Merge pull request #40186 from BerriAI/litellm_lit_5546_count_tokens_offload

fix(token_counter): release the GIL for HuggingFace counts and cap exact counting per string

litellm/constants.py

@@ -398,6 +398,18 @@ TIKTOKEN_ENCODE_CHUNK_SIZE_CHARS: Final = get_env_int_in_range(
     minimum=1,
     maximum=TIKTOKEN_ENCODE_MAX_CHUNK_SIZE_CHARS,
 )
+TOKEN_COUNTER_MAX_EXACT_CHARS: Final = get_env_int_in_range(
+    "TOKEN_COUNTER_MAX_EXACT_CHARS",
+    default=4_000_000,
+    minimum=1,
+    maximum=1_000_000_000,
+)
+TOKEN_COUNTER_MAX_CONCURRENT_COUNTS: Final = get_env_int_in_range(
+    "TOKEN_COUNTER_MAX_CONCURRENT_COUNTS",
+    default=4,
+    minimum=1,
+    maximum=256,
+)
 MAX_TILE_WIDTH: Final = int(os.getenv("MAX_TILE_WIDTH", 512))
 MAX_TILE_HEIGHT: Final = int(os.getenv("MAX_TILE_HEIGHT", 512))
 OPENAI_FILE_SEARCH_COST_PER_1K_CALLS: Final = float(os.getenv("OPENAI_FILE_SEARCH_COST_PER_1K_CALLS", 2.5 / 1000))

litellm/litellm_core_utils/token_counter.py

@@ -3,11 +3,15 @@
 import base64
 import io
 import struct
-from collections.abc import Callable, Iterable, Mapping, Sequence
+from collections.abc import Awaitable, Callable, Iterable, Mapping, Sequence
 from typing import Final, Literal, cast
 
+import anyio
+import anyio.lowlevel
 import httpx
 import tiktoken
+from tokenizers import Tokenizer
+from typing_extensions import ParamSpec, TypeVar
 
 import litellm
 from litellm import verbose_logger
@@ -21,7 +25,10 @@ from litellm.constants import (
     MAX_TILE_HEIGHT,
     MAX_TILE_WIDTH,
     TIKTOKEN_ENCODE_CHUNK_SIZE_CHARS,
+    TOKEN_COUNTER_MAX_CONCURRENT_COUNTS,
+    TOKEN_COUNTER_MAX_EXACT_CHARS,
 )
+from litellm.litellm_core_utils.asyncify import asyncify
 from litellm.litellm_core_utils.default_encoding import encoding as default_encoding
 from litellm.litellm_core_utils.url_utils import safe_get
 from litellm.llms.custom_httpx.http_handler import _get_httpx_client
@@ -317,6 +324,32 @@ TokenCounterFunction = Callable[[str], int]
 Type for a function that counts tokens in a string.
 """
 
+EXTRAPOLATION_SAMPLES: Final = 16
+T_ParamSpec: Final = ParamSpec("T_ParamSpec")
+T_Retval = TypeVar("T_Retval")
+_COUNT_OFFLOAD_LIMITER: Final = anyio.lowlevel.RunVar[anyio.CapacityLimiter]("litellm_count_offload_limiter")
+
+
+def _count_offload_limiter_for_this_loop() -> anyio.CapacityLimiter:
+    existing: Final = _COUNT_OFFLOAD_LIMITER.get(None)
+    if existing is not None:
+        return existing
+    created: Final = anyio.CapacityLimiter(TOKEN_COUNTER_MAX_CONCURRENT_COUNTS)
+    _COUNT_OFFLOAD_LIMITER.set(created)
+    return created
+
+
+def offload_token_count(
+    function: Callable[T_ParamSpec, T_Retval],
+) -> Callable[T_ParamSpec, Awaitable[T_Retval]]:
+    async def offloaded(
+        *args: T_ParamSpec.args,
+        **kwargs: T_ParamSpec.kwargs,  # kwargs-ok: ParamSpec keeps the wrapped function's own keyword contract
+    ) -> T_Retval:
+        return await asyncify(function, limiter=_count_offload_limiter_for_this_loop())(*args, **kwargs)
+
+    return offloaded
+
 
 def _get_tiktoken_count_function(
     encode_length: Callable[[str], int],
@@ -538,9 +571,40 @@ def _count_extra(
     return num_tokens
 
