vllm.model_executor.models.gpt_neox ¶

class GPTNeoXAttention(nn.Module):
    def __init__(
        self,
        config: GPTNeoXConfig,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.total_num_heads = config.num_attention_heads
        self.hidden_size = config.hidden_size
        self.head_size = self.hidden_size // self.total_num_heads
        self.bias = getattr(config, "attention_bias", True)

        tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
        assert self.total_num_heads % tensor_model_parallel_world_size == 0
        self.num_heads = self.total_num_heads // tensor_model_parallel_world_size

        self.query_key_value = QKVParallelLinear(
            config.hidden_size,
            self.head_size,
            self.total_num_heads,
            bias=self.bias,
            quant_config=quant_config,
        )
        self.dense = RowParallelLinear(
            config.hidden_size,
            config.hidden_size,
            bias=self.bias,
            quant_config=quant_config,
        )
        scaling = self.head_size**-0.5
        rotary_dim = int(self.head_size * config.rotary_pct)
        assert rotary_dim % 2 == 0
        rope_theta = getattr(config, "rope_theta", 10000)
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
        self.rotary_emb = get_rope(
            self.head_size,
            rotary_dim=rotary_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
        )
        self.attn = Attention(
            self.num_heads,
            self.head_size,
            scaling,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.query_key_value(hidden_states)
        q, k, v = qkv.chunk(chunks=3, dim=-1)
        q, k = self.rotary_emb(position_ids, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.dense(attn_output)
        return output

class GPTNeoXForCausalLM(nn.Module, SupportsPP):
    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.gpt_neox = GPTNeoXModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "gpt_neox")
        )
        self.embed_out = ParallelLMHead(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=maybe_prefix(prefix, "embed_out"),
        )
        if self.config.tie_word_embeddings:
            self.embed_out.weight = self.gpt_neox.embed_in.weight
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.gpt_neox.make_empty_intermediate_tensors
        )

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
        hidden_states = self.gpt_neox(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor | None:
        logits = self.logits_processor(self.embed_out, hidden_states)
        return logits

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

class GPTNeoXLayer(nn.Module):
    def __init__(
        self,
        config: GPTNeoXConfig,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.use_parallel_residual = config.use_parallel_residual
        self.input_layernorm = nn.LayerNorm(
            config.hidden_size, eps=config.layer_norm_eps
        )
        self.post_attention_layernorm = nn.LayerNorm(
            config.hidden_size, eps=config.layer_norm_eps
        )
        self.attention = GPTNeoXAttention(
            config, cache_config, quant_config, prefix=f"{prefix}.attention"
        )
        self.mlp = GPTNeoXMLP(config, quant_config)

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        attn_input = self.input_layernorm(hidden_states)
        attn_output = self.attention(
            position_ids=position_ids,
            hidden_states=attn_input,
        )

        if self.use_parallel_residual:
            # pseudocode:
            # x = x + attn(ln1(x)) + mlp(ln2(x))
            mlp_input = self.post_attention_layernorm(hidden_states)
            mlp_output = self.mlp(mlp_input)
            hidden_states = mlp_output + attn_output + hidden_states
        else:
            # pseudocode:
            # x = x + attn(ln1(x))
            # x = x + mlp(ln2(x))
            attn_output = attn_output + hidden_states
            mlp_input = self.post_attention_layernorm(attn_output)
            mlp_output = self.mlp(mlp_input)
            hidden_states = mlp_output + attn_output
        return hidden_states

class GPTNeoXMLP(nn.Module):
    def __init__(
        self,
        config: GPTNeoXConfig,
        quant_config: QuantizationConfig | None = None,
    ):
        super().__init__()
        self.dense_h_to_4h = ColumnParallelLinear(
            config.hidden_size,
            config.intermediate_size,
            quant_config=quant_config,
        )
        self.dense_4h_to_h = RowParallelLinear(
            config.intermediate_size,
            config.hidden_size,
            quant_config=quant_config,
        )
        self.act = get_act_fn(config.hidden_act)

    def forward(self, hidden_states):
        hidden_states, _ = self.dense_h_to_4h(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.dense_4h_to_h(hidden_states)
        return hidden_states

@support_torch_compile
class GPTNeoXModel(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_in = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: GPTNeoXLayer(
                config, cache_config, quant_config, prefix=prefix
            ),
            prefix=f"{prefix}.layers",
        )
        self.final_layer_norm = nn.LayerNorm(
            config.hidden_size, eps=config.layer_norm_eps
        )
        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_in(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> 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)
        else:
            hidden_states = intermediate_tensors["hidden_states"]
        for layer in islice(self.layers, self.start_layer, self.end_layer):
            hidden_states = layer(position_ids, hidden_states)
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({"hidden_states": hidden_states})
        hidden_states = self.final_layer_norm(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 (
                "attention.bias" in name
                or "attention.masked_bias" in name
                or "rotary_emb.inv_freq" in name
            ):
                continue
            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
                # Models trained using OpenRLHF may include
                # these tensors in the checkpoint. Skip them.
                continue
            if is_pp_missing_parameter(name, self):
                continue
            param = params_dict[name]

            if "query_key_value" in name:
                # NOTE: GPT-NeoX's fused QKV's output_dim has the shape of
                # (num_heads * 3 * head_size), while the
                # required shape is (3 * num_heads * head_size).
                # Thus, we need weight conversion.
                output_dim = getattr(param, "output_dim", None)
                num_heads = self.config.num_attention_heads
                if output_dim is not None:
                    loaded_weight_shape = loaded_weight.shape
                    loaded_weight = loaded_weight.view(
                        loaded_weight_shape[:output_dim]
                        + (num_heads, 3, -1)
                        + loaded_weight_shape[output_dim + 1 :]
                    )
                    loaded_weight = loaded_weight.transpose(output_dim, output_dim + 1)
                    loaded_weight = loaded_weight.reshape(loaded_weight_shape)

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