vllm.model_executor.models.granitemoehybrid

Inference-only GraniteMoeHybrid model.

ALL_DECODER_LAYER_TYPES module-attribute

ALL_DECODER_LAYER_TYPES = {
    "attention": GraniteMoeHybridAttentionDecoderLayer,
    "mamba": GraniteMoeHybridMambaDecoderLayer,
}

GraniteMoeHybridAttention

Bases: Module

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

class GraniteMoeHybridAttention(nn.Module):

    def __init__(
        self,
        config: GraniteMoeHybridConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.causal = True
        self.hidden_size = config.hidden_size
        self.attention_bias = config.attention_bias
        self.attention_multiplier = config.attention_multiplier
        self.total_num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.total_num_heads
        self.total_num_kv_heads = config.num_key_value_heads

        # TensorParallel logic
        tp_size = get_tensor_model_parallel_world_size()
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        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_key_value_heads = max(1, self.total_num_kv_heads // tp_size)

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

        self.o_proj = RowParallelLinear(self.hidden_size,
                                        self.hidden_size,
                                        bias=self.attention_bias,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")

        if config.position_embedding_type == "rope":
            self.rotary_emb = get_rope(
                self.head_dim,
                rotary_dim=self.head_dim,
                max_position=config.max_position_embeddings,
                base=int(config.rope_theta),
                rope_scaling=config.rope_scaling \
                    if hasattr(config, "rope_scaling") \
                    and config.rope_scaling is not None else None,
                is_neox_style=True,
            )
        else:
            self.rotary_emb = None

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

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:

        qkv, _ = self.qkv_proj(hidden_states)
        query, key, value = qkv.split([
            self.num_heads * self.head_dim, self.num_key_value_heads *
            self.head_dim, self.num_key_value_heads * self.head_dim
        ],
                                      dim=-1)

        if self.rotary_emb is not None:
            query, key = self.rotary_emb(positions, query, key)

        hidden_states = self.attn(query, key, value)
        del query, key, value

        hidden_states = self.o_proj(hidden_states)[0]
        return hidden_states

attention_bias instance-attribute

attention_bias = attention_bias

attention_multiplier instance-attribute

attention_multiplier = attention_multiplier

attn instance-attribute

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

causal instance-attribute

causal = True

head_dim instance-attribute

head_dim = hidden_size // total_num_heads

hidden_size instance-attribute

hidden_size = hidden_size

num_heads instance-attribute

num_heads = total_num_heads // tp_size

num_key_value_heads instance-attribute

num_key_value_heads = max(1, total_num_kv_heads // tp_size)

o_proj instance-attribute

o_proj = RowParallelLinear(
    hidden_size,
    hidden_size,
    bias=attention_bias,
    quant_config=quant_config,
    prefix=f"{prefix}.o_proj",
)

qkv_proj instance-attribute

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

rotary_emb instance-attribute

rotary_emb = get_rope(
    head_dim,
    rotary_dim=head_dim,
    max_position=max_position_embeddings,
    base=int(rope_theta),
    rope_scaling=rope_scaling
    if hasattr(config, "rope_scaling")
    and rope_scaling is not None
    else None,
    is_neox_style=True,
)

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: GraniteMoeHybridConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None

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

def __init__(
    self,
    config: GraniteMoeHybridConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.causal = True
    self.hidden_size = config.hidden_size
    self.attention_bias = config.attention_bias
    self.attention_multiplier = config.attention_multiplier
    self.total_num_heads = config.num_attention_heads
    self.head_dim = self.hidden_size // self.total_num_heads
    self.total_num_kv_heads = config.num_key_value_heads

    # TensorParallel logic
    tp_size = get_tensor_model_parallel_world_size()
    assert self.total_num_heads % tp_size == 0
    self.num_heads = self.total_num_heads // tp_size
    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_key_value_heads = max(1, self.total_num_kv_heads // tp_size)