 
+def _get_extrapolating_count_function(
+    count_exactly: TokenCounterFunction,
+    max_exact_chars: int = TOKEN_COUNTER_MAX_EXACT_CHARS,
+) -> TokenCounterFunction:
+    def count_tokens(text: str) -> int:
+        if len(text) <= max_exact_chars:
+            return count_exactly(text)
+        samples: Final = _evenly_spaced_samples(text, max_exact_chars)
+        sampled_chars: Final = sum(len(sample) for sample in samples)
+        return round(sum(count_exactly(sample) for sample in samples) * len(text) / sampled_chars)
+
+    return count_tokens
+
+
+def _evenly_spaced_samples(text: str, total_chars: int) -> tuple[str, ...]:
+    sample_count: Final = min(EXTRAPOLATION_SAMPLES, total_chars)
+    sample_chars: Final = total_chars // sample_count
+    last_start: Final = len(text) - sample_chars
+    return tuple(
+        text[start : start + sample_chars]
+        for start in (last_start * index // max(sample_count - 1, 1) for index in range(sample_count))
+    )
+
+
 def _get_count_function(
     model: str | None,
     custom_tokenizer: dict | SelectTokenizerResponse | None = None,
+) -> TokenCounterFunction:
+    return _get_extrapolating_count_function(_get_exact_count_function(model, custom_tokenizer))
+
+
+def _get_exact_count_function(
+    model: str | None,
+    custom_tokenizer: dict | SelectTokenizerResponse | None = None,
 ) -> TokenCounterFunction:
     """
     Get the function to count tokens based on the model and custom tokenizer."""
@@ -549,10 +613,10 @@ def _get_count_function(
     if model is not None or custom_tokenizer is not None:
         tokenizer_json: Final = custom_tokenizer or _select_tokenizer(model)
         if tokenizer_json["type"] == "huggingface_tokenizer":
+            tokenizer: Final[Tokenizer] = tokenizer_json["tokenizer"]
 
             def count_tokens(text: str) -> int:
-                enc: Final = tokenizer_json["tokenizer"].encode(text)
-                return len(enc.ids)
+                return len(tokenizer.encode_batch_fast([text])[0])
 
             return count_tokens
         elif tokenizer_json["type"] == "openai_tokenizer":

litellm/proxy/hooks/parallel_request_limiter_v3.py

@@ -31,6 +31,7 @@ from litellm.integrations.custom_logger import CustomLogger
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     get_str_from_messages,
 )
+from litellm.litellm_core_utils.token_counter import offload_token_count
 from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.auth.auth_utils import (
     ESTIMATED_OUTPUT_TOKENS_FIELD,
@@ -3307,7 +3308,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
             min_configured_tpm_limit=min_configured_otpm_limit,
             call_type=call_type,
         )
-        raw_estimated_input_tokens: Final = self._estimate_precise_input_tokens(
+        raw_estimated_input_tokens: Final = await offload_token_count(self._estimate_precise_input_tokens)(
             data=data, model=requested_model, call_type=call_type
         )
         estimated_input_tokens: Final = max(raw_estimated_input_tokens, 1)

litellm/proxy/proxy_server.py

@@ -63,11 +63,13 @@ from litellm.constants import (
     LITELLM_UI_SESSION_DURATION,
     RUNTIME_UPDATABLE_ROUTER_SETTINGS,
 )
+from litellm.litellm_core_utils.asyncify import asyncify
 from litellm.litellm_core_utils.litellm_logging import (
     _init_custom_logger_compatible_class,
 )
 from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
 from litellm.litellm_core_utils.safe_json_loads import safe_json_loads
+from litellm.litellm_core_utils.token_counter import offload_token_count
 from litellm.proxy._types import (
     UI_TEAM_ID,
     CallbackDelete,
@@ -272,7 +274,6 @@ from litellm.integrations.SlackAlerting.slack_alerting import SlackAlerting
 from litellm.litellm_core_utils.agentic_loop_settings import (
     validated_max_agentic_loops,
 )
-from litellm.litellm_core_utils.asyncify import asyncify
 from litellm.litellm_core_utils.audio_utils.utils import resolve_speech_media_type
 from litellm.litellm_core_utils.core_helpers import (
     _get_parent_otel_span_from_kwargs,
@@ -12816,7 +12817,9 @@ async def token_counter(request: TokenCountRequest, call_endpoint: bool = False)
             CustomHuggingfaceTokenizer | None,
             model_info.get("custom_tokenizer", None),
         )
-    _tokenizer_used: Final = litellm.utils._select_tokenizer(model=model_to_use, custom_tokenizer=custom_tokenizer)
+    _tokenizer_used: Final = await asyncify(litellm.utils._select_tokenizer)(
+        model=model_to_use, custom_tokenizer=custom_tokenizer
+    )
 
     tokenizer_used: Final = str(_tokenizer_used["type"])
     system_message: Final = _system_message(system)
@@ -12829,7 +12832,7 @@ async def token_counter(request: TokenCountRequest, call_endpoint: bool = False)
     counted_tools: Final = cast(  # cast-ok: raw OpenAI or Anthropic tool dicts, both of which token_counter formats
         list[ChatCompletionToolParam] | None, tools if counted_messages is not None else None
     )
-    total_tokens: Final = await asyncify(litellm.token_counter)(
+    total_tokens: Final = await offload_token_count(litellm.token_counter)(
         model=model_to_use,
         text=prompt,
         messages=counted_messages,

