vllm.model_executor.models.nemotron_h

Inference-only NemotronH model.

ALL_DECODER_LAYER_TYPES module-attribute

ALL_DECODER_LAYER_TYPES = {
    "M": NemotronHMambaDecoderLayer,
    "-": NemotronHMLPDecoderLayer,
    "*": NemotronHAttentionDecoderLayer,
}

NemotronHAttention

Bases: Module

Source code in vllm/model_executor/models/nemotron_h.py

class NemotronHAttention(nn.Module):

    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = 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)
        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

attn instance-attribute

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

head_dim instance-attribute

head_dim = hidden_size // total_num_heads

hidden_size instance-attribute

hidden_size = hidden_size

kv_size instance-attribute

kv_size = num_kv_heads * head_dim

num_heads instance-attribute

num_heads = total_num_heads // tp_size

num_kv_heads instance-attribute

num_kv_heads = max(1, total_num_kv_heads // tp_size)

o_proj instance-attribute

o_proj = RowParallelLinear(
    total_num_heads * head_dim,
    hidden_size,
    bias=False,
    quant_config=quant_config,
    prefix=f"{prefix}.o_proj",
)

q_size instance-attribute

q_size = num_heads * head_dim

qkv_proj instance-attribute

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

scaling instance-attribute

scaling = head_dim ** -0.5

total_num_heads instance-attribute

total_num_heads = num_attention_heads

total_num_kv_heads instance-attribute

total_num_kv_heads = num_key_value_heads

__init__

__init__(
    config: NemotronHConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None

Source code in vllm/model_executor/models/nemotron_h.py

def __init__(
    self,
    config: NemotronHConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = 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)
    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",
    )

forward

forward(hidden_states: Tensor, **kwargs) -> Tensor

Source code in vllm/model_executor/models/nemotron_h.py

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

NemotronHAttentionDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/nemotron_h.py

class NemotronHAttentionDecoderLayer(nn.Module):

    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()

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

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

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        **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

mixer instance-attribute

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

norm instance-attribute

norm = RMSNorm(hidden_size, eps=rms_norm_eps)

__init__

__init__(
    config: NemotronHConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None

Source code in vllm/model_executor/models/nemotron_h.py

def __init__(
    self,
    config: NemotronHConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()

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

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

forward

forward(
    positions: Tensor,
    hidden_states: Tensor,
    residual: Optional[Tensor],
    **kwargs,
)

Source code in vllm/model_executor/models/nemotron_h.py

def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
    **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

NemotronHForCausalLM

Bases: Module, HasInnerState, SupportsLoRA, SupportsPP, IsHybrid, SupportsQuant

Source code in vllm/model_executor/models/nemotron_h.py

class NemotronHForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
                           IsHybrid, SupportsQuant):
    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"]

    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
        cache_config = vllm_config.cache_config
        lora_config = vllm_config.lora_config
        scheduler_config = vllm_config.scheduler_config
        assert not cache_config.enable_prefix_caching, \
            "NemotronH currently does not support prefix caching"

        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,
        )
        # Used to track and store by the Mamba cache between steps.
        self.mamba_cache: Optional[MambaCacheManager] = None

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

        self.make_empty_intmd_tensors = (self.model.make_empty_intmd_tensors)

    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: Optional[IntermediateTensors] = None,
                inputs_embeds: Optional[torch.Tensor] = None,
                **kwargs):

        mamba_cache_params = None
        if not envs.VLLM_USE_V1:
            if self.mamba_cache is None:

                num_mamba_layers = \
                    self.model_config.get_num_layers_by_block_type(
                        self.vllm_config.parallel_config,
                        LayerBlockType.mamba
                    )

                self.mamba_cache = MambaCacheManager(
                    self.vllm_config, self.lm_head.weight.dtype,
                    num_mamba_layers, *self._get_mamba_cache_shape())

            mamba_cache_params = self.mamba_cache.current_run_tensors(**kwargs)

        hidden_states = self.model(input_ids, positions, mamba_cache_params,
                                   intermediate_tensors, inputs_embeds)

        return hidden_states

    def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs):
        return self.mamba_cache.copy_inputs_before_cuda_graphs(
            input_buffers, **kwargs)

    def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
        return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size)

    def _get_mamba_cache_shape(
            self) -> tuple[tuple[int, int], tuple[int, int]]:
        world_size = get_tensor_model_parallel_world_size()
        hidden_size = self.config.hidden_size

        conv_state_shape, temporal_state_shape = None, None

        intermediate_size = self.config.expand * hidden_size

        # if n_groups is not divisible by world_size, need to extend the shards
        # to ensure all groups needed by a head is sharded along with it
        n_groups = (
            self.config.n_groups +
            extra_groups_for_head_shards(self.config.n_groups, world_size))

