vllm.model_executor.models.nemotron_h ¶

class NemotronHAttention(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ) -> None:
        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.total_num_kv_heads = config.num_key_value_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)
        if hasattr(config, "head_dim") and config.head_dim is not None:
            self.head_dim = config.head_dim
        else:
            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.qkv_proj = QKVParallelLinear(
            config.hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            config.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )

        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            prefix=f"{prefix}.attn",
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        **kwargs,
    ) -> 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 NemotronHAttentionDecoderLayer(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        parallel_config: ParallelConfig | None = None,
        prefix: str = "",
    ) -> None:
        super().__init__()

        self.mixer = NemotronHAttention(
            config,
            layer_idx,
            model_config,
            cache_config,
            quant_config,
            prefix=f"{prefix}.mixer",
        )

        self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
        **kwargs,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.norm(hidden_states)
        else:
            hidden_states, residual = self.norm(hidden_states, residual)

        hidden_states = self.mixer(hidden_states=hidden_states)
        return hidden_states, residual

class NemotronHForCausalLM(
    nn.Module,
    HasInnerState,
    SupportsLoRA,
    SupportsPP,
    IsHybrid,
    SupportsQuant,
    MixtureOfExperts,
):
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={"backbone": "model"},
        orig_to_new_substr={"A_log": "A", "embeddings": "embed_tokens"},
    )

    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }

    # LoRA specific attributes
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }
    embedding_padding_modules = ["lm_head"]

    @classmethod
    def get_mamba_state_dtype_from_config(
        cls,
        vllm_config: "VllmConfig",
    ) -> tuple[torch.dtype, torch.dtype]:
        return MambaStateDtypeCalculator.mamba2_state_dtype(
            vllm_config.model_config.dtype,
            vllm_config.cache_config.mamba_cache_dtype,
            vllm_config.cache_config.mamba_ssm_cache_dtype,
        )

    @classmethod
    def get_mamba_state_shape_from_config(
        cls,
        vllm_config: "VllmConfig",
    ) -> tuple[tuple[int, int], tuple[int, int, int]]:
        """Calculate shapes for Mamba's convolutional and state caches.

        Args:
            vllm_config: vLLM config

        Returns:
            Tuple containing:
            - conv_state_shape: Shape for convolutional state cache
            - temporal_state_shape: Shape for state space model cache
        """
        parallel_config = vllm_config.parallel_config
        hf_config = vllm_config.model_config.hf_config
        intermediate_size = hf_config.mamba_num_heads * hf_config.mamba_head_dim

        return MambaStateShapeCalculator.mamba2_state_shape(
            intermediate_size=intermediate_size,
            tp_world_size=parallel_config.tensor_parallel_size,
            n_groups=hf_config.n_groups,
            num_heads=hf_config.mamba_num_heads,
            head_dim=hf_config.mamba_head_dim,
            state_size=hf_config.ssm_state_size,
            conv_kernel=hf_config.conv_kernel,
        )

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        config = vllm_config.model_config.hf_config
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        lora_config = vllm_config.lora_config
        scheduler_config = vllm_config.scheduler_config

        self.quant_config = vllm_config.quant_config

        super().__init__()
        self.config = config
        self.scheduler_config = scheduler_config
        self.model = NemotronHModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
        self.unpadded_vocab_size = config.vocab_size
        if lora_config:
            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
        self.lm_head = ParallelLMHead(
            self.unpadded_vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            padding_size=DEFAULT_VOCAB_PADDING_SIZE
            # We need bigger padding if using lora for kernel
            # compatibility
            if not lora_config
            else lora_config.lora_vocab_padding_size,
            prefix=maybe_prefix(prefix, "lm_head"),
        )

        self.logits_processor = LogitsProcessor(
            self.unpadded_vocab_size, config.vocab_size
        )

        self.make_empty_intmd_tensors = self.model.make_empty_intmd_tensors

        # Set MoE hyperparameters
        if self.model.has_moe:
            self.expert_weights = []
            self.num_expert_groups = config.n_group

            self.moe_layers: list[SharedFusedMoE] = []
            example_moe = None
            for layer in self.model.layers:
                if isinstance(layer, NemotronHMoEDecoderLayer):
                    # Pick last one layer since the first ones
                    # may be dense layers.
                    example_moe = layer.mixer
                    self.moe_layers.append(layer.mixer.experts)

            self.num_moe_layers = len(self.moe_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  # noqa: E501
            self.num_routed_experts = example_moe.n_routed_experts
            self.num_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, NemotronHMoEDecoderLayer):
                moe = layer.mixer
                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 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,
        **kwargs,
    ):
        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

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

class NemotronHMLP(nn.Module):
    def __init__(
        self,
        config: NemotronHConfig,
        intermediate_size: int,
        quant_config: QuantizationConfig | None = None,
        bias: bool = False,
        reduce_results: bool = True,
        is_sequence_parallel: bool = False,
        prefix: str = "",
    ) -> None:
        super().__init__()

        self.up_proj = ColumnParallelLinear(
            input_size=config.hidden_size,
            output_size=intermediate_size,
            bias=bias,
            quant_config=quant_config,
            disable_tp=is_sequence_parallel,
            prefix=f"{prefix}.up_proj",
        )
        self.down_proj = RowParallelLinear(
            input_size=intermediate_size,
            output_size=config.hidden_size,
            bias=bias,
            quant_config=quant_config,
            reduce_results=reduce_results,
            disable_tp=is_sequence_parallel,
            prefix=f"{prefix}.down_proj",
        )
        self.act_fn = ReLUSquaredActivation()

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