vllm.attention

class Attention(nn.Module):
    """Attention layer.

    This class takes query, key, and value tensors as input. The input tensors
    can either contain prompt tokens or generation tokens.
    The class does the following:

    1. Store the input key and value tensors in the KV cache.
    2. Perform (multi-head/multi-query/grouped-query) attention.
    3. Return the output tensor.
    """

    def __init__(
        self,
        num_heads: int,
        head_size: int,
        scale: float,
        num_kv_heads: Optional[int] = None,
        alibi_slopes: Optional[List[float]] = None,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        blocksparse_params: Optional[Dict[str, Any]] = None,
        logits_soft_cap: Optional[float] = None,
        per_layer_sliding_window: Optional[int] = None,
        use_mla: bool = False,
        prefix: str = "",
        attn_type: str = AttentionType.DECODER,
        **extra_impl_args,
    ) -> None:
        """
        The KV cache is stored inside this class and is accessed via
        `self.kv_cache`.
        """
        super().__init__()
        if per_layer_sliding_window is not None:
            # per-layer sliding window
            sliding_window = per_layer_sliding_window
        elif cache_config is not None:
            # model-level sliding window
            sliding_window = cache_config.sliding_window
        else:
            sliding_window = None

        if cache_config is not None:
            kv_cache_dtype = cache_config.cache_dtype
            block_size = cache_config.block_size
            is_attention_free = cache_config.is_attention_free
            calculate_kv_scales = cache_config.calculate_kv_scales
        else:
            kv_cache_dtype = "auto"
            block_size = 16
            is_attention_free = False
            calculate_kv_scales = False
        if num_kv_heads is None:
            num_kv_heads = num_heads

        # The default k/v_scale is set to 1.0. This is ignored
        # when kv-cache is not fp8, and should be used with
        # kv-cache in fp8_e5m2. For kv-cache in fp8_e4m3, we
        # expect the pre-quantized k/v_scale to be loaded along
        # with the model weights.
        self.kv_cache_dtype = kv_cache_dtype
        self.calculate_kv_scales = calculate_kv_scales
        self._k_scale = torch.tensor(1.0, dtype=torch.float32)
        self._v_scale = torch.tensor(1.0, dtype=torch.float32)
        # FlashAttn doesn't support quantizing the kv-cache only
        # but requires q to be quantized as well.
        self._q_scale = torch.tensor(1.0, dtype=torch.float32)
        self._prob_scale = torch.tensor(1.0, dtype=torch.float32)

        # We also keep the float32 versions of k/v_scale for attention
        # backends that don't support tensors (Flashinfer)
        self._k_scale_float = 1.0
        self._v_scale_float = 1.0

        self.use_mla = use_mla
        self.num_heads = num_heads
        self.head_size = head_size
        self.num_kv_heads = num_kv_heads
        self.sliding_window = sliding_window

        quant_method = quant_config.get_quant_method(
            self, prefix=prefix) if quant_config else None
        if quant_method is not None and not isinstance(
                quant_method, UnquantizedLinearMethod):
            assert isinstance(quant_method, BaseKVCacheMethod)
            # TODO (mgoin): kv cache dtype should be specified in the FP8
            # checkpoint config and become the "auto" behavior
            if self.kv_cache_dtype == "fp8_e5m2":
                raise ValueError("fp8_e5m2 kv-cache is not supported with "
                                 "fp8 checkpoints.")
            # If quantization is enabled, we make "k_scale" and "v_scale"
            # parameters so that it can be loaded from the model checkpoint.
            # The k/v_scale will then be converted back to native float32
            # values after weight loading.
            self.quant_method = quant_method
            self.quant_method.create_weights(self)

        # During model initialization, the default dtype is set as the model
        # weight and activation dtype.
        dtype = torch.get_default_dtype()
        attn_backend = get_attn_backend(head_size,
                                        dtype,
                                        kv_cache_dtype,
                                        block_size,
                                        is_attention_free,
                                        blocksparse_params is not None,
                                        use_mla=use_mla)
        impl_cls = attn_backend.get_impl_cls()
        self.impl = impl_cls(num_heads, head_size, scale, num_kv_heads,
                             alibi_slopes, sliding_window, kv_cache_dtype,
                             blocksparse_params, logits_soft_cap, attn_type,
                             **extra_impl_args)
        self.backend = backend_name_to_enum(attn_backend.get_name())
        self.dtype = dtype

