vllm.model_executor.models.phi3_small

HeadMajorColumnParallelLinear

Bases: MergedColumnParallelLinear

Source code in vllm/model_executor/models/phi3_small.py

class HeadMajorColumnParallelLinear(MergedColumnParallelLinear):

    def weight_loader(self, param: torch.nn.Parameter,
                      loaded_weight: torch.Tensor):
        return load_column_parallel_weight(param, loaded_weight)

weight_loader

weight_loader(param: Parameter, loaded_weight: Tensor)

Source code in vllm/model_executor/models/phi3_small.py

def weight_loader(self, param: torch.nn.Parameter,
                  loaded_weight: torch.Tensor):
    return load_column_parallel_weight(param, loaded_weight)

HeadMajorQKVParallelLinear

Bases: QKVParallelLinear

Source code in vllm/model_executor/models/phi3_small.py

class HeadMajorQKVParallelLinear(QKVParallelLinear):

    def weight_loader(self, param: torch.nn.Parameter,
                      loaded_weight: torch.Tensor):
        return load_column_parallel_weight(param, loaded_weight)

weight_loader

weight_loader(param: Parameter, loaded_weight: Tensor)

Source code in vllm/model_executor/models/phi3_small.py

def weight_loader(self, param: torch.nn.Parameter,
                  loaded_weight: torch.Tensor):
    return load_column_parallel_weight(param, loaded_weight)

Phi3SmallDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/phi3_small.py

class Phi3SmallDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        layer_idx: int,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = Phi3SmallSelfAttention(config,
                                                layer_idx,
                                                cache_config=cache_config,
                                                quant_config=quant_config,
                                                prefix=f"{prefix}.self_attn")
        self.mlp = Phi3SmallMLP(config, quant_config)

        self.input_layernorm = nn.LayerNorm(config.hidden_size,
                                            eps=config.layer_norm_epsilon)
        self.post_attention_layernorm = nn.LayerNorm(
            config.hidden_size, eps=config.layer_norm_epsilon)

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

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

        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states

hidden_size instance-attribute

hidden_size = hidden_size

input_layernorm instance-attribute

input_layernorm = LayerNorm(
    hidden_size, eps=layer_norm_epsilon
)

mlp instance-attribute

mlp = Phi3SmallMLP(config, quant_config)

post_attention_layernorm instance-attribute

post_attention_layernorm = LayerNorm(
    hidden_size, eps=layer_norm_epsilon
)

self_attn instance-attribute

self_attn = Phi3SmallSelfAttention(
    config,
    layer_idx,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.self_attn",
)

__init__

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

Source code in vllm/model_executor/models/phi3_small.py

def __init__(
    self,
    config: PretrainedConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.hidden_size = config.hidden_size
    self.self_attn = Phi3SmallSelfAttention(config,
                                            layer_idx,
                                            cache_config=cache_config,
                                            quant_config=quant_config,
                                            prefix=f"{prefix}.self_attn")
    self.mlp = Phi3SmallMLP(config, quant_config)

    self.input_layernorm = nn.LayerNorm(config.hidden_size,
                                        eps=config.layer_norm_epsilon)
    self.post_attention_layernorm = nn.LayerNorm(
        config.hidden_size, eps=config.layer_norm_epsilon)

forward

forward(positions: Tensor, hidden_states: Tensor) -> Tensor

Source code in vllm/model_executor/models/phi3_small.py

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

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

    residual = hidden_states
    hidden_states = self.post_attention_layernorm(hidden_states)
    hidden_states = self.mlp(hidden_states)
    hidden_states = residual + hidden_states
    return hidden_states

Phi3SmallForCausalLM

Bases: Module, SupportsPP

Source code in vllm/model_executor/models/phi3_small.py

class Phi3SmallForCausalLM(nn.Module, SupportsPP):
    _tied_weights_keys = ["lm_head.weight"]

    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_suffix={"rotary_emb.inv_freq": None})

