vllm.model_executor.models.openpangu ¶

class OpenPanguDecoderLayer(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        prefix: str,
        vllm_config: VllmConfig,
    ) -> None:
        super().__init__()

        if config is None:
            config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        parallel_config = vllm_config.parallel_config

        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)

        layer_idx = int(prefix.split(sep=".")[-1])
        self.layer_idx = layer_idx

        self.use_mla = (
            hasattr(config, "qk_nope_head_dim")
            and hasattr(config, "qk_rope_head_dim")
            and hasattr(config, "v_head_dim")
            and hasattr(config, "kv_lora_rank")
        )
        if self.use_mla:
            self.self_attn = OpenPanguMLAAttention(
                config=config,
                hidden_size=self.hidden_size,
                num_heads=config.num_attention_heads,
                qk_nope_head_dim=config.qk_nope_head_dim,
                qk_rope_head_dim=config.qk_rope_head_dim,
                v_head_dim=config.v_head_dim,
                q_lora_rank=(
                    config.q_lora_rank if hasattr(config, "q_lora_rank") else None
                ),
                kv_lora_rank=config.kv_lora_rank,
                rope_theta=rope_theta,
                max_position_embeddings=max_position_embeddings,
                cache_config=cache_config,
                quant_config=quant_config,
                prefix=f"{prefix}.self_attn",
            )
        else:
            attention_bias = getattr(config, "attention_bias", False) or getattr(
                config, "bias", False
            )
            bias_o_proj = attention_bias
            if hasattr(config, "qkv_bias"):
                attention_bias = config.qkv_bias
            # By default, PanguEmbedded uses causal attention
            # as it is a decoder-only model.
            # You can override the HF config with `is_causal=False` to enable
            # bidirectional attention, which is used in some embedding models
            if getattr(config, "is_causal", True):
                attn_type = AttentionType.DECODER
            else:
                attn_type = AttentionType.ENCODER_ONLY
            self.self_attn = OpenPanguEmbeddedAttention(
                config=config,
                hidden_size=self.hidden_size,
                num_heads=config.num_attention_heads,
                num_kv_heads=getattr(
                    config, "num_key_value_heads", config.num_attention_heads
                ),
                rope_theta=rope_theta,
                rope_scaling=getattr(config, "rope_scaling", None),
                max_position_embeddings=max_position_embeddings,
                quant_config=quant_config,
                bias=attention_bias,
                bias_o_proj=bias_o_proj,
                cache_config=cache_config,
                prefix=f"{prefix}.self_attn",
                attn_type=attn_type,
            )

        if (
            getattr(config, "n_routed_experts", None) is not None
            and layer_idx >= config.first_k_dense_replace
        ):
            self.mlp = OpenPanguMoE(
                config=config,
                parallel_config=parallel_config,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )
        else:
            self.mlp = OpenPanguMLP(
                hidden_size=self.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                bias=getattr(config, "mlp_bias", False),
                prefix=f"{prefix}.mlp",
            )
        self.routed_scaling_factor = getattr(config, "routed_scaling_factor", None)
        self.num_hidden_layers = config.num_hidden_layers
        self.first_k_dense_replace = getattr(
            config, "first_k_dense_replace", self.num_hidden_layers
        )

        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )
        self.tp_group = get_tp_group().device_group
        if getattr(config, "sandwich_norm", False):
            self.sandwich_norm = True
            self.pre_mlp_layernorm = RMSNorm(
                config.hidden_size, eps=config.rms_norm_eps
            )
            self.post_mlp_layernorm = RMSNorm(
                config.hidden_size, eps=config.rms_norm_eps
            )
        else:
            self.sandwich_norm = False

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
    ) -> torch.Tensor:
        if residual is None:
            residual = hidden_states.clone()
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(hidden_states, residual)

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

        if (
            self.routed_scaling_factor is not None
            and hidden_states.dtype == torch.float16
        ):
            # Fix FP16 overflow
            # We scale both hidden_states and residual before
            # rmsnorm, and rmsnorm result would not affect by scale.
            hidden_states *= 1.0 / self.routed_scaling_factor
            if self.layer_idx == 0:
                # The residual is shared by all layers, we only scale it on
                # first layer.
                residual *= 1.0 / self.routed_scaling_factor

        if self.sandwich_norm:
            hidden_states = self.post_attention_layernorm(hidden_states)
            hidden_states, residual = self.pre_mlp_layernorm(hidden_states, residual)
        else:
            hidden_states, residual = self.post_attention_layernorm(
                hidden_states, residual
            )