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

    self.o_proj = RowParallelLinear(self.hidden_size,
                                    self.hidden_size,
                                    bias=self.attention_bias,
                                    quant_config=quant_config,
                                    prefix=f"{prefix}.o_proj")

    if config.position_embedding_type == "rope":
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=config.max_position_embeddings,
            base=int(config.rope_theta),
            rope_scaling=config.rope_scaling \
                if hasattr(config, "rope_scaling") \
                and config.rope_scaling is not None else None,
            is_neox_style=True,
        )
    else:
        self.rotary_emb = None

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

forward

forward(positions: Tensor, hidden_states: Tensor) -> Tensor

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

def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:

    qkv, _ = self.qkv_proj(hidden_states)
    query, key, value = qkv.split([
        self.num_heads * self.head_dim, self.num_key_value_heads *
        self.head_dim, self.num_key_value_heads * self.head_dim
    ],
                                  dim=-1)

    if self.rotary_emb is not None:
        query, key = self.rotary_emb(positions, query, key)

    hidden_states = self.attn(query, key, value)
    del query, key, value

    hidden_states = self.o_proj(hidden_states)[0]
    return hidden_states

GraniteMoeHybridAttentionDecoderLayer

Bases: Module

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

class GraniteMoeHybridAttentionDecoderLayer(nn.Module):

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

        self.self_attn = GraniteMoeHybridAttention(
            config,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn")

        self.block_sparse_moe = None
        if getattr(config, "num_local_experts", 0) > 0:
            self.block_sparse_moe = GraniteMoeMoE(
                num_experts=config.num_local_experts,
                top_k=config.num_experts_per_tok,
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                quant_config=quant_config,
                prefix=f"{prefix}.block_sparse_moe")

        self.shared_mlp = None if \
            getattr(config, 'shared_intermediate_size', 0) == 0 \
            else GraniteMoeSharedMLP(
                config,
                quant_config=quant_config,
                prefix=f"{prefix}.shared_mlp"
            )

        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)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        mamba_cache_params: MambaCacheParams,
        mamba2_metadata: Mamba2Metadata,
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)

        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )
        hidden_states = residual + hidden_states * self.residual_multiplier

        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        if self.shared_mlp is None:
            if self.block_sparse_moe is not None:
                hidden_states = self.block_sparse_moe(hidden_states)
            # else: skip
        else:
            # create a copy since block_sparse_moe modifies in-place
            if self.block_sparse_moe is not None:
                moe_hidden_states = hidden_states.clone()
                moe_hidden_states = self.block_sparse_moe(moe_hidden_states)
                hidden_states = moe_hidden_states + self.shared_mlp(
                    hidden_states)
                del moe_hidden_states
            else:
                hidden_states = self.shared_mlp(hidden_states)
        hidden_states = residual + hidden_states * self.residual_multiplier

        return hidden_states, residual

block_sparse_moe instance-attribute

block_sparse_moe = None

hidden_size instance-attribute

hidden_size = hidden_size

input_layernorm instance-attribute

input_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)

post_attention_layernorm instance-attribute

post_attention_layernorm = RMSNorm(
    hidden_size, eps=rms_norm_eps
)

residual_multiplier instance-attribute

residual_multiplier = residual_multiplier

self_attn instance-attribute

self_attn = GraniteMoeHybridAttention(
    config,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.self_attn",
)

shared_mlp instance-attribute

shared_mlp = (
    None
    if getattr(config, "shared_intermediate_size", 0) == 0
    else GraniteMoeSharedMLP(
        config,
        quant_config=quant_config,
        prefix=f"{prefix}.shared_mlp",
    )
)

__init__

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

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

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

    self.self_attn = GraniteMoeHybridAttention(
        config,
        cache_config=cache_config,
        quant_config=quant_config,
        prefix=f"{prefix}.self_attn")

    self.block_sparse_moe = None
    if getattr(config, "num_local_experts", 0) > 0:
        self.block_sparse_moe = GraniteMoeMoE(
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            quant_config=quant_config,
            prefix=f"{prefix}.block_sparse_moe")

    self.shared_mlp = None if \
        getattr(config, 'shared_intermediate_size', 0) == 0 \
        else GraniteMoeSharedMLP(
            config,
            quant_config=quant_config,
            prefix=f"{prefix}.shared_mlp"
        )

