vllm.v1.sample.logits_processor

@dataclass(frozen=True)
class BatchUpdate:
    """Persistent batch state change info for logitsprocs"""
    batch_size: int  # Current num reqs in batch

    # Metadata for requests added to, removed from, and moved
    # within the persistent batch.
    #
    # Key assumption: the `output_tok_ids` list (which is an element of each
    # tuple in `added`) is a reference to the request's running output tokens
    # list; via this reference, the logits processors always see the latest
    # list of generated output tokens
    removed: Sequence[RemovedRequest]
    moved: Sequence[MovedRequest]
    added: Sequence[AddedRequest]

class BatchUpdateBuilder:
    """Helps track persistent batch state changes and build
    a batch update data structure for logitsprocs
    Assumptions:
    * All information about requests removed from persistent batch
      during a step is aggregated in self._removed through calls to
      self.removed_append() at the beginning of a step. This must happen
      before the first time that self.removed, self.pop_removed()
      or self.peek_removed() are invoked in a given step
    * After the first time that self.removed, self.pop_removed()
      or self.peek_removed() are read in a step, no new removals
      are registered using self.removed_append()
    * Elements of self._removed are never directly modified, added or
      removed (i.e. modification is only via self.removed_append() and
      self.pop_removed())
    Guarantees under above assumptions:
    * self.removed is always sorted in descending order
    * self.pop_removed() and self.peek_removed() both return
      the lowest removed request index in the current step
    """

    _removed: list[RemovedRequest]
    _is_removed_sorted: bool
    moved: list[MovedRequest]
    added: list[AddedRequest]

    def __init__(
        self,
        removed: Optional[list[RemovedRequest]] = None,
        moved: Optional[list[MovedRequest]] = None,
        added: Optional[list[AddedRequest]] = None,
    ) -> None:
        self._removed = removed or []
        self.moved = moved or []
        self.added = added or []
        self._is_removed_sorted = False

        # Used to track changes in the pooling case
        # where we don't populate the added list.
        self.batch_changed = False

    def _ensure_removed_sorted(self) -> None:
        """Sort removed request indices in
        descending order.
        Idempotent after first call in a
        given step, until reset.
        """
        if not self._is_removed_sorted:
            self._removed.sort(reverse=True)
            self._is_removed_sorted = True

    @property
    def removed(self) -> list[RemovedRequest]:
        """Removed request indices sorted in
        descending order"""
        self._ensure_removed_sorted()
        return self._removed

    def removed_append(self, index: int) -> None:
        """Register the removal of a request from the persistent batch.

        Must not be called after the first time self.removed,
        self.pop_removed() or self.peek_removed() are invoked.

        Args:
          index: request index
        """
        if self._is_removed_sorted:
            raise RuntimeError("Cannot register new removed request after"
                               " self.removed has been read.")
        self._removed.append(index)
        self.batch_changed = True

    def has_removed(self) -> bool:
        return bool(self._removed)

    def peek_removed(self) -> Optional[int]:
        """Return lowest removed request index"""
        if self.has_removed():
            self._ensure_removed_sorted()
            return self._removed[-1]
        return None

    def pop_removed(self) -> Optional[int]:
        """Pop lowest removed request index"""
        if self.has_removed():
            self._ensure_removed_sorted()
            return self._removed.pop()
        return None

    def reset(self) -> bool:
        """Returns True if there were any changes to the batch."""
        self._is_removed_sorted = False
        self._removed.clear()
        self.moved.clear()
        self.added.clear()
        batch_changed = self.batch_changed
        self.batch_changed = False
        return batch_changed

    def get_and_reset(self, batch_size: int) -> Optional[BatchUpdate]:
        """Generate a logitsprocs batch update data structure and reset
        internal batch update builder state.