    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 = Phi3SmallModel(vllm_config=vllm_config,
                                    prefix=maybe_prefix(prefix, "model"))
        self.vocab_size = config.vocab_size
        self.mup_width_multiplier = config.mup_width_multiplier
        self.lm_head = ParallelLMHead(
            self.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            padding_size=DEFAULT_VOCAB_PADDING_SIZE,
            quant_config=quant_config,
        )
        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)

        # tokens in tiktoken but not used
        if hasattr(config, 'dummy_token_indices'):
            device = self.lm_head.weight.device
            self.register_buffer('dummy_token_indices',
                                 torch.LongTensor(
                                     config.dummy_token_indices).to(device),
                                 persistent=False)
        else:
            self.dummy_token_indices = None

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

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, value):
        self.lm_head = value

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model

    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)
        if self.dummy_token_indices is not None and logits is not None:
            logits.index_fill_(-1, self.dummy_token_indices, -torch.inf)
        logits = logits / self.mup_width_multiplier
        return logits

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

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=(["lm_head.weight"]
                           if self.config.tie_word_embeddings else None))
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

_tied_weights_keys class-attribute instance-attribute

_tied_weights_keys = ['lm_head.weight']

config instance-attribute

config = config

dummy_token_indices instance-attribute

dummy_token_indices = None

hf_to_vllm_mapper class-attribute instance-attribute

hf_to_vllm_mapper = WeightsMapper(
    orig_to_new_suffix={"rotary_emb.inv_freq": None}
)

lm_head instance-attribute

lm_head = ParallelLMHead(
    vocab_size,
    hidden_size,
    org_num_embeddings=vocab_size,
    padding_size=DEFAULT_VOCAB_PADDING_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 = Phi3SmallModel(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "model"),
)

mup_width_multiplier instance-attribute

mup_width_multiplier = mup_width_multiplier

quant_config instance-attribute

quant_config = quant_config

vocab_size instance-attribute

vocab_size = vocab_size

__init__

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

Source code in vllm/model_executor/models/phi3_small.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 = Phi3SmallModel(vllm_config=vllm_config,
                                prefix=maybe_prefix(prefix, "model"))
    self.vocab_size = config.vocab_size
    self.mup_width_multiplier = config.mup_width_multiplier
    self.lm_head = ParallelLMHead(
        self.vocab_size,
        config.hidden_size,
        org_num_embeddings=config.vocab_size,
        padding_size=DEFAULT_VOCAB_PADDING_SIZE,
        quant_config=quant_config,
    )
    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)

    # tokens in tiktoken but not used
    if hasattr(config, 'dummy_token_indices'):
        device = self.lm_head.weight.device
        self.register_buffer('dummy_token_indices',
                             torch.LongTensor(
                                 config.dummy_token_indices).to(device),
                             persistent=False)
    else:
        self.dummy_token_indices = None

compute_logits

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]

Source code in vllm/model_executor/models/phi3_small.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)
    if self.dummy_token_indices is not None and logits is not None:
        logits.index_fill_(-1, self.dummy_token_indices, -torch.inf)
    logits = logits / self.mup_width_multiplier
    return logits

forward

forward(
    input_ids: LongTensor,
    positions: Optional[LongTensor],
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
) -> Union[Tensor, IntermediateTensors]

Source code in vllm/model_executor/models/phi3_small.py

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

get_decoder

get_decoder()

Source code in vllm/model_executor/models/phi3_small.py

def get_decoder(self):
    return self.model

get_input_embeddings

get_input_embeddings(input_ids: Tensor) -> Tensor

Source code in vllm/model_executor/models/phi3_small.py

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

get_output_embeddings

get_output_embeddings()

Source code in vllm/model_executor/models/phi3_small.py

def get_output_embeddings(self):
    return self.lm_head

load_weights

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

Source code in vllm/model_executor/models/phi3_small.py

def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(
        self,
        skip_prefixes=(["lm_head.weight"]
                       if self.config.tie_word_embeddings else None))
    return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

set_decoder

set_decoder(decoder)

Source code in vllm/model_executor/models/phi3_small.py

def set_decoder(self, decoder):
    self.model = decoder

set_input_embeddings

set_input_embeddings(value)

