hf

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
        self.cache = 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.cache = TokenTrie(None, self.cache.logprobs)

    def clear_kv_cache(self):
        """Clear any key and value vectors from the cache."""
        self.cache.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):
        """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.
        """
        result = self.model(torch.tensor([prompt_tokens]).to(self.device))
        node = self.cache.extend_cache(1, prompt_tokens, result.logits[0], 0)
        node.past_key_values = result.past_key_values

    @torch.no_grad()
    def batch_evaluate_queries(self):
        """
        Process a batch of queued language model queries.

        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

        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), False)
        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:
            pasts = [
                [
                    torch.cat(
                        (
                            *(
                                q.past_padded(
                                    layer,
                                    j,
                                    max_past_length,
                                    past_example[0][0].dtype,
                                    self.device,
                                    past_example[0][0].shape,
                                )
                                for q in unique_queries
                            ),
                        ),
                        dim=0,
                    )
                    for j in range(2)
                ]
                for layer in range(len(past_example))
            ]
        else:
            pasts = None

        results = self.model(
            input_ids,
            attention_mask=attn_masks,
            position_ids=posn_ids,
            past_key_values=pasts,
            use_cache=pasts is not None,
        )

        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):
        """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 (list[tuple[torch.Tensor]]|None): Past key/value states from previous evaluation,
                or None if this is a new query
        """
        self.queries.append(Query(query, future, past))

        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):
        """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

        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
                - list[tuple[torch.Tensor]]|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.cache
        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):
        """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.

        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)

        # 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)
        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):
        """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.

        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)

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

        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):
        """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.

        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():
            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)

__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
    self.cache = 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)

add_query(query, future, past)

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	list[tuple[Tensor]] | None	Past key/value states from previous evaluation, or None if this is a new query	required

Source code in genlm_backend/llm/hf.py

@torch.no_grad()
def add_query(self, query, future, past):
    """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 (list[tuple[torch.Tensor]]|None): Past key/value states from previous evaluation,
            or None if this is a new query
    """
    self.queries.append(Query(query, future, past))

    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()
        )

batch_evaluate_queries()

Process a batch of queued language model queries.

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.

    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

    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), False)
    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:
        pasts = [
            [
                torch.cat(
                    (
                        *(
                            q.past_padded(
                                layer,
                                j,
                                max_past_length,
                                past_example[0][0].dtype,
                                self.device,
                                past_example[0][0].shape,
                            )
                            for q in unique_queries
                        ),
                    ),
                    dim=0,
                )
                for j in range(2)
            ]
            for layer in range(len(past_example))
        ]
    else:
        pasts = None

    results = self.model(
        input_ids,
        attention_mask=attn_masks,
        position_ids=posn_ids,
        past_key_values=pasts,
        use_cache=pasts is not None,
    )

    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)

cache_kv(prompt_tokens)

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

Source code in genlm_backend/llm/hf.py

@torch.no_grad()
def cache_kv(self, prompt_tokens):
    """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.
    """
    result = self.model(torch.tensor([prompt_tokens]).to(self.device))
    node = self.cache.extend_cache(1, prompt_tokens, result.logits[0], 0)
    node.past_key_values = result.past_key_values

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.cache = TokenTrie(None, self.cache.logprobs)

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."""
    self.cache.clear_kv_cache()

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)

next_token_logprobs(token_ids) 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

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):
    """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.

    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)

    # 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)
    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)

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

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):
    """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.

    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)

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

    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)

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

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):
    """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.

    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():
        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)

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 = []

walk_cache(token_ids)

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

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 list[tuple[torch.Tensor]]|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):
    """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

    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
            - list[tuple[torch.Tensor]]|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.cache
    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

Query

A query to a language model, waiting to be batched.

Source code in genlm_backend/llm/hf.py

class Query:
    """A query to a language model, waiting to be batched."""

    def __init__(self, prompt, future, past=None):
        self.prompt = prompt
        self.future = future
        self.past = past

        if self.past is not None:
            self.past_len = past[
                0
            ][
                0
            ].shape[
                2
            ]  # layers, key or value, batch size, num heads, num tokens, head repr length
        else:
            self.past_len = 0

    @torch.no_grad()
    def past_padded(self, layer, j, to_length, dtype, device, past_shape):
        if self.past is not None:
            return torch.cat(
                (
                    self.past[layer][j],
                    torch.zeros(
                        1,
                        past_shape[1],
                        to_length - self.past_len,
                        past_shape[3],
                        dtype=dtype,
                        device=device,
                    ),
                ),
                dim=2,
            )
        else:
            return torch.zeros(
                1, past_shape[1], to_length, past_shape[3], dtype=dtype, device=device
            )

    def prompt_padded(self, pad_token, to_length):
        return [*self.prompt, *[pad_token for _ in range(to_length - len(self.prompt))]]

    def attention_mask(self, total_past_length, total_seq_length):
        return [
            *[1 for _ in range(self.past_len)],
            *[0 for _ in range(total_past_length - self.past_len)],
            *[1 for _ in range(len(self.prompt))],
            *[0 for _ in range(total_seq_length - len(self.prompt))],
        ]

    def position_ids(self, total_past_length, total_seq_length):
        return [
            *range(self.past_len, self.past_len + len(self.prompt)),
            *[0 for _ in range(total_seq_length - len(self.prompt))],
        ]