backend

AsyncVirtualLM

Bases: AsyncLM

Async language model using vLLM v1 with live engine requests.

Concurrent calls are auto-batched. A one-token extension of an idle context appends to that context's engine request, and identical contexts in one batch share one row.

Source code in genlm/backend/llm/vllm.py

class AsyncVirtualLM(AsyncLM):  # pragma: no cover
    """Async language model using vLLM v1 with live engine requests.

    Concurrent calls are auto-batched. A one-token extension of an idle
    context appends to that context's engine request, and identical
    contexts in one batch share one row.
    """

    def __init__(self, llm_engine):
        """Initialize an `AsyncVirtualLM` instance.

        Args:
            llm_engine (LLM): The vLLM engine instance.
        """
        self.llm_engine = llm_engine
        self.tokenizer = llm_engine.get_tokenizer()
        self._params = SamplingParams(
            n=1, max_tokens=1 << 20, ignore_eos=True, detokenize=False
        )

        # name -> LoRARequest. vLLM caches adapter weights by int id, so ids
        # are never reused. Loop-side only; the engine thread never reads this map.
        self._lora_requests = {}
        self._next_lora_id = 1

        # Only the engine thread reads or writes the state below.
        self._requests = {}  # rid -> (tuple(ids), lora_name), recency order
        self._by_content = {}  # (tuple(ids), lora_name) -> rid
        self._pending = {}  # rid -> [(future, loop)] awaiting a row or an exception
        self._served = 0  # rows resolved; the engine thread's progress signal
        self._next_rid = 0
        self._max_requests = 1 << 30  # tightened by from_name

        self._work = queue.SimpleQueue()
        self._engine_thread = None  # started by from_name once the engine is wired
        self._sched = None
        self._core = (
            None  # in-process engine-core client, stepped by the engine thread
        )
        self._block_hasher = None

        super().__init__(tokenizer=self.tokenizer)

    @classmethod
    def from_name(cls, model_name, engine_opts=None, **kwargs):
        """Create a `AsyncVirtualLM` instance from a model name.

        Args:
            model_name (str): Name of the model to load.
            engine_opts (dict): Additional options to pass to the `LLM` engine.
            **kwargs: Additional arguments passed to `AsyncVirtualLM` constructor.

        Returns:
            (AsyncVirtualLM): An `AsyncVirtualLM` instance.
        """
        if not HAS_VLLM:
            raise ImportError(  # pragma: no cover
                "vLLM not available. Install vLLM or use AsyncTransformer instead."
            )

        engine_opts = {
            "enable_prefix_caching": True,
            "disable_log_stats": True,
            "gpu_memory_utilization": 0.9,
            "async_scheduling": True,
            "scheduler_cls": BackendScheduler,
            **(engine_opts or {}),
        }

        global _PENDING_CAPTURE
        capture = {"armed": False, "resolve": None}
        _PENDING_CAPTURE = capture
        try:
            llm = LLM(model=model_name, tokenizer=model_name, **engine_opts)
        finally:
            _PENDING_CAPTURE = None
        if not capture["armed"]:
            raise RuntimeError(
                "capture sampler was not installed: vLLM's "
                "ModelState.custom_sampler hook has moved"
            )

        inst = cls(llm, **kwargs)
        capture["resolve"] = _weak_resolve(inst)
        inst._capture = capture

        # The engine thread drives the engine core directly: LLMEngine's output
        # processor would reject requests it never registered.
        inst._core = llm.llm_engine.engine_core
        engine_core = inst._core.engine_core
        sched = engine_core.scheduler
        if not isinstance(sched, BackendScheduler):
            raise RuntimeError(
                f"engine scheduler is {type(sched).__name__}, not BackendScheduler"
            )
        inst._sched = sched
        inst._block_hasher = engine_core.request_block_hasher
        inst._max_requests = max(
            8, llm.llm_engine.vllm_config.scheduler_config.max_num_seqs - 8
        )
        inst._engine_thread = threading.Thread(
            target=_engine_loop,
            args=(inst._work, weakref.ref(inst)),
            name="genlm-engine",
            daemon=True,
        )
        inst._engine_thread.start()
        return inst

    @property
    def underlying_model(self):
        """Access the underlying model for advanced use cases."""
        engine_core = self.llm_engine.llm_engine.engine_core.engine_core
        return engine_core.model_executor.driver_worker.worker.model_runner.model

    # -- LoRA -------------------------------------------------------------

    def add_new_lora(self, lora_path, lora_name="lora_1"):
        """Register a LoRA adapter under ``lora_name``.

        Re-registering an existing name rebinds it to the weights at
        ``lora_path``; forwards still pending under the old weights fail.
        Forwards select the adapter per call via ``lora_name=``.

        Args:
            lora_path (str): Path to the adapter weights directory or identifier in HuggingFace's model hub.
            lora_name (str): Name to assign to the loaded adapter.
        """
        if lora_name in self._lora_requests:
            del self._lora_requests[lora_name]
            self._work.put(("abort_adapter", lora_name, None, None))
        lid = self._next_lora_id
        self._next_lora_id += 1
        self._lora_requests[lora_name] = LoRARequest(lora_name, lid, lora_path)

    def remove_lora(self, lora_name):
        """Unregister ``lora_name`` and evict its weights. Forwards still pending
        under it fail.

        Args:
            lora_name (str): Name of the adapter to remove.
        """
        req = self._lora_requests.pop(lora_name)
        self._work.put(("remove_lora", (lora_name, req.lora_int_id), None, None))

    def _abort_adapter(self, lora_name):
        """(engine thread) Abort every request under ``lora_name``."""
        self._abort(
            [rid for rid, (_, name) in self._requests.items() if name == lora_name],
            RuntimeError(
                f"adapter {lora_name!r} was rebound or removed mid-forward"
            ),
        )

    # -- the batch --------------------------------------------------------

    async def next_token_logprobs(self, token_ids, lora_name=None):
        """Request log probabilities of next token asynchronously with auto-batching.

        Args:
            token_ids (list[int]): A list of token IDs, representing a prompt to the language model.
            lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

        Returns:
            result (torch.Tensor): Normalized log probability tensor.
        """
        (future,) = await self._enqueue([token_ids], lora_name)
        return (await future).clone()

    async def batch_next_token_logprobs(self, token_ids_list, lora_name=None):
        """Batch request log probabilities for multiple token sequences asynchronously.

        Concurrent callers, batched or single, are dispatched together as
        one batch.

        Args:
            token_ids_list (list[list[int]]): A list of token ID lists.
            lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

        Returns:
            (torch.Tensor): A ``[N, vocab]`` tensor of normalized log probabilities.
        """
        futures = await self._enqueue(token_ids_list, lora_name)
        return torch.stack(await asyncio.gather(*futures))

    async def _enqueue(self, token_ids_list, lora_name):
        """Queue requests whose futures share captured rows; public returns copy once."""
        self._check_alive()
        if lora_name is not None and lora_name not in self._lora_requests:
            raise ValueError(UNKNOWN_ADAPTER.format(lora_name))
        if any(not token_ids for token_ids in token_ids_list):
            raise ValueError("token_ids must not be empty")

        loop = asyncio.get_running_loop()
        futures = [loop.create_future() for _ in token_ids_list]
        batch = await self._join_batch(
            [
                ((tuple(token_ids), lora_name), loop, future)
                for token_ids, future in zip(token_ids_list, futures)
            ]
        )
        if batch is not None:
            queries = self._bind_adapters(batch)
            if queries:
                self._work.put(("queries", queries, None, None))
        return futures

    def _check_alive(self):
        if self._engine_thread is None:
            raise RuntimeError(
                "engine thread not running: this model was cleaned up"
                if getattr(self, "_engine_cleaned", False)
                else "engine thread not running; construct via from_name()"
            )

    def _bind_adapters(self, batch):
        """Resolve each query's adapter as the batch is dispatched.

        Nothing may yield between this and the ``_work`` put, or a rebind
        could land between them.
        """
        bound = []
        for key, loop, future in batch:
            lora_name = key[1]
            if lora_name is None:
                bound.append((key, None, loop, future))
            elif lora_name in self._lora_requests:
                bound.append((key, self._lora_requests[lora_name], loop, future))
            else:  # removed while this query sat in the batch
                self._resolve(
                    loop, future, exc=ValueError(UNKNOWN_ADAPTER.format(lora_name))
                )
        return bound

    # -- the engine thread --------------------------------------------------

    def _engine_step(self, item):
        """(engine thread) Handle one work item, then step the engine until no
        row is pending. A failure goes to every pending future. Returns whether
        the thread should stop."""
        stopping = item is _STOP
        try:
            if not stopping:
                self._handle(item)
            stalled = 0
            steps = 0
            while not stopping and self._pending:
                served = self._served
                self._core.get_output()
                steps += 1
                while True:
                    try:
                        nxt = self._work.get_nowait()
                    except queue.Empty:
                        break
                    if nxt is _STOP:
                        stopping = True
                        break
                    self._handle(nxt)
                    stalled = 0
                if stopping:
                    break
                if self._served == served:
                    stalled += 1
                    if stalled > 4096:
                        raise RuntimeError(
                            "engine made no progress on pending rows: "
                            f"{list(self._pending)}"
                        )
                else:
                    stalled = 0
        except BaseException as exc:
            self._fail_pending(exc)
        if stopping:
            self._fail_pending(RuntimeError("backend was shut down"))
        return stopping

    def _handle(self, item):
        """(engine thread) Execute one work item. A barrier item resolves its
        own future with the result or the failure; a query batch leaves its
        futures in ``_pending`` for ``_resolve`` or ``_fail_pending``."""
        kind, arg, future, loop = item
        if kind == "queries":
            self._reconcile(arg)
            return
        try:
            if kind == "abort_adapter":
                self._abort_adapter(arg)
            elif kind == "release":
                self._evict_idle(len(self._requests))
            elif kind == "remove_lora":
                lora_name, lora_int_id = arg
                self._abort_adapter(lora_name)
                # One step flushes the finished ids through the runner
                # before the weights they used disappear.
                self._core.get_output()
                self.llm_engine.llm_engine.remove_lora(lora_int_id)
        except BaseException as exc:
            if future is not None:
                self._resolve(loop, future, exc=exc)
            raise
        if future is not None:
            self._resolve(loop, future, None)

    def _reconcile(self, batch):
        """(engine thread) Reconcile one batch against the request table, each
        context before its extensions. A failure fails every query in the batch."""
        try:
            self._reconcile_inner(batch)
        except BaseException as exc:
            for _, _, loop, future in batch:
                self._resolve(loop, future, exc=exc)
            raise

    def _reconcile_inner(self, batch):
        self._evict_under_pressure()
        grouped = {}
        for key, lora_request, loop, future in batch:
            entry = grouped.setdefault(key, (lora_request, []))
            entry[1].append((future, loop))
        for key in sorted(grouped, key=lambda k: len(k[0])):
            lora_request, waiters = grouped[key]
            ids, lora_name = key
            rid = None
            # A request is extendable only while no row is pending on it: one
            # request cannot serve both a context and its extension.
            cand = self._by_content.get((ids[:-1], lora_name))
            if cand is not None and cand not in self._pending:
                request = self._sched.requests.get(cand)
                if request is None:
                    self._forget(cand)  # engine dropped it (e.g. preempt races)
                else:
                    rid = cand
                    self._sched.feed_token(request, ids[-1])
                    self._rekey(rid, key)
            if rid is None:
                rid = self._new_rid()
                self._sched.add_request(
                    Request(
                        request_id=rid,
                        prompt_token_ids=list(ids),
                        sampling_params=self._params,
                        pooling_params=None,
                        lora_request=lora_request,
                        block_hasher=self._block_hasher,
                    )
                )
                self._admit(rid, key)
            self._pending[rid] = waiters

        self._evict_over_cap()

    def _new_rid(self):
        self._next_rid += 1
        return f"{_REQ_PREFIX}{self._next_rid}"

    # -- delivery -----------------------------------------------------------

    def _receive_rows(self, rows):
        """(engine thread) Capture callback: resolve the forward's pending
        futures, one callback per event loop."""
        by_loop = {}
        for req_id, row in rows.items():
            waiters = self._pending.pop(req_id, None)
            if not waiters:
                continue
            self._served += 1
            for future, loop in waiters:
                by_loop.setdefault(loop, []).append((future, row))
        for loop, items in by_loop.items():

            def _set(items=items):
                for future, result in items:
                    if not future.done():
                        future.set_result(result)

            try:
                loop.call_soon_threadsafe(_set)
            except RuntimeError:
                pass  # the loop closed; the rows have no reader

    def _fail_pending(self, exc):
        """(engine thread) Abort every request with a pending row."""
        self._abort(list(self._pending), exc)

    def _fail_waiters(self, waiters, exc):
        for future, loop in waiters:
            self._resolve(loop, future, exc=exc)

    @staticmethod
    def _resolve(loop, future, result=None, exc=None):
        def _set():
            if not future.done():
                if exc is not None:
                    future.set_exception(exc)
                else:
                    future.set_result(result)

        try:
            loop.call_soon_threadsafe(_set)
        except RuntimeError:
            pass  # the loop closed; the result has no reader

    async def _submit(self, kind, arg=None):
        self._check_alive()
        loop = asyncio.get_running_loop()
        future = loop.create_future()
        self._work.put((kind, arg, future, loop))
        return await future

    # -- the request table ------------------------------------------------

    def _abort(self, rids, exc):
        """(engine thread) Fail the futures pending on ``rids`` and drop the
        requests from the table and the engine.