Source code in vllm/model_executor/models/phi3_small.py

def set_input_embeddings(self, value):
    self.model.embed_tokens = value

set_output_embeddings

set_output_embeddings(value)

Source code in vllm/model_executor/models/phi3_small.py

def set_output_embeddings(self, value):
    self.lm_head = value

Phi3SmallMLP

Bases: Module

Source code in vllm/model_executor/models/phi3_small.py

class Phi3SmallMLP(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
    ) -> None:
        super().__init__()
        self.config = config
        assert (self.config.hidden_act == "gegelu"
                ), "Only `gegelu` is supported for the 4.7 series of models .."
        self.hidden_size = config.hidden_size
        self.gegelu_limit = config.gegelu_limit
        self.intermediate_size = config.intermediate_size

        self.up_proj = HeadMajorColumnParallelLinear(
            self.hidden_size,
            2 * [self.intermediate_size],
            bias=True,
            quant_config=quant_config,
        )
        self.down_proj = RowParallelLinear(
            self.intermediate_size,
            self.hidden_size,
            bias=True,
            quant_config=quant_config,
        )

    def forward(self, x):
        gate_up, _ = self.up_proj(x)
        x = gegelu(gate_up)
        x, _ = self.down_proj(x)
        return x

config instance-attribute

config = config

down_proj instance-attribute

down_proj = RowParallelLinear(
    intermediate_size,
    hidden_size,
    bias=True,
    quant_config=quant_config,
)

gegelu_limit instance-attribute

gegelu_limit = gegelu_limit

hidden_size instance-attribute

hidden_size = hidden_size

intermediate_size instance-attribute

intermediate_size = intermediate_size

up_proj instance-attribute

up_proj = HeadMajorColumnParallelLinear(
    hidden_size,
    2 * [intermediate_size],
    bias=True,
    quant_config=quant_config,
)

__init__

__init__(
    config: PretrainedConfig,
    quant_config: Optional[QuantizationConfig] = None,
) -> None

Source code in vllm/model_executor/models/phi3_small.py

def __init__(
    self,
    config: PretrainedConfig,
    quant_config: Optional[QuantizationConfig] = None,
) -> None:
    super().__init__()
    self.config = config
    assert (self.config.hidden_act == "gegelu"
            ), "Only `gegelu` is supported for the 4.7 series of models .."
    self.hidden_size = config.hidden_size
    self.gegelu_limit = config.gegelu_limit
    self.intermediate_size = config.intermediate_size

    self.up_proj = HeadMajorColumnParallelLinear(
        self.hidden_size,
        2 * [self.intermediate_size],
        bias=True,
        quant_config=quant_config,
    )
    self.down_proj = RowParallelLinear(
        self.intermediate_size,
        self.hidden_size,
        bias=True,
        quant_config=quant_config,
    )

forward

forward(x)

Source code in vllm/model_executor/models/phi3_small.py

def forward(self, x):
    gate_up, _ = self.up_proj(x)
    x = gegelu(gate_up)
    x, _ = self.down_proj(x)
    return x

Phi3SmallModel

Bases: Module

Source code in vllm/model_executor/models/phi3_small.py

class Phi3SmallModel(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.config = config
        self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
                                                   config.hidden_size)
        self.mup_embedding_multiplier = config.mup_embedding_multiplier
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: Phi3SmallDecoderLayer(config,
                                                 int(prefix.split('.')[-1]),
                                                 cache_config,
                                                 quant_config,
                                                 prefix=prefix),
            prefix=f"{prefix}.layers")

        self.final_layernorm = nn.LayerNorm(config.hidden_size,
                                            eps=config.layer_norm_epsilon)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(["hidden_states"],
                                                    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.LongTensor,
        positions: Optional[torch.LongTensor],
        intermediate_tensors: Optional[IntermediateTensors],
        inputs_embeds: Optional[torch.Tensor],
    ) -> 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)
            if (self.mup_embedding_multiplier is not None
                    and self.mup_embedding_multiplier > 0.0):
                hidden_states = hidden_states * self.mup_embedding_multiplier
        else:
            assert intermediate_tensors
            hidden_states = intermediate_tensors["hidden_states"]
        for layer in self.layers[self.start_layer:self.end_layer]:
            hidden_states = layer(positions, hidden_states)
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({"hidden_states": hidden_states})
        hidden_states = self.final_layernorm(hidden_states)
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if name.endswith(".bias") and name not in params_dict:
                continue
            if is_pp_missing_parameter(name, self):
                continue
            param = params_dict[name]
            weight_loader = 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
)