        # - heads and n_groups are TP-ed
        conv_dim = (intermediate_size +
                    2 * n_groups * self.config.ssm_state_size)
        conv_state_shape = (
            divide(conv_dim, world_size),
            self.config.conv_kernel - 1,
        )

        # These are not TP-ed as they depend on A, dt_bias, D
        # - they are typically small
        #   e.g., (h_heads, d_head, d_state) = (128, 64, 128)
        temporal_state_shape = (
            divide(self.config.mamba_num_heads, world_size),
            self.config.mamba_head_dim,
            self.config.ssm_state_size,
        )
        return conv_state_shape, temporal_state_shape

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states,
                                       sampling_metadata)
        return logits

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        # update name in weights before passing to loader
        updated_weights = []
        for name, loaded_weight in weights:
            name = name.replace("backbone", "model")
            updated_weights.append((name, loaded_weight))
        loader = AutoWeightsLoader(self)
        return loader.load_weights(updated_weights)

config instance-attribute

config = config

embedding_modules class-attribute instance-attribute

embedding_modules = {
    "embed_tokens": "input_embeddings",
    "lm_head": "output_embeddings",
}

embedding_padding_modules class-attribute instance-attribute

embedding_padding_modules = ['lm_head']

lm_head instance-attribute

lm_head = ParallelLMHead(
    unpadded_vocab_size,
    hidden_size,
    org_num_embeddings=vocab_size,
    padding_size=DEFAULT_VOCAB_PADDING_SIZE
    if not lora_config
    else lora_vocab_padding_size,
)

logits_processor instance-attribute

logits_processor = LogitsProcessor(
    unpadded_vocab_size, vocab_size
)

make_empty_intmd_tensors instance-attribute

make_empty_intmd_tensors = make_empty_intmd_tensors

mamba_cache instance-attribute

mamba_cache: Optional[MambaCacheManager] = None

model instance-attribute

model = NemotronHModel(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "model"),
)

model_config instance-attribute

model_config = model_config

packed_modules_mapping class-attribute instance-attribute

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

quant_config instance-attribute

quant_config = quant_config

scheduler_config instance-attribute

scheduler_config = scheduler_config

unpadded_vocab_size instance-attribute

unpadded_vocab_size = vocab_size

vllm_config instance-attribute

vllm_config = vllm_config

__init__

__init__(*, vllm_config: VllmConfig, prefix: str = '')

Source code in vllm/model_executor/models/nemotron_h.py

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
    cache_config = vllm_config.cache_config
    lora_config = vllm_config.lora_config
    scheduler_config = vllm_config.scheduler_config
    assert not cache_config.enable_prefix_caching, \
        "NemotronH currently does not support prefix caching"

    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,
    )
    # Used to track and store by the Mamba cache between steps.
    self.mamba_cache: Optional[MambaCacheManager] = None

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

    self.make_empty_intmd_tensors = (self.model.make_empty_intmd_tensors)

_get_mamba_cache_shape

_get_mamba_cache_shape() -> tuple[
    tuple[int, int], tuple[int, int]
]

Source code in vllm/model_executor/models/nemotron_h.py

def _get_mamba_cache_shape(
        self) -> tuple[tuple[int, int], tuple[int, int]]:
    world_size = get_tensor_model_parallel_world_size()
    hidden_size = self.config.hidden_size

    conv_state_shape, temporal_state_shape = None, None

    intermediate_size = self.config.expand * hidden_size

    # if n_groups is not divisible by world_size, need to extend the shards
    # to ensure all groups needed by a head is sharded along with it
    n_groups = (
        self.config.n_groups +
        extra_groups_for_head_shards(self.config.n_groups, world_size))

    # - heads and n_groups are TP-ed
    conv_dim = (intermediate_size +
                2 * n_groups * self.config.ssm_state_size)
    conv_state_shape = (
        divide(conv_dim, world_size),
        self.config.conv_kernel - 1,
    )