        # Fully Connected
        hidden_states = self.mlp(hidden_states)

        if (
            self.routed_scaling_factor is not None
            and isinstance(self.mlp, OpenPanguMLP)
            and hidden_states.dtype == torch.float16
        ):
            hidden_states *= 1.0 / self.routed_scaling_factor

        if self.sandwich_norm:
            hidden_states = self.post_mlp_layernorm(hidden_states)

        return hidden_states, residual

class OpenPanguEmbeddedAttention(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: dict[str, Any] | None = None,
        max_position_embeddings: int = 8192,
        quant_config: QuantizationConfig | None = None,
        bias: bool = False,
        bias_o_proj: bool = False,
        cache_config: CacheConfig | None = None,
        prefix: str = "",
        attn_type: str = AttentionType.DECODER,
    ) -> None:
        super().__init__()
        layer_idx = extract_layer_index(prefix)
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        if self.total_num_heads % tp_size != 0:
            raise ValueError(
                f"total_num_heads {self.total_num_heads} "
                f"is not divisible by tp_size {tp_size}."
            )
        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
            and self.total_num_kv_heads % tp_size != 0
        ):
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel ranks.
            raise ValueError(
                "Number of KV heads is greater than TP size, "
                f"but total_num_kv_heads {self.total_num_kv_heads} "
                f"is not divisible by tp_size {tp_size}."
            )
        elif (
            self.total_num_kv_heads < tp_size and tp_size % self.total_num_kv_heads != 0
        ):
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel ranks.
            raise ValueError(
                f"Number of KV heads is less than TP size, but tp_size {tp_size} "
                f"is not divisible by total_num_kv_heads {self.total_num_kv_heads}."
            )
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        head_dim = getattr(config, "head_dim", None)
        if head_dim is None:
            head_dim = self.hidden_size // self.total_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.max_position_embeddings = max_position_embeddings

        self.qkv_proj = QKVParallelLinear(
            hidden_size=hidden_size,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )

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

        self._init_rotary_emb(
            config, rope_scaling=rope_scaling, quant_config=quant_config
        )

        if hasattr(config, "interleaved_sliding_window"):
            interleaved_sliding_window = config.interleaved_sliding_window
            if isinstance(interleaved_sliding_window, int):
                sliding_window = interleaved_sliding_window
            elif isinstance(interleaved_sliding_window, list):
                sw_idx = layer_idx % len(interleaved_sliding_window)
                sliding_window = interleaved_sliding_window[sw_idx]
            else:
                raise ValueError(
                    f"{type(interleaved_sliding_window)} "
                    "for interleaved_sliding_window is not supported."
                )
        else:
            sliding_window = None

        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,
            per_layer_sliding_window=sliding_window,
            attn_type=attn_type,
            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

    def _init_rotary_emb(
        self,
        config: PretrainedConfig,
        rope_scaling: dict[str, Any] | None,
        quant_config: QuantizationConfig | None,
    ) -> None:
        is_neox_style = True
        is_gguf = quant_config and quant_config.get_name() == "gguf"
        if is_gguf and config.model_type == "Pangu":
            is_neox_style = False

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

class OpenPanguForCausalLM(nn.Module, SupportsPP, MixtureOfExperts, SupportsLoRA):
    packed_modules_mapping = {
        "gate_up_proj": ["gate_proj", "up_proj"],
    }

    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.fuse_qkv_a_proj = (
            hasattr(config, "q_lora_rank") and config.q_lora_rank is not None
        )
        if self.fuse_qkv_a_proj:
            self.packed_modules_mapping["fused_qkv_a_proj"] = [
                "q_a_proj",
                "kv_a_proj_with_mqa",
            ]

        self.model = OpenPanguModel(
            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"),
            )
        else:
            self.lm_head = PPMissingLayer()
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )

        self.expert_weights = []

        # Set MoE hyperparameters
        self.num_moe_layers = config.num_hidden_layers - config.first_k_dense_replace
        self.num_expert_groups = 1

        self.moe_layers: list[SharedFusedMoE] = []
        example_moe = None
        for layer in self.model.layers:
            if isinstance(layer, PPMissingLayer):
                continue

            assert isinstance(layer, OpenPanguDecoderLayer)
            if isinstance(layer.mlp, OpenPanguMoE):
                # Pick last one layer since the first ones may be dense layers.
                example_moe = layer.mlp
                self.moe_layers.append(layer.mlp.experts)

        if example_moe is None:
            raise RuntimeError("No MOE layer found in model.layers.")