    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)

forward

forward(
    positions: Tensor,
    hidden_states: Tensor,
    residual: Optional[Tensor],
    mamba_cache_params: MambaCacheParams,
    mamba2_metadata: Mamba2Metadata,
) -> Tensor

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

def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
    mamba_cache_params: MambaCacheParams,
    mamba2_metadata: Mamba2Metadata,
) -> torch.Tensor:
    residual = hidden_states
    hidden_states = self.input_layernorm(hidden_states)

    hidden_states = self.self_attn(
        positions=positions,
        hidden_states=hidden_states,
    )
    hidden_states = residual + hidden_states * self.residual_multiplier

    residual = hidden_states
    hidden_states = self.post_attention_layernorm(hidden_states)
    if self.shared_mlp is None:
        if self.block_sparse_moe is not None:
            hidden_states = self.block_sparse_moe(hidden_states)
        # else: skip
    else:
        # create a copy since block_sparse_moe modifies in-place
        if self.block_sparse_moe is not None:
            moe_hidden_states = hidden_states.clone()
            moe_hidden_states = self.block_sparse_moe(moe_hidden_states)
            hidden_states = moe_hidden_states + self.shared_mlp(
                hidden_states)
            del moe_hidden_states
        else:
            hidden_states = self.shared_mlp(hidden_states)
    hidden_states = residual + hidden_states * self.residual_multiplier

    return hidden_states, residual

GraniteMoeHybridForCausalLM

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

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

class GraniteMoeHybridForCausalLM(nn.Module, HasInnerState, SupportsLoRA,
                                  SupportsPP, IsHybrid, SupportsQuant):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }
    embedding_padding_modules = ["lm_head"]

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

        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
        if cache_config.enable_prefix_caching:
            raise RuntimeError(
                "GraniteMoeHybrid currently does not support prefix caching")

        self.quant_config = vllm_config.quant_config
        self.config = config
        self.scheduler_config = scheduler_config
        self.model = GraniteMoeHybridModel(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,
            quant_config=self.quant_config,
            prefix=maybe_prefix(prefix, "lm_head"))
        if config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size,
                                                scale=1 /
                                                self.config.logits_scaling)

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

        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_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.model_config.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.mamba_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.mamba_n_groups + extra_groups_for_head_shards(
            self.config.mamba_n_groups, world_size))

        # - heads and n_groups are TP-ed
        conv_dim = (intermediate_size +
                    2 * n_groups * self.config.mamba_d_state)
        conv_state_shape = (
            divide(conv_dim, world_size),
            self.config.mamba_d_conv - 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_n_heads, world_size),
            self.config.mamba_d_head,
            self.config.mamba_d_state,
        )
        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]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(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,
    quant_config=quant_config,
    prefix=maybe_prefix(prefix, "lm_head"),
)

logits_processor instance-attribute

logits_processor = LogitsProcessor(
    unpadded_vocab_size,
    vocab_size,
    scale=1 / logits_scaling,
)

make_empty_intermediate_tensors instance-attribute

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

mamba_cache instance-attribute

mamba_cache: Optional[MambaCacheManager] = None

model instance-attribute

model = GraniteMoeHybridModel(
    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/granitemoehybrid.py

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

    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
    if cache_config.enable_prefix_caching:
        raise RuntimeError(
            "GraniteMoeHybrid currently does not support prefix caching")

    self.quant_config = vllm_config.quant_config
    self.config = config
    self.scheduler_config = scheduler_config
    self.model = GraniteMoeHybridModel(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,
        quant_config=self.quant_config,
        prefix=maybe_prefix(prefix, "lm_head"))
    if config.tie_word_embeddings:
        self.lm_head.weight = self.model.embed_tokens.weight
    self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                            config.vocab_size,
                                            scale=1 /
                                            self.config.logits_scaling)