    # These are not TP-ed as they depend on A, dt_bias, D
    # - they are typically small
    #   e.g., (h_heads, d_head, d_state) = (128, 64, 128)
    temporal_state_shape = (
        divide(self.config.mamba_num_heads, world_size),
        self.config.mamba_head_dim,
        self.config.ssm_state_size,
    )
    return conv_state_shape, temporal_state_shape

compute_logits

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]

Source code in vllm/model_executor/models/nemotron_h.py

def compute_logits(
    self,
    hidden_states: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[torch.Tensor]:
    logits = self.logits_processor(self.lm_head, hidden_states,
                                   sampling_metadata)
    return logits

copy_inputs_before_cuda_graphs

copy_inputs_before_cuda_graphs(input_buffers, **kwargs)

Source code in vllm/model_executor/models/nemotron_h.py

def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs):
    return self.mamba_cache.copy_inputs_before_cuda_graphs(
        input_buffers, **kwargs)

forward

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    **kwargs,
)

Source code in vllm/model_executor/models/nemotron_h.py

def forward(self,
            input_ids: torch.Tensor,
            positions: torch.Tensor,
            intermediate_tensors: Optional[IntermediateTensors] = None,
            inputs_embeds: Optional[torch.Tensor] = None,
            **kwargs):

    mamba_cache_params = None
    if not envs.VLLM_USE_V1:
        if self.mamba_cache is None:

            num_mamba_layers = \
                self.model_config.get_num_layers_by_block_type(
                    self.vllm_config.parallel_config,
                    LayerBlockType.mamba
                )

            self.mamba_cache = MambaCacheManager(
                self.vllm_config, self.lm_head.weight.dtype,
                num_mamba_layers, *self._get_mamba_cache_shape())

        mamba_cache_params = self.mamba_cache.current_run_tensors(**kwargs)

    hidden_states = self.model(input_ids, positions, mamba_cache_params,
                               intermediate_tensors, inputs_embeds)

    return hidden_states

get_input_embeddings

get_input_embeddings(input_ids: Tensor) -> Tensor

Source code in vllm/model_executor/models/nemotron_h.py

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

get_seqlen_agnostic_capture_inputs

get_seqlen_agnostic_capture_inputs(batch_size: int)

Source code in vllm/model_executor/models/nemotron_h.py

def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
    return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size)

load_weights

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]

Source code in vllm/model_executor/models/nemotron_h.py

def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    # update name in weights before passing to loader
    updated_weights = []
    for name, loaded_weight in weights:
        name = name.replace("backbone", "model")
        updated_weights.append((name, loaded_weight))
    loader = AutoWeightsLoader(self)
    return loader.load_weights(updated_weights)

NemotronHMLP

Bases: Module

Source code in vllm/model_executor/models/nemotron_h.py

class NemotronHMLP(nn.Module):

    def __init__(
        self,
        config: NemotronHConfig,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.up_proj = ColumnParallelLinear(
            input_size=config.hidden_size,
            output_size=config.intermediate_size,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.up_proj",
        )
        self.down_proj = RowParallelLinear(
            input_size=config.intermediate_size,
            output_size=config.hidden_size,
            bias=bias,
            quant_config=quant_config,
            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

act_fn instance-attribute

act_fn = ReLUSquaredActivation()

down_proj instance-attribute

down_proj = RowParallelLinear(
    input_size=intermediate_size,
    output_size=hidden_size,
    bias=bias,
    quant_config=quant_config,
    prefix=f"{prefix}.down_proj",
)

up_proj instance-attribute

up_proj = ColumnParallelLinear(
    input_size=hidden_size,
    output_size=intermediate_size,
    bias=bias,
    quant_config=quant_config,
    prefix=f"{prefix}.up_proj",
)

__init__

__init__(
    config: NemotronHConfig,
    quant_config: Optional[QuantizationConfig] = None,
    bias: bool = False,
    prefix: str = "",
) -> None