        Never raises: it runs in the engine thread's failure handler, where an
        exception would kill the thread.
        """
        rids = list(rids)
        for rid in rids:
            waiters = self._pending.pop(rid, None)
            if waiters:
                self._fail_waiters(waiters, exc)
            self._forget(rid)
        if rids:
            try:
                self._sched.finish_requests(rids, RequestStatus.FINISHED_ABORTED)
            except BaseException:  # pragma: no cover
                logging.getLogger(__name__).exception(
                    "could not finish aborted requests %s", rids
                )

    def _admit(self, rid, key):
        """Register a live request at ``key``. A newcomer takes the content
        index; a displaced incumbent stays in the table as an idle row."""
        self._requests[rid] = key
        self._by_content[key] = rid

    def _rekey(self, rid, key):
        """Re-key an extended request (re-insertion keeps recency order)."""
        self._forget(rid)
        self._admit(rid, key)

    def _forget(self, rid):
        """Drop ``rid`` from the table and, if it still holds it, the index."""
        key = self._requests.pop(rid, None)
        if key is not None and self._by_content.get(key) == rid:
            del self._by_content[key]

    def _evict_idle(self, n):
        """(engine thread) Finish up to ``n`` oldest requests with no pending row."""
        victims = []
        for rid in self._requests:
            if len(victims) >= n:
                break
            if rid not in self._pending:
                victims.append(rid)
        for rid in victims:
            self._forget(rid)
        if victims:
            self._sched.finish_requests(victims, RequestStatus.FINISHED_ABORTED)

    def _evict_over_cap(self):
        over = len(self._requests) - self._max_requests
        if over > 0:
            self._evict_idle(over)

    def _evict_under_pressure(self):
        """(engine thread) Evict the oldest idle requests when KV-cache usage reaches 90%."""
        if self._sched.kv_cache_manager.usage >= 0.9:
            self._evict_idle(max(1, len(self._requests) // 8))

    async def release_all(self):
        """Evict every idle request (end of an inference run)."""
        await self._submit("release")

    # -- sync paths -----------------------------------------------------

    def next_token_logprobs_sync(self, token_ids, lora_name=None):
        """Request log probabilities of next token synchronously.

        Does not support auto-batching. For batched sync calls, use
        ``batch_next_token_logprobs_sync`` instead.

        Args:
            token_ids (list[int]): A list of token IDs, representing a prompt to the language model.
            lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

        Returns:
            (torch.Tensor): Normalized log probability tensor.
        """
        return asyncio.run(self.next_token_logprobs(token_ids, lora_name=lora_name))

    def batch_next_token_logprobs_sync(self, token_ids_list, lora_name=None):
        """
        Request log probabilities of next tokens in a batch synchronously.

        Args:
            token_ids_list (list[list[int]]): A list of token ID lists, each representing a prompt to the language model.
            lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

        Returns:
            (torch.Tensor): A tensor of normalized log probability tensors, one for each prompt in the input list.
        """
        return asyncio.run(
            self.batch_next_token_logprobs(token_ids_list, lora_name=lora_name)
        )

    def cleanup(self):
        """Explicitly clean up GPU resources. Call this when done with the model."""
        self._cleanup_engine()

    def __enter__(self):
        return self

    def __exit__(self, exc_type, exc_val, exc_tb):
        self.cleanup()
        return False

    async def __aenter__(self):
        return self

    async def __aexit__(self, exc_type, exc_val, exc_tb):
        self.cleanup()
        return False

    def __del__(self):
        """Clean up resources on deletion."""
        self._cleanup_engine()

    def _cleanup_engine(self):
        """Clean up the vLLM engine and associated resources.

        This is invoked from both :meth:`cleanup` (explicit, during normal
        program flow) and :meth:`__del__` (implicit, possibly at
        interpreter shutdown). The narrow exception classes below cover
        the races and idempotency issues we know about:

        * ``ImportError`` / ``AttributeError`` arise when ``__del__`` runs
          after ``sys.meta_path`` is already torn down during interpreter
          shutdown.
        * ``AssertionError`` is raised by vLLM's
          ``destroy_distributed_environment`` if it's called twice.
        * ``RuntimeError`` can surface from CUDA when the driver is
          already being torn down.

        Anything else is re-raised so real bugs are not swallowed.
        """
        if getattr(self, "_engine_cleaned", False):
            return
        self._engine_cleaned = True
        thread = getattr(self, "_engine_thread", None)
        if thread is not None and thread.is_alive():
            self._work.put(_STOP)
            thread.join(timeout=10)
        self._engine_thread = None
        # Only release what the engine thread can no longer reach: a thread
        # that outlived its join would fault on the handles this drops.
        released = thread is None or not thread.is_alive()
        try:
            import gc

            capture = getattr(self, "_capture", None)
            if capture is not None:
                capture["resolve"] = _drop_rows
            if released:
                self._sched = self._core = self._block_hasher = None
                self.llm_engine = None
            # vLLM's internals are cyclic; only a collection frees them.
            gc.collect()
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
                torch.cuda.synchronize()
            destroy_model_parallel()
            destroy_distributed_environment()
        except (ImportError, AttributeError, AssertionError, RuntimeError):
            pass

__init__(llm_engine)

Initialize an AsyncVirtualLM instance.

Parameters:

Name	Type	Description	Default
llm_engine	LLM	The vLLM engine instance.	required

Source code in genlm/backend/llm/vllm.py

def __init__(self, llm_engine):
    """Initialize an `AsyncVirtualLM` instance.

    Args:
        llm_engine (LLM): The vLLM engine instance.
    """
    self.llm_engine = llm_engine
    self.tokenizer = llm_engine.get_tokenizer()
    self._params = SamplingParams(
        n=1, max_tokens=1 << 20, ignore_eos=True, detokenize=False
    )

    # name -> LoRARequest. vLLM caches adapter weights by int id, so ids
    # are never reused. Loop-side only; the engine thread never reads this map.
    self._lora_requests = {}
    self._next_lora_id = 1

    # Only the engine thread reads or writes the state below.
    self._requests = {}  # rid -> (tuple(ids), lora_name), recency order
    self._by_content = {}  # (tuple(ids), lora_name) -> rid
    self._pending = {}  # rid -> [(future, loop)] awaiting a row or an exception
    self._served = 0  # rows resolved; the engine thread's progress signal
    self._next_rid = 0
    self._max_requests = 1 << 30  # tightened by from_name

    self._work = queue.SimpleQueue()
    self._engine_thread = None  # started by from_name once the engine is wired
    self._sched = None
    self._core = (
        None  # in-process engine-core client, stepped by the engine thread
    )
    self._block_hasher = None

    super().__init__(tokenizer=self.tokenizer)

from_name(model_name, engine_opts=None, **kwargs) classmethod

Create a AsyncVirtualLM instance from a model name.

Parameters:

Name	Type	Description	Default
model_name	str	Name of the model to load.	required
engine_opts	dict	Additional options to pass to the LLM engine.	None
**kwargs		Additional arguments passed to AsyncVirtualLM constructor.	{}

Returns:

Type	Description
AsyncVirtualLM	An AsyncVirtualLM instance.

Source code in genlm/backend/llm/vllm.py

@classmethod
def from_name(cls, model_name, engine_opts=None, **kwargs):
    """Create a `AsyncVirtualLM` instance from a model name.

    Args:
        model_name (str): Name of the model to load.
        engine_opts (dict): Additional options to pass to the `LLM` engine.
        **kwargs: Additional arguments passed to `AsyncVirtualLM` constructor.

    Returns:
        (AsyncVirtualLM): An `AsyncVirtualLM` instance.
    """
    if not HAS_VLLM:
        raise ImportError(  # pragma: no cover
            "vLLM not available. Install vLLM or use AsyncTransformer instead."
        )

    engine_opts = {
        "enable_prefix_caching": True,
        "disable_log_stats": True,
        "gpu_memory_utilization": 0.9,
        "async_scheduling": True,
        "scheduler_cls": BackendScheduler,
        **(engine_opts or {}),
    }

    global _PENDING_CAPTURE
    capture = {"armed": False, "resolve": None}
    _PENDING_CAPTURE = capture
    try:
        llm = LLM(model=model_name, tokenizer=model_name, **engine_opts)
    finally:
        _PENDING_CAPTURE = None
    if not capture["armed"]:
        raise RuntimeError(
            "capture sampler was not installed: vLLM's "
            "ModelState.custom_sampler hook has moved"
        )

    inst = cls(llm, **kwargs)
    capture["resolve"] = _weak_resolve(inst)
    inst._capture = capture

    # The engine thread drives the engine core directly: LLMEngine's output
    # processor would reject requests it never registered.
    inst._core = llm.llm_engine.engine_core
    engine_core = inst._core.engine_core
    sched = engine_core.scheduler
    if not isinstance(sched, BackendScheduler):
        raise RuntimeError(
            f"engine scheduler is {type(sched).__name__}, not BackendScheduler"
        )
    inst._sched = sched
    inst._block_hasher = engine_core.request_block_hasher
    inst._max_requests = max(
        8, llm.llm_engine.vllm_config.scheduler_config.max_num_seqs - 8
    )
    inst._engine_thread = threading.Thread(
        target=_engine_loop,
        args=(inst._work, weakref.ref(inst)),
        name="genlm-engine",
        daemon=True,
    )
    inst._engine_thread.start()
    return inst

underlying_model property

Access the underlying model for advanced use cases.

add_new_lora(lora_path, lora_name='lora_1')

Register a LoRA adapter under lora_name.

Re-registering an existing name rebinds it to the weights at lora_path; forwards still pending under the old weights fail. Forwards select the adapter per call via lora_name=.

Parameters:

Name	Type	Description	Default
lora_path	str	Path to the adapter weights directory or identifier in HuggingFace's model hub.	required
lora_name	str	Name to assign to the loaded adapter.	'lora_1'

Source code in genlm/backend/llm/vllm.py

def add_new_lora(self, lora_path, lora_name="lora_1"):
    """Register a LoRA adapter under ``lora_name``.

    Re-registering an existing name rebinds it to the weights at
    ``lora_path``; forwards still pending under the old weights fail.
    Forwards select the adapter per call via ``lora_name=``.

    Args:
        lora_path (str): Path to the adapter weights directory or identifier in HuggingFace's model hub.
        lora_name (str): Name to assign to the loaded adapter.
    """
    if lora_name in self._lora_requests:
        del self._lora_requests[lora_name]
        self._work.put(("abort_adapter", lora_name, None, None))
    lid = self._next_lora_id
    self._next_lora_id += 1
    self._lora_requests[lora_name] = LoRARequest(lora_name, lid, lora_path)

remove_lora(lora_name)

Unregister lora_name and evict its weights. Forwards still pending under it fail.

Parameters:

Name	Type	Description	Default
lora_name	str	Name of the adapter to remove.	required

Source code in genlm/backend/llm/vllm.py

def remove_lora(self, lora_name):
    """Unregister ``lora_name`` and evict its weights. Forwards still pending
    under it fail.

    Args:
        lora_name (str): Name of the adapter to remove.
    """
    req = self._lora_requests.pop(lora_name)
    self._work.put(("remove_lora", (lora_name, req.lora_int_id), None, None))

next_token_logprobs(token_ids, lora_name=None) async

Request log probabilities of next token asynchronously with auto-batching.

Parameters:

Name	Type	Description	Default
token_ids	list[int]	A list of token IDs, representing a prompt to the language model.	required
lora_name	str	Name of the LoRA adapter to use. Defaults to None (the base model).	None

Returns:

Name	Type	Description
result	Tensor	Normalized log probability tensor.

Source code in genlm/backend/llm/vllm.py

async def next_token_logprobs(self, token_ids, lora_name=None):
    """Request log probabilities of next token asynchronously with auto-batching.

    Args:
        token_ids (list[int]): A list of token IDs, representing a prompt to the language model.
        lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

    Returns:
        result (torch.Tensor): Normalized log probability tensor.
    """
    (future,) = await self._enqueue([token_ids], lora_name)
    return (await future).clone()

batch_next_token_logprobs(token_ids_list, lora_name=None) async

Batch request log probabilities for multiple token sequences asynchronously.

Concurrent callers, batched or single, are dispatched together as one batch.

