vllm.model_executor.layers.fused_moe.routed_experts_capturer ¶

class RoutedExpertsCapturer:
    """
    Capturer for routed experts with device and optional shared memory buffer.

    This class captures expert routing decisions during model forward passes
    and optionally stores them in shared memory for cross-process access.
    """

    _instance: RoutedExpertsCapturer | None = None

    def __init__(self) -> None:
        self._device_buffer: torch.Tensor | None = None
        self._shm: shared_memory.SharedMemory | None = None
        self._host_buffer_view: np.ndarray | None = None
        self._lock_file: str | None = None
        self._shm_name: str | None = None

    @classmethod
    def create(cls) -> RoutedExpertsCapturer:
        """Create a global singleton instance."""
        global _global_experts_capturer
        if _global_experts_capturer is not None:
            raise RuntimeError("Experts capturer already created.")

        _global_experts_capturer = cls()
        return _global_experts_capturer

    @staticmethod
    def get_instance() -> RoutedExpertsCapturer | None:
        """Get the global singleton instance."""
        return _global_experts_capturer

    def init_buffer(
        self,
        max_num_batched_tokens: int,
        max_num_kv_tokens: int,
        model_config: ModelConfig,
        instance_id: str,
    ) -> None:
        """
        Initialize the device buffer and optionally shared memory buffer.

        Args:
            max_num_batched_tokens: Maximum number of tokens in a batch.
            max_num_kv_tokens: Maximum number of KV tokens for shared memory.
            model_config: Model configuration containing layer and expert info.
            instance_id: Unique identifier for the shared memory buffer.
        """

        if self._device_buffer is not None:
            raise RuntimeError("Device buffer has already been initialized")

        hf_config = model_config.hf_text_config
        num_layers = hf_config.num_hidden_layers
        num_experts_per_tok = hf_config.num_experts_per_tok

        # Initialize device buffer
        self._device_buffer = torch.zeros(
            (max_num_batched_tokens, num_layers, num_experts_per_tok),
            dtype=torch.int32,
            device="cuda",
        )

        if get_tensor_model_parallel_rank() != 0:
            return

        # Initialize shared memory
        shape = (max_num_kv_tokens, num_layers, num_experts_per_tok)
        buffer_size = int(np.prod(shape)) * np.dtype(np.int32).itemsize

        self._lock_file = f"{_LOCK_FILE_PREFIX}_{instance_id}.lock"
        self._shm_name = f"{_BUFFER_PREFIX}_{instance_id}"

        self._shm = _create_or_attach_shared_memory(
            self._shm_name, buffer_size, self._lock_file
        )
        self._host_buffer_view = np.ndarray(shape, dtype=np.int32, buffer=self._shm.buf)
        self._host_buffer_view.fill(0)

        logger.debug(
            "Created shared memory buffer '%s' with shape %s",
            self._shm.name,
            shape,
        )

    def capture(self, layer_id: int, topk_ids: torch.Tensor) -> None:
        """
        Capture expert routing decisions for a specific layer.

        Args:
            layer_id: The layer index.
            topk_ids: Tensor of shape (batch_size, num_routed_experts).
        """
        if self._device_buffer is None:
            raise RuntimeError("Buffer not initialized. Call init_buffer() first.")

        if layer_id >= self._device_buffer.shape[1]:
            return

        batch_size = topk_ids.shape[0]
        self._device_buffer[:batch_size, layer_id, :] = topk_ids

    def clear_buffer(self) -> None:
        """Clear the device buffer."""
        if self._device_buffer is not None:
            self._device_buffer.zero_()

    def save_captured_experts(self, indices: np.ndarray) -> None:
        """
        Save captured experts from device buffer to shared memory.