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

    self.make_empty_intermediate_tensors = (
        self.model.make_empty_intermediate_tensors)

_get_mamba_cache_shape

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

Source code in vllm/model_executor/models/granitemoehybrid.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.mamba_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.mamba_n_groups + extra_groups_for_head_shards(
        self.config.mamba_n_groups, world_size))

    # - heads and n_groups are TP-ed
    conv_dim = (intermediate_size +
                2 * n_groups * self.config.mamba_d_state)
    conv_state_shape = (
        divide(conv_dim, world_size),
        self.config.mamba_d_conv - 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_n_heads, world_size),
        self.config.mamba_d_head,
        self.config.mamba_d_state,
    )
    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/granitemoehybrid.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/granitemoehybrid.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/granitemoehybrid.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.model_config.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/granitemoehybrid.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/granitemoehybrid.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/granitemoehybrid.py

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

GraniteMoeHybridMambaDecoderLayer

Bases: Module

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

class GraniteMoeHybridMambaDecoderLayer(nn.Module):

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

        self.mamba = MambaMixer2(hidden_size= config.hidden_size,
                                ssm_state_size = config.mamba_d_state,
                                conv_kernel_size = config.mamba_d_conv,
                                intermediate_size = config.mamba_expand *\
                                                    config.hidden_size,
                                use_conv_bias = config.mamba_conv_bias,
                                use_bias = config.mamba_proj_bias,
                                n_groups=config.mamba_n_groups,
                                num_heads=config.mamba_n_heads,
                                head_dim=config.mamba_d_head,
                                rms_norm_eps=config.rms_norm_eps,
                                activation=config.hidden_act,
                                quant_config=quant_config,
                                prefix=f"{prefix}.mixer",
                                chunk_size=config.mamba_chunk_size)

        self.block_sparse_moe = None
        if getattr(config, "num_local_experts", 0) > 0:
            self.block_sparse_moe = GraniteMoeMoE(
                num_experts=config.num_local_experts,
                top_k=config.num_experts_per_tok,
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                quant_config=quant_config,
                prefix=f"{prefix}.block_sparse_moe")

        self.shared_mlp = None if \
            getattr(config, 'shared_intermediate_size', 0) == 0 \
            else GraniteMoeSharedMLP(
                config,
                quant_config=quant_config,
                prefix=f"{prefix}.shared_mlp"
            )

        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)

    def forward(
        self,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        mamba_cache_params: MambaCacheParams,
        mamba2_metadata: Mamba2Metadata,
        **kwargs,
    ):
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.mamba(hidden_states, mamba_cache_params,
                                   mamba2_metadata)
        hidden_states = residual + hidden_states * self.residual_multiplier

        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        if self.shared_mlp is None:
            if self.block_sparse_moe is not None:
                hidden_states = self.block_sparse_moe(hidden_states)
            # else: skip
        else:
            # create a copy since block_sparse_moe modifies in-place
            if self.block_sparse_moe is not None:
                moe_hidden_states = hidden_states.clone()
                moe_hidden_states = self.block_sparse_moe(moe_hidden_states)
                hidden_states = moe_hidden_states + self.shared_mlp(
                    hidden_states)
                del moe_hidden_states
            else:
                hidden_states = self.shared_mlp(hidden_states)
        hidden_states = residual + hidden_states * self.residual_multiplier

        return hidden_states, residual

block_sparse_moe instance-attribute

block_sparse_moe = None

config instance-attribute

config = config

hidden_size instance-attribute

hidden_size = hidden_size

input_layernorm instance-attribute

input_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)

mamba instance-attribute

mamba = MambaMixer2(
    hidden_size=hidden_size,
    ssm_state_size=mamba_d_state,
    conv_kernel_size=mamba_d_conv,
    intermediate_size=mamba_expand * hidden_size,
    use_conv_bias=mamba_conv_bias,
    use_bias=mamba_proj_bias,
    n_groups=mamba_n_groups,
    num_heads=mamba_n_heads,
    head_dim=mamba_d_head,
    rms_norm_eps=rms_norm_eps,
    activation=hidden_act,
    quant_config=quant_config,
    prefix=f"{prefix}.mixer",
    chunk_size=mamba_chunk_size,
)

