vllm.model_executor.models.commandr

PyTorch Cohere model.

CohereAttention

Bases: Module

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

class CohereAttention(nn.Module):

    def __init__(
        self,
        config: CohereConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        tp_size = get_tensor_model_parallel_world_size()
        self.config = config
        self.attention_dropout = config.attention_dropout
        self.hidden_size = config.hidden_size
        self.total_num_heads = config.num_attention_heads
        self.num_heads = self.total_num_heads // tp_size
        self.head_dim = self.hidden_size // self.total_num_heads
        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.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.max_position_embeddings = getattr(
            config, "model_max_length", None) or getattr(
                config, "max_position_embeddings", 8192)
        self.rope_theta = config.rope_theta
        self.rope_scaling = getattr(config, "rope_scaling", None)
        self.use_qk_norm = getattr(config, "use_qk_norm", False)
        self.qkv_proj = QKVParallelLinear(
            self.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,
            self.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=self.max_position_embeddings,
            base=self.rope_theta,
            rope_scaling=self.rope_scaling,
            is_neox_style=False,
        )

        # Model v2 has interleaved sliding windows, v1 does not
        interleaved_sliding_window = getattr(config,
                                             "interleaved_sliding_window",
                                             None)
        self.v1 = interleaved_sliding_window is None

        layer_idx = extract_layer_index(prefix)
        layer_has_sliding_window = (
            getattr(config, "sliding_window_pattern", False)
            and (layer_idx + 1) % self.config.sliding_window_pattern != 0)

        self.sliding_window = (interleaved_sliding_window
                               if layer_has_sliding_window else None)

        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
                              num_kv_heads=self.num_kv_heads,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              per_layer_sliding_window=self.sliding_window,
                              prefix=f"{prefix}.attn")
        if self.use_qk_norm:
            self.q_norm = LayerNorm(param_shape=(self.num_heads,
                                                 self.head_dim),
                                    eps=config.layer_norm_eps)
            self.k_norm = LayerNorm(param_shape=(self.num_kv_heads,
                                                 self.head_dim),
                                    eps=config.layer_norm_eps)

    def _apply_qk_norm(self, q, k):
        q = q.view(*q.shape[:-1], -1, self.head_dim)
        k = k.view(*k.shape[:-1], -1, self.head_dim)
        q, _ = self.q_norm(q)
        k, _ = self.k_norm(k)
        q = q.view(*q.shape[:-2], -1)
        k = k.view(*k.shape[:-2], -1)
        return q, k

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        if self.use_qk_norm:
            q, k = self._apply_qk_norm(q, k)
        if self.v1 or self.sliding_window:
            q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

attention_dropout instance-attribute

attention_dropout = attention_dropout

attn instance-attribute

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

config instance-attribute

config = config

head_dim instance-attribute

head_dim = hidden_size // total_num_heads

hidden_size instance-attribute

hidden_size = hidden_size

k_norm instance-attribute

k_norm = LayerNorm(
    param_shape=(num_kv_heads, head_dim), eps=layer_norm_eps
)

kv_size instance-attribute

kv_size = num_kv_heads * head_dim

max_position_embeddings instance-attribute

max_position_embeddings = getattr(
    config, "model_max_length", None
) or getattr(config, "max_position_embeddings", 8192)

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_norm instance-attribute

q_norm = LayerNorm(
    param_shape=(num_heads, head_dim), eps=layer_norm_eps
)

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",
)

rope_scaling instance-attribute

rope_scaling = getattr(config, 'rope_scaling', None)

rope_theta instance-attribute

rope_theta = rope_theta

rotary_emb instance-attribute

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

scaling instance-attribute

scaling = head_dim ** -0.5

sliding_window instance-attribute

sliding_window = (
    interleaved_sliding_window
    if layer_has_sliding_window
    else None
)