        Args:
            indices: Array of indices indicating where to store the data.
        """
        if get_tensor_model_parallel_rank() != 0:
            return
        if self._lock_file is None:
            raise RuntimeError("Shared memory not initialized.")
        if self._host_buffer_view is None:
            return
        if self._device_buffer is None:
            raise RuntimeError("Device buffer not initialized.")

        num_tokens = len(indices)
        data = self._device_buffer[:num_tokens, :, :].cpu().numpy()

        with _file_lock(self._lock_file):
            self._host_buffer_view[indices, :, :] = data

    def cleanup(self) -> None:
        """Explicitly clean up shared memory resources."""
        if self._shm is not None:
            try:
                self._shm.close()
                self._shm.unlink()
            except Exception:
                logger.debug("Exception during cleanup for capturer", exc_info=True)
            finally:
                self._shm = None

    def __del__(self) -> None:
        """Clean up shared memory on destruction."""
        self.cleanup()

class RoutedExpertsReader:
    """
    Reader for routed experts from shared memory.

    This class attaches to shared memory created by RoutedExpertsCapturer
    and reads expert routing decisions.
    """

    _instance: RoutedExpertsReader | None = None

    def __init__(self) -> None:
        self._shm: shared_memory.SharedMemory | None = None
        self._host_buffer_view: np.ndarray | None = None
        self._lock_file: str | None = None

    @classmethod
    def create(cls) -> RoutedExpertsReader:
        """Create a global singleton instance."""
        global _global_experts_reader
        if _global_experts_reader is not None:
            raise RuntimeError("Experts reader already created.")

        _global_experts_reader = cls()
        return _global_experts_reader

    @staticmethod
    def get_instance() -> RoutedExpertsReader | None:
        """Get the global singleton instance."""
        if _global_experts_reader is None:
            logger.info("Experts reader not initialized.")
        return _global_experts_reader

    def attach_buffer(
        self,
        max_num_kv_tokens: int,
        model_config: ModelConfig,
        instance_id: str,
    ) -> None:
        """
        Attach to an existing shared memory buffer.

        Args:
            max_num_kv_tokens: Maximum number of KV tokens.
            model_config: Model configuration.
            instance_id: Unique identifier for the shared memory buffer.
        """
        if self._shm is not None:
            logger.warning("Already attached to shared memory buffer.")
            return  # Already attached

        hf_config = model_config.hf_text_config
        shape = (
            max_num_kv_tokens,
            hf_config.num_hidden_layers,
            hf_config.num_experts_per_tok,
        )

        self._lock_file = f"{_LOCK_FILE_PREFIX}_{instance_id}.lock"
        shm_name = f"{_BUFFER_PREFIX}_{instance_id}"

        with _file_lock(self._lock_file, mode="rb+"):
            # Avoid resource_tracker registering the shared memory
            with patch(
                "multiprocessing.resource_tracker.register",
                lambda *args, **kwargs: None,
            ):
                self._shm = shared_memory.SharedMemory(name=shm_name)

            self._host_buffer_view = np.ndarray(
                shape, dtype=np.int32, buffer=self._shm.buf
            )

    def get_routed_experts(self, indices: np.ndarray) -> np.ndarray:
        """
        Read routed expert data from shared memory.

        Args:
            indices: Array of indices to read.

        Returns:
            Copy of the expert routing data for the given indices.
        """
        if self._host_buffer_view is None:
            raise RuntimeError("Buffer not attached. Call attach_buffer() first.")
        if self._lock_file is None:
            raise RuntimeError("Lock file not initialized.")

        with _file_lock(self._lock_file, mode="rb+"):
            return self._host_buffer_view[indices, :, :].copy()

    def cleanup(self) -> None:
        """Explicitly clean up resources (close without unlink)."""
        if self._shm is not None:
            try:
                self._shm.close()
            except Exception:
                logger.debug("Exception during cleanup for reader", exc_info=True)
            finally:
                self._shm = None

    def __del__(self) -> None:
        """Close shared memory on destruction (do not unlink)."""
        self.cleanup()