Parameters:

Name	Type	Description	Default
token_ids_list	list[list[int]]	A list of token ID lists.	required
lora_name	str	Name of the LoRA adapter to use. Defaults to None (the base model).	None

Returns:

Type	Description
Tensor	A [N, vocab] tensor of normalized log probabilities.

Source code in genlm/backend/llm/vllm.py

async def batch_next_token_logprobs(self, token_ids_list, lora_name=None):
    """Batch request log probabilities for multiple token sequences asynchronously.

    Concurrent callers, batched or single, are dispatched together as
    one batch.

    Args:
        token_ids_list (list[list[int]]): A list of token ID lists.
        lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

    Returns:
        (torch.Tensor): A ``[N, vocab]`` tensor of normalized log probabilities.
    """
    futures = await self._enqueue(token_ids_list, lora_name)
    return torch.stack(await asyncio.gather(*futures))

release_all() async

Evict every idle request (end of an inference run).

Source code in genlm/backend/llm/vllm.py

async def release_all(self):
    """Evict every idle request (end of an inference run)."""
    await self._submit("release")

next_token_logprobs_sync(token_ids, lora_name=None)

Request log probabilities of next token synchronously.

Does not support auto-batching. For batched sync calls, use batch_next_token_logprobs_sync instead.

Parameters:

Name	Type	Description	Default
token_ids	list[int]	A list of token IDs, representing a prompt to the language model.	required
lora_name	str	Name of the LoRA adapter to use. Defaults to None (the base model).	None

Returns:

Type	Description
Tensor	Normalized log probability tensor.

Source code in genlm/backend/llm/vllm.py

def next_token_logprobs_sync(self, token_ids, lora_name=None):
    """Request log probabilities of next token synchronously.

    Does not support auto-batching. For batched sync calls, use
    ``batch_next_token_logprobs_sync`` instead.

    Args:
        token_ids (list[int]): A list of token IDs, representing a prompt to the language model.
        lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

    Returns:
        (torch.Tensor): Normalized log probability tensor.
    """
    return asyncio.run(self.next_token_logprobs(token_ids, lora_name=lora_name))

batch_next_token_logprobs_sync(token_ids_list, lora_name=None)

Request log probabilities of next tokens in a batch synchronously.

Parameters:

Name	Type	Description	Default
token_ids_list	list[list[int]]	A list of token ID lists, each representing a prompt to the language model.	required
lora_name	str	Name of the LoRA adapter to use. Defaults to None (the base model).	None

Returns:

Type	Description
Tensor	A tensor of normalized log probability tensors, one for each prompt in the input list.

Source code in genlm/backend/llm/vllm.py

def batch_next_token_logprobs_sync(self, token_ids_list, lora_name=None):
    """
    Request log probabilities of next tokens in a batch synchronously.

    Args:
        token_ids_list (list[list[int]]): A list of token ID lists, each representing a prompt to the language model.
        lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

    Returns:
        (torch.Tensor): A tensor of normalized log probability tensors, one for each prompt in the input list.
    """
    return asyncio.run(
        self.batch_next_token_logprobs(token_ids_list, lora_name=lora_name)
    )

cleanup()

Explicitly clean up GPU resources. Call this when done with the model.

Source code in genlm/backend/llm/vllm.py

def cleanup(self):
    """Explicitly clean up GPU resources. Call this when done with the model."""
    self._cleanup_engine()

__del__()

Clean up resources on deletion.

Source code in genlm/backend/llm/vllm.py

def __del__(self):
    """Clean up resources on deletion."""
    self._cleanup_engine()

AsyncTransformer

Bases: AsyncLM

Asynchronous wrapper around a HuggingFace causal language model with caching support.

This class provides an asynchronous interface to HuggingFace language models with automatic batching and caching (output and KV) for improved efficiency.

Source code in genlm/backend/llm/hf.py

class AsyncTransformer(AsyncLM):
    """Asynchronous wrapper around a HuggingFace causal language model with caching support.

    This class provides an asynchronous interface to HuggingFace language models with automatic batching
    and caching (output and KV) for improved efficiency.
    """

    @classmethod
    def from_name(cls, model_id, bitsandbytes_opts=None, hf_opts=None, **kwargs):
        """Create an AsyncTransformer instance from a pretrained HuggingFace model.

        Args:
            model_id (str): Model identifier in HuggingFace's model hub.
            bitsandbytes_opts (dict, optional): Additional configuration options for bitsandbytes quantization.
                Defaults to None.
            hf_opts (dict, optional): Additional configuration options for loading the HuggingFace model.
                Defaults to None.
            **kwargs: Additional arguments passed to the `AsyncTransformer` constructor

        Returns:
            (AsyncTransformer): An initialized `AsyncTransformer` instance.
        """
        if bitsandbytes_opts:
            bnb_config = BitsAndBytesConfig(**bitsandbytes_opts)
        else:
            bnb_config = None

        _hf_opts = {
            "device_map": "auto",
            "torch_dtype": "auto",
        }
        if hf_opts:
            _hf_opts.update(hf_opts)

        tok = AutoTokenizer.from_pretrained(model_id)
        mod = AutoModelForCausalLM.from_pretrained(
            model_id, quantization_config=bnb_config, **_hf_opts
        )

        return cls(mod, tok, **kwargs)

    @torch.no_grad()
    def __init__(self, hf_model, hf_tokenizer, batch_size=20, timeout=0.02):
        """Initialize an AsyncTransformer instance.

        Args:
            hf_model: A HuggingFace CausalLM model instance.
            hf_tokenizer: A HuggingFace Tokenizer.
            batch_size (int, optional): Maximum queries to process in one batch during auto-batching.
                Defaults to 20.
            timeout (float, optional): Seconds to wait since last query before processing current batch.
                Defaults to 0.02.
        """
        self.model = hf_model
        self.tokenizer = hf_tokenizer
        self.device = hf_model.device
        # One cache trie per LoRA adapter (None = base); tries must never mix.
        self._caches = defaultdict(TokenTrie)

        # Queries to be batched. Each query is a sequence of tokens,
        # and a Future to be called when the query is resolved.
        self.queries = []
        self.batch_size = batch_size
        self.timeout = timeout
        self.timer = None

        self.model.eval()

        super().__init__(tokenizer=self.tokenizer)

    def clear_cache(self):
        """Clear the cache of log probabilities and key/value pairs."""
        self._caches = defaultdict(TokenTrie)

    def clear_kv_cache(self):
        """Clear any key and value vectors from the cache."""
        for trie in self._caches.values():
            trie.clear_kv_cache()

    def reset_async_queries(self):
        """Clear any pending language model queries from the queue. Use this method when an exception prevented an inference algorithm from executing
        to completion."""
        self.queries = []

    @torch.no_grad()
    def cache_kv(self, prompt_tokens, lora_name=None):
        """Cache the key and value vectors for a prompt. Future queries that have this prompt as a prefix will only run the LLM on new tokens.

        Args:
            prompt_tokens (list[int]): token ids for the prompt to cache.
            lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).
        """
        self._activate(lora_name)
        result = self.model(torch.tensor([prompt_tokens]).to(self.device))
        node = self._caches[lora_name].extend_cache(
            0, prompt_tokens, result.logits[0], 0
        )
        node.past_key_values = result.past_key_values

    def _has_adapter(self, lora_name):
        return (
            getattr(self.model, "_hf_peft_config_loaded", False)
            and lora_name in self.model.peft_config
        )

    def add_new_lora(self, lora_path, lora_name="lora_1"):
        """Load a LoRA adapter into the base model.

        Loading under an existing name replaces that adapter. Select it per call with `lora_name`.

        Args:
            lora_path (str): Path to the adapter weights directory or identifier in HuggingFace's model hub.
            lora_name (str): Name to assign to the loaded adapter.
        """
        if self._has_adapter(lora_name):
            self.remove_lora(lora_name)
        self.model.load_adapter(lora_path, lora_name)

    def remove_lora(self, lora_name):
        """Remove a previously loaded LoRA adapter and its cache."""
        self.model.delete_adapter(lora_name)
        self._caches.pop(lora_name, None)

    def _activate(self, lora_name):
        """Activate `lora_name` on the model, or disable adapters if it is None.

        Adapters outside the transformers peft mixin are left untouched under None.

        Raises:
            ValueError: If `lora_name` has not been loaded.
        """
        if lora_name is None:
            if getattr(self.model, "_hf_peft_config_loaded", False):
                self.model.disable_adapters()
        else:
            if not self._has_adapter(lora_name):
                raise ValueError(UNKNOWN_ADAPTER.format(lora_name))
            self.model.set_adapter(lora_name)
            self.model.enable_adapters()

    @torch.no_grad()
    def batch_evaluate_queries(self):
        """
        Process a batch of queued language model queries, one forward per LoRA adapter.

        This method is called internally when the `batch_size` has been met or the `timeout` has expired.
        """

        queries, self.queries = self.queries, []
        if len(queries) == 0:
            return

        by_lora = defaultdict(list)
        for query in queries:
            by_lora[query.lora_name].append(query)
        for lora_name, group in by_lora.items():
            # An exception escaping here would leave every awaiting caller hanging.
            try:
                self._activate(lora_name)
                self._evaluate_queries(group)
            except Exception as exc:
                for query in group:
                    if not query.future.done():
                        query.future.set_exception(exc)

    @torch.no_grad()
    def _evaluate_queries(self, queries):
        """Run one padded forward over `queries` under the currently active adapter."""
        query_groups = defaultdict(list)
        for query in queries:
            key = tuple(query.prompt)  # XXX: cache based on past_len too?
            query_groups[key].append(query)

        # Use one representative query from each group
        unique_queries = [group[0] for group in query_groups.values()]

        past_example = next(
            (q.past for q in unique_queries if q.past is not None), None
        )
        max_past_length = max(q.past_len for q in unique_queries)
        max_query_length = max(len(q.prompt) for q in unique_queries)

        padding_token_id = (
            self.tokenizer.pad_token_id
            if self.tokenizer.pad_token_id is not None
            else 0
        )

        input_ids = torch.tensor(
            [
                q.prompt_padded(padding_token_id, max_query_length)
                for q in unique_queries
            ]
        ).to(self.device)
        attn_masks = torch.tensor(
            [
                q.attention_mask(max_past_length, max_query_length)
                for q in unique_queries
            ]
        ).to(self.device)
        posn_ids = torch.tensor(
            [q.position_ids(max_past_length, max_query_length) for q in unique_queries]
        ).to(self.device)
        if past_example is not None:
            example = past_example.layers[0].keys
            pasts = [
                [
                    torch.cat(
                        (
                            *(
                                q.past_padded(
                                    layer,
                                    side,
                                    max_past_length,
                                    example.dtype,
                                    self.device,
                                    example.shape,
                                )
                                for q in unique_queries
                            ),
                        ),
                        dim=0,
                    )
                    for side in ("keys", "values")
                ]
                for layer in range(len(past_example))
            ]
        else:
            pasts = None

        pasts = DynamicCache(None if pasts is None else [tuple(kv) for kv in pasts])

        results = self.model(
            input_ids,
            attention_mask=attn_masks,
            position_ids=posn_ids,
            past_key_values=pasts,
            use_cache=True,
        )

        assert len(results.logits) == len(unique_queries)

        for i, q in enumerate(unique_queries):
            result = results.logits[i]
            for dup_query in query_groups[tuple(q.prompt)]:
                dup_query.future.set_result(result)

    @torch.no_grad()
    def add_query(self, query, future, past, lora_name=None):
        """Add a query to be evaluated in the next batch.

        This method is called internally when a `next_token_logprobs` request is made.

        Args:
            query (list[int]): Token IDs representing the query prompt
            future (asyncio.Future): Future to store the result in
            past (DynamicCache|None): Past key/value states from previous evaluation,
                or None if this is a new query
            lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).
        """
        self.queries.append(Query(query, future, past, lora_name))

        if self.timer:
            self.timer.cancel()
            self.timer = None
        if len(self.queries) >= self.batch_size:
            self.batch_evaluate_queries()
        else:
            self.timer = asyncio.get_running_loop().call_later(
                self.timeout, lambda: self.batch_evaluate_queries()
            )

    def walk_cache(self, token_ids, lora_name=None):
        """Walk the cache tree to find the deepest node matching a sequence of tokens.

        Args:
            token_ids (list[int]): Sequence of token IDs to follow in the cache tree
            lora_name (str, optional): Which adapter's cache tree to walk. Defaults to None (the base model).

        Returns:
            tuple:
                - CacheNode: The deepest node in the cache tree that matches the token sequence
                - int: Number of tokens matched from the start of token_ids
                - DynamicCache|None: Past key/value states from the deepest cached node,
                    or None if no cached states were found
                - int: Base index indicating where the past states start in token_ids
        """
        # Walk while tokens can be found
        node = self._caches[lora_name]
        next_token_index = 0

        past = None
        base = 0
        while next_token_index < len(token_ids):
            if node.past_key_values is not None:
                past = node.past_key_values
                base = next_token_index
            if node.has_token(token_ids[next_token_index]):
                node = node.get_token(token_ids[next_token_index])
                next_token_index += 1
            else:
                break

        return node, next_token_index, past, base

    @torch.no_grad()
    async def next_token_logprobs(self, token_ids, lora_name=None):
        """Request log probabilities of next token. This version is asynchronous because it automatically batches concurrent requests; use with `await`.