final_layernorm instance-attribute

final_layernorm = LayerNorm(
    hidden_size, eps=layer_norm_epsilon
)

make_empty_intermediate_tensors instance-attribute

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

mup_embedding_multiplier instance-attribute

mup_embedding_multiplier = mup_embedding_multiplier

__init__

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

Source code in vllm/model_executor/models/phi3_small.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.config = config
    self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
                                               config.hidden_size)
    self.mup_embedding_multiplier = config.mup_embedding_multiplier
    self.start_layer, self.end_layer, self.layers = make_layers(
        config.num_hidden_layers,
        lambda prefix: Phi3SmallDecoderLayer(config,
                                             int(prefix.split('.')[-1]),
                                             cache_config,
                                             quant_config,
                                             prefix=prefix),
        prefix=f"{prefix}.layers")

    self.final_layernorm = nn.LayerNorm(config.hidden_size,
                                        eps=config.layer_norm_epsilon)
    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(["hidden_states"],
                                                config.hidden_size))

forward

forward(
    input_ids: LongTensor,
    positions: Optional[LongTensor],
    intermediate_tensors: Optional[IntermediateTensors],
    inputs_embeds: Optional[Tensor],
) -> Union[Tensor, IntermediateTensors]

Source code in vllm/model_executor/models/phi3_small.py

def forward(
    self,
    input_ids: torch.LongTensor,
    positions: Optional[torch.LongTensor],
    intermediate_tensors: Optional[IntermediateTensors],
    inputs_embeds: Optional[torch.Tensor],
) -> 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)
        if (self.mup_embedding_multiplier is not None
                and self.mup_embedding_multiplier > 0.0):
            hidden_states = hidden_states * self.mup_embedding_multiplier
    else:
        assert intermediate_tensors
        hidden_states = intermediate_tensors["hidden_states"]
    for layer in self.layers[self.start_layer:self.end_layer]:
        hidden_states = layer(positions, hidden_states)
    if not get_pp_group().is_last_rank:
        return IntermediateTensors({"hidden_states": hidden_states})
    hidden_states = self.final_layernorm(hidden_states)
    return hidden_states

get_input_embeddings

get_input_embeddings(input_ids: Tensor) -> Tensor

Source code in vllm/model_executor/models/phi3_small.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/phi3_small.py

def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if name.endswith(".bias") and name not in params_dict:
            continue
        if is_pp_missing_parameter(name, self):
            continue
        param = params_dict[name]
        weight_loader = getattr(param, "weight_loader",
                                default_weight_loader)
        weight_loader(param, loaded_weight)
        loaded_params.add(name)
    return loaded_params

Phi3SmallSelfAttention

Bases: Module

Source code in vllm/model_executor/models/phi3_small.py

class Phi3SmallSelfAttention(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        layer_idx: int,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.layer_idx = layer_idx
        self.config = config
        self.sparse_block_size = config.blocksparse_block_size
        self.homo_heads = config.blocksparse_homo_head_pattern
        self.local_blocks = config.blocksparse_num_local_blocks
        self.vert_stride = config.blocksparse_vert_stride

        assert (config.blocksparse_block_size ==
                config.blocksparse_triton_kernel_block_size)

        self.hidden_size = config.hidden_size
        # Number of Query Heads
        self.num_heads = config.num_attention_heads

        self.head_dim = self.hidden_size // self.num_heads
        self.tp_size = get_tensor_model_parallel_world_size()
        # Number of total Key Value Heads before tensor parallel
        self.num_key_value_heads = config.num_key_value_heads
        self.num_q_per_kv = self.num_heads // self.num_key_value_heads
        if self.tp_size > 1:
            assert self.num_key_value_heads % self.tp_size == 0
        self.num_kv_heads_per_partition = max(
            1, self.num_key_value_heads // self.tp_size)
        self.num_heads_per_partition = self.num_heads // self.tp_size