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

use_qk_norm instance-attribute

use_qk_norm = getattr(config, 'use_qk_norm', False)

v1 instance-attribute

v1 = interleaved_sliding_window is None

__init__

__init__(
    config: CohereConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)

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

def __init__(
    self,
    config: CohereConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    tp_size = get_tensor_model_parallel_world_size()
    self.config = config
    self.attention_dropout = config.attention_dropout
    self.hidden_size = config.hidden_size
    self.total_num_heads = config.num_attention_heads
    self.num_heads = self.total_num_heads // tp_size
    self.head_dim = self.hidden_size // self.total_num_heads
    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.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.max_position_embeddings = getattr(
        config, "model_max_length", None) or getattr(
            config, "max_position_embeddings", 8192)
    self.rope_theta = config.rope_theta
    self.rope_scaling = getattr(config, "rope_scaling", None)
    self.use_qk_norm = getattr(config, "use_qk_norm", False)
    self.qkv_proj = QKVParallelLinear(
        self.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,
        self.hidden_size,
        bias=False,
        quant_config=quant_config,
        prefix=f"{prefix}.o_proj",
    )
    self.rotary_emb = get_rope(
        self.head_dim,
        rotary_dim=self.head_dim,
        max_position=self.max_position_embeddings,
        base=self.rope_theta,
        rope_scaling=self.rope_scaling,
        is_neox_style=False,
    )

    # Model v2 has interleaved sliding windows, v1 does not
    interleaved_sliding_window = getattr(config,
                                         "interleaved_sliding_window",
                                         None)
    self.v1 = interleaved_sliding_window is None

    layer_idx = extract_layer_index(prefix)
    layer_has_sliding_window = (
        getattr(config, "sliding_window_pattern", False)
        and (layer_idx + 1) % self.config.sliding_window_pattern != 0)

    self.sliding_window = (interleaved_sliding_window
                           if layer_has_sliding_window else None)

    self.attn = Attention(self.num_heads,
                          self.head_dim,
                          self.scaling,
                          num_kv_heads=self.num_kv_heads,
                          cache_config=cache_config,
                          quant_config=quant_config,
                          per_layer_sliding_window=self.sliding_window,
                          prefix=f"{prefix}.attn")
    if self.use_qk_norm:
        self.q_norm = LayerNorm(param_shape=(self.num_heads,
                                             self.head_dim),
                                eps=config.layer_norm_eps)
        self.k_norm = LayerNorm(param_shape=(self.num_kv_heads,
                                             self.head_dim),
                                eps=config.layer_norm_eps)

_apply_qk_norm

_apply_qk_norm(q, k)

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

def _apply_qk_norm(self, q, k):
    q = q.view(*q.shape[:-1], -1, self.head_dim)
    k = k.view(*k.shape[:-1], -1, self.head_dim)
    q, _ = self.q_norm(q)
    k, _ = self.k_norm(k)
    q = q.view(*q.shape[:-2], -1)
    k = k.view(*k.shape[:-2], -1)
    return q, k

forward

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

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

def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    qkv, _ = self.qkv_proj(hidden_states)
    q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
    if self.use_qk_norm:
        q, k = self._apply_qk_norm(q, k)
    if self.v1 or self.sliding_window:
        q, k = self.rotary_emb(positions, q, k)
    attn_output = self.attn(q, k, v)
    output, _ = self.o_proj(attn_output)
    return output

CohereDecoderLayer

Bases: Module

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

class CohereDecoderLayer(nn.Module):

    def __init__(self,
                 config: CohereConfig,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.hidden_size = config.hidden_size

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

        self.mlp = CohereMLP(config,
                             quant_config=quant_config,
                             prefix=f"{prefix}.mlp")
        self.input_layernorm = LayerNorm(param_shape=(config.hidden_size),
                                         eps=config.layer_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Self Attention
        residual = hidden_states
        hidden_states, residual = self.input_layernorm(hidden_states, residual)
        hidden_states_attention = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )
        hidden_states_mlp = self.mlp(hidden_states)
        # Add everything together
        hidden_states = residual + hidden_states_attention + hidden_states_mlp

        return hidden_states, residual

hidden_size instance-attribute

hidden_size = hidden_size

input_layernorm instance-attribute

input_layernorm = LayerNorm(
    param_shape=hidden_size, eps=layer_norm_eps
)

mlp instance-attribute

mlp = CohereMLP(
    config,
    quant_config=quant_config,
    prefix=f"{prefix}.mlp",
)

self_attn instance-attribute

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

__init__

__init__(
    config: CohereConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)

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

def __init__(self,
             config: CohereConfig,
             cache_config: Optional[CacheConfig] = None,
             quant_config: Optional[QuantizationConfig] = None,
             prefix: str = ""):
    super().__init__()
    self.hidden_size = config.hidden_size