        Args:
            token_ids (list[int]): a list of token ids, representing a prompt to the language model.
            lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

        Returns:
            logprobs (torch.Tensor): a tensor of with the language model's log (normalized) probabilities for the next token following the prompt.
        """
        if not token_ids:
            raise ValueError("Token ids must not be empty")

        node, next_token_index, past, base = self.walk_cache(token_ids, lora_name)

        # If we processed all tokens, then we're done.
        if next_token_index == len(token_ids):
            return node.logprobs

        # Create a future with the prompt
        future = asyncio.get_running_loop().create_future()
        self.add_query(token_ids[base:], future, past, lora_name)
        logits = await future

        # Create new nodes
        node = node.extend_cache(next_token_index, token_ids, logits, base)

        return node.logprobs

    @torch.no_grad()
    def next_token_logprobs_sync(self, token_ids, lora_name=None):
        """Request log probabilities of next token. Not asynchronous, and does not support auto-batching.

        Args:
            token_ids (list[int]): a list of token ids, representing a prompt to the language model.
            lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

        Returns:
            logprobs (torch.Tensor): a tensor with the language model's log (normalized) probabilities for the next token following the prompt.
        """
        if not token_ids:
            raise ValueError("Token ids must not be empty")

        # Walk while tokens can be found
        node, next_token_index, past, base = self.walk_cache(token_ids, lora_name)

        if next_token_index == len(token_ids):
            return node.logprobs

        self._activate(lora_name)
        logits = self.model(
            torch.tensor([token_ids[base:]]).to(self.device),
            past_key_values=node.past_key_values,
            use_cache=node.past_key_values is not None,
        ).logits[0]

        node = node.extend_cache(next_token_index, token_ids, logits, base)

        return node.logprobs

    def next_token_logprobs_uncached(self, token_ids, lora_name=None):
        """Request log probabilities of next token. No KV or output caching, and does not support auto-batching.

        Args:
            token_ids (list[int]): a list of token ids, representing a prompt to the language model.
            lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

        Returns:
            logprobs (torch.Tensor): a tensor with the language model's log (normalized) probabilities for the next token following the prompt.
        """
        if not token_ids:
            raise ValueError("Token ids must not be empty")

        with torch.no_grad():
            self._activate(lora_name)
            logits = self.model(
                torch.tensor([token_ids]).to(self.device),
                past_key_values=None,
                use_cache=False,
            ).logits[0]
            return torch.log_softmax(logits[-1], dim=0, dtype=torch.float32)

from_name(model_id, bitsandbytes_opts=None, hf_opts=None, **kwargs) classmethod

Create an AsyncTransformer instance from a pretrained HuggingFace model.

Parameters:

Name	Type	Description	Default
model_id	str	Model identifier in HuggingFace's model hub.	required
bitsandbytes_opts	dict	Additional configuration options for bitsandbytes quantization. Defaults to None.	None
hf_opts	dict	Additional configuration options for loading the HuggingFace model. Defaults to None.	None
**kwargs		Additional arguments passed to the AsyncTransformer constructor	{}

Returns:

Type	Description
AsyncTransformer	An initialized AsyncTransformer instance.

Source code in genlm/backend/llm/hf.py

@classmethod
def from_name(cls, model_id, bitsandbytes_opts=None, hf_opts=None, **kwargs):
    """Create an AsyncTransformer instance from a pretrained HuggingFace model.

    Args:
        model_id (str): Model identifier in HuggingFace's model hub.
        bitsandbytes_opts (dict, optional): Additional configuration options for bitsandbytes quantization.
            Defaults to None.
        hf_opts (dict, optional): Additional configuration options for loading the HuggingFace model.
            Defaults to None.
        **kwargs: Additional arguments passed to the `AsyncTransformer` constructor

    Returns:
        (AsyncTransformer): An initialized `AsyncTransformer` instance.
    """
    if bitsandbytes_opts:
        bnb_config = BitsAndBytesConfig(**bitsandbytes_opts)
    else:
        bnb_config = None

    _hf_opts = {
        "device_map": "auto",
        "torch_dtype": "auto",
    }
    if hf_opts:
        _hf_opts.update(hf_opts)

    tok = AutoTokenizer.from_pretrained(model_id)
    mod = AutoModelForCausalLM.from_pretrained(
        model_id, quantization_config=bnb_config, **_hf_opts
    )

    return cls(mod, tok, **kwargs)

__init__(hf_model, hf_tokenizer, batch_size=20, timeout=0.02)

Initialize an AsyncTransformer instance.

Parameters:

Name	Type	Description	Default
hf_model		A HuggingFace CausalLM model instance.	required
hf_tokenizer		A HuggingFace Tokenizer.	required
batch_size	int	Maximum queries to process in one batch during auto-batching. Defaults to 20.	20
timeout	float	Seconds to wait since last query before processing current batch. Defaults to 0.02.	0.02

Source code in genlm/backend/llm/hf.py

@torch.no_grad()
def __init__(self, hf_model, hf_tokenizer, batch_size=20, timeout=0.02):
    """Initialize an AsyncTransformer instance.

    Args:
        hf_model: A HuggingFace CausalLM model instance.
        hf_tokenizer: A HuggingFace Tokenizer.
        batch_size (int, optional): Maximum queries to process in one batch during auto-batching.
            Defaults to 20.
        timeout (float, optional): Seconds to wait since last query before processing current batch.
            Defaults to 0.02.
    """
    self.model = hf_model
    self.tokenizer = hf_tokenizer
    self.device = hf_model.device
    # One cache trie per LoRA adapter (None = base); tries must never mix.
    self._caches = defaultdict(TokenTrie)

    # Queries to be batched. Each query is a sequence of tokens,
    # and a Future to be called when the query is resolved.
    self.queries = []
    self.batch_size = batch_size
    self.timeout = timeout
    self.timer = None

    self.model.eval()

    super().__init__(tokenizer=self.tokenizer)

clear_cache()

Clear the cache of log probabilities and key/value pairs.

Source code in genlm/backend/llm/hf.py

def clear_cache(self):
    """Clear the cache of log probabilities and key/value pairs."""
    self._caches = defaultdict(TokenTrie)

clear_kv_cache()

Clear any key and value vectors from the cache.

Source code in genlm/backend/llm/hf.py

def clear_kv_cache(self):
    """Clear any key and value vectors from the cache."""
    for trie in self._caches.values():
        trie.clear_kv_cache()

reset_async_queries()

Clear any pending language model queries from the queue. Use this method when an exception prevented an inference algorithm from executing to completion.

Source code in genlm/backend/llm/hf.py

def reset_async_queries(self):
    """Clear any pending language model queries from the queue. Use this method when an exception prevented an inference algorithm from executing
    to completion."""
    self.queries = []

cache_kv(prompt_tokens, lora_name=None)

Cache the key and value vectors for a prompt. Future queries that have this prompt as a prefix will only run the LLM on new tokens.

Parameters:

Name	Type	Description	Default
prompt_tokens	list[int]	token ids for the prompt to cache.	required
lora_name	str	Name of the LoRA adapter to use. Defaults to None (the base model).	None

Source code in genlm/backend/llm/hf.py

@torch.no_grad()
def cache_kv(self, prompt_tokens, lora_name=None):
    """Cache the key and value vectors for a prompt. Future queries that have this prompt as a prefix will only run the LLM on new tokens.

    Args:
        prompt_tokens (list[int]): token ids for the prompt to cache.
        lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).
    """
    self._activate(lora_name)
    result = self.model(torch.tensor([prompt_tokens]).to(self.device))
    node = self._caches[lora_name].extend_cache(
        0, prompt_tokens, result.logits[0], 0
    )
    node.past_key_values = result.past_key_values

add_new_lora(lora_path, lora_name='lora_1')

Load a LoRA adapter into the base model.

Loading under an existing name replaces that adapter. Select it per call with lora_name.

Parameters:

Name	Type	Description	Default
lora_path	str	Path to the adapter weights directory or identifier in HuggingFace's model hub.	required
lora_name	str	Name to assign to the loaded adapter.	'lora_1'

Source code in genlm/backend/llm/hf.py

def add_new_lora(self, lora_path, lora_name="lora_1"):
    """Load a LoRA adapter into the base model.

    Loading under an existing name replaces that adapter. Select it per call with `lora_name`.

    Args:
        lora_path (str): Path to the adapter weights directory or identifier in HuggingFace's model hub.
        lora_name (str): Name to assign to the loaded adapter.
    """
    if self._has_adapter(lora_name):
        self.remove_lora(lora_name)
    self.model.load_adapter(lora_path, lora_name)

remove_lora(lora_name)

Remove a previously loaded LoRA adapter and its cache.

Source code in genlm/backend/llm/hf.py

def remove_lora(self, lora_name):
    """Remove a previously loaded LoRA adapter and its cache."""
    self.model.delete_adapter(lora_name)
    self._caches.pop(lora_name, None)

batch_evaluate_queries()

Process a batch of queued language model queries, one forward per LoRA adapter.

This method is called internally when the batch_size has been met or the timeout has expired.

Source code in genlm/backend/llm/hf.py

@torch.no_grad()
def batch_evaluate_queries(self):
    """
    Process a batch of queued language model queries, one forward per LoRA adapter.

    This method is called internally when the `batch_size` has been met or the `timeout` has expired.
    """

    queries, self.queries = self.queries, []
    if len(queries) == 0:
        return

    by_lora = defaultdict(list)
    for query in queries:
        by_lora[query.lora_name].append(query)
    for lora_name, group in by_lora.items():
        # An exception escaping here would leave every awaiting caller hanging.
        try:
            self._activate(lora_name)
            self._evaluate_queries(group)
        except Exception as exc:
            for query in group:
                if not query.future.done():
                    query.future.set_exception(exc)

add_query(query, future, past, lora_name=None)

Add a query to be evaluated in the next batch.

This method is called internally when a next_token_logprobs request is made.

Parameters:

Name	Type	Description	Default
query	list[int]	Token IDs representing the query prompt	required
future	Future	Future to store the result in	required
past	DynamicCache | None	Past key/value states from previous evaluation, or None if this is a new query	required
lora_name	str	Name of the LoRA adapter to use. Defaults to None (the base model).	None

Source code in genlm/backend/llm/hf.py

@torch.no_grad()
def add_query(self, query, future, past, lora_name=None):
    """Add a query to be evaluated in the next batch.

    This method is called internally when a `next_token_logprobs` request is made.

    Args:
        query (list[int]): Token IDs representing the query prompt
        future (asyncio.Future): Future to store the result in
        past (DynamicCache|None): Past key/value states from previous evaluation,
            or None if this is a new query
        lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).
    """
    self.queries.append(Query(query, future, past, lora_name))

    if self.timer:
        self.timer.cancel()
        self.timer = None
    if len(self.queries) >= self.batch_size:
        self.batch_evaluate_queries()
    else:
        self.timer = asyncio.get_running_loop().call_later(
            self.timeout, lambda: self.batch_evaluate_queries()
        )

walk_cache(token_ids, lora_name=None)

Walk the cache tree to find the deepest node matching a sequence of tokens.

Parameters:

Name	Type	Description	Default
token_ids	list[int]	Sequence of token IDs to follow in the cache tree	required
lora_name	str	Which adapter's cache tree to walk. Defaults to None (the base model).	None

Returns:

Name	Type	Description
tuple		CacheNode: The deepest node in the cache tree that matches the token sequence int: Number of tokens matched from the start of token_ids DynamicCache|None: Past key/value states from the deepest cached node, or None if no cached states were found int: Base index indicating where the past states start in token_ids

Source code in genlm/backend/llm/hf.py

def walk_cache(self, token_ids, lora_name=None):
    """Walk the cache tree to find the deepest node matching a sequence of tokens.

    Args:
        token_ids (list[int]): Sequence of token IDs to follow in the cache tree
        lora_name (str, optional): Which adapter's cache tree to walk. Defaults to None (the base model).

    Returns:
        tuple:
            - CacheNode: The deepest node in the cache tree that matches the token sequence
            - int: Number of tokens matched from the start of token_ids
            - DynamicCache|None: Past key/value states from the deepest cached node,
                or None if no cached states were found
            - int: Base index indicating where the past states start in token_ids
    """
    # Walk while tokens can be found
    node = self._caches[lora_name]
    next_token_index = 0

    past = None
    base = 0
    while next_token_index < len(token_ids):
        if node.past_key_values is not None:
            past = node.past_key_values
            base = next_token_index
        if node.has_token(token_ids[next_token_index]):
            node = node.get_token(token_ids[next_token_index])
            next_token_index += 1
        else:
            break

    return node, next_token_index, past, base

next_token_logprobs(token_ids, lora_name=None) async

Request log probabilities of next token. This version is asynchronous because it automatically batches concurrent requests; use with await.