litellm/proxy/utils.py

@@ -101,6 +101,7 @@ from litellm.litellm_core_utils.core_helpers import (
 from litellm.litellm_core_utils.litellm_logging import Logging
 from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
 from litellm.litellm_core_utils.safe_json_loads import safe_json_loads
+from litellm.litellm_core_utils.token_counter import offload_token_count
 from litellm.llms import load_guardrail_translation_mappings
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
 from litellm.proxy._types import (
@@ -2900,7 +2901,7 @@ class ProxyLogging:
                 original_exception=original_exception,
             )
 
-        request_data.update(_failure_fields_to_lift(request_data))
+        request_data.update(await offload_token_count(_failure_fields_to_lift)(request_data))
 
         # Remove before callbacks iterate — not serialisable
         request_data.pop("litellm_logging_obj", None)

litellm/router.py

@@ -67,7 +67,7 @@ from litellm.constants import (
     SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
 )
 from litellm.integrations.custom_logger import CustomLogger
-from litellm.litellm_core_utils.asyncify import asyncify, run_async_function
+from litellm.litellm_core_utils.asyncify import run_async_function
 from litellm.litellm_core_utils.core_helpers import (
     _get_parent_otel_span_from_kwargs,
     coerce_token_limit,
@@ -98,6 +98,7 @@ from litellm.litellm_core_utils.sensitive_data_masker import (
     mask_credentials_in_payload,
     mask_sensitive_structure,
 )
+from litellm.litellm_core_utils.token_counter import offload_token_count
 from litellm.llms.base_llm.vector_store.transformation import (
     RouterVectorStoreEmbeddingExecutor,
     vector_store_request_metadata,
@@ -12113,7 +12114,7 @@ class Router:
         try:
             if not self._pre_call_checks_need_token_count(model, healthy_deployments):
                 return None
-            return await asyncify(self._count_pre_call_check_tokens)(
+            return await offload_token_count(self._count_pre_call_check_tokens)(
                 messages=cast(list[dict[str, str]] | None, messages),  # cast-ok: forwarded to the sync counter
                 input=cast(str | list | None, input),  # cast-ok: forwarded to the sync counter
                 request_kwargs=request_kwargs,

litellm/router_strategy/complexity_router/complexity_router.py

@@ -2568,14 +2568,14 @@ class ComplexityRouter(CustomLogger):
         """Real-tokenizer count of the resolved messages plus the out-of-band carriers, off the
         event loop; None when counting fails, and the gate then leaves the placement alone."""
         import litellm
-        from litellm.litellm_core_utils.asyncify import asyncify
+        from litellm.litellm_core_utils.token_counter import offload_token_count
 
         out_of_band: Final = self._out_of_band_request_text(request_kwargs)
         try:
-            counted: Final = await asyncify(litellm.token_counter)(
+            counted: Final = await offload_token_count(litellm.token_counter)(
                 messages=cast(list, resolved_messages)  # cast-ok: token_counter only iterates the sequence
             )
-            return counted + (await asyncify(litellm.token_counter)(text=out_of_band) if out_of_band else 0)
+            return counted + (await offload_token_count(litellm.token_counter)(text=out_of_band) if out_of_band else 0)
         except Exception as e:  # noqa: BLE001  # best-effort: an uncountable prompt must not fail the request
             verbose_router_logger.debug("ComplexityRouter: context-window token count failed. Got - %s", e)
             return None

litellm/router_utils/pre_call_checks/io_token_rate_limit_check.py

@@ -21,6 +21,7 @@ import litellm
 from litellm import token_counter
 from litellm._logging import verbose_router_logger
 from litellm.caching.dual_cache import DualCache
+from litellm.litellm_core_utils.token_counter import offload_token_count
 from litellm.types.router import RouterCacheEnum, RouterErrors
 from litellm.utils import get_utc_datetime
 