Source code in vllm/model_executor/models/nemotron_h.py

def __init__(
    self,
    config: NemotronHConfig,
    quant_config: Optional[QuantizationConfig] = None,
    bias: bool = False,
    prefix: str = "",
) -> None:
    super().__init__()
    self.up_proj = ColumnParallelLinear(
        input_size=config.hidden_size,
        output_size=config.intermediate_size,
        bias=bias,
        quant_config=quant_config,
        prefix=f"{prefix}.up_proj",
    )
    self.down_proj = RowParallelLinear(
        input_size=config.intermediate_size,
        output_size=config.hidden_size,
        bias=bias,
        quant_config=quant_config,
        prefix=f"{prefix}.down_proj",
    )
    self.act_fn = ReLUSquaredActivation()

forward

forward(x: Tensor)

Source code in vllm/model_executor/models/nemotron_h.py

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

NemotronHMLPDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/nemotron_h.py

class NemotronHMLPDecoderLayer(nn.Module):

    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config

        self.mixer = NemotronHMLP(
            config,
            quant_config=quant_config,
            bias=config.mlp_bias,
            prefix=f"{prefix}.mixer",
        )

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        **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)
        return hidden_states, residual

config instance-attribute

config = config

mixer instance-attribute

mixer = NemotronHMLP(
    config,
    quant_config=quant_config,
    bias=mlp_bias,
    prefix=f"{prefix}.mixer",
)

norm instance-attribute

norm = RMSNorm(hidden_size, eps=rms_norm_eps)

__init__

__init__(
    config: NemotronHConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None

Source code in vllm/model_executor/models/nemotron_h.py

def __init__(
    self,
    config: NemotronHConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.config = config

    self.mixer = NemotronHMLP(
        config,
        quant_config=quant_config,
        bias=config.mlp_bias,
        prefix=f"{prefix}.mixer",
    )

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

forward

forward(
    hidden_states: Tensor,
    residual: Optional[Tensor],
    **kwargs,
)

Source code in vllm/model_executor/models/nemotron_h.py

def forward(
    self,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
    **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)
    return hidden_states, residual

NemotronHMambaDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/nemotron_h.py

class NemotronHMambaDecoderLayer(nn.Module):

    def __init__(
        self,
        config: NemotronHConfig,
        layer_idx: int,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.mixer = MambaMixer2(
            hidden_size=config.hidden_size,
            ssm_state_size=config.ssm_state_size,
            conv_kernel_size=config.conv_kernel,
            intermediate_size=config.expand * config.hidden_size,
            use_conv_bias=config.use_conv_bias,
            use_bias=config.use_bias,
            n_groups=config.n_groups,
            num_heads=config.mamba_num_heads,
            head_dim=config.mamba_head_dim,
            rms_norm_eps=config.rms_norm_eps,
            activation=config.mamba_hidden_act,
            quant_config=quant_config,
            prefix=f"{prefix}.mixer",
            chunk_size=config.chunk_size,
        )

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        mamba_cache_params: MambaCacheParams,
        mamba2_metadata: Mamba2Metadata,
        **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, mamba_cache_params,
                                   mamba2_metadata)
        return hidden_states, residual

config instance-attribute

config = config

mixer instance-attribute

mixer = MambaMixer2(
    hidden_size=hidden_size,
    ssm_state_size=ssm_state_size,
    conv_kernel_size=conv_kernel,
    intermediate_size=expand * hidden_size,
    use_conv_bias=use_conv_bias,
    use_bias=use_bias,
    n_groups=n_groups,
    num_heads=mamba_num_heads,
    head_dim=mamba_head_dim,
    rms_norm_eps=rms_norm_eps,
    activation=mamba_hidden_act,
    quant_config=quant_config,
    prefix=f"{prefix}.mixer",
    chunk_size=chunk_size,
)

norm instance-attribute

norm = RMSNorm(hidden_size, eps=rms_norm_eps)

__init__

__init__(
    config: NemotronHConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None

Source code in vllm/model_executor/models/nemotron_h.py

def __init__(
    self,
    config: NemotronHConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.config = config
    self.mixer = MambaMixer2(
        hidden_size=config.hidden_size,
        ssm_state_size=config.ssm_state_size,
        conv_kernel_size=config.conv_kernel,
        intermediate_size=config.expand * config.hidden_size,
        use_conv_bias=config.use_conv_bias,
        use_bias=config.use_bias,
        n_groups=config.n_groups,
        num_heads=config.mamba_num_heads,
        head_dim=config.mamba_head_dim,
        rms_norm_eps=config.rms_norm_eps,
        activation=config.mamba_hidden_act,
        quant_config=quant_config,
        prefix=f"{prefix}.mixer",
        chunk_size=config.chunk_size,
    )