        self.num_logical_experts = example_moe.n_logical_experts
        self.num_physical_experts = example_moe.n_physical_experts
        self.num_local_physical_experts = example_moe.n_local_physical_experts
        self.n_routed_experts = example_moe.n_routed_experts
        self.n_shared_experts = example_moe.n_shared_experts
        self.num_redundant_experts = example_moe.n_redundant_experts

    def set_eplb_state(
        self,
        expert_load_view: torch.Tensor,
        logical_to_physical_map: torch.Tensor,
        logical_replica_count: torch.Tensor,
    ) -> None:
        for layer_idx, layer in enumerate(self.moe_layers):
            # Register the expert weights.
            self.expert_weights.append(layer.get_expert_weights())
            layer.set_eplb_state(
                moe_layer_idx=layer_idx,
                expert_load_view=expert_load_view,
                logical_to_physical_map=logical_to_physical_map,
                logical_replica_count=logical_replica_count,
            )

    def update_physical_experts_metadata(
        self,
        num_physical_experts: int,
        num_local_physical_experts: int,
    ) -> None:
        assert self.num_local_physical_experts == num_local_physical_experts
        self.num_physical_experts = num_physical_experts
        self.num_local_physical_experts = num_local_physical_experts
        self.num_redundant_experts = num_physical_experts - self.num_logical_experts
        for layer in self.model.layers:
            if isinstance(layer.mlp, OpenPanguMoE):
                moe = layer.mlp
                moe.n_local_physical_experts = num_local_physical_experts
                moe.n_physical_experts = num_physical_experts
                moe.n_redundant_experts = self.num_redundant_experts
                moe.experts.update_expert_map()

    def load_attn_mlp_weight(
        self,
        attn_mlp_replace_mapping: list[tuple[str, str, int]],
        params_dict: dict[str, Any],
        weight_name: str,
        loaded_weight: torch.Tensor,
        loaded_params: set[str],
    ) -> bool:
        for param_name, origin_name, shard_id in attn_mlp_replace_mapping:
            if origin_name not in weight_name or (
                ("mlp.experts." in weight_name) and weight_name not in params_dict
            ):
                continue
            weight_name_mapped = weight_name.replace(origin_name, param_name)
            if (
                param_name == "fused_qkv_a_proj"
                and weight_name_mapped not in params_dict
            ):
                continue
            else:
                weight_name = weight_name_mapped
            if weight_name.endswith(".bias") and weight_name not in params_dict:
                continue
            if is_pp_missing_parameter(weight_name, self):
                continue

            param = params_dict[weight_name]
            weight_loader = param.weight_loader
            weight_loader(param, loaded_weight, shard_id)
            loaded_params.add(weight_name)
            return True
        return False

    def load_expert_weight(
        self,
        expert_merge_mapping: list[tuple[str, str, int, str]],
        params_dict: dict[str, Any],
        weight_name: str,
        loaded_weight: torch.Tensor,
        loaded_params: set[str],
        flag_dict: dict[str, bool],
    ) -> bool:
        for mapping in expert_merge_mapping:
            param_name, origin_name, expert_id, shard_id = mapping
            if origin_name not in weight_name:
                continue
            flag_dict["is_expert_weight"] = True
            weight_name_mapped = weight_name.replace(origin_name, param_name)
            if is_pp_missing_parameter(weight_name_mapped, self):
                continue
            param = params_dict[weight_name_mapped]
            weight_loader = typing.cast(Callable[..., bool], param.weight_loader)
            success = weight_loader(
                param,
                loaded_weight,
                weight_name_mapped,
                shard_id=shard_id,
                expert_id=expert_id,
                return_success=True,
            )
            if success:
                weight_name = weight_name_mapped
                loaded_params.add(weight_name_mapped)
                return True
        return False

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        attn_mlp_replace_mapping = [
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
            (".fused_qkv_a_proj", ".q_a_proj", 0),
            (".fused_qkv_a_proj", ".kv_a_proj_with_mqa", 1),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]
        has_experts = hasattr(self.config, "n_routed_experts")
        if has_experts:
            expert_merge_mapping = SharedFusedMoE.make_expert_params_mapping(
                ckpt_gate_proj_name="gate_proj",
                ckpt_down_proj_name="down_proj",
                ckpt_up_proj_name="up_proj",
                num_experts=self.config.n_routed_experts,
                num_redundant_experts=self.num_redundant_experts,
            )

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue

            if (
                "layers" in name
                and hasattr(self.config, "num_nextn_predict_layers")
                and (self.config.num_nextn_predict_layers > 0)
            ):
                layer_idx = int(name.split("layers.")[-1].split(".")[0])
                mtp_idx = layer_idx - self.config.num_hidden_layers
                if mtp_idx >= 0 and mtp_idx < self.config.num_nextn_predict_layers:
                    continue  # skip spec decode layers for main model

            flag_dict = {"is_expert_weight": False}
            if (
                self.load_attn_mlp_weight(
                    attn_mlp_replace_mapping,
                    params_dict,
                    name,
                    loaded_weight,
                    loaded_params,
                )
                or has_experts
                and self.load_expert_weight(
                    expert_merge_mapping,
                    params_dict,
                    name,
                    loaded_weight,
                    loaded_params,
                    flag_dict,
                )
            ):
                continue
            else:
                if flag_dict["is_expert_weight"]:
                    continue
                if name.endswith(".bias") and name not in params_dict:
                    continue
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    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)
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        return loaded_params

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