        Args:
          batch_size: current persistent batch size

        Returns:
          Frozen logitsprocs batch update instance; `None` if no updates
        """
        # Reset removal-sorting logic
        self._is_removed_sorted = False
        self.batch_changed = False
        if not any((self._removed, self.moved, self.added)):
            # No update; short-circuit
            return None
        # Build batch state update
        batch_update = BatchUpdate(
            batch_size=batch_size,
            removed=self._removed,
            moved=self.moved,
            added=self.added,
        )
        self._removed = []
        self.moved = []
        self.added = []
        return batch_update

class LogitBiasLogitsProcessor(LogitsProcessor):

    def __init__(self, _, device: torch.device, is_pin_memory: bool):
        self.device = device
        self.pin_memory = is_pin_memory
        self.biases: dict[int, dict[int, float]] = {}

        self.bias_tensor: torch.Tensor = torch.tensor(())
        self.logits_slice = (self._device_tensor([], torch.int32),
                             self._device_tensor([], torch.int32))

    def is_argmax_invariant(self) -> bool:
        """Logit bias can rebalance token probabilities and change the
        outcome of argmax in greedy sampling."""
        return False

    def update_state(self, batch_update: Optional[BatchUpdate]):
        if not batch_update:
            return

        needs_update: bool = False
        # Process added requests.
        for index, params, _, _ in batch_update.added:
            if lb := params.logit_bias:
                self.biases[index] = lb
                needs_update = True
            else:
                # Drop biases metadata at batch index
                if self.biases.pop(index, None) is not None:
                    # If a new request replaces an old request which
                    # specified biases, we should update processor tensors
                    needs_update = True

        if self.biases:
            # Process removed requests.
            for index in batch_update.removed:
                if self.biases.pop(index, None):
                    needs_update = True

            # Process moved requests, unidirectional (a->b) and swap (a<->b)
            for a_index, b_index, direct in batch_update.moved:
                if direct == MoveDirectionality.UNIDIRECTIONAL:
                    if (a_entry := self.biases.pop(a_index, None)) is None:
                        if self.biases.pop(b_index, None) is not None:
                            needs_update = True
                    else:
                        self.biases[b_index] = a_entry
                        needs_update = True
                else:
                    a_entry = self.biases.pop(a_index, None)
                    if (b_entry := self.biases.pop(b_index, None)) is not None:
                        self.biases[a_index] = b_entry
                        needs_update = True
                    if a_entry is not None:
                        self.biases[b_index] = a_entry
                        needs_update = True

        # Update tensors if needed.
        if needs_update:
            reqs, tok_ids, biases = [], [], []
            for req, lb in self.biases.items():
                reqs.extend([req] * len(lb))
                tok_ids.extend(lb.keys())
                biases.extend(lb.values())

            self.bias_tensor = self._device_tensor(biases, torch.float32)
            self.logits_slice = (self._device_tensor(reqs, torch.int32),
                                 self._device_tensor(tok_ids, torch.int32))

    def _device_tensor(self, data: list, dtype: torch.dtype) -> torch.Tensor:
        return (torch.tensor(data,
                             device="cpu",
                             dtype=dtype,
                             pin_memory=self.pin_memory).to(device=self.device,
                                                            non_blocking=True))

    def apply(self, logits: torch.Tensor) -> torch.Tensor:
        if self.biases:
            logits[self.logits_slice] += self.bias_tensor
        return logits

class LogitsProcessor(ABC):

    @abstractmethod
    def __init__(self, vllm_config: "VllmConfig", device: torch.device,
                 is_pin_memory: bool) -> None:
        raise NotImplementedError

    @abstractmethod
    def apply(self, logits: torch.Tensor) -> torch.Tensor:
        raise NotImplementedError

    @abstractmethod
    def is_argmax_invariant(self) -> bool:
        """True if logits processor has no impact on the
        argmax computation in greedy sampling.
        NOTE: may or may not have the same value for all
        instances of a given LogitsProcessor subclass,
        depending on subclass implementation.
        """
        raise NotImplementedError

    @abstractmethod
    def update_state(
        self,
        batch_update: Optional["BatchUpdate"],
    ) -> None:
        """Called when there are new output tokens, prior
        to each forward pass.

