vllm.model_executor.models.step1 ¶

class Step1ForCausalLM(nn.Module, SupportsPP):
    packed_modules_mapping = STEP_PACKED_MODULES_MAPPING

    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 = StepDecoderModel(
            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,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
            if getattr(config, "tie_word_embeddings", True):
                self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
            self.logits_processor = LogitsProcessor(config.vocab_size)
        else:
            self.lm_head = PPMissingLayer()
            self.logits_processor = None  # type: ignore[assignment]

        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )

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

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

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor | None:
        if not get_pp_group().is_last_rank:
            return None
        return self.logits_processor(self.lm_head, hidden_states)

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

class StepAttention(nn.Module):
    def __init__(
        self,
        config,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = config.hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = config.num_attention_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.head_dim = self.hidden_size // self.total_num_heads

        total_num_kv_heads = getattr(
            config, "num_attention_groups", getattr(config, "num_key_value_heads", 1)
        )
        if total_num_kv_heads is None or total_num_kv_heads <= 0:
            total_num_kv_heads = 1
        self.total_num_kv_heads = total_num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            assert self.total_num_kv_heads % tp_size == 0
        else:
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)

        self.qkv_proj = QKVParallelLinear(
            hidden_size=self.hidden_size,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=getattr(config, "attention_bias", False),
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim

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

        tp_rank = get_tensor_model_parallel_rank()
        head_start = tp_rank * self.num_heads
        head_end = (tp_rank + 1) * self.num_heads
        alibi_slopes = _get_step_alibi_slopes(self.total_num_heads)[head_start:head_end]
        alibi_slopes = alibi_slopes.tolist()

        self.scale = self.head_dim**-0.5
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scale,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            alibi_slopes=alibi_slopes,
            prefix=f"{prefix}.attn",
            use_alibi_sqrt=True,
            attn_type=AttentionType.DECODER,
        )

    def forward(
        self,
        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)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

class StepDecoderLayer(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.hidden_size = config.hidden_size
        self.self_attn = StepAttention(
            config=config,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
        )
        self.mlp = StepMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            quant_config=quant_config,
            prefix=f"{prefix}.mlp",
            bias=getattr(config, "mlp_bias", False),
        )
        self.input_layernorm = RMSNorm(
            self.hidden_size,
            eps=config.rms_norm_eps,
        )
        self.post_attention_layernorm = RMSNorm(
            self.hidden_size,
            eps=config.rms_norm_eps,
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        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(hidden_states=hidden_states)
        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            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
                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:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader", default_weight_loader)  # type: ignore[name-defined]
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

class StepDecoderModel(nn.Module):
    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
        # Need embed_tokens on first rank, and also on last rank if tie_word_embeddings
        if get_pp_group().is_first_rank or (
            config.tie_word_embeddings and get_pp_group().is_last_rank
        ):
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
            )
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: StepDecoderLayer(vllm_config=vllm_config, prefix=prefix),
            prefix=maybe_prefix(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.aux_hidden_state_layers: tuple[int, ...] = getattr(
            config, "aux_hidden_state_layers", ()
        )
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"],
            config.hidden_size,
        )

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

    def forward(
        self,
        input_ids: torch.Tensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                assert input_ids is not None
                hidden_states = self.embed_input_ids(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        aux_hidden_states = []
        for idx, layer in enumerate(self.layers[self.start_layer : self.end_layer]):
            if idx in self.aux_hidden_state_layers:
                if residual is None:
                    aux_hidden_states.append(hidden_states)
                else:
                    aux_hidden_states.append(hidden_states + residual)
            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)
        if aux_hidden_states:
            return hidden_states, aux_hidden_states
        return hidden_states

class StepMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
        bias: bool = False,
    ):
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            input_size=hidden_size,
            output_sizes=[intermediate_size, intermediate_size],
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            input_size=intermediate_size,
            output_size=hidden_size,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.down_proj",
        )
        self.act_fn = SiluAndMul()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x, _ = self.gate_up_proj(x)
        x = self.act_fn(x)
        x, _ = self.down_proj(x)
        return x