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

forward

forward(
    hidden_states: Tensor,
    residual: Optional[Tensor],
    mamba_cache_params: MambaCacheParams,
    mamba2_metadata: Mamba2Metadata,
    **kwargs,
)

Source code in vllm/model_executor/models/nemotron_h.py

def forward(
    self,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
    mamba_cache_params: MambaCacheParams,
    mamba2_metadata: Mamba2Metadata,
    **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, mamba_cache_params,
                               mamba2_metadata)
    return hidden_states, residual

NemotronHModel

Bases: Module

Source code in vllm/model_executor/models/nemotron_h.py

class NemotronHModel(nn.Module):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config: NemotronHConfig = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
        lora_vocab = ((lora_config.lora_extra_vocab_size *
                       (lora_config.max_loras or 1)) if lora_config else 0)
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            self.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
        )

        def get_layer(prefix: str):
            layer_idx = int(prefix.rsplit(".", 1)[1])
            layer_class = ALL_DECODER_LAYER_TYPES[
                config.hybrid_override_pattern[layer_idx]]
            return layer_class(
                config,
                layer_idx,
                cache_config,
                quant_config=quant_config,
                prefix=prefix,
            )

        self.start_layer, self.end_layer, self.layers = make_layers(
            len(config.hybrid_override_pattern),
            get_layer,
            prefix=f"{prefix}.layers")
        self.make_empty_intmd_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size)

        self.norm_f = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        mamba_cache_params: MambaCacheParams,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:

        attn_metadata = get_forward_context().attn_metadata

        if not envs.VLLM_USE_V1:
            mamba2_metadata = prepare_mamba2_metadata(
                chunk_size=self.config.chunk_size,
                attn_metadata=attn_metadata,
            )
        else:
            # v1 get mamba2_metadata from forward_context
            mamba2_metadata = None

        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)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        residual = None
        num_non_mamba_layers = 0
        for i in range(len(self.layers)):
            layer = self.layers[i]
            layer_mamba_cache_params = None
            if isinstance(layer,
                          NemotronHMambaDecoderLayer) and mamba_cache_params:
                layer_mamba_cache_params = mamba_cache_params.at_layer_idx(
                    i - num_non_mamba_layers)
            else:
                num_non_mamba_layers += 1

            hidden_states, residual = layer(
                positions=positions,
                hidden_states=hidden_states,
                residual=residual,
                mamba_cache_params=layer_mamba_cache_params,
                mamba2_metadata=mamba2_metadata,
            )

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
        hidden_states, _ = self.norm_f(hidden_states, residual)
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        attb_params_mapping = {
            "q_proj": "q",
            "k_proj": "k",
            "v_proj": "v",
        }

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "embeddings" in name:
                name = name.replace("embeddings", "embed_tokens")

            if "A_log" in name:
                name = name.replace("A_log", "A")
                loaded_weight = loaded_weight.to(torch.float32)

            if "D" in name:
                loaded_weight = loaded_weight.to(torch.float32)

            if "dt_bias" in name:
                loaded_weight = loaded_weight.to(torch.float32)

            # load attn params
            if any(proj in name for proj in ["q_proj", "k_proj", "v_proj"]):
                weight_name = next(proj
                                   for proj in ["q_proj", "k_proj", "v_proj"]
                                   if proj in name)
                name = name.replace(weight_name, "qkv_proj")
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight,
                              attb_params_mapping[weight_name])
            # load other params
            else:
                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

config instance-attribute

config = config

embed_tokens instance-attribute

embed_tokens = VocabParallelEmbedding(
    vocab_size, hidden_size, org_num_embeddings=vocab_size
)

make_empty_intmd_tensors instance-attribute

make_empty_intmd_tensors = (
    make_empty_intermediate_tensors_factory(
        ["hidden_states", "residual"], hidden_size
    )
)

norm_f instance-attribute

norm_f = RMSNorm(hidden_size, eps=rms_norm_eps)

org_vocab_size instance-attribute

org_vocab_size = vocab_size

vocab_size instance-attribute

vocab_size = vocab_size + lora_vocab

__init__

__init__(*, vllm_config: VllmConfig, prefix: str = '')

Source code in vllm/model_executor/models/nemotron_h.py

def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()

