vllm.model_executor.models.cohere2_moe

Classes:

• Cohere2Moe –

  Tensor-parallel MoE block for Cohere2Moe with shared experts.

• Cohere2MoeAttention –

  Cohere MoE attention with sliding-window interleave.

• Cohere2MoeDecoderLayer –
• Cohere2MoeMLP –

  Cohere MLP used as shared experts in the MoE block.

• Cohere2MoeModel –

  Transformer decoder for Cohere2Moe.

Functions:

• is_prefix_dense_layer –

  True when layer_idx lies in the contiguous dense MLP prefix.

• make_norm –

  RMSNorm when config.rms_norm_eps is set, otherwise Cohere LayerNorm with

• token_choice_with_bias –

  Sigmoid -> top-k (-> renormalize) custom routing for Cohere2Moe.

Cohere2Moe

Bases: Module

Tensor-parallel MoE block for Cohere2Moe with shared experts.

Source code in vllm/model_executor/models/cohere2_moe.py

class Cohere2Moe(nn.Module):
    """Tensor-parallel MoE block for Cohere2Moe with shared experts."""

    def __init__(
        self,
        config: CohereConfig,
        params_dtype: torch.dtype | None = None,
        quant_config: QuantizationConfig | None = None,
        tp_size: int | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.tp_size = get_tensor_model_parallel_world_size()

        if self.tp_size > config.num_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {config.num_experts}."
            )

        if (
            hasattr(config, "expert_selection_fn")
            and config.expert_selection_fn == "sigmoid"
        ):
            self.custom_routing_function = token_choice_with_bias
        else:
            self.custom_routing_function = None

        self.gate = GateLinear(
            config.hidden_size,
            config.num_experts,
            params_dtype=params_dtype,
            prefix=f"{prefix}.gate",
        )

        self.shared_experts: Cohere2MoeMLP | None
        if hasattr(config, "num_shared_experts") and config.num_shared_experts > 0:
            self.shared_experts = Cohere2MoeMLP(
                config=config,
                intermediate_size=config.intermediate_size * config.num_shared_experts,
                quant_config=quant_config,
                prefix=f"{prefix}.shared_experts",
            )
            self.shared_expert_combination_strategy = getattr(
                config, "shared_expert_combination_strategy", "sum"
            )
            assert self.shared_expert_combination_strategy in ("average", "sum"), (
                "shared_expert_combination_strategy must be one of ['average', 'sum']"
            )
        else:
            self.shared_experts = None
            self.shared_expert_combination_strategy = None

        # Routed-only MoE leaves its TP all-reduce to the decoder layer, which
        # reduces the attention and MoE outputs together.
        defer_all_reduce = self.tp_size > 1 and self.shared_experts is None
        self.experts = FusedMoEFactory(
            num_experts=config.num_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            params_dtype=params_dtype,
            renormalize=getattr(config, "norm_topk_prob", True),
            quant_config=quant_config,
            tp_size=tp_size,
            prefix=f"{prefix}.experts",
            custom_routing_function=self.custom_routing_function,
            shared_experts=self.shared_experts,
            reduce_results=not defer_all_reduce,
        )
        # The deferral is dropped when the MoE runner must reduce itself
        # (e.g. EP all2all or sequence parallel).
        self.output_is_reduced = not (
            defer_all_reduce and self.experts.moe_config.skip_final_all_reduce
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        orig_shape = hidden_states.shape
        hidden_states = hidden_states.view(-1, self.hidden_size)
        router_logits, _ = self.gate(hidden_states)
        # FusedMoEFactory handles shared expert overlap internally and returns
        # shared_output + routed_output when shared_experts is set.
        final_hidden_states = self.experts(hidden_states, router_logits)
        if self.shared_expert_combination_strategy == "average":
            final_hidden_states = final_hidden_states / 2
        return final_hidden_states.view(orig_shape)

Cohere2MoeAttention

Bases: Module

Cohere MoE attention with sliding-window interleave.

Source code in vllm/model_executor/models/cohere2_moe.py

class Cohere2MoeAttention(nn.Module):
    """Cohere MoE attention with sliding-window interleave."""

    def __init__(
        self,
        config: CohereConfig,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        tp_size = get_tensor_model_parallel_world_size()
        self.config = config
        self.layer_idx = extract_layer_index(prefix)
        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 = getattr(
            config, "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:
            assert self.total_num_kv_heads % tp_size == 0
        else:
            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)
        assert isinstance(self.max_position_embeddings, int)
        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,
            reduce_results=False,
            prefix=f"{prefix}.o_proj",
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            max_position=self.max_position_embeddings,
            rope_parameters=config.rope_parameters,
            is_neox_style=False,
        )

        self.sliding_window = None
        layer_types = getattr(config, "layer_types", None)
        if (
            layer_types is not None
            and layer_types[self.layer_idx] == "sliding_attention"
        ):
            self.sliding_window = config.sliding_window + 1