Parameters:

Name	Type	Description	Default
token_ids	list[int]	a list of token ids, representing a prompt to the language model.	required
lora_name	str	Name of the LoRA adapter to use. Defaults to None (the base model).	None

Returns:

Name	Type	Description
logprobs	Tensor	a tensor of with the language model's log (normalized) probabilities for the next token following the prompt.

Source code in genlm/backend/llm/hf.py

@torch.no_grad()
async def next_token_logprobs(self, token_ids, lora_name=None):
    """Request log probabilities of next token. This version is asynchronous because it automatically batches concurrent requests; use with `await`.

    Args:
        token_ids (list[int]): a list of token ids, representing a prompt to the language model.
        lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

    Returns:
        logprobs (torch.Tensor): a tensor of with the language model's log (normalized) probabilities for the next token following the prompt.
    """
    if not token_ids:
        raise ValueError("Token ids must not be empty")

    node, next_token_index, past, base = self.walk_cache(token_ids, lora_name)

    # If we processed all tokens, then we're done.
    if next_token_index == len(token_ids):
        return node.logprobs

    # Create a future with the prompt
    future = asyncio.get_running_loop().create_future()
    self.add_query(token_ids[base:], future, past, lora_name)
    logits = await future

    # Create new nodes
    node = node.extend_cache(next_token_index, token_ids, logits, base)

    return node.logprobs

next_token_logprobs_sync(token_ids, lora_name=None)

Request log probabilities of next token. Not asynchronous, and does not support auto-batching.

Parameters:

Name	Type	Description	Default
token_ids	list[int]	a list of token ids, representing a prompt to the language model.	required
lora_name	str	Name of the LoRA adapter to use. Defaults to None (the base model).	None

Returns:

Name	Type	Description
logprobs	Tensor	a tensor with the language model's log (normalized) probabilities for the next token following the prompt.

Source code in genlm/backend/llm/hf.py

@torch.no_grad()
def next_token_logprobs_sync(self, token_ids, lora_name=None):
    """Request log probabilities of next token. Not asynchronous, and does not support auto-batching.

    Args:
        token_ids (list[int]): a list of token ids, representing a prompt to the language model.
        lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

    Returns:
        logprobs (torch.Tensor): a tensor with the language model's log (normalized) probabilities for the next token following the prompt.
    """
    if not token_ids:
        raise ValueError("Token ids must not be empty")

    # Walk while tokens can be found
    node, next_token_index, past, base = self.walk_cache(token_ids, lora_name)

    if next_token_index == len(token_ids):
        return node.logprobs

    self._activate(lora_name)
    logits = self.model(
        torch.tensor([token_ids[base:]]).to(self.device),
        past_key_values=node.past_key_values,
        use_cache=node.past_key_values is not None,
    ).logits[0]

    node = node.extend_cache(next_token_index, token_ids, logits, base)

    return node.logprobs

next_token_logprobs_uncached(token_ids, lora_name=None)

Request log probabilities of next token. No KV or output caching, and does not support auto-batching.

Parameters:

Name	Type	Description	Default
token_ids	list[int]	a list of token ids, representing a prompt to the language model.	required
lora_name	str	Name of the LoRA adapter to use. Defaults to None (the base model).	None

Returns:

Name	Type	Description
logprobs	Tensor	a tensor with the language model's log (normalized) probabilities for the next token following the prompt.

Source code in genlm/backend/llm/hf.py

def next_token_logprobs_uncached(self, token_ids, lora_name=None):
    """Request log probabilities of next token. No KV or output caching, and does not support auto-batching.

    Args:
        token_ids (list[int]): a list of token ids, representing a prompt to the language model.
        lora_name (str, optional): Name of the LoRA adapter to use. Defaults to None (the base model).

    Returns:
        logprobs (torch.Tensor): a tensor with the language model's log (normalized) probabilities for the next token following the prompt.
    """
    if not token_ids:
        raise ValueError("Token ids must not be empty")

    with torch.no_grad():
        self._activate(lora_name)
        logits = self.model(
            torch.tensor([token_ids]).to(self.device),
            past_key_values=None,
            use_cache=False,
        ).logits[0]
        return torch.log_softmax(logits[-1], dim=0, dtype=torch.float32)

load_model_by_name(name, backend=None, llm_opts=None)

Load a language model by name.

Parameters:

Name	Type	Description	Default
name	str	Hugging Face model name (e.g. "openai-community/gpt2", "meta-llama/Llama-3.2-1B-Instruct")	required
backend	str	Backend to use for inference. Can be "vllm", "hf", "mlx", "sgl", or "mock". If None, defaults to "vllm" if CUDA is available, otherwise "hf".	None
llm_opts	dict	Additional options to pass to the backend constructor. See AsyncVirtualLM and AsyncTransformer documentation for details.	None

Returns:

Type	Description
AsyncLM	An asynchronous language model.

Raises:

Type	Description
ValueError	If an invalid backend is specified.

Source code in genlm/backend/llm/__init__.py

def load_model_by_name(name, backend=None, llm_opts=None):
    """Load a language model by name.

    Args:
        name (str): Hugging Face model name (e.g. "openai-community/gpt2", "meta-llama/Llama-3.2-1B-Instruct")
        backend (str, optional): Backend to use for inference. Can be "vllm", "hf", "mlx", "sgl", or "mock".
            If None, defaults to "vllm" if CUDA is available, otherwise "hf".
        llm_opts (dict, optional): Additional options to pass to the backend constructor.
            See AsyncVirtualLM and AsyncTransformer documentation for details.

    Returns:
        (AsyncLM): An asynchronous language model.

    Raises:
        (ValueError): If an invalid backend is specified.
    """
    if backend is None:
        backend = "vllm" if torch.cuda.is_available() else "hf"

    if llm_opts is None:
        llm_opts = {}

    if backend == "vllm":
        return AsyncVirtualLM.from_name(name, **llm_opts)
    elif backend == "hf":
        return AsyncTransformer.from_name(name, **llm_opts)
    elif backend == "mock":
        return MockAsyncLM.from_name(name, **llm_opts)
    elif backend == "mlx":
        return AsyncMlxLM.from_name(name, **llm_opts)
    elif backend == "sgl":
        return AsyncSGLTransformer.from_name(name, **llm_opts)
    else:
        raise ValueError(f"Invalid backend: {backend}")

decode_vocab(tokenizer, byte2str_fallback='tokenizer')

Convert tokenizer vocabulary into byte and string representations.

Warning

The byte representation is the canonical form. Each element in byte_vocab is a Token object that contains both the token_id and byte_string. The string representation is provided for convenience but may not decode properly for all tokens, especially those containing invalid UTF-8 sequences.

Parameters:

Name	Type	Description	Default
tokenizer		A Hugging Face tokenizer instance	required
byte2str_fallback	str	Strategy for converting invalid UTF-8 bytes to strings. Options: 'tokenizer': Use tokenizer's convert_ids_to_tokens (default) 'latin1': Decode using latin1 encoding 'replace': Use Unicode replacement character '�'	'tokenizer'

Returns:

Type	Description
tuple	(byte_vocab, str_vocab) where byte_vocab is a list of Token objects and str_vocab is a list of strings

Source code in genlm/backend/tokenization/vocab.py

def decode_vocab(tokenizer, byte2str_fallback="tokenizer"):
    """Convert tokenizer vocabulary into byte and string representations.

    Warning:
        The byte representation is the canonical form. Each element in byte_vocab is a Token object that
        contains both the token_id and byte_string. The string representation is provided for convenience
        but may not decode properly for all tokens, especially those containing invalid UTF-8 sequences.

    Args:
        tokenizer: A Hugging Face tokenizer instance
        byte2str_fallback (str): Strategy for converting invalid UTF-8 bytes to strings. Options:\n
            - 'tokenizer': Use tokenizer's `convert_ids_to_tokens` (default)
            - 'latin1': Decode using latin1 encoding
            - 'replace': Use Unicode replacement character '�'

    Returns:
        (tuple): (byte_vocab, str_vocab) where byte_vocab is a list of Token objects
            and str_vocab is a list of strings
    """
    if byte2str_fallback not in ["latin1", "tokenizer", "replace"]:
        raise ValueError(f"Unknown byte2str_fallback strategy: {byte2str_fallback}")

    try:
        raw_byte_vocab = get_byte_vocab(tokenizer)
    except ByteVocabError as e:
        raise ValueError(
            f"Could not decode byte representation of token vocabuary from tokenizer {tokenizer.name_or_path}"
        ) from e

    # Create Token objects for byte_vocab.
    # Assumption: token_id == position index in the vocabulary. This is relied upon
    # by the trie (idx_to_leaf) and weight arrays (ws[i] corresponds to decode[i]).
    byte_vocab = [
        Token(token_id=i, byte_string=b) for i, b in enumerate(raw_byte_vocab)
    ]
    str_vocab = bytes_to_strs(tokenizer, raw_byte_vocab, byte2str_fallback)

    return byte_vocab, str_vocab

Token

Bases: bytes

A vocabulary token carrying both a token ID and its byte representation.

Subclasses bytes so that existing code using byte operations (b"".join, len, indexing, .decode()) continues to work. Equality and hashing between Token objects use token_id, not byte content.

Parameters:

Name	Type	Description	Default
token_id	int	The unique identifier for this token in the vocabulary.	required
byte_string	bytes	The byte representation of this token.	required

Source code in genlm/backend/tokenization/token.py

class Token(bytes):
    """A vocabulary token carrying both a token ID and its byte representation.

    Subclasses ``bytes`` so that existing code using byte operations (``b"".join``,
    ``len``, indexing, ``.decode()``) continues to work. Equality and hashing
    between Token objects use ``token_id``, not byte content.

    Args:
        token_id (int): The unique identifier for this token in the vocabulary.
        byte_string (bytes): The byte representation of this token.
    """

    def __new__(cls, token_id: int, byte_string: bytes):
        if not isinstance(token_id, int):
            raise TypeError(f"token_id must be an int, got {type(token_id)}")
        if not isinstance(byte_string, bytes):
            raise TypeError(f"byte_string must be bytes, got {type(byte_string)}")
        obj = super().__new__(cls, byte_string)
        obj.token_id = token_id
        return obj

    @property
    def byte_string(self):
        """The byte representation of this token (as plain bytes)."""
        return bytes(self)

    def __repr__(self):
        return f"Token(token_id={self.token_id}, byte_string={bytes(self)!r})"

    # -- Equality / hashing: by token_id between Tokens, by content vs bytes --

    def __eq__(self, other):
        if isinstance(other, Token):
            return self.token_id == other.token_id
        return NotImplemented

    def __ne__(self, other):
        if isinstance(other, Token):
            return self.token_id != other.token_id
        return NotImplemented

    def __hash__(self):
        return hash(self.token_id)

    # -- Ordering: by token_id --

    def __lt__(self, other):
        if not isinstance(other, Token):
            return NotImplemented
        return self.token_id < other.token_id

    def __le__(self, other):
        if not isinstance(other, Token):
            return NotImplemented
        return self.token_id <= other.token_id

    def __gt__(self, other):
        if not isinstance(other, Token):
            return NotImplemented
        return self.token_id > other.token_id

    def __ge__(self, other):
        if not isinstance(other, Token):
            return NotImplemented
        return self.token_id >= other.token_id

    # -- Helpers --

    @staticmethod
    def as_bytes(x):
        """Extract byte string from a Token or pass through plain bytes."""
        return x.byte_string if isinstance(x, Token) else x

    @staticmethod
    def is_plain_bytes(x):
        """Check if x is plain bytes (not a Token)."""
        return isinstance(x, bytes) and not isinstance(x, Token)

    # -- Pickle / deepcopy support --

    def __reduce__(self):
        return (Token, (self.token_id, bytes(self)))

byte_string property

The byte representation of this token (as plain bytes).

as_bytes(x) staticmethod

Extract byte string from a Token or pass through plain bytes.

Source code in genlm/backend/tokenization/token.py

@staticmethod
def as_bytes(x):
    """Extract byte string from a Token or pass through plain bytes."""
    return x.byte_string if isinstance(x, Token) else x

is_plain_bytes(x) staticmethod

Check if x is plain bytes (not a Token).

Source code in genlm/backend/tokenization/token.py

@staticmethod
def is_plain_bytes(x):
    """Check if x is plain bytes (not a Token)."""
    return isinstance(x, bytes) and not isinstance(x, Token)

TokenCharacterTrie

A trie data structure for efficient token-to-character mapping.

Source code in genlm/backend/trie/base.py

class TokenCharacterTrie:
    """A trie data structure for efficient token-to-character mapping."""

    def __init__(self, decode):
        """Initialize a `TokenCharacterTrie`.