        # For cuda-alike (CUDA and ROCM) and cpu platforms, we control how
        # torch.compile works by registering the attention as one giant
        # opaque custom op. For other platforms, we directly call them
        # and let torch.compile handle them.
        self.use_direct_call = not current_platform.is_cuda_alike(
        ) and not current_platform.is_cpu()

        self.use_output = attn_backend.accept_output_buffer
        compilation_config = get_current_vllm_config().compilation_config
        if prefix in compilation_config.static_forward_context:
            raise ValueError(f"Duplicate layer name: {prefix}")
        compilation_config.static_forward_context[prefix] = self
        self.layer_name = prefix
        self.attn_type = attn_type
        # use a placeholder kv cache tensor during init, which will be replaced
        # by bind_kv_cache
        # this variable will not be accessed if use_direct_call is True
        self.kv_cache = [
            torch.tensor([]) for _ in range(get_current_vllm_config(
            ).parallel_config.pipeline_parallel_size)
        ]

        self.q_range = torch.tensor(envs.Q_SCALE_CONSTANT, dtype=torch.float32)
        self.k_range = torch.tensor(envs.K_SCALE_CONSTANT, dtype=torch.float32)
        self.v_range = torch.tensor(envs.V_SCALE_CONSTANT, dtype=torch.float32)

    def forward(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        # For some alternate attention backends like MLA the attention output
        # shape does not match the query shape, so we optionally let the model
        # definition specify the output tensor shape.
        output_shape: Optional[torch.Size] = None,
    ) -> torch.Tensor:
        """
        The KV cache is stored inside this class and is accessed via
        `self.kv_cache`.

        Attention metadata (`attn_metadata`) is set using a context manager in
        the model runner's `execute_model` method. It is accessed via forward
        context using
        `vllm.forward_context.get_forward_context().attn_metadata`.
        """
        if self.calculate_kv_scales:
            attn_metadata = get_forward_context().attn_metadata
            if attn_metadata.enable_kv_scales_calculation:
                self.calc_kv_scales(query, key, value)
        if self.use_output:
            output_shape = (output_shape
                            if output_shape is not None else query.shape)
            output = torch.empty(output_shape,
                                 dtype=query.dtype,
                                 device=query.device)
            hidden_size = output_shape[-1]
            # We skip reshaping query, key and value tensors for the MLA
            # backend since these tensors have different semantics and are
            # processed differently.
            if not self.use_mla:
                # Reshape the query, key, and value tensors.
                # NOTE(woosuk): We do this outside the custom op to minimize the
                # CPU overheads from the non-CUDA-graph regions.
                query = query.view(-1, self.num_heads, self.head_size)
                output = output.view(-1, self.num_heads, self.head_size)
                if key is not None:
                    key = key.view(-1, self.num_kv_heads, self.head_size)
                if value is not None:
                    value = value.view(-1, self.num_kv_heads, self.head_size)
            if self.use_direct_call:
                forward_context: ForwardContext = get_forward_context()
                attn_metadata = forward_context.attn_metadata
                if isinstance(attn_metadata, dict):
                    attn_metadata = attn_metadata[self.layer_name]
                self_kv_cache = self.kv_cache[forward_context.virtual_engine]
                self.impl.forward(self,
                                  query,
                                  key,
                                  value,
                                  self_kv_cache,
                                  attn_metadata,
                                  output=output)
            else:
                torch.ops.vllm.unified_attention_with_output(
                    query, key, value, output, self.layer_name)
            return output.view(-1, hidden_size)
        else:
            if self.use_direct_call:
                forward_context = get_forward_context()
                attn_metadata = forward_context.attn_metadata
                if isinstance(attn_metadata, dict):
                    attn_metadata = attn_metadata[self.layer_name]
                self_kv_cache = self.kv_cache[forward_context.virtual_engine]
                return self.impl.forward(self, query, key, value,
                                         self_kv_cache, attn_metadata)
            else:
                return torch.ops.vllm.unified_attention(
                    query, key, value, self.layer_name)

    def calc_kv_scales(self, query, key, value):
        self._q_scale.copy_(torch.abs(query).max() / self.q_range)
        self._k_scale.copy_(torch.abs(key).max() / self.k_range)
        self._v_scale.copy_(torch.abs(value).max() / self.v_range)
        self._k_scale_float = self._k_scale.item()
        self._v_scale_float = self._v_scale.item()
        # We only calculate the scales once
        self.calculate_kv_scales = False

    def extra_repr(self) -> str:
        s = f"head_size={self.impl.head_size}"  # type: ignore
        s += f", num_heads={self.impl.num_heads}"  # type: ignore
        s += f", num_kv_heads={self.impl.num_kv_heads}"  # type: ignore
        s += f", scale={self.impl.scale}"  # type: ignore
        s += f", backend={self.impl.__class__.__name__}"
        return s

    def process_weights_after_loading(self, act_dtype: torch.dtype):
        if hasattr(self.impl, "process_weights_after_loading"):
            self.impl.process_weights_after_loading(act_dtype)

class AttentionBackend(ABC):
    """Abstract class for attention backends."""
    # For some attention backends, we allocate an output tensor before
    # calling the custom op. When piecewise cudagraph is enabled, this
    # makes sure the output tensor is allocated inside the cudagraph.
    accept_output_buffer: bool = False