        self.max_position_embeddings = config.max_position_embeddings
        self.rope_embedding_base = config.rope_embedding_base
        self.rope_position_scale = config.rope_position_scale
        self.is_causal = True

        norm_factor = None
        if config.mup_use_scaling:
            norm_factor = self.head_dim / config.mup_attn_multiplier
        else:
            norm_factor = math.sqrt(self.head_dim)
        self.scale = 1 / norm_factor

        self.query_key_value = HeadMajorQKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            self.num_heads,
            self.num_key_value_heads,
            bias=True,
            quant_config=quant_config,
        )

        self.dense = RowParallelLinear(self.hidden_size,
                                       self.hidden_size,
                                       bias=True,
                                       quant_config=quant_config)

        if getattr(self.config, "rope_scaling", None) is not None:
            rope_scaling = self.config.rope_scaling
            for key in rope_scaling:
                if isinstance(rope_scaling[key], list):
                    rope_scaling[key] = tuple(rope_scaling[key])

            if "factor" not in rope_scaling:
                rope_scaling["factor"] = self.rope_position_scale
        else:
            rope_scaling = {
                "rope_type": "linear",
                "factor": self.rope_position_scale,
            }

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=self.max_position_embeddings,
            base=self.rope_embedding_base,
            rope_scaling=rope_scaling,
        )

        # blocksparse params
        self.blocksparse_block_size = config.blocksparse_block_size
        self.blocksparse_num_local_blocks = config.blocksparse_num_local_blocks
        self.blocksparse_vert_stride = config.blocksparse_vert_stride

        use_dense_attn = (getattr(self.config,
                                  "dense_attention_every_n_layers", None)
                          and (self.layer_idx + 1) %
                          self.config.dense_attention_every_n_layers == 0)

        bs_params = None
        if not use_dense_attn:
            bs_params = {
                'max_seqlen': self.max_position_embeddings,
                'num_heads': self.num_heads_per_partition,
                "num_kv_heads": self.num_kv_heads_per_partition,
                "block_size": self.sparse_block_size,
                "local_blocks": self.local_blocks,
                "vert_stride": self.vert_stride,
                "homo_head": self.homo_heads
            }

        self.attn = Attention(self.num_heads_per_partition,
                              self.head_dim,
                              self.scale,
                              num_kv_heads=self.num_kv_heads_per_partition,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              blocksparse_params=bs_params,
                              prefix=f"{prefix}.attn")

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> tuple[torch.Tensor, Optional[torch.Tensor],
               Optional[tuple[torch.Tensor]]]:
        qkv, _ = self.query_key_value(hidden_states)

        qkv = qkv.view(qkv.shape[:-1] +
                       (-1, (self.num_q_per_kv + 2), self.head_dim))
        q, k, v = qkv.split([self.num_q_per_kv, 1, 1], dim=-2)

        # NOTE: this is required by RotaryEmbed, which indeed does not have to
        # TODO: allow 3D QK for rotary forward
        q = q.reshape(-1, self.head_dim * self.num_heads_per_partition)
        k = k.reshape(-1, self.head_dim * self.num_kv_heads_per_partition)
        v = v.reshape(-1, self.head_dim * self.num_kv_heads_per_partition)

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

        return output

attn instance-attribute

attn = Attention(
    num_heads_per_partition,
    head_dim,
    scale,
    num_kv_heads=num_kv_heads_per_partition,
    cache_config=cache_config,
    quant_config=quant_config,
    blocksparse_params=bs_params,
    prefix=f"{prefix}.attn",
)

blocksparse_block_size instance-attribute

blocksparse_block_size = blocksparse_block_size

blocksparse_num_local_blocks instance-attribute

blocksparse_num_local_blocks = blocksparse_num_local_blocks

blocksparse_vert_stride instance-attribute

blocksparse_vert_stride = blocksparse_vert_stride

config instance-attribute

config = config

dense instance-attribute

dense = RowParallelLinear(
    hidden_size,
    hidden_size,
    bias=True,
    quant_config=quant_config,
)