        Args:
            batch_update is non-None iff there have been
            changes to the batch makeup.
        """
        raise NotImplementedError

class LogitsProcessors:
    """Encapsulates initialized logitsproc objects."""

    def __init__(
            self,
            logitsprocs: Optional[Iterator["LogitsProcessor"]] = None) -> None:
        self.argmax_invariant: list[LogitsProcessor] = []
        self.non_argmax_invariant: list[LogitsProcessor] = []
        if logitsprocs:
            for logitproc in logitsprocs:
                (self.argmax_invariant if logitproc.is_argmax_invariant() else
                 self.non_argmax_invariant).append(logitproc)

    @property
    def all(self) -> Iterator["LogitsProcessor"]:
        """Iterator over all logits processors."""
        return chain(self.argmax_invariant, self.non_argmax_invariant)

class MinPLogitsProcessor(LogitsProcessor):

    def __init__(self, vllm_config: "VllmConfig", device: torch.device,
                 is_pin_memory: bool):
        max_num_reqs = vllm_config.scheduler_config.max_num_seqs
        self.min_p_count: int = 0

        self.min_p_cpu_tensor = torch.zeros((max_num_reqs, ),
                                            dtype=torch.float32,
                                            device="cpu",
                                            pin_memory=is_pin_memory)
        self.min_p_cpu = self.min_p_cpu_tensor.numpy()

        self.use_double_tensor = torch.device(device).type != "cpu"

        if self.use_double_tensor:
            # Pre-allocated device tensor
            self.min_p_device: torch.Tensor = torch.empty((max_num_reqs, ),
                                                          dtype=torch.float32,
                                                          device=device)
        else:
            self.min_p_device = self.min_p_cpu_tensor
        # Current slice of the device tensor
        self.min_p: torch.Tensor = self.min_p_device[:0]

    def is_argmax_invariant(self) -> bool:
        """Min-p never impacts greedy sampling"""
        return True

    def get_min_p_by_index(self, index: int) -> float:
        return float(self.min_p_cpu[index])

    def update_state(self, batch_update: Optional[BatchUpdate]):
        if not batch_update:
            return

        needs_update = False
        # Process added requests.
        for index, params, _, _ in batch_update.added:
            min_p = params.min_p
            min_p_before = self.min_p_cpu[index]
            if min_p_before != min_p:
                needs_update = True
                self.min_p_cpu[index] = min_p
                if min_p and not min_p_before:
                    self.min_p_count += 1
                elif not min_p and min_p_before:
                    self.min_p_count -= 1

        if self.min_p_count:
            # Process removed requests.
            if batch_update.removed:
                needs_update = True
                for index in batch_update.removed:
                    if self.min_p_cpu[index]:
                        self.min_p_cpu[index] = 0
                        self.min_p_count -= 1

            # Process moved requests, unidirectional (a->b) and swap (a<->b).
            for adx, bdx, direct in batch_update.moved:
                min_p_a, min_p_b = self.min_p_cpu[adx], self.min_p_cpu[bdx]
                if min_p_a != min_p_b:
                    needs_update = True
                    self.min_p_cpu[bdx] = min_p_a
                    if direct == MoveDirectionality.SWAP:
                        self.min_p_cpu[adx] = min_p_b
                if direct == MoveDirectionality.UNIDIRECTIONAL:
                    if min_p_a:
                        self.min_p_cpu[adx] = 0
                    if min_p_b:
                        self.min_p_count -= 1

        # Update tensors if needed.
        size = batch_update.batch_size
        if self.min_p_count and (needs_update or self.min_p.shape[0] != size):
            self.min_p = self.min_p_device[:size]
            if self.use_double_tensor:
                self.min_p.copy_(self.min_p_cpu_tensor[:size],
                                 non_blocking=True)
            self.min_p.unsqueeze_(1)

    def apply(self, logits: torch.Tensor) -> torch.Tensor:
        if not self.min_p_count:
            return logits

        # Convert logits to probability distribution
        probability_values = torch.nn.functional.softmax(logits, dim=-1)
        # Calculate maximum probabilities per sequence
        max_probabilities = torch.amax(probability_values,
                                       dim=-1,
                                       keepdim=True)
        # Adjust min_p
        adjusted_min_p = max_probabilities.mul_(self.min_p)
        # Identify valid tokens using threshold comparison
        invalid_token_mask = probability_values < adjusted_min_p
        # Apply mask using boolean indexing
        logits[invalid_token_mask] = -float('inf')
        return logits