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

    self.mlp = CohereMLP(config,
                         quant_config=quant_config,
                         prefix=f"{prefix}.mlp")
    self.input_layernorm = LayerNorm(param_shape=(config.hidden_size),
                                     eps=config.layer_norm_eps)

forward

forward(
    positions: Tensor,
    hidden_states: Tensor,
    residual: Optional[Tensor],
) -> tuple[Tensor, Tensor]

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

def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
) -> tuple[torch.Tensor, torch.Tensor]:
    # Self Attention
    residual = hidden_states
    hidden_states, residual = self.input_layernorm(hidden_states, residual)
    hidden_states_attention = self.self_attn(
        positions=positions,
        hidden_states=hidden_states,
    )
    hidden_states_mlp = self.mlp(hidden_states)
    # Add everything together
    hidden_states = residual + hidden_states_attention + hidden_states_mlp

    return hidden_states, residual

CohereForCausalLM

Bases: Module, SupportsLoRA, SupportsPP, SupportsQuant

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

class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }
    # LoRA specific attributes
    embedding_modules = {"embed_tokens": "input_embeddings"}

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
        self.config = config
        # currently all existing command R models have `tie_word_embeddings`
        # enabled
        assert config.tie_word_embeddings
        self.unpadded_vocab_size = config.vocab_size
        if lora_config:
            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
        self.quant_config = quant_config
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size,
                                                scale=config.logit_scale)
        self.model = CohereModel(vllm_config=vllm_config,
                                 prefix=maybe_prefix(prefix, "model"))
        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)

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds)
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        is_not_lora = hasattr(self.model.embed_tokens, 'weight')
        if is_not_lora:
            logits = self.logits_processor(self.model.embed_tokens,
                                           hidden_states, sampling_metadata)
        else:
            logits = self.logits_processor(self.model.embed_tokens.base_layer,
                                           hidden_states, sampling_metadata)

        return logits

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

config instance-attribute

config = config

embedding_modules class-attribute instance-attribute

embedding_modules = {'embed_tokens': 'input_embeddings'}

logits_processor instance-attribute

logits_processor = LogitsProcessor(
    unpadded_vocab_size, vocab_size, scale=logit_scale
)

make_empty_intermediate_tensors instance-attribute

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

model instance-attribute

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

packed_modules_mapping class-attribute instance-attribute

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

quant_config instance-attribute

quant_config = quant_config

unpadded_vocab_size instance-attribute

unpadded_vocab_size = vocab_size

__init__

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

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

def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    quant_config = vllm_config.quant_config
    lora_config = vllm_config.lora_config
    self.config = config
    # currently all existing command R models have `tie_word_embeddings`
    # enabled
    assert config.tie_word_embeddings
    self.unpadded_vocab_size = config.vocab_size
    if lora_config:
        self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
    self.quant_config = quant_config
    self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                            config.vocab_size,
                                            scale=config.logit_scale)
    self.model = CohereModel(vllm_config=vllm_config,
                             prefix=maybe_prefix(prefix, "model"))
    self.make_empty_intermediate_tensors = (
        self.model.make_empty_intermediate_tensors)

compute_logits

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

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

def compute_logits(
    self,
    hidden_states: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[torch.Tensor]:
    is_not_lora = hasattr(self.model.embed_tokens, 'weight')
    if is_not_lora:
        logits = self.logits_processor(self.model.embed_tokens,
                                       hidden_states, sampling_metadata)
    else:
        logits = self.logits_processor(self.model.embed_tokens.base_layer,
                                       hidden_states, sampling_metadata)