post_attention_layernorm instance-attribute

post_attention_layernorm = RMSNorm(
    hidden_size, eps=rms_norm_eps
)

residual_multiplier instance-attribute

residual_multiplier = residual_multiplier

shared_mlp instance-attribute

shared_mlp = (
    None
    if getattr(config, "shared_intermediate_size", 0) == 0
    else GraniteMoeSharedMLP(
        config,
        quant_config=quant_config,
        prefix=f"{prefix}.shared_mlp",
    )
)

__init__

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

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

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

    self.mamba = MambaMixer2(hidden_size= config.hidden_size,
                            ssm_state_size = config.mamba_d_state,
                            conv_kernel_size = config.mamba_d_conv,
                            intermediate_size = config.mamba_expand *\
                                                config.hidden_size,
                            use_conv_bias = config.mamba_conv_bias,
                            use_bias = config.mamba_proj_bias,
                            n_groups=config.mamba_n_groups,
                            num_heads=config.mamba_n_heads,
                            head_dim=config.mamba_d_head,
                            rms_norm_eps=config.rms_norm_eps,
                            activation=config.hidden_act,
                            quant_config=quant_config,
                            prefix=f"{prefix}.mixer",
                            chunk_size=config.mamba_chunk_size)

    self.block_sparse_moe = None
    if getattr(config, "num_local_experts", 0) > 0:
        self.block_sparse_moe = GraniteMoeMoE(
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            quant_config=quant_config,
            prefix=f"{prefix}.block_sparse_moe")

    self.shared_mlp = None if \
        getattr(config, 'shared_intermediate_size', 0) == 0 \
        else GraniteMoeSharedMLP(
            config,
            quant_config=quant_config,
            prefix=f"{prefix}.shared_mlp"
        )

    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)

forward

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

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

def forward(
    self,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
    mamba_cache_params: MambaCacheParams,
    mamba2_metadata: Mamba2Metadata,
    **kwargs,
):
    residual = hidden_states
    hidden_states = self.input_layernorm(hidden_states)
    hidden_states = self.mamba(hidden_states, mamba_cache_params,
                               mamba2_metadata)
    hidden_states = residual + hidden_states * self.residual_multiplier

    residual = hidden_states
    hidden_states = self.post_attention_layernorm(hidden_states)
    if self.shared_mlp is None:
        if self.block_sparse_moe is not None:
            hidden_states = self.block_sparse_moe(hidden_states)
        # else: skip
    else:
        # create a copy since block_sparse_moe modifies in-place
        if self.block_sparse_moe is not None:
            moe_hidden_states = hidden_states.clone()
            moe_hidden_states = self.block_sparse_moe(moe_hidden_states)
            hidden_states = moe_hidden_states + self.shared_mlp(
                hidden_states)
            del moe_hidden_states
        else:
            hidden_states = self.shared_mlp(hidden_states)
    hidden_states = residual + hidden_states * self.residual_multiplier

    return hidden_states, residual

GraniteMoeHybridModel

Bases: Module

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

class GraniteMoeHybridModel(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
        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,
        )
        self.embedding_multiplier = config.embedding_multiplier

        def get_layer(prefix: str):
            layer_idx = int(prefix.rsplit(".", 1)[1])
            layer_class = ALL_DECODER_LAYER_TYPES[
                config.layer_types[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(
            config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers")
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

        self.norm = 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.mamba_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)
                hidden_states = hidden_states * self.embedding_multiplier
            residual = None
        else:
            if intermediate_tensors is None:
                raise RuntimeError('Intermediate tensors may not be None!')
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        num_attn = 0
        for i in range(len(self.layers)):
            layer = self.layers[i]
            if isinstance(layer, GraniteMoeHybridAttentionDecoderLayer):
                num_attn += 1

            layer_mamba_cache_params = None
            if isinstance(
                    layer,
                    GraniteMoeHybridMambaDecoderLayer) and mamba_cache_params:
                layer_mamba_cache_params = mamba_cache_params.at_layer_idx(
                    i - num_attn)

