization scales
                param = params_dict[scale_name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                                 loaded_weight[0])
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue

            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # 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
                    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") 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

embed_tokens instance-attribute

embed_tokens = VocabParallelEmbedding(
    vocab_size, hidden_size, org_num_embeddings=vocab_size
)

make_empty_intermediate_tensors instance-attribute

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

norm instance-attribute

norm = LayerNorm(
    hidden_size, eps=rms_norm_eps, elementwise_affine=True
)

org_vocab_size instance-attribute

org_vocab_size = vocab_size

quant_config instance-attribute

quant_config = quant_config

vocab_size instance-attribute

vocab_size = vocab_size + lora_vocab

__init__

__init__(*, vllm_config: VllmConfig, prefix: str = '')

Source code in vllm/model_executor/models/phimoe.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
    lora_config = vllm_config.lora_config

    lora_vocab = ((lora_config.lora_extra_vocab_size *
                   (lora_config.max_loras or 1)) if lora_config else 0)
    self.vocab_size = config.vocab_size + lora_vocab
    self.org_vocab_size = config.vocab_size
    self.config = config
    self.quant_config = quant_config

    self.embed_tokens = VocabParallelEmbedding(
        self.vocab_size,
        config.hidden_size,
        org_num_embeddings=config.vocab_size,
    )
    self.start_layer, self.end_layer, self.layers = make_layers(
        config.num_hidden_layers,
        lambda prefix: PhiMoEDecoderLayer(
            config, cache_config, quant_config, prefix=prefix),
        prefix=f"{prefix}.layers")
    self.norm = nn.LayerNorm(config.hidden_size,
                             eps=config.rms_norm_eps,
                             elementwise_affine=True)

    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],
    inputs_embeds: Optional[Tensor] = None,
) -> Union[Tensor, IntermediateTensors]

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

def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors],
    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 layer in self.layers[self.start_layer:self.end_layer]:
        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)
    return hidden_states

get_input_embeddings

get_input_embeddings(input_ids: Tensor) -> Tensor

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

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

load_weights

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]

Source code in vllm/model_executor/models/phimoe.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"),
    ]

    expert_params_mapping = FusedMoE.make_expert_params_mapping(
        ckpt_gate_proj_name="w1",
        ckpt_down_proj_name="w2",
        ckpt_up_proj_name="w3",
        num_experts=self.config.num_local_experts)

    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if (self.quant_config is not None and
            (scale_name := self.quant_config.get_cache_scale(name))):
            # Loading kv cache quantization scales
            param = params_dict[scale_name]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                             loaded_weight[0])
            weight_loader(param, loaded_weight)
            loaded_params.add(scale_name)
            continue

        for param_name, weight_name, shard_id in stacked_params_mapping:
            if weight_name not in name:
                continue
            name = name.replace(weight_name, param_name)
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue
            # 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
                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") 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

mp

Bases: Function

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

class mp(torch.autograd.Function):

    @staticmethod
    def forward(
        ctx,
        scores: torch.Tensor,
        multiplier: torch.Tensor,
        selected_experts: torch.Tensor,
        masked_gates: torch.Tensor,
        mask_for_one: torch.Tensor,
    ):
        ctx.save_for_backward(multiplier, selected_experts, masked_gates)
        return multiplier * mask_for_one

    @staticmethod
    def backward(
        ctx,
        grad_at_output: torch.Tensor,
    ):
        multiplier, selected_experts, masked_gates = ctx.saved_tensors

        grad_at_output = grad_at_output * multiplier

        grad_at_scores_expanded = masked_gates * grad_at_output.mul(-1)
        grad_at_scores_expanded.scatter_add_(
            dim=-1,
            index=selected_experts,
            src=grad_at_output,
        )

        return (
            grad_at_scores_expanded,
            None,
            None,
            None,
            None,
        )

backward staticmethod

backward(ctx, grad_at_output: Tensor)

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

@staticmethod
def backward(
    ctx,
    grad_at_output: torch.Tensor,
):
    multiplier, selected_experts, masked_gates = ctx.saved_tensors

    grad_at_output = grad_at_output * multiplier

    grad_at_scores_expanded = masked_gates * grad_at_output.mul(-1)
    grad_at_scores_expanded.scatter_add_(
        dim=-1,
        index=selected_experts,
        src=grad_at_output,
    )

    return (
        grad_at_scores_expanded,
        None,
        None,
        None,
        None,
    )

forward staticmethod

forward(
    ctx,
    scores: Tensor,
    multiplier: Tensor,
    selected_experts: Tensor,
    masked_gates: Tensor,
    mask_for_one: Tensor,
)

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

@staticmethod
def forward(
    ctx,
    scores: torch.Tensor,
    multiplier: torch.Tensor,
    selected_experts: torch.Tensor,
    masked_gates: torch.Tensor,
    mask_for_one: torch.Tensor,
):
    ctx.save_for_backward(multiplier, selected_experts, masked_gates)
    return multiplier * mask_for_one

phimoe_routing_function

phimoe_routing_function(
    hidden_states: Tensor,
    gating_output: Tensor,
    topk: int,
    renormalize: bool,
)

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

def phimoe_routing_function(
    hidden_states: torch.Tensor,
    gating_output: torch.Tensor,
    topk: int,
    renormalize: bool,
):
    assert hidden_states.shape[0] == gating_output.shape[0], (
        "Number of tokens mismatch")
    assert topk == 2, "Only top-2 routing is supported"
    assert renormalize is False, "Renormalization is not supported"

    topk_weights, topk_ids = sparsemixer(gating_output)
    return topk_weights, topk_ids

sparsemixer

sparsemixer(scores, jitter_eps=0.01)

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

def sparsemixer(scores, jitter_eps=0.01):
    ################ first expert ################

    with torch.no_grad():
        # compute mask for sparsity
        mask_logits_threshold, max_ind = scores.max(dim=-1, keepdim=True)
        factor = scores.abs().clamp(min=mask_logits_threshold)
        mask_logits_threshold = ((mask_logits_threshold - scores) /
                                 factor) > (2 * jitter_eps)

    # apply mask
    masked_gates = scores.masked_fill(mask_logits_threshold, float("-inf"))
    selected_experts = max_ind

    # compute scores for gradients
    masked_gates = torch.softmax(masked_gates, dim=-1)
    multiplier_o = masked_gates.gather(dim=-1, index=selected_experts)

    multiplier = multiplier_o

    # masked out first expert
    masked_scores = torch.scatter(
        scores,
        -1,
        selected_experts,
        float("-inf"),
    )
    with torch.no_grad():
        # compute mask for sparsity
        mask_logits_threshold, max_ind = masked_scores.max(dim=-1,
                                                           keepdim=True)
        factor = scores.abs().clamp(min=mask_logits_threshold)
        mask_logits_threshold = ((mask_logits_threshold - scores) /
                                 factor) > (2 * jitter_eps)

    # apply mask
    masked_gates_top2 = masked_scores.masked_fill(mask_logits_threshold,
                                                  float("-inf"))
    selected_experts_top2 = max_ind
    # compute scores for gradients
    masked_gates_top2 = torch.softmax(masked_gates_top2, dim=-1)
    multiplier_top2 = masked_gates_top2.gather(dim=-1,
                                               index=selected_experts_top2)

    multiplier = torch.concat((multiplier, multiplier_top2), dim=-1)
    selected_experts = torch.concat((selected_experts, selected_experts_top2),
                                    dim=-1)

    return (
        multiplier,
        selected_experts,
    )