    return logits

forward

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

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

@torch.no_grad()
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
    hidden_states = self.model(input_ids, positions, intermediate_tensors,
                               inputs_embeds)
    return hidden_states

get_input_embeddings

get_input_embeddings(input_ids: Tensor) -> Tensor

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

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

load_weights

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

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

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

CohereMLP

Bases: Module

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

class CohereMLP(nn.Module):

    def __init__(
        self,
        config: CohereConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_up_proj = MergedColumnParallelLinear(
            self.hidden_size,
            [self.intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            self.intermediate_size,
            self.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.down_proj",
        )
        self.act_fn = SiluAndMul()

    def forward(self, x):
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x

act_fn instance-attribute

act_fn = SiluAndMul()

config instance-attribute

config = config

down_proj instance-attribute

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

gate_up_proj instance-attribute

gate_up_proj = MergedColumnParallelLinear(
    hidden_size,
    [intermediate_size] * 2,
    bias=False,
    quant_config=quant_config,
    prefix=f"{prefix}.gate_up_proj",
)

hidden_size instance-attribute

hidden_size = hidden_size

intermediate_size instance-attribute

intermediate_size = intermediate_size

__init__

__init__(
    config: CohereConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)

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

def __init__(
    self,
    config: CohereConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.config = config
    self.hidden_size = config.hidden_size
    self.intermediate_size = config.intermediate_size
    self.gate_up_proj = MergedColumnParallelLinear(
        self.hidden_size,
        [self.intermediate_size] * 2,
        bias=False,
        quant_config=quant_config,
        prefix=f"{prefix}.gate_up_proj",
    )
    self.down_proj = RowParallelLinear(
        self.intermediate_size,
        self.hidden_size,
        bias=False,
        quant_config=quant_config,
        prefix=f"{prefix}.down_proj",
    )
    self.act_fn = SiluAndMul()

forward

forward(x)

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

def forward(self, x):
    gate_up, _ = self.gate_up_proj(x)
    x = self.act_fn(gate_up)
    x, _ = self.down_proj(x)
    return x

CohereModel

Bases: Module

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

@support_torch_compile
class CohereModel(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.quant_config = quant_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(config.vocab_size,
                                                   config.hidden_size)
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: CohereDecoderLayer(
                config, cache_config, quant_config, prefix=prefix),
            prefix=f"{prefix}.layers")
        self.norm = LayerNorm(param_shape=(config.hidden_size),
                              eps=config.layer_norm_eps)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

    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,
        intermediate_tensors: Optional[IntermediateTensors],
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
        for layer in self.layers[self.start_layer:self.end_layer]:
            hidden_states, residual = layer(
                positions,
                hidden_states,
                residual,
            )
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
        hidden_states, _ = self.norm(hidden_states, residual)
        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"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if (self.quant_config is not None and
                (scale_name := self.quant_config.get_cache_scale(name))):
                # Loading kv cache quantization scales
                param = params_dict[scale_name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                                 loaded_weight[0])
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue

            for param_name, shard_name, shard_id in stacked_params_mapping:
                if shard_name not in name:
                    continue
                name = name.replace(shard_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue

                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

config instance-attribute

config = config

embed_tokens instance-attribute

embed_tokens = VocabParallelEmbedding(
    vocab_size, hidden_size
)

make_empty_intermediate_tensors instance-attribute

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

norm instance-attribute

norm = LayerNorm(
    param_shape=hidden_size, eps=layer_norm_eps
)

org_vocab_size instance-attribute

org_vocab_size = vocab_size

quant_config instance-attribute

quant_config = quant_config

vocab_size instance-attribute

vocab_size = vocab_size + lora_vocab

__init__

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

Source code in vllm/model_executor/models/commandr.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.quant_config = quant_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(config.vocab_size,
                                               config.hidden_size)
    self.start_layer, self.end_layer, self.layers = make_layers(
        config.num_hidden_layers,
        lambda prefix: CohereDecoderLayer(
            config, cache_config, quant_config, prefix=prefix),
        prefix=f"{prefix}.layers")
    self.norm = LayerNorm(param_shape=(config.hidden_size),
                          eps=config.layer_norm_eps)
    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size))