        # Prefix-dense layers have full attention (no sliding window). When
        # prefix_dense_sliding_window_pattern == 1, they keep RoPE even though
        # they are not sliding-window layers.
        prefix_dense_sliding_window_pattern = getattr(
            config, "prefix_dense_sliding_window_pattern", 1
        )
        self.force_rope = bool(
            is_prefix_dense_layer(config, self.layer_idx)
            and prefix_dense_sliding_window_pattern == 1
        )

        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",
        )

    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.sliding_window or self.force_rope:
            q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

Cohere2MoeDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/cohere2_moe.py

class Cohere2MoeDecoderLayer(nn.Module):
    def __init__(
        self,
        config: CohereConfig,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
        *,
        moe_stream: torch.cuda.Stream | None = None,
    ):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.layer_idx = extract_layer_index(prefix)

        self.self_attn = Cohere2MoeAttention(
            config,
            cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
        )

        if config.mlp_layer_types[self.layer_idx] == "dense":
            self.mlp = Cohere2MoeMLP(
                config=config,
                intermediate_size=getattr(
                    config, "prefix_dense_intermediate_size", config.intermediate_size
                ),
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )
        else:
            self.mlp = Cohere2Moe(
                config=config, quant_config=quant_config, prefix=f"{prefix}.mlp"
            )
        self.mlp_output_is_reduced = (
            isinstance(self.mlp, Cohere2Moe) and self.mlp.output_is_reduced
        )

        self.input_layernorm = make_norm(config)
        self.uses_rms_norm = isinstance(self.input_layernorm, RMSNorm)

        if moe_stream is not None and not self._allows_dual_stream():
            logger.info_once(
                "Cohere2MoE MoE layers that reduce internally under TP (shared "
                "experts, EP all2all or sequence parallel) run single-stream."
            )
            moe_stream = None
        self.dual_stream = (
            _DualStreamState(
                stream=moe_stream,
                token_threshold=envs.VLLM_COHERE2_ATTN_MOE_DUAL_STREAM_TOKEN_THRESHOLD,
            )
            if moe_stream is not None
            else None
        )
        self.layer_name = prefix
        if self.dual_stream is not None:
            static_forward_context = (
                get_current_vllm_config().compilation_config.static_forward_context
            )
            if prefix in static_forward_context:
                raise ValueError(f"Duplicate layer name: {prefix}")
            static_forward_context[prefix] = self

    def _allows_dual_stream(self) -> bool:
        """With the dual stream the only TP collective in a layer is the
        all-reduce of attention + MLP after the join. A MoE that reduces
        internally would let the attention all-reduce run while the MoE stream
        is still communicating.

        EP all2all kernels stay single-stream at any TP size (including DP + EP
        at TP1): all2all on the side stream, and next to DBO, is not
        validated."""
        if (
            isinstance(self.mlp, Cohere2Moe)
            and self.mlp.experts.moe_config.moe_parallel_config.use_all2all_kernels
        ):
            return False
        return (
            get_tensor_model_parallel_world_size() == 1
            or not self.mlp_output_is_reduced
        )

    def _fork_mlp(self, hidden_states: torch.Tensor) -> torch.Tensor:
        state = self.dual_stream
        assert state is not None
        assert not state.join_pending, "Cohere2MoE dual-stream fork was not joined"
        # A fork must not span a piecewise CUDA graph boundary: the segment
        # captured before attention would end with an unjoined stream. Prefill
        # and mixed batches use piecewise graphs, so they run the MLP inline.
        if (
            hidden_states.shape[0] > state.token_threshold
            or get_forward_context().cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
            or BreakableCUDAGraphCapture.is_active()
        ):
            return self.mlp(hidden_states)

        logger.info_once("Cohere2MoE attention/MoE dual stream is active.")
        # Inductor may reuse the storage of hidden_states once this op returns
        # (bypassing the caching allocator), so the MoE stream reads a copy.
        mlp_input = hidden_states.clone()
        state.fork_event.record(torch.cuda.current_stream())
        with torch.cuda.stream(state.stream):
            state.fork_event.wait()
            mlp_output = self.mlp(mlp_input)
            state.join_event.record()
        mlp_input.record_stream(state.stream)
        # mlp_output needs no record_stream: its block can only be reused by a
        # side-stream allocation, and side-stream work always runs after a later
        # fork_event, recorded on the main stream after every use of mlp_output.
        # Keep it that way: no side-stream work outside a fork.
        state.join_pending = True
        return mlp_output

    def _join_mlp(
        self, attn_output: torch.Tensor, mlp_output: torch.Tensor
    ) -> torch.Tensor:
        state = self.dual_stream
        assert state is not None
        if state.join_pending:
            torch.cuda.current_stream().wait_event(state.join_event)
            state.join_pending = False
        return attn_output + mlp_output