    @staticmethod
    @abstractmethod
    def get_name() -> str:
        raise NotImplementedError

    @staticmethod
    @abstractmethod
    def get_impl_cls() -> Type["AttentionImpl"]:
        raise NotImplementedError

    @staticmethod
    @abstractmethod
    def get_metadata_cls() -> Type["AttentionMetadata"]:
        raise NotImplementedError

    @staticmethod
    @abstractmethod
    def get_state_cls() -> Type["AttentionState"]:
        raise NotImplementedError

    @classmethod
    def make_metadata(cls, *args, **kwargs) -> "AttentionMetadata":
        return cls.get_metadata_cls()(*args, **kwargs)

    @staticmethod
    @abstractmethod
    def get_builder_cls() -> Type["AttentionMetadataBuilder"]:
        raise NotImplementedError

    @staticmethod
    @abstractmethod
    def get_kv_cache_shape(
        num_blocks: int,
        block_size: int,
        num_kv_heads: int,
        head_size: int,
    ) -> Tuple[int, ...]:
        raise NotImplementedError

    @staticmethod
    def get_kv_cache_stride_order() -> Tuple[int, ...]:
        raise NotImplementedError

    @staticmethod
    @abstractmethod
    def swap_blocks(
        src_kv_cache: torch.Tensor,
        dst_kv_cache: torch.Tensor,
        src_to_dst: torch.Tensor,
    ) -> None:
        raise NotImplementedError

    @staticmethod
    @abstractmethod
    def copy_blocks(
        kv_caches: List[torch.Tensor],
        src_to_dists: torch.Tensor,
    ) -> None:
        raise NotImplementedError

    def advance_step(self, model_input: "ModelRunnerInputBase",
                     sampled_token_ids: Optional[torch.Tensor],
                     block_size: int, num_seqs: int, num_queries: int) -> None:
        raise NotImplementedError

@dataclass
class AttentionMetadata:
    """Attention metadata for prefill and decode batched together."""
    # Total number of prefill requests.
    num_prefills: int
    # Number of prefill tokens.
    num_prefill_tokens: int
    # Number of decode tokens. Note that it is equivalent to the number of
    # decode requests.
    num_decode_tokens: int
    # (num_tokens,). The indices of the token slots that input tokens will be
    # stored into. E.g., if `slot_mapping` is [35, 2, 17] and the block size
    # is 16, the three tokens are stored in the 3rd slot in block 2, 2nd slot
    # in block 0, and 1st slot in block 1, respectively.
    slot_mapping: torch.Tensor

    # The index maps that relate multi-modal embeddings to the corresponding
    # placeholders.
    #
    # N.B. These aren't really related to attention and don't belong on this
    # type -- this is just a temporary solution to make them available to
    # `model_executable`.
    multi_modal_placeholder_index_maps: Optional[Dict[
        str, MultiModalPlaceholderMap.IndexMap]]

    # Enable/disable KV scales calculation. This is so that we can disable the
    # calculation until after prefill and cuda graph capture.
    enable_kv_scales_calculation: bool

    @property
    @abstractmethod
    def prefill_metadata(self) -> Optional["AttentionMetadata"]:
        """Return the attention metadata that's required to run prefill
        attention."""
        pass

    @property
    @abstractmethod
    def decode_metadata(self) -> Optional["AttentionMetadata"]:
        """Return the attention metadata that's required to run decode
        attention."""
        pass

    def asdict_zerocopy(self,
                        skip_fields: Optional[Set[str]] = None
                        ) -> Dict[str, Any]:
        """Similar to dataclasses.asdict, but avoids deepcopying."""
        if skip_fields is None:
            skip_fields = set()
        # Note that if we add dataclasses as fields, they will need
        # similar handling.
        return {
            field.name: getattr(self, field.name)
            for field in fields(self) if field.name not in skip_fields
        }

class AttentionMetadataBuilder(ABC, Generic[T]):
    """Abstract class for attention metadata builders."""

    @abstractmethod
    def __init__(self, input_builder: "ModelRunnerInputBuilderBase") -> None:
        """Create the builder, remember some configuration and parameters."""
        raise NotImplementedError

    @abstractmethod
    def prepare(self) -> None:
        """Prepare for one batch."""
        raise NotImplementedError

    @abstractmethod
    def build(self, seq_lens: List[int], query_lens: List[int],
              cuda_graph_pad_size: int, batch_size: int) -> T:
        """Build attention metadata with on-device tensors."""
        raise NotImplementedError