            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(hidden_states)
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()

        def _load(n, p):
            param = params_dict[n]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, p)
            loaded_params.add(n)

        def _load_shard(n, p, shard_id):
            # Skip layers on other devices.
            if not is_pp_missing_parameter(n, self):
                param = params_dict[n]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, p, shard_id)
                loaded_params.add(n)

        def _load_expert(n, p, name, shard_id, expert_id):
            param = params_dict[n]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param,
                          p,
                          name,
                          shard_id=shard_id,
                          expert_id=expert_id)
            loaded_params.add(n)

        for n, p in weights:
            if "A_log" in n:
                n = n.replace("A_log", "A")

            # Logic analogous to: https://github.com/vllm-project/vllm/blob/f49e5aff11c986ed4d45202b1716c5d74786efa9/vllm/model_executor/models/granitemoeshared.py#L215
            # Mapping different experts' layout:
            #  from HF (input_linear, output_linear, router)
            #  to vLLM (experts_w13({e}.w1, {e}.w2), experts_w3({e}.w3), gate)
            if n.endswith('.block_sparse_moe.input_linear.weight'):
                for e in range(p.size(0)):
                    w1_name = n.replace(
                        '.block_sparse_moe.input_linear.weight',
                        f".block_sparse_moe.experts.{e}.w1.weight")
                    w3_name = n.replace(
                        '.block_sparse_moe.input_linear.weight',
                        f".block_sparse_moe.experts.{e}.w3.weight")
                    w1_param, w3_param = p[e].chunk(2, dim=0)
                    _load_expert(n.replace('.input_linear.', '.experts.w13_'),
                                 w1_param,
                                 w1_name,
                                 shard_id='w1',
                                 expert_id=e)
                    _load_expert(n.replace('.input_linear.', '.experts.w13_'),
                                 w3_param,
                                 w3_name,
                                 shard_id='w3',
                                 expert_id=e)
            elif n.endswith('.block_sparse_moe.output_linear.weight'):
                for e in range(p.size(0)):
                    w2_name = n.replace(
                        '.block_sparse_moe.output_linear.weight',
                        f".block_sparse_moe.experts.{e}.w2.weight")
                    w2_param = p[e]
                    _load_expert(n.replace('.output_linear.', '.experts.w2_'),
                                 w2_param,
                                 w2_name,
                                 shard_id='w2',
                                 expert_id=e)
            elif n.endswith('.block_sparse_moe.router.layer.weight'):
                gate_name = n.replace('.block_sparse_moe.router.layer.weight',
                                      ".block_sparse_moe.gate.weight")
                _load(gate_name, p)
            else:
                loaded = False
                for param_name, weight_name, shard_id in stacked_params_mapping:
                    if weight_name in n:
                        _load_shard(n.replace(weight_name, param_name),
                                    p,
                                    shard_id=shard_id)
                        loaded = True
                if not loaded:
                    _load(n, p)

        return loaded_params

config instance-attribute

config = config

embed_tokens instance-attribute

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

embedding_multiplier instance-attribute

embedding_multiplier = embedding_multiplier

make_empty_intermediate_tensors instance-attribute

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

norm instance-attribute

norm = 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/granitemoehybrid.py

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
    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,
    )
    self.embedding_multiplier = config.embedding_multiplier

    def get_layer(prefix: str):
        layer_idx = int(prefix.rsplit(".", 1)[1])
        layer_class = ALL_DECODER_LAYER_TYPES[
            config.layer_types[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(
        config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers")
    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size))