head_dim instance-attribute

head_dim = hidden_size // num_heads

hidden_size instance-attribute

hidden_size = hidden_size

homo_heads instance-attribute

homo_heads = blocksparse_homo_head_pattern

is_causal instance-attribute

is_causal = True

layer_idx instance-attribute

layer_idx = layer_idx

local_blocks instance-attribute

local_blocks = blocksparse_num_local_blocks

max_position_embeddings instance-attribute

max_position_embeddings = max_position_embeddings

num_heads instance-attribute

num_heads = num_attention_heads

num_heads_per_partition instance-attribute

num_heads_per_partition = num_heads // tp_size

num_key_value_heads instance-attribute

num_key_value_heads = num_key_value_heads

num_kv_heads_per_partition instance-attribute

num_kv_heads_per_partition = max(
    1, num_key_value_heads // tp_size
)

num_q_per_kv instance-attribute

num_q_per_kv = num_heads // num_key_value_heads

query_key_value instance-attribute

query_key_value = HeadMajorQKVParallelLinear(
    hidden_size,
    head_dim,
    num_heads,
    num_key_value_heads,
    bias=True,
    quant_config=quant_config,
)

rope_embedding_base instance-attribute

rope_embedding_base = rope_embedding_base

rope_position_scale instance-attribute

rope_position_scale = rope_position_scale

rotary_emb instance-attribute

rotary_emb = get_rope(
    head_dim,
    rotary_dim=head_dim,
    max_position=max_position_embeddings,
    base=rope_embedding_base,
    rope_scaling=rope_scaling,
)

scale instance-attribute

scale = 1 / norm_factor

sparse_block_size instance-attribute

sparse_block_size = blocksparse_block_size

tp_size instance-attribute

tp_size = get_tensor_model_parallel_world_size()

vert_stride instance-attribute

vert_stride = blocksparse_vert_stride

__init__

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

Source code in vllm/model_executor/models/phi3_small.py

def __init__(
    self,
    config: PretrainedConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.layer_idx = layer_idx
    self.config = config
    self.sparse_block_size = config.blocksparse_block_size
    self.homo_heads = config.blocksparse_homo_head_pattern
    self.local_blocks = config.blocksparse_num_local_blocks
    self.vert_stride = config.blocksparse_vert_stride

    assert (config.blocksparse_block_size ==
            config.blocksparse_triton_kernel_block_size)

    self.hidden_size = config.hidden_size
    # Number of Query Heads
    self.num_heads = config.num_attention_heads

    self.head_dim = self.hidden_size // self.num_heads
    self.tp_size = get_tensor_model_parallel_world_size()
    # Number of total Key Value Heads before tensor parallel
    self.num_key_value_heads = config.num_key_value_heads
    self.num_q_per_kv = self.num_heads // self.num_key_value_heads
    if self.tp_size > 1:
        assert self.num_key_value_heads % self.tp_size == 0
    self.num_kv_heads_per_partition = max(
        1, self.num_key_value_heads // self.tp_size)
    self.num_heads_per_partition = self.num_heads // self.tp_size

    self.max_position_embeddings = config.max_position_embeddings
    self.rope_embedding_base = config.rope_embedding_base
    self.rope_position_scale = config.rope_position_scale
    self.is_causal = True

    norm_factor = None
    if config.mup_use_scaling:
        norm_factor = self.head_dim / config.mup_attn_multiplier
    else:
        norm_factor = math.sqrt(self.head_dim)
    self.scale = 1 / norm_factor

    self.query_key_value = HeadMajorQKVParallelLinear(
        self.hidden_size,
        self.head_dim,
        self.num_heads,
        self.num_key_value_heads,
        bias=True,
        quant_config=quant_config,
    )

    self.dense = RowParallelLinear(self.hidden_size,
                                   self.hidden_size,
                                   bias=True,
                                   quant_config=quant_config)

    if getattr(self.config, "rope_scaling", None) is not None:
        rope_scaling = self.config.rope_scaling
        for key in rope_scaling:
            if isinstance(rope_scaling[key], list):
                rope_scaling[key] = tuple(rope_scaling[key])

        if "factor" not in rope_scaling:
            rope_scaling["factor"] = self.rope_position_scale
    else:
        rope_scaling = {
            "rope_type": "linear",
            "factor": self.rope_position_scale,
        }

    self.rotary_emb = get_rope(
        self.head_dim,
        rotary_dim=self.head_dim,
        max_position=self.max_position_embeddings,
        base=self.rope_embedding_base,
        rope_scaling=rope_scaling,
    )

