vllm.model_executor.models.chatglm ¶

class ChatGLMBaseModel(nn.Module):
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_substr={".word_embeddings": ""},
    )

    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        transformer_type: type[ChatGLMModel] = ChatGLMModel,
    ) -> None:
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
        multimodal_config = vllm_config.model_config.multimodal_config
        self.config = config
        self.lora_config = lora_config
        self.multimodal_config = multimodal_config

        self.quant_config = quant_config
        self.max_position_embeddings = getattr(config, "max_sequence_length", 8192)
        self.transformer = transformer_type(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "transformer")
        )
        if self.config.tie_word_embeddings:
            self.transformer.output_layer.weight = self.transformer.embedding.weight
        self.lm_head = self.transformer.output_layer
        self.logits_processor = LogitsProcessor(config.padded_vocab_size)
        self.make_empty_intermediate_tensors = (
            self.transformer.make_empty_intermediate_tensors
        )

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

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

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

class ChatGLMForCausalLM(ChatGLMBaseModel, SupportsLoRA, SupportsPP, SupportsQuant):
    packed_modules_mapping = {
        "query_key_value": ["query_key_value"],
        "dense_h_to_4h": ["dense_h_to_4h"],
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        config = vllm_config.model_config.hf_config
        if hasattr(config, "vision_config"):
            hf_overrides = {"architectures": ["GLM4VForCausalLM"]}
            raise RuntimeError(
                "The configuration of this model indicates that it supports "
                "vision inputs, but you instantiated the text-only version "
                "of this model. Please use the vision model by setting "
                f"`--hf-overrides '{json.dumps(hf_overrides)}'`"
            )

        super().__init__(vllm_config=vllm_config, prefix=prefix)

    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.transformer(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
        return hidden_states

@support_torch_compile
class ChatGLMModel(nn.Module, SupportsQuant):
    packed_modules_mapping = {
        "linear_proj.merged_proj": [
            "linear_proj.gate_proj",
            "linear_proj.dense_h_to_4h",
        ]
    }

    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.embedding = VocabParallelEmbedding(
            config.padded_vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=f"{prefix}.embedding",
        )

        self.num_layers = config.num_layers
        self.multi_query_group_num = config.multi_query_group_num
        self.kv_channels = config.kv_channels
        self.encoder = GLMTransformer(
            config, cache_config, quant_config, prefix=f"{prefix}.encoder"
        )

        self.output_layer = ParallelLMHead(
            config.padded_vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=f"{prefix}.output_layer",
        )

        self.make_empty_intermediate_tensors = (
            self.encoder.make_empty_intermediate_tensors
        )

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
        **kwargs: object,
    ) -> 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:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]

        # Run encoder.
        hidden_states = self.encoder(
            hidden_states=hidden_states,
            position_ids=positions,
        )

        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("linear_proj.merged_proj", "linear_proj.gate_proj", 0),
            ("linear_proj.merged_proj", "linear_proj.dense_h_to_4h", 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:
                if "rotary_pos_emb.inv_freq" in name:
                    continue
                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

class GLMAttention(nn.Module):
    def __init__(
        self,
        config: ChatGLMConfig,
        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.multi_query_attention = config.multi_query_attention
        self.total_num_kv_heads = (
            config.multi_query_group_num
            if config.multi_query_attention
            else config.num_attention_heads
        )
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = config.hidden_size // self.total_num_heads
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5

        self.query_key_value = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=config.add_bias_linear or config.add_qkv_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.query_key_value",
        )
        self.dense = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            config.hidden_size,
            bias=config.add_bias_linear,
            quant_config=quant_config,
            prefix=f"{prefix}.dense",
        )

        # https://huggingface.co/zai-org/chatglm3-6b-32k/blob/e210410255278dd9d74463cf396ba559c0ef801c/modeling_chatglm.py#L141
        rope_ratio = getattr(config, "rope_ratio", 1.0)
        max_positions = getattr(config, "seq_length", 8192)
        # NOTE: zai-org/cogagent-9b-20241220 uses original_rope=False,
        # which is equivalent to is_neox_style=True
        is_neox_style = not config.original_rope
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim // 2,
            max_position=max_positions,
            base=10000 * rope_ratio,
            is_neox_style=is_neox_style,
        )
        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,
        hidden_states: torch.Tensor,
        position_ids: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.query_key_value(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(position_ids, q, k)
        context_layer = self.attn(q, k, v)
        attn_output, _ = self.dense(context_layer)
        return attn_output

class GLMBlock(nn.Module):
    """A single transformer layer.

    Transformer layer takes input with size [s, b, h] and returns an
    output of the same size.
    """

    def __init__(
        self,
        config: ChatGLMConfig,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.apply_residual_connection_post_layernorm = (
            config.apply_residual_connection_post_layernorm
        )

        self.fp32_residual_connection = config.fp32_residual_connection

        layer_norm_func = RMSNorm if config.rmsnorm else LayerNorm
        # Layernorm on the input data.
        self.input_layernorm = layer_norm_func(
            config.hidden_size, eps=config.layernorm_epsilon
        )

        # Self attention.
        self.self_attention = GLMAttention(
            config, cache_config, quant_config, prefix=f"{prefix}.self_attention"
        )
        self.hidden_dropout = config.hidden_dropout

        # Layernorm on the attention output
        self.post_attention_layernorm = layer_norm_func(
            config.hidden_size, eps=config.layernorm_epsilon
        )

        # MLP
        self.mlp = GLMMLP(config, quant_config, prefix=f"{prefix}.mlp")

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: torch.Tensor,
    ) -> torch.Tensor:
        # hidden_states: [num_tokens, h]
        # Layer norm at the beginning of the transformer layer.
        layernorm_output = self.input_layernorm(hidden_states)
        # Self attention.
        attention_output = self.self_attention(
            hidden_states=layernorm_output,
            position_ids=position_ids,
        )

        # Residual connection.
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = hidden_states

        layernorm_input = residual + attention_output

        # Layer norm post the self attention.
        layernorm_output = self.post_attention_layernorm(layernorm_input)

        # Second residual connection.
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = layernorm_input

        output = self.mlp(layernorm_output) + residual

        return output