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if not self.uses_rms_norm:
            residual = hidden_states
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
        elif residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(hidden_states, residual)

        if self.dual_stream is not None:
            layer_name = _encode_layer_name(self.layer_name)
            hidden_states_mlp = torch.ops.vllm.cohere2_mlp_fork(
                hidden_states, layer_name
            )
            hidden_states_attention = self.self_attn(
                positions=positions, hidden_states=hidden_states
            )
            # Only layers that reduce attention + MLP together keep a
            # dual_stream (see __init__), so the join can return the sum.
            parallel_block_output = tensor_model_parallel_all_reduce(
                torch.ops.vllm.cohere2_mlp_join(
                    hidden_states_attention, hidden_states_mlp, layer_name
                )
            )
        else:
            hidden_states_attention = self.self_attn(
                positions=positions,
                hidden_states=hidden_states,
            )
            hidden_states_mlp = self.mlp(hidden_states)
            if self.mlp_output_is_reduced:
                parallel_block_output = (
                    tensor_model_parallel_all_reduce(hidden_states_attention)
                    + hidden_states_mlp
                )
            else:
                parallel_block_output = tensor_model_parallel_all_reduce(
                    hidden_states_attention + hidden_states_mlp
                )

        if self.uses_rms_norm:
            # The residual add is fused into the next RMSNorm (or the final norm).
            return parallel_block_output, residual
        return residual + parallel_block_output, residual

_allows_dual_stream()

With the dual stream the only TP collective in a layer is the all-reduce of attention + MLP after the join. A MoE that reduces internally would let the attention all-reduce run while the MoE stream is still communicating.

EP all2all kernels stay single-stream at any TP size (including DP + EP at TP1): all2all on the side stream, and next to DBO, is not validated.

Source code in vllm/model_executor/models/cohere2_moe.py

def _allows_dual_stream(self) -> bool:
    """With the dual stream the only TP collective in a layer is the
    all-reduce of attention + MLP after the join. A MoE that reduces
    internally would let the attention all-reduce run while the MoE stream
    is still communicating.

    EP all2all kernels stay single-stream at any TP size (including DP + EP
    at TP1): all2all on the side stream, and next to DBO, is not
    validated."""
    if (
        isinstance(self.mlp, Cohere2Moe)
        and self.mlp.experts.moe_config.moe_parallel_config.use_all2all_kernels
    ):
        return False
    return (
        get_tensor_model_parallel_world_size() == 1
        or not self.mlp_output_is_reduced
    )

Cohere2MoeMLP

Bases: Module

Cohere MLP used as shared experts in the MoE block.

Source code in vllm/model_executor/models/cohere2_moe.py

class Cohere2MoeMLP(nn.Module):
    """Cohere MLP used as shared experts in the MoE block."""

    def __init__(
        self,
        config: CohereConfig,
        intermediate_size: int | None = None,
        quant_config: QuantizationConfig | None = None,
        reduce_results: bool = False,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = (
            intermediate_size
            if intermediate_size is not None
            else 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,
            reduce_results=reduce_results,
            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

Cohere2MoeModel

Bases: Module, EagleModelMixin

Transformer decoder for Cohere2Moe.

Source code in vllm/model_executor/models/cohere2_moe.py

@support_torch_compile
class Cohere2MoeModel(nn.Module, EagleModelMixin):
    """Transformer decoder for Cohere2Moe."""

    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

        self.config = config
        self.quant_config = quant_config
        self.vocab_size = config.vocab_size
        self.org_vocab_size = config.vocab_size
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size, config.hidden_size
        )

        self._moe_stream = _make_dual_stream(vllm_config)

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: Cohere2MoeDecoderLayer(
                config,
                cache_config,
                quant_config,
                prefix=prefix,
                moe_stream=self._moe_stream,
            ),
            prefix=f"{prefix}.layers",
        )
        self.norm = make_norm(config)
        self.uses_rms_norm = isinstance(self.norm, RMSNorm)
        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: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]:
        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"]

        aux_hidden_states: list[torch.Tensor] = []
        if self.start_layer in self.aux_hidden_state_layers:
            start_hidden_states = hidden_states
            if self.uses_rms_norm and residual is None:
                # The first layer carries hidden_states as the residual, which
                # later fused RMSNorm calls update in place.
                start_hidden_states = hidden_states.clone()
            aux_hidden_states = self._maybe_add_hidden_state(
                aux_hidden_states,
                self.start_layer,
                start_hidden_states,
                residual if self.uses_rms_norm else None,
            )
        for layer_idx, layer in enumerate(
            islice(self.layers, self.start_layer, self.end_layer),
            start=self.start_layer,
        ):
            hidden_states, residual = layer(positions, hidden_states, residual)
            aux_hidden_states = self._maybe_add_hidden_state(
                aux_hidden_states,
                layer_idx + 1,
                hidden_states,
                residual if self.uses_rms_norm else None,
            )
        if not get_pp_group().is_last_rank:
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
        hidden_states, _ = self.norm(hidden_states, residual)
        if aux_hidden_states:
            return hidden_states, aux_hidden_states
        return hidden_states