        Args:
            decode (list): List representing the token vocabulary.
                Each element must be iterable. Token objects use byte_string for iteration,
                other iterables (bytes, EndOfSequence) are iterated directly.
        """
        self.decode = decode
        self.word2leaf = {}
        self.children = [{}]  # First node is root
        self.root = 0
        # Maps position index in decode to leaf node: idx_to_leaf[k] = (idx, leaf_node)
        self.idx_to_leaf = []

        _warned_bytes = False
        for idx, item in enumerate(self.decode):
            # Get the word (bytes to iterate) and a unique key for word2leaf
            if isinstance(item, Token):
                word = item.byte_string
                word_key = (item.byte_string, item.token_id)
            elif Token.is_plain_bytes(item):
                if not _warned_bytes:
                    warnings.warn(
                        "Passing plain bytes to TokenCharacterTrie is deprecated. "
                        "Use Token objects from decode_vocab() instead.",
                        DeprecationWarning,
                        stacklevel=2,
                    )
                    _warned_bytes = True
                word = item
                word_key = item
            else:
                # For other iterables (e.g. EndOfSequence), iterate directly
                word = item
                word_key = item

            curr = self.root
            for letter in word:
                if letter not in self.children[curr]:
                    self.children[curr][letter] = len(self.children)
                    self.children.append({})
                curr = self.children[curr][letter]

            # Each item gets its own leaf
            leaf_edge_key = (None, idx)  # Use position index for uniqueness

            self.children[curr][leaf_edge_key] = last = len(self.children)
            self.children.append({})

            if word_key in self.word2leaf:
                raise ValueError(f"Duplicate word in vocabulary: {word_key}")
            self.word2leaf[word_key] = last

            # Use position index for weight array indexing
            self.idx_to_leaf.append((idx, last))

        self.leaf2word = dict(zip(self.word2leaf.values(), self.word2leaf.keys()))
        self.jump = List(
            [np.array(sorted(x.values()), dtype=np.int32) for x in self.children]
        )
        self.ordering = np.array(list(self._order(self.root)), np.int32)

        # Renumber the states of the trie so that they are named by a contiguous
        # range of integers and those integers respect the are topologically
        # ordering of the trie topology.  This improves the efficiency of the
        # updating the trie as it improves memory locality.
        ordering = {}
        for i, x in enumerate(self._order_full(self.root)):
            ordering[x] = i
        self._rename(f=lambda x: ordering[x])

        node2prefix = {self.root: []}
        for x in reversed(range(len(self.children))):
            for letter, y in self.children[x].items():
                # Leaf edges are tuples of (None, token_id)
                if isinstance(letter, tuple) and letter[0] is None:
                    node2prefix[y] = node2prefix[x]
                else:
                    node2prefix[y] = node2prefix[x] + [letter]
        self.node2prefix = node2prefix

    def _rename(self, f):
        """Rename all node indices in the trie using the provided mapping function.

        Args:
            f (callable): Function that maps old node indices to new node indices
        """
        N = len(self.children)

        new_children = [{} for _ in range(N)]
        nodes = range(N)

        for x in nodes:
            for letter, y in self.children[x].items():
                new_children[f(x)][letter] = f(y)

        self.root = f(self.root)
        self.children = new_children
        self.word2leaf = {w: f(x) for w, x in self.word2leaf.items()}
        self.leaf2word = dict(zip(self.word2leaf.values(), self.word2leaf.keys()))

        self.idx_to_leaf = np.array(
            [(i, f(x)) for i, x in self.idx_to_leaf], dtype=np.int32
        )

        self.ordering = np.array([f(x) for x in self.ordering])
        self.jump = List(
            [np.array(sorted(x.values()), dtype=np.int32) for x in new_children]
        )

    def _alloc_weights(self):
        """Allocate an array to store weight values for all nodes.

        Returns:
            np.ndarray: Zero-initialized array for storing weight values
        """
        return np.zeros(len(self.children), dtype=np.float64)

    def _preprocess_ws(self, ws):
        """Preprocess the weight vector to ensure it is a numpy array and on the correct device.

        Args:
            ws (torch.Tensor|np.ndarray): Token weights over the vocabulary of shape `(len(self.decode),)`

        Returns:
            (np.ndarray): Weight vector
        """
        if isinstance(ws, torch.Tensor):
            if ws.device.type != "cpu":
                ws = ws.cpu()
            ws = ws.numpy()
        return ws

    def weight_sum(self, ws):
        """Compute weight sum for each node in the trie.

        For each node in the trie, this computes the sum of weights of all leaf nodes (tokens)
        that are descendants of that node.

        Args:
            ws (torch.Tensor|np.ndarray): Token weights indexed by position in `self.decode`,
                i.e., `ws[i]` is the weight for `self.decode[i]`. Shape: `(len(self.decode),)`

        Returns:
            (np.ndarray): Summed weights for each node in the trie.
        """
        ws = self._preprocess_ws(ws)
        node_ws = self._alloc_weights()
        _update_trie_numba_sum(
            node_ws=node_ws,
            ws=ws,
            idx_to_leaf=self.idx_to_leaf,
            jump=self.jump,
            ordering=self.ordering,
        )
        return node_ws

    def weight_max(self, ws):
        """Compute weight max for each node in the trie.

        For each node in the trie, this computes the maximum weight among all leaf nodes (tokens)
        that are descendants of that node.

        Args:
            ws (torch.Tensor|np.ndarray): Token weights indexed by position in `self.decode`,
                i.e., `ws[i]` is the weight for `self.decode[i]`. Shape: `(len(self.decode),)`

        Returns:
            (np.ndarray): Weight max values for each node in the trie.
        """
        ws = self._preprocess_ws(ws)
        node_ws = self._alloc_weights()
        _update_trie_numba_max(
            node_ws=node_ws,
            ws=ws,
            idx_to_leaf=self.idx_to_leaf,
            jump=self.jump,
            ordering=self.ordering,
        )
        return node_ws

    def batch_weight_sum(self, ws):
        """Batched equivalent of `weight_sum`.

        Args:
            ws (list[torch.Tensor|np.ndarray]): Batch of token weights, each indexed by
                position in `self.decode`. Shape of each: `(len(self.decode),)`

        Returns:
            (np.ndarray): Batch of weight values of `len(ws)` for each node in the trie
        """
        return np.array([self.weight_sum(w) for w in ws])

    def batch_weight_max(self, ws):
        """Batched equivalent of `weight_max`.

        Args:
            ws (list[torch.Tensor|np.ndarray]): Batch of token weights, each indexed by
                position in `self.decode`. Shape of each: `(len(self.decode),)`

        Returns:
            (np.ndarray): Batch of weight max values of `len(ws)` for each node in the trie
        """
        return np.array([self.weight_max(w) for w in ws])

    def _order(self, node):
        """Generate a topological ordering of nodes beneath the given node.

        Args:
            node (int): Starting node index

        Yields:
            int: Node indices in topological order
        """
        for a in self.children[node]:
            # Skip leaf edges (tuples of (None, token_id))
            if isinstance(a, tuple) and a[0] is None:
                pass
            else:
                yield from self._order(self.children[node][a])
        yield node

    def _order_full(self, node):
        """Generate a complete topological ordering including all child nodes.

        Args:
            node (int): Starting node index

        Yields:
            (int): Node indices in complete topological order
        """
        for a in self.children[node]:
            yield from self._order_full(self.children[node][a])
        yield node

    def visualize(self, ws=None):
        """Visualize the trie structure using Graphviz.

        Args:
            ws (np.ndarray|None): Optional weight vector to display at each node.
                                Should be of length `len(self.children)`.

        Returns:
            (graphviz.Digraph): The generated graph object
        """
        try:
            import graphviz
        except ImportError:  # pragma: no cover
            raise ImportError(
                "Please install graphviz: pip install graphviz"
            )  # pragma: no cover

        if ws is not None and len(ws) != len(self.children):
            raise ValueError(
                f"Weight vector length ({len(ws)}) must match number of nodes ({len(self.children)})"
            )

        dot = graphviz.Digraph(comment="Token Character Trie")
        dot.attr(rankdir="LR")

        # Create a subgraph for the legend
        with dot.subgraph(name="cluster_legend") as legend:
            legend.attr(label="Legend", fontsize="10")
            legend.attr("node", fontsize="7", width="0.1", height="0.1")

            # Example internal node
            legend.node(
                "legend_internal",
                "Internal Node ID\n'Prefix'\nWeight (if provided)",
                shape="circle",
            )

            # Example leaf node
            legend.node("legend_leaf", "Complete Token", shape="doublecircle")

            legend.edge(
                "legend_internal",
                "legend_leaf",
                label="Token item",
                fontsize="10",
            )

            # Align legend horizontally
            legend.attr(rankdir="TB")
            legend.attr(rank="same")

        # Add the main trie nodes and edges
        for node_id in range(len(self.children)):
            prefix = self.node2prefix[node_id]

            if ws is not None:
                label = f"{node_id}\n'{prefix}'\n{ws[node_id]:.4f}"
            else:
                label = f"{node_id}\n'{prefix}'"

            # Color nodes based on mass if provided
            if ws is not None:
                max_ws = ws.max()
                if max_ws > 0:
                    intensity = int(255 * (1 - ws[node_id] / max_ws))
                    color = f"#{intensity:02x}{255:02x}{intensity:02x}"
                else:
                    color = "#ffffff"  # white for zero mass
            else:
                color = "#ffffff"  # default white

            if node_id in self.leaf2word:
                dot.node(
                    str(node_id),
                    label,
                    shape="doublecircle",
                    style="filled",
                    fillcolor=color,
                )
            else:
                dot.node(
                    str(node_id), label, shape="circle", style="filled", fillcolor=color
                )

        for node_id, children in enumerate(self.children):
            for char, child_id in children.items():
                # Leaf edges are tuples of (None, token_id)
                if isinstance(char, tuple) and char[0] is None:
                    edge_label = f"End-of-Token (ID: {char[1]})"
                else:
                    edge_label = str(char)

                dot.edge(str(node_id), str(child_id), label=edge_label)

        return dot

__init__(decode)

Initialize a TokenCharacterTrie.

Parameters:

Name	Type	Description	Default
decode	list	List representing the token vocabulary. Each element must be iterable. Token objects use byte_string for iteration, other iterables (bytes, EndOfSequence) are iterated directly.	required

Source code in genlm/backend/trie/base.py

def __init__(self, decode):
    """Initialize a `TokenCharacterTrie`.

    Args:
        decode (list): List representing the token vocabulary.
            Each element must be iterable. Token objects use byte_string for iteration,
            other iterables (bytes, EndOfSequence) are iterated directly.
    """
    self.decode = decode
    self.word2leaf = {}
    self.children = [{}]  # First node is root
    self.root = 0
    # Maps position index in decode to leaf node: idx_to_leaf[k] = (idx, leaf_node)
    self.idx_to_leaf = []

    _warned_bytes = False
    for idx, item in enumerate(self.decode):
        # Get the word (bytes to iterate) and a unique key for word2leaf
        if isinstance(item, Token):
            word = item.byte_string
            word_key = (item.byte_string, item.token_id)
        elif Token.is_plain_bytes(item):
            if not _warned_bytes:
                warnings.warn(
                    "Passing plain bytes to TokenCharacterTrie is deprecated. "
                    "Use Token objects from decode_vocab() instead.",
                    DeprecationWarning,
                    stacklevel=2,
                )
                _warned_bytes = True
            word = item
            word_key = item
        else:
            # For other iterables (e.g. EndOfSequence), iterate directly
            word = item
            word_key = item

        curr = self.root
        for letter in word:
            if letter not in self.children[curr]:
                self.children[curr][letter] = len(self.children)
                self.children.append({})
            curr = self.children[curr][letter]

        # Each item gets its own leaf
        leaf_edge_key = (None, idx)  # Use position index for uniqueness

        self.children[curr][leaf_edge_key] = last = len(self.children)
        self.children.append({})

        if word_key in self.word2leaf:
            raise ValueError(f"Duplicate word in vocabulary: {word_key}")
        self.word2leaf[word_key] = last

        # Use position index for weight array indexing
        self.idx_to_leaf.append((idx, last))

    self.leaf2word = dict(zip(self.word2leaf.values(), self.word2leaf.keys()))
    self.jump = List(
        [np.array(sorted(x.values()), dtype=np.int32) for x in self.children]
    )
    self.ordering = np.array(list(self._order(self.root)), np.int32)

    # Renumber the states of the trie so that they are named by a contiguous
    # range of integers and those integers respect the are topologically
    # ordering of the trie topology.  This improves the efficiency of the
    # updating the trie as it improves memory locality.
    ordering = {}
    for i, x in enumerate(self._order_full(self.root)):
        ordering[x] = i
    self._rename(f=lambda x: ordering[x])

    node2prefix = {self.root: []}
    for x in reversed(range(len(self.children))):
        for letter, y in self.children[x].items():
            # Leaf edges are tuples of (None, token_id)
            if isinstance(letter, tuple) and letter[0] is None:
                node2prefix[y] = node2prefix[x]
            else:
                node2prefix[y] = node2prefix[x] + [letter]
    self.node2prefix = node2prefix

weight_sum(ws)

Compute weight sum for each node in the trie.