forward

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

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

def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors],
    inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
    if get_pp_group().is_first_rank:
        if inputs_embeds is not None:
            hidden_states = inputs_embeds
        else:
            hidden_states = self.get_input_embeddings(input_ids)
        residual = None
    else:
        assert intermediate_tensors is not None
        hidden_states = intermediate_tensors["hidden_states"]
        residual = intermediate_tensors["residual"]
    for layer in self.layers[self.start_layer:self.end_layer]:
        hidden_states, residual = layer(
            positions,
            hidden_states,
            residual,
        )
    if not get_pp_group().is_last_rank:
        return IntermediateTensors({
            "hidden_states": hidden_states,
            "residual": residual
        })
    hidden_states, _ = self.norm(hidden_states, residual)
    return hidden_states

get_input_embeddings

get_input_embeddings(input_ids: Tensor) -> Tensor

Source code in vllm/model_executor/models/commandr.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/commandr.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"),
        ("gate_up_proj", "gate_proj", 0),
        ("gate_up_proj", "up_proj", 1),
    ]
    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if (self.quant_config is not None and
            (scale_name := self.quant_config.get_cache_scale(name))):
            # Loading kv cache quantization scales
            param = params_dict[scale_name]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                             loaded_weight[0])
            weight_loader(param, loaded_weight)
            loaded_params.add(scale_name)
            continue

        for param_name, shard_name, shard_id in stacked_params_mapping:
            if shard_name not in name:
                continue
            name = name.replace(shard_name, param_name)
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue
            if is_pp_missing_parameter(name, self):
                continue
            param = params_dict[name]
            weight_loader = param.weight_loader
            weight_loader(param, loaded_weight, shard_id)
            break
        else:
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue
            # Remapping the name of FP8 kv-scale.
            name = maybe_remap_kv_scale_name(name, params_dict)
            if name is None:
                continue

            if is_pp_missing_parameter(name, self):
                continue
            param = params_dict[name]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, loaded_weight)
        loaded_params.add(name)
    return loaded_params

LayerNorm

Bases: Module

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

class LayerNorm(nn.Module):

    def __init__(self, param_shape=None, eps=1e-5):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(param_shape))
        self.variance_epsilon = eps
        set_weight_attrs(self.weight,
                         {"weight_loader": row_parallel_weight_loader})

    def forward(self, hidden_states, residuals=None):
        hidden_states = layer_norm_func(hidden_states, self.weight,
                                        self.variance_epsilon)
        return hidden_states, residuals

variance_epsilon instance-attribute

variance_epsilon = eps

weight instance-attribute

weight = Parameter(ones(param_shape))

__init__

__init__(param_shape=None, eps=1e-05)

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

def __init__(self, param_shape=None, eps=1e-5):
    super().__init__()
    self.weight = nn.Parameter(torch.ones(param_shape))
    self.variance_epsilon = eps
    set_weight_attrs(self.weight,
                     {"weight_loader": row_parallel_weight_loader})

forward

forward(hidden_states, residuals=None)

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

def forward(self, hidden_states, residuals=None):
    hidden_states = layer_norm_func(hidden_states, self.weight,
                                    self.variance_epsilon)
    return hidden_states, residuals

layer_norm_func

layer_norm_func(hidden_states, weight, variance_epsilon)

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

@torch.compile(backend=current_platform.simple_compile_backend)
def layer_norm_func(hidden_states, weight, variance_epsilon):
    input_dtype = hidden_states.dtype
    hidden_states = hidden_states.to(torch.float32)
    mean = hidden_states.mean(-1, keepdim=True)
    variance = (hidden_states - mean).pow(2).mean(-1, keepdim=True)
    hidden_states = (hidden_states - mean) * torch.rsqrt(variance +
                                                         variance_epsilon)
    hidden_states = weight.to(torch.float32) * hidden_states
    return hidden_states.to(input_dtype)