    self.norm = 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/granitemoehybrid.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.mamba_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)
            hidden_states = hidden_states * self.embedding_multiplier
        residual = None
    else:
        if intermediate_tensors is None:
            raise RuntimeError('Intermediate tensors may not be None!')
        hidden_states = intermediate_tensors["hidden_states"]
        residual = intermediate_tensors["residual"]

    num_attn = 0
    for i in range(len(self.layers)):
        layer = self.layers[i]
        if isinstance(layer, GraniteMoeHybridAttentionDecoderLayer):
            num_attn += 1

        layer_mamba_cache_params = None
        if isinstance(
                layer,
                GraniteMoeHybridMambaDecoderLayer) and mamba_cache_params:
            layer_mamba_cache_params = mamba_cache_params.at_layer_idx(
                i - num_attn)

        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(hidden_states)
    return hidden_states

get_input_embeddings

get_input_embeddings(input_ids: Tensor) -> Tensor

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

def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    stacked_params_mapping = [
        # (param_name, shard_name, shard_id)
        (".qkv_proj", ".q_proj", "q"),
        (".qkv_proj", ".k_proj", "k"),
        (".qkv_proj", ".v_proj", "v"),
    ]
    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()

    def _load(n, p):
        param = params_dict[n]
        weight_loader = getattr(param, "weight_loader",
                                default_weight_loader)
        weight_loader(param, p)
        loaded_params.add(n)

    def _load_shard(n, p, shard_id):
        # Skip layers on other devices.
        if not is_pp_missing_parameter(n, self):
            param = params_dict[n]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, p, shard_id)
            loaded_params.add(n)

    def _load_expert(n, p, name, shard_id, expert_id):
        param = params_dict[n]
        weight_loader = getattr(param, "weight_loader",
                                default_weight_loader)
        weight_loader(param,
                      p,
                      name,
                      shard_id=shard_id,
                      expert_id=expert_id)
        loaded_params.add(n)

    for n, p in weights:
        if "A_log" in n:
            n = n.replace("A_log", "A")

        # Logic analogous to: https://github.com/vllm-project/vllm/blob/f49e5aff11c986ed4d45202b1716c5d74786efa9/vllm/model_executor/models/granitemoeshared.py#L215
        # Mapping different experts' layout:
        #  from HF (input_linear, output_linear, router)
        #  to vLLM (experts_w13({e}.w1, {e}.w2), experts_w3({e}.w3), gate)
        if n.endswith('.block_sparse_moe.input_linear.weight'):
            for e in range(p.size(0)):
                w1_name = n.replace(
                    '.block_sparse_moe.input_linear.weight',
                    f".block_sparse_moe.experts.{e}.w1.weight")
                w3_name = n.replace(
                    '.block_sparse_moe.input_linear.weight',
                    f".block_sparse_moe.experts.{e}.w3.weight")
                w1_param, w3_param = p[e].chunk(2, dim=0)
                _load_expert(n.replace('.input_linear.', '.experts.w13_'),
                             w1_param,
                             w1_name,
                             shard_id='w1',
                             expert_id=e)
                _load_expert(n.replace('.input_linear.', '.experts.w13_'),
                             w3_param,
                             w3_name,
                             shard_id='w3',
                             expert_id=e)
        elif n.endswith('.block_sparse_moe.output_linear.weight'):
            for e in range(p.size(0)):
                w2_name = n.replace(
                    '.block_sparse_moe.output_linear.weight',
                    f".block_sparse_moe.experts.{e}.w2.weight")
                w2_param = p[e]
                _load_expert(n.replace('.output_linear.', '.experts.w2_'),
                             w2_param,
                             w2_name,
                             shard_id='w2',
                             expert_id=e)
        elif n.endswith('.block_sparse_moe.router.layer.weight'):
            gate_name = n.replace('.block_sparse_moe.router.layer.weight',
                                  ".block_sparse_moe.gate.weight")
            _load(gate_name, p)
        else:
            loaded = False
            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name in n:
                    _load_shard(n.replace(weight_name, param_name),
                                p,
                                shard_id=shard_id)
                    loaded = True
            if not loaded:
                _load(n, p)

    return loaded_params