For each node in the trie, this computes the sum of weights of all leaf nodes (tokens) that are descendants of that node.

Parameters:

Name	Type	Description	Default
ws	Tensor | ndarray	Token weights indexed by position in self.decode, i.e., ws[i] is the weight for self.decode[i]. Shape: (len(self.decode),)	required

Returns:

Type	Description
ndarray	Summed weights for each node in the trie.

Source code in genlm/backend/trie/base.py

def weight_sum(self, ws):
    """Compute weight sum for each node in the trie.

    For each node in the trie, this computes the sum of weights of all leaf nodes (tokens)
    that are descendants of that node.

    Args:
        ws (torch.Tensor|np.ndarray): Token weights indexed by position in `self.decode`,
            i.e., `ws[i]` is the weight for `self.decode[i]`. Shape: `(len(self.decode),)`

    Returns:
        (np.ndarray): Summed weights for each node in the trie.
    """
    ws = self._preprocess_ws(ws)
    node_ws = self._alloc_weights()
    _update_trie_numba_sum(
        node_ws=node_ws,
        ws=ws,
        idx_to_leaf=self.idx_to_leaf,
        jump=self.jump,
        ordering=self.ordering,
    )
    return node_ws

weight_max(ws)

Compute weight max for each node in the trie.

For each node in the trie, this computes the maximum weight among all leaf nodes (tokens) that are descendants of that node.

Parameters:

Name	Type	Description	Default
ws	Tensor | ndarray	Token weights indexed by position in self.decode, i.e., ws[i] is the weight for self.decode[i]. Shape: (len(self.decode),)	required

Returns:

Type	Description
ndarray	Weight max values for each node in the trie.

Source code in genlm/backend/trie/base.py

def weight_max(self, ws):
    """Compute weight max for each node in the trie.

    For each node in the trie, this computes the maximum weight among all leaf nodes (tokens)
    that are descendants of that node.

    Args:
        ws (torch.Tensor|np.ndarray): Token weights indexed by position in `self.decode`,
            i.e., `ws[i]` is the weight for `self.decode[i]`. Shape: `(len(self.decode),)`

    Returns:
        (np.ndarray): Weight max values for each node in the trie.
    """
    ws = self._preprocess_ws(ws)
    node_ws = self._alloc_weights()
    _update_trie_numba_max(
        node_ws=node_ws,
        ws=ws,
        idx_to_leaf=self.idx_to_leaf,
        jump=self.jump,
        ordering=self.ordering,
    )
    return node_ws

batch_weight_sum(ws)

Batched equivalent of weight_sum.

Parameters:

Name	Type	Description	Default
ws	list[Tensor | ndarray]	Batch of token weights, each indexed by position in self.decode. Shape of each: (len(self.decode),)	required

Returns:

Type	Description
ndarray	Batch of weight values of len(ws) for each node in the trie

Source code in genlm/backend/trie/base.py

def batch_weight_sum(self, ws):
    """Batched equivalent of `weight_sum`.

    Args:
        ws (list[torch.Tensor|np.ndarray]): Batch of token weights, each indexed by
            position in `self.decode`. Shape of each: `(len(self.decode),)`

    Returns:
        (np.ndarray): Batch of weight values of `len(ws)` for each node in the trie
    """
    return np.array([self.weight_sum(w) for w in ws])

batch_weight_max(ws)

Batched equivalent of weight_max.

Parameters:

Name	Type	Description	Default
ws	list[Tensor | ndarray]	Batch of token weights, each indexed by position in self.decode. Shape of each: (len(self.decode),)	required

Returns:

Type	Description
ndarray	Batch of weight max values of len(ws) for each node in the trie

Source code in genlm/backend/trie/base.py

def batch_weight_max(self, ws):
    """Batched equivalent of `weight_max`.

    Args:
        ws (list[torch.Tensor|np.ndarray]): Batch of token weights, each indexed by
            position in `self.decode`. Shape of each: `(len(self.decode),)`

    Returns:
        (np.ndarray): Batch of weight max values of `len(ws)` for each node in the trie
    """
    return np.array([self.weight_max(w) for w in ws])

visualize(ws=None)

Visualize the trie structure using Graphviz.

Parameters:

Name	Type	Description	Default
ws	ndarray | None	Optional weight vector to display at each node. Should be of length len(self.children).	None

Returns:

Type	Description
Digraph	The generated graph object

Source code in genlm/backend/trie/base.py

def visualize(self, ws=None):
    """Visualize the trie structure using Graphviz.

    Args:
        ws (np.ndarray|None): Optional weight vector to display at each node.
                            Should be of length `len(self.children)`.

    Returns:
        (graphviz.Digraph): The generated graph object
    """
    try:
        import graphviz
    except ImportError:  # pragma: no cover
        raise ImportError(
            "Please install graphviz: pip install graphviz"
        )  # pragma: no cover

    if ws is not None and len(ws) != len(self.children):
        raise ValueError(
            f"Weight vector length ({len(ws)}) must match number of nodes ({len(self.children)})"
        )

    dot = graphviz.Digraph(comment="Token Character Trie")
    dot.attr(rankdir="LR")

    # Create a subgraph for the legend
    with dot.subgraph(name="cluster_legend") as legend:
        legend.attr(label="Legend", fontsize="10")
        legend.attr("node", fontsize="7", width="0.1", height="0.1")

        # Example internal node
        legend.node(
            "legend_internal",
            "Internal Node ID\n'Prefix'\nWeight (if provided)",
            shape="circle",
        )

        # Example leaf node
        legend.node("legend_leaf", "Complete Token", shape="doublecircle")

        legend.edge(
            "legend_internal",
            "legend_leaf",
            label="Token item",
            fontsize="10",
        )

        # Align legend horizontally
        legend.attr(rankdir="TB")
        legend.attr(rank="same")

    # Add the main trie nodes and edges
    for node_id in range(len(self.children)):
        prefix = self.node2prefix[node_id]

        if ws is not None:
            label = f"{node_id}\n'{prefix}'\n{ws[node_id]:.4f}"
        else:
            label = f"{node_id}\n'{prefix}'"

        # Color nodes based on mass if provided
        if ws is not None:
            max_ws = ws.max()
            if max_ws > 0:
                intensity = int(255 * (1 - ws[node_id] / max_ws))
                color = f"#{intensity:02x}{255:02x}{intensity:02x}"
            else:
                color = "#ffffff"  # white for zero mass
        else:
            color = "#ffffff"  # default white

        if node_id in self.leaf2word:
            dot.node(
                str(node_id),
                label,
                shape="doublecircle",
                style="filled",
                fillcolor=color,
            )
        else:
            dot.node(
                str(node_id), label, shape="circle", style="filled", fillcolor=color
            )

    for node_id, children in enumerate(self.children):
        for char, child_id in children.items():
            # Leaf edges are tuples of (None, token_id)
            if isinstance(char, tuple) and char[0] is None:
                edge_label = f"End-of-Token (ID: {char[1]})"
            else:
                edge_label = str(char)

            dot.edge(str(node_id), str(child_id), label=edge_label)

    return dot

ParallelTokenCharacterTrie

Bases: TokenCharacterTrie

A GPU-optimized version of TokenCharacterTrie that performs weight sum and max operations in parallel.

Source code in genlm/backend/trie/parallel.py

class ParallelTokenCharacterTrie(TokenCharacterTrie):
    """A GPU-optimized version of `TokenCharacterTrie` that performs weight sum and max operations in parallel."""

    def __init__(self, decode, device=None, **kwargs):
        super().__init__(decode, **kwargs)

        self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
        if self.device not in ["cpu", "cuda"]:
            raise ValueError(f"Invalid device: {device}. Must be 'cpu', 'cuda' or None")

        self._build_reachability_matrix()
        # Position indices for weight array indexing
        self.positions = torch.tensor(
            self.idx_to_leaf[:, 0], dtype=torch.long, device=self.device
        )

    def _build_parent_map(self):
        """Builds a mapping from each node to its parent node in the trie.

        Returns:
            (dict): A dictionary where keys are child nodes and values are their parent nodes.
        """
        parent = {}
        for node in range(len(self.children)):
            for child in self.jump[node]:
                parent[child] = node
        return parent

    def _build_reachability_matrix(self):
        """Constructs a sparse reachability matrix for efficient weight propagation.

        The matrix M is constructed such that M[i,j] = 1 if node j is either:
        - The leaf node i itself (self-connection)
        - An ancestor of leaf node i in the trie
        """
        leaf_indices = self.idx_to_leaf[:, 1]
        parent = self._build_parent_map()

        rows, cols = [], []
        for i, node in enumerate(leaf_indices):
            # self connections
            rows.append(i)
            cols.append(node)

            current = node
            while current in parent:  # Walk up to root
                ancestor = parent[current]
                rows.append(i)
                cols.append(ancestor)
                current = ancestor

        self.src_indices = torch.tensor(rows, dtype=torch.long, device=self.device)
        self.dst_indices = torch.tensor(cols, dtype=torch.long, device=self.device)

        indices = torch.tensor([rows, cols], dtype=torch.long, device=self.device)
        values = torch.ones(len(rows), device=self.device)

        self.M = torch.sparse_coo_tensor(
            indices, values, (len(leaf_indices), len(self.children))
        ).to_sparse_csr()

    def _preprocess_ws(self, batch_ws):
        processed_batch_ws = []
        for ws in batch_ws:
            if not isinstance(ws, torch.Tensor):
                ws = torch.tensor(ws, device=self.device, dtype=torch.float32)
            elif ws.device != self.device or ws.dtype != torch.float32:
                ws = ws.to(device=self.device, dtype=torch.float32)
            assert ws.shape[0] == len(self.decode), [ws.shape[0], len(self.decode)]
            processed_batch_ws.append(ws)
        return torch.stack(processed_batch_ws)

    def weight_sum(self, ws):
        """Computes weight sums given token weights.

        For each node in the trie, this computes the sum of weights of all leaf nodes (tokens)
        that are descendants of that node. This is efficiently implemented using sparse matrix multiplication
        with a pre-computed reachability matrix.

        Args:
            ws (torch.Tensor): Token weights, shape (`len(self.decode)`,).

        Returns:
            (numpy.ndarray): Summed weights for each node in the trie, shape (`len(self.decode)`,).
        """
        return self.batch_weight_sum(self._preprocess_ws([ws]))[0]

    def batch_weight_sum(self, ws):
        """Batch version of `weight_sum`.

        Args:
            ws (torch.Tensor): Batch of token weights, shape (batch_size × `len(self.decode)`).

        Returns:
            numpy.ndarray: Summed weights for each node in the trie, shape (batch_size × num_nodes).
        """
        ws = self._preprocess_ws(ws)
        masses = torch.sparse.mm(ws[:, self.positions], self.M)
        return masses.cpu().numpy()

    def weight_max(self, ws):
        """Computes the max weights given the token weights.

        For each node in the trie, this computes the maximum weight among all leaf nodes (tokens)
        that are descendants of that node. This is efficiently implemented using parallel scatter_reduce
        operations on GPU.

        Args:
            ws (torch.Tensor): Token weights, shape (`len(self.decode)`,).

        Returns:
            (numpy.ndarray): Maximum weights for each node in the trie, shape (`len(self.decode)`,).
        """
        return self.batch_weight_max(self._preprocess_ws([ws]))[0]

    def batch_weight_max(self, ws):
        """Batch version of `weight_max`.

        Args:
            ws (torch.Tensor): Batch of token weights, shape (batch_size × `len(self.decode)`).

        Returns:
            (numpy.ndarray): Maximum weights for each node in the trie, shape (batch_size × num_nodes).
        """
        ws = self._preprocess_ws(ws)

        # Get leaf weights (indexed by position in decode)
        leaf_weights = ws[:, self.positions]  # shape: (batch_size × num_leafs)
        batch_size = leaf_weights.shape[0]

        # Use scatter_reduce to propagate maximum values in parallel
        result = torch.zeros((batch_size, len(self.children)), device=self.device)
        result.scatter_reduce_(
            dim=1,
            index=self.dst_indices.expand(batch_size, -1),
            src=leaf_weights[:, self.src_indices],
            reduce="amax",
            include_self=False,
        )

        return result.cpu().numpy()

weight_sum(ws)

Computes weight sums given token weights.

For each node in the trie, this computes the sum of weights of all leaf nodes (tokens) that are descendants of that node. This is efficiently implemented using sparse matrix multiplication with a pre-computed reachability matrix.

Parameters:

Name	Type	Description	Default
ws	Tensor	Token weights, shape (len(self.decode),).	required

Returns:

Type	Description
ndarray	Summed weights for each node in the trie, shape (len(self.decode),).