    config: NemotronHConfig = vllm_config.model_config.hf_config
    cache_config = vllm_config.cache_config
    quant_config = vllm_config.quant_config
    lora_config = vllm_config.lora_config

    self.config = config
    lora_vocab = ((lora_config.lora_extra_vocab_size *
                   (lora_config.max_loras or 1)) if lora_config else 0)
    self.vocab_size = config.vocab_size + lora_vocab
    self.org_vocab_size = config.vocab_size

    self.embed_tokens = VocabParallelEmbedding(
        self.vocab_size,
        config.hidden_size,
        org_num_embeddings=config.vocab_size,
    )

    def get_layer(prefix: str):
        layer_idx = int(prefix.rsplit(".", 1)[1])
        layer_class = ALL_DECODER_LAYER_TYPES[
            config.hybrid_override_pattern[layer_idx]]
        return layer_class(
            config,
            layer_idx,
            cache_config,
            quant_config=quant_config,
            prefix=prefix,
        )

    self.start_layer, self.end_layer, self.layers = make_layers(
        len(config.hybrid_override_pattern),
        get_layer,
        prefix=f"{prefix}.layers")
    self.make_empty_intmd_tensors = make_empty_intermediate_tensors_factory(
        ["hidden_states", "residual"], config.hidden_size)

    self.norm_f = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

forward

forward(
    input_ids: Tensor,
    positions: Tensor,
    mamba_cache_params: MambaCacheParams,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
) -> Tensor

Source code in vllm/model_executor/models/nemotron_h.py

def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    mamba_cache_params: MambaCacheParams,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:

    attn_metadata = get_forward_context().attn_metadata

    if not envs.VLLM_USE_V1:
        mamba2_metadata = prepare_mamba2_metadata(
            chunk_size=self.config.chunk_size,
            attn_metadata=attn_metadata,
        )
    else:
        # v1 get mamba2_metadata from forward_context
        mamba2_metadata = None

    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)
        residual = None
    else:
        assert intermediate_tensors is not None
        hidden_states = intermediate_tensors["hidden_states"]
        residual = intermediate_tensors["residual"]

    residual = None
    num_non_mamba_layers = 0
    for i in range(len(self.layers)):
        layer = self.layers[i]
        layer_mamba_cache_params = None
        if isinstance(layer,
                      NemotronHMambaDecoderLayer) and mamba_cache_params:
            layer_mamba_cache_params = mamba_cache_params.at_layer_idx(
                i - num_non_mamba_layers)
        else:
            num_non_mamba_layers += 1

        hidden_states, residual = layer(
            positions=positions,
            hidden_states=hidden_states,
            residual=residual,
            mamba_cache_params=layer_mamba_cache_params,
            mamba2_metadata=mamba2_metadata,
        )

    if not get_pp_group().is_last_rank:
        return IntermediateTensors({
            "hidden_states": hidden_states,
            "residual": residual
        })
    hidden_states, _ = self.norm_f(hidden_states, residual)
    return hidden_states

get_input_embeddings

get_input_embeddings(input_ids: Tensor) -> Tensor

Source code in vllm/model_executor/models/nemotron_h.py

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

load_weights

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]

Source code in vllm/model_executor/models/nemotron_h.py

def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    attb_params_mapping = {
        "q_proj": "q",
        "k_proj": "k",
        "v_proj": "v",
    }

    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if "embeddings" in name:
            name = name.replace("embeddings", "embed_tokens")

        if "A_log" in name:
            name = name.replace("A_log", "A")
            loaded_weight = loaded_weight.to(torch.float32)

        if "D" in name:
            loaded_weight = loaded_weight.to(torch.float32)

        if "dt_bias" in name:
            loaded_weight = loaded_weight.to(torch.float32)

        # load attn params
        if any(proj in name for proj in ["q_proj", "k_proj", "v_proj"]):
            weight_name = next(proj
                               for proj in ["q_proj", "k_proj", "v_proj"]
                               if proj in name)
            name = name.replace(weight_name, "qkv_proj")
            param = params_dict[name]
            weight_loader = param.weight_loader
            weight_loader(param, loaded_weight,
                          attb_params_mapping[weight_name])
        # load other params
        else:
            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