    # blocksparse params
    self.blocksparse_block_size = config.blocksparse_block_size
    self.blocksparse_num_local_blocks = config.blocksparse_num_local_blocks
    self.blocksparse_vert_stride = config.blocksparse_vert_stride

    use_dense_attn = (getattr(self.config,
                              "dense_attention_every_n_layers", None)
                      and (self.layer_idx + 1) %
                      self.config.dense_attention_every_n_layers == 0)

    bs_params = None
    if not use_dense_attn:
        bs_params = {
            'max_seqlen': self.max_position_embeddings,
            'num_heads': self.num_heads_per_partition,
            "num_kv_heads": self.num_kv_heads_per_partition,
            "block_size": self.sparse_block_size,
            "local_blocks": self.local_blocks,
            "vert_stride": self.vert_stride,
            "homo_head": self.homo_heads
        }

    self.attn = Attention(self.num_heads_per_partition,
                          self.head_dim,
                          self.scale,
                          num_kv_heads=self.num_kv_heads_per_partition,
                          cache_config=cache_config,
                          quant_config=quant_config,
                          blocksparse_params=bs_params,
                          prefix=f"{prefix}.attn")

forward

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

Source code in vllm/model_executor/models/phi3_small.py

def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> tuple[torch.Tensor, Optional[torch.Tensor],
           Optional[tuple[torch.Tensor]]]:
    qkv, _ = self.query_key_value(hidden_states)

    qkv = qkv.view(qkv.shape[:-1] +
                   (-1, (self.num_q_per_kv + 2), self.head_dim))
    q, k, v = qkv.split([self.num_q_per_kv, 1, 1], dim=-2)

    # NOTE: this is required by RotaryEmbed, which indeed does not have to
    # TODO: allow 3D QK for rotary forward
    q = q.reshape(-1, self.head_dim * self.num_heads_per_partition)
    k = k.reshape(-1, self.head_dim * self.num_kv_heads_per_partition)
    v = v.reshape(-1, self.head_dim * self.num_kv_heads_per_partition)

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

    return output

gegelu

gegelu(input, limit: Optional[float] = None)

Source code in vllm/model_executor/models/phi3_small.py

@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
def gegelu(input, limit: Optional[float] = None):
    a_gelu, a_linear = input[..., ::2], input[..., 1::2]
    if limit is not None:
        a_gelu = torch.where(torch.isinf(a_gelu), a_gelu,
                             a_gelu.clamp(min=None, max=limit))
        a_linear = torch.where(
            torch.isinf(a_linear),
            a_linear,
            a_linear.clamp(min=-limit, max=limit),
        )
    out_gelu = quick_gelu(a_gelu)
    return out_gelu * (a_linear + 1)

load_column_parallel_weight

load_column_parallel_weight(
    param: Parameter, loaded_weight: Tensor
)

Source code in vllm/model_executor/models/phi3_small.py

def load_column_parallel_weight(param: torch.nn.Parameter,
                                loaded_weight: torch.Tensor):
    tp = get_tensor_model_parallel_world_size()
    rk = get_tensor_model_parallel_rank()
    assert param.size(0) * tp == loaded_weight.size(0)
    s = rk * param.size(0)
    e = (rk + 1) * param.size(0)
    loaded_weight = loaded_weight[s:e]
    assert param.shape == loaded_weight.shape
    param.data.copy_(loaded_weight)

quick_gelu

quick_gelu(x)

Source code in vllm/model_executor/models/phi3_small.py

@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
def quick_gelu(x):
    return x * torch.sigmoid(1.702 * x)