Source code in genlm/backend/trie/parallel.py

def weight_sum(self, ws):
    """Computes weight sums given token weights.

    For each node in the trie, this computes the sum of weights of all leaf nodes (tokens)
    that are descendants of that node. This is efficiently implemented using sparse matrix multiplication
    with a pre-computed reachability matrix.

    Args:
        ws (torch.Tensor): Token weights, shape (`len(self.decode)`,).

    Returns:
        (numpy.ndarray): Summed weights for each node in the trie, shape (`len(self.decode)`,).
    """
    return self.batch_weight_sum(self._preprocess_ws([ws]))[0]

batch_weight_sum(ws)

Batch version of weight_sum.

Parameters:

Name	Type	Description	Default
ws	Tensor	Batch of token weights, shape (batch_size × len(self.decode)).	required

Returns:

Type	Description
	numpy.ndarray: Summed weights for each node in the trie, shape (batch_size × num_nodes).

Source code in genlm/backend/trie/parallel.py

def batch_weight_sum(self, ws):
    """Batch version of `weight_sum`.

    Args:
        ws (torch.Tensor): Batch of token weights, shape (batch_size × `len(self.decode)`).

    Returns:
        numpy.ndarray: Summed weights for each node in the trie, shape (batch_size × num_nodes).
    """
    ws = self._preprocess_ws(ws)
    masses = torch.sparse.mm(ws[:, self.positions], self.M)
    return masses.cpu().numpy()

weight_max(ws)

Computes the max weights given the token weights.

For each node in the trie, this computes the maximum weight among all leaf nodes (tokens) that are descendants of that node. This is efficiently implemented using parallel scatter_reduce operations on GPU.

Parameters:

Name	Type	Description	Default
ws	Tensor	Token weights, shape (len(self.decode),).	required

Returns:

Type	Description
ndarray	Maximum weights for each node in the trie, shape (len(self.decode),).

Source code in genlm/backend/trie/parallel.py

def weight_max(self, ws):
    """Computes the max weights given the token weights.

    For each node in the trie, this computes the maximum weight among all leaf nodes (tokens)
    that are descendants of that node. This is efficiently implemented using parallel scatter_reduce
    operations on GPU.

    Args:
        ws (torch.Tensor): Token weights, shape (`len(self.decode)`,).

    Returns:
        (numpy.ndarray): Maximum weights for each node in the trie, shape (`len(self.decode)`,).
    """
    return self.batch_weight_max(self._preprocess_ws([ws]))[0]

batch_weight_max(ws)

Batch version of weight_max.

Parameters:

Name	Type	Description	Default
ws	Tensor	Batch of token weights, shape (batch_size × len(self.decode)).	required

Returns:

Type	Description
ndarray	Maximum weights for each node in the trie, shape (batch_size × num_nodes).

Source code in genlm/backend/trie/parallel.py

def batch_weight_max(self, ws):
    """Batch version of `weight_max`.

    Args:
        ws (torch.Tensor): Batch of token weights, shape (batch_size × `len(self.decode)`).

    Returns:
        (numpy.ndarray): Maximum weights for each node in the trie, shape (batch_size × num_nodes).
    """
    ws = self._preprocess_ws(ws)

    # Get leaf weights (indexed by position in decode)
    leaf_weights = ws[:, self.positions]  # shape: (batch_size × num_leafs)
    batch_size = leaf_weights.shape[0]

    # Use scatter_reduce to propagate maximum values in parallel
    result = torch.zeros((batch_size, len(self.children)), device=self.device)
    result.scatter_reduce_(
        dim=1,
        index=self.dst_indices.expand(batch_size, -1),
        src=leaf_weights[:, self.src_indices],
        reduce="amax",
        include_self=False,
    )

    return result.cpu().numpy()

AsyncTokenCharacterTrie

An asynchronous wrapper for TokenCharacterTrie implementations that provides automatic request batching.

Source code in genlm/backend/trie/async_impl.py

class AsyncTokenCharacterTrie:
    """An asynchronous wrapper for TokenCharacterTrie implementations that provides automatic request batching."""

    def __init__(self, trie):
        """Initialize an `AsyncTokenCharacterTrie`.

        Args:
            trie (TokenCharacterTrie|ParallelTokenCharacterTrie): The underlying `TokenCharacterTrie` or `ParallelTokenCharacterTrie` instance
        """
        self.trie = trie
        self._queue = None
        self._task = None

    @classmethod
    def from_vocab(cls, vocab, backend="parallel", **kwargs):
        """Creates an `AsyncTokenCharacterTrie` from a vocabulary.

        Args:
            vocab (list[Token]): The vocabulary over which the trie will be defined.
            backend (str, optional): The trie implementation to use - either 'sequential' or 'parallel'.
                    Defaults to 'parallel' which uses GPU acceleration when available.
            **kwargs: Additional arguments passed to the trie constructor

        Returns:
            (AsyncTokenCharacterTrie): The initialized asynchronous trie instance.
        """
        if backend == "sequential":
            trie = TokenCharacterTrie(decode=vocab, **kwargs)
        elif backend == "parallel":
            trie = ParallelTokenCharacterTrie(decode=vocab, **kwargs)
        else:
            raise ValueError(
                f"Unknown backend: {backend}. Must be one of ['sequential', 'parallel']"
            )
        return cls(trie)

    async def _queue_request(self, request, op):
        if not self._task or self._task.done():
            self.start()

        future = asyncio.Future()
        await self._queue.put((request, future, op))
        return future

    async def weight_sum(self, ws):
        """Queue a `weight_sum` request. Multiple concurrent calls will be automatically batched
        together.

        Args:
            ws (torch.Tensor): Token weights, shape (`len(self.trie.decode)`,).

        Returns:
            (np.ndarray): The calculated mass sums for the given distribution.
        """
        future = await self._queue_request(ws, "sum")
        result = await future
        return result

    async def weight_max(self, ws):
        """Queue a `weight_max` request. Multiple concurrent calls will be automatically batched
        together.

        Args:
            ws (torch.Tensor): Token weights, shape (`len(self.trie.decode)`,).

        Returns:
            (np.ndarray): The calculated max weights for the given distribution.
        """
        future = await self._queue_request(ws, "max")
        result = await future
        return result

    def start(self):
        """Start the background processing task if not already running."""
        if not self._task or self._task.done():
            self._queue = (
                asyncio.Queue()
            )  # Create a new queue so that it is bound to the current event loop
            self._task = asyncio.create_task(self._background_loop())

    def _do_weight_sums(self, batch_weights):
        return self.trie.batch_weight_sum(batch_weights)

    def _do_weight_maxs(self, batch_weights):
        return self.trie.batch_weight_max(batch_weights)

    async def _background_loop(self):
        """Background task that processes queued weight sum and max requests.

        Continuously monitors the queue for new requests and processes them in batches
        using the underlying trie implementation.

        Raises:
            Exception: If any error occurs during processing, it is propagated to all
                      pending futures in the current batch.
        """
        while True:
            try:
                op_groups = defaultdict(list)

                request, future, op = await self._queue.get()
                op_groups[op].append((request, future))

                while not self._queue.empty():
                    request, future, op = await self._queue.get()
                    op_groups[op].append((request, future))

                for op, group in op_groups.items():
                    requests, futures = zip(*group)

                    if op == "sum":
                        logger.debug(f"processing {len(requests)} sum requests")
                        results = self._do_weight_sums(requests)
                    elif op == "max":
                        logger.debug(f"processing {len(requests)} max requests")
                        results = self._do_weight_maxs(requests)
                    else:
                        raise ValueError(f"Unknown operation: {op}")

                    for future, result in zip(futures, results):
                        future.set_result(result)

            except Exception as e:
                for group in op_groups.values():
                    for _, future in group:
                        if not future.done():
                            future.set_exception(e)
                raise

    async def cleanup(self):
        """Async cleanup - preferred method"""
        if self._task and not self._task.done():
            self._task.cancel()
            try:
                await self._task
            except asyncio.CancelledError:
                pass
            self._task = None

    def shutdown(self):
        """Stop the background processing task and cleanup resources."""
        if self._task is not None:
            try:
                self._task.cancel()
            except RuntimeError:
                # Ignore runtime errors that might occur if event loop is closed
                pass
            self._task = None

    def __del__(self):
        self.shutdown()

__init__(trie)

Initialize an AsyncTokenCharacterTrie.

Parameters:

Name	Type	Description	Default
trie	TokenCharacterTrie | ParallelTokenCharacterTrie	The underlying TokenCharacterTrie or ParallelTokenCharacterTrie instance	required

Source code in genlm/backend/trie/async_impl.py

def __init__(self, trie):
    """Initialize an `AsyncTokenCharacterTrie`.

    Args:
        trie (TokenCharacterTrie|ParallelTokenCharacterTrie): The underlying `TokenCharacterTrie` or `ParallelTokenCharacterTrie` instance
    """
    self.trie = trie
    self._queue = None
    self._task = None

from_vocab(vocab, backend='parallel', **kwargs) classmethod

Creates an AsyncTokenCharacterTrie from a vocabulary.

Parameters:

Name	Type	Description	Default
vocab	list[Token]	The vocabulary over which the trie will be defined.	required
backend	str	The trie implementation to use - either 'sequential' or 'parallel'. Defaults to 'parallel' which uses GPU acceleration when available.	'parallel'
**kwargs		Additional arguments passed to the trie constructor	{}

Returns:

Type	Description
AsyncTokenCharacterTrie	The initialized asynchronous trie instance.

Source code in genlm/backend/trie/async_impl.py

@classmethod
def from_vocab(cls, vocab, backend="parallel", **kwargs):
    """Creates an `AsyncTokenCharacterTrie` from a vocabulary.

    Args:
        vocab (list[Token]): The vocabulary over which the trie will be defined.
        backend (str, optional): The trie implementation to use - either 'sequential' or 'parallel'.
                Defaults to 'parallel' which uses GPU acceleration when available.
        **kwargs: Additional arguments passed to the trie constructor

    Returns:
        (AsyncTokenCharacterTrie): The initialized asynchronous trie instance.
    """
    if backend == "sequential":
        trie = TokenCharacterTrie(decode=vocab, **kwargs)
    elif backend == "parallel":
        trie = ParallelTokenCharacterTrie(decode=vocab, **kwargs)
    else:
        raise ValueError(
            f"Unknown backend: {backend}. Must be one of ['sequential', 'parallel']"
        )
    return cls(trie)

weight_sum(ws) async

Queue a weight_sum request. Multiple concurrent calls will be automatically batched together.

Parameters:

Name	Type	Description	Default
ws	Tensor	Token weights, shape (len(self.trie.decode),).	required

Returns:

Type	Description
ndarray	The calculated mass sums for the given distribution.

Source code in genlm/backend/trie/async_impl.py

async def weight_sum(self, ws):
    """Queue a `weight_sum` request. Multiple concurrent calls will be automatically batched
    together.

    Args:
        ws (torch.Tensor): Token weights, shape (`len(self.trie.decode)`,).

    Returns:
        (np.ndarray): The calculated mass sums for the given distribution.
    """
    future = await self._queue_request(ws, "sum")
    result = await future
    return result

weight_max(ws) async

Queue a weight_max request. Multiple concurrent calls will be automatically batched together.

Parameters:

Name	Type	Description	Default
ws	Tensor	Token weights, shape (len(self.trie.decode),).	required

Returns:

Type	Description
ndarray	The calculated max weights for the given distribution.

Source code in genlm/backend/trie/async_impl.py

async def weight_max(self, ws):
    """Queue a `weight_max` request. Multiple concurrent calls will be automatically batched
    together.

    Args:
        ws (torch.Tensor): Token weights, shape (`len(self.trie.decode)`,).

    Returns:
        (np.ndarray): The calculated max weights for the given distribution.
    """
    future = await self._queue_request(ws, "max")
    result = await future
    return result

start()

Start the background processing task if not already running.

Source code in genlm/backend/trie/async_impl.py

def start(self):
    """Start the background processing task if not already running."""
    if not self._task or self._task.done():
        self._queue = (
            asyncio.Queue()
        )  # Create a new queue so that it is bound to the current event loop
        self._task = asyncio.create_task(self._background_loop())

cleanup() async

Async cleanup - preferred method

Source code in genlm/backend/trie/async_impl.py

async def cleanup(self):
    """Async cleanup - preferred method"""
    if self._task and not self._task.done():
        self._task.cancel()
        try:
            await self._task
        except asyncio.CancelledError:
            pass
        self._task = None

shutdown()

Stop the background processing task and cleanup resources.

Source code in genlm/backend/trie/async_impl.py

def shutdown(self):
    """Stop the background processing task and cleanup resources."""
    if self._task is not None:
        try:
            self._task.cancel()
        except RuntimeError:
            # Ignore runtime errors that might occur if event loop is closed
            pass
        self._task = None