_DualStreamState dataclass

Dual-stream state of one decoder layer.

stream is the model's single side stream, shared by every layer; the events and join_pending are per layer, since a layer has at most one fork in flight. DBO would need one state per ubatch.

Source code in vllm/model_executor/models/cohere2_moe.py

@dataclasses.dataclass
class _DualStreamState:
    """Dual-stream state of one decoder layer.

    ``stream`` is the model's single side stream, shared by every layer; the
    events and ``join_pending`` are per layer, since a layer has at most one
    fork in flight. DBO would need one state per ubatch."""

    stream: torch.cuda.Stream
    token_threshold: int
    fork_event: torch.cuda.Event = dataclasses.field(default_factory=torch.cuda.Event)
    join_event: torch.cuda.Event = dataclasses.field(default_factory=torch.cuda.Event)
    join_pending: bool = False

_make_dual_stream(vllm_config)

Side stream shared by all decoder layers, or None when the dual stream is disabled (token threshold 0) or unsupported.

Source code in vllm/model_executor/models/cohere2_moe.py

def _make_dual_stream(vllm_config: VllmConfig) -> torch.cuda.Stream | None:
    """Side stream shared by all decoder layers, or None when the dual stream
    is disabled (token threshold 0) or unsupported."""
    token_threshold = envs.VLLM_COHERE2_ATTN_MOE_DUAL_STREAM_TOKEN_THRESHOLD
    if token_threshold <= 0:
        return None
    if not current_platform.is_cuda():
        logger.info_once(
            "Cohere2MoE attention/MoE dual stream is CUDA-only; running single-stream."
        )
        return None
    if vllm_config.parallel_config.use_ubatching:
        # Each layer holds a single fork/join state, so microbatches would race.
        logger.info_once(
            "Cohere2MoE attention/MoE dual stream does not support "
            "microbatching (DBO); running single-stream."
        )
        return None
    logger.info_once(
        "Cohere2MoE attention/MoE dual stream enabled (token threshold=%d).",
        token_threshold,
    )
    # Dedicated stream: the global aux_stream() is used by the shared-expert
    # and LoRA overlaps inside the MoE branch.
    return torch.cuda.Stream()

is_prefix_dense_layer(config, layer_idx)

True when layer_idx lies in the contiguous dense MLP prefix.

Source code in vllm/model_executor/models/cohere2_moe.py

def is_prefix_dense_layer(config: CohereConfig, layer_idx: int) -> bool:
    """True when layer_idx lies in the contiguous dense MLP prefix."""
    if layer_idx >= len(config.mlp_layer_types):
        return False
    return all(t == "dense" for t in config.mlp_layer_types[: layer_idx + 1])

make_norm(config)

RMSNorm when config.rms_norm_eps is set, otherwise Cohere LayerNorm with config.layer_norm_eps.

Source code in vllm/model_executor/models/cohere2_moe.py

def make_norm(config: CohereConfig) -> nn.Module:
    """RMSNorm when config.rms_norm_eps is set, otherwise Cohere LayerNorm with
    config.layer_norm_eps."""
    rms_eps = getattr(config, "rms_norm_eps", None)
    if rms_eps is not None:
        return RMSNorm(config.hidden_size, eps=rms_eps)
    return LayerNorm(param_shape=(config.hidden_size,), eps=config.layer_norm_eps)

token_choice_with_bias(hidden_states, gating_output, topk, renormalize)

Sigmoid -> top-k (-> renormalize) custom routing for Cohere2Moe.

Source code in vllm/model_executor/models/cohere2_moe.py

@torch.compile(backend=current_platform.simple_compile_backend)
def token_choice_with_bias(
    hidden_states: torch.Tensor,
    gating_output: torch.Tensor,
    topk: int,
    renormalize: bool,
):
    """Sigmoid -> top-k (-> renormalize) custom routing for Cohere2Moe."""
    assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"

    scores = gating_output.float().sigmoid()
    topk_weights, topk_ids = torch.topk(scores, k=topk, dim=-1, sorted=False)

    if renormalize:
        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)

    return topk_weights.to(torch.float32), topk_ids.to(torch.int32)