@@ -466,7 +467,7 @@ async def async_io_token_pre_call_check(
 
     request_kwargs: Final = get_io_token_rate_limit_request_kwargs()
     _model: Final = (deployment.get("litellm_params") or {}).get("model") or ""
-    estimated_input: Final = _estimate_input_tokens(request_kwargs, model=_model)
+    estimated_input: Final = await offload_token_count(_estimate_input_tokens)(request_kwargs, model=_model)
     max_tokens: Final = _resolve_max_tokens(request_kwargs, deployment)
 
     dt: Final = get_utc_datetime()

litellm/router_utils/pre_call_checks/prompt_caching_deployment_check.py

@@ -14,6 +14,7 @@ from litellm.integrations.anthropic_cache_control_hook import (
     AnthropicCacheControlHook,
 )
 from litellm.integrations.custom_logger import CustomLogger, Span
+from litellm.litellm_core_utils.token_counter import offload_token_count
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.utils import CallTypes, StandardLoggingPayload
 from litellm.utils import get_prompt_cache_min_tokens, is_prompt_caching_valid_prompt
@@ -61,7 +62,7 @@ class PromptCachingDeploymentCheck(CustomLogger):
         if request_kwargs is not None and request_kwargs.get("_target_order") is not None:
             return healthy_deployments
 
-        if messages is not None and is_prompt_caching_valid_prompt(
+        if messages is not None and await offload_token_count(is_prompt_caching_valid_prompt)(
             messages=messages,
             model=model,
             min_token_count=_get_min_token_count_for_deployments(healthy_deployments),
@@ -139,7 +140,7 @@ class PromptCachingDeploymentCheck(CustomLogger):
             return
 
         ## PROMPT CACHING - cache model id, if prompt caching valid prompt + provider
-        if is_prompt_caching_valid_prompt(
+        if await offload_token_count(is_prompt_caching_valid_prompt)(
             model=model,
             messages=cast(list[AllMessageValues], messages),
         ):

litellm/utils.py

@@ -2293,15 +2293,7 @@ def create_pretrained_tokenizer(identifier: str, revision="main", auth_token: st
     dict: A dictionary with the tokenizer and its type.
     """
 
-    try:
-        tokenizer = Tokenizer.from_pretrained(
-            identifier,
-            revision=revision,
-            auth_token=auth_token,
-        )
-    except Exception as e:
-        verbose_logger.error("Error creating pretrained tokenizer: %s. Defaulting to version without 'auth_token'.", e)
-        tokenizer = Tokenizer.from_pretrained(identifier, revision=revision)
+    tokenizer: Final = Tokenizer.from_pretrained(identifier, revision=revision, token=auth_token)
     return {"type": "huggingface_tokenizer", "tokenizer": tokenizer}
 
 

tests/proxy_unit_tests/test_custom_tokenizer_bug.py

@@ -23,9 +23,9 @@ from litellm.proxy.proxy_server import token_counter
 
 def _fake_hf_tokenizer(num_tokens: int) -> MagicMock:
     encoding = MagicMock()
-    encoding.ids = list(range(num_tokens))
+    encoding.__len__.return_value = num_tokens
     tokenizer = MagicMock()
-    tokenizer.encode.return_value = encoding
+    tokenizer.encode_batch_fast.return_value = [encoding]
     return tokenizer
 
 
@@ -68,13 +68,11 @@ async def test_custom_tokenizer_from_model_info_is_used(monkeypatch):
             )
         )
 
-    mock_tokenizer_cls.from_pretrained.assert_called_once_with(
-        "my-org/custom-tokenizer", revision="v2", auth_token=None
-    )
+    mock_tokenizer_cls.from_pretrained.assert_called_once_with("my-org/custom-tokenizer", revision="v2", token=None)
     assert response.tokenizer_type == "huggingface_tokenizer"
     assert response.request_model == "my-embedding-model"
     assert response.model_used == "self-hosted-embedder"
-    assert response.total_tokens > 0
+    assert response.total_tokens >= 7
 
 
 @pytest.mark.asyncio

tests/test_litellm/litellm_core_utils/event_loop_lag.py

@@ -0,0 +1,40 @@
+import asyncio
+import time
+from collections.abc import Awaitable, Callable
+from typing import Final, TypeVar
+
+import litellm
+
+T = TypeVar("T")
+
+
+def warm_tokenizer(model: str) -> None:
+    litellm.token_counter(model=model, text="load the tokenizer before anything is timed")
+
+
+async def loop_wake_lags(until: asyncio.Event) -> tuple[float, ...]:
+    async def wake_lag() -> float:
+        started: Final = time.perf_counter()
+        await asyncio.sleep(0.001)
+        return time.perf_counter() - started - 0.001
+
+    return tuple([await wake_lag() for _ in iter(until.is_set, True)])
+
+
+async def timed_with_loop_lags(run: Callable[[], Awaitable[T]]) -> tuple[T, float, tuple[float, ...]]:
+    finished: Final = asyncio.Event()
+
+    async def timed() -> tuple[T, float]:
+        await asyncio.sleep(0)
+        started: Final = time.perf_counter()
+        try:
+            return await run(), time.perf_counter() - started
+        finally:
+            finished.set()
+
+    (result, took), lags = await asyncio.gather(timed(), loop_wake_lags(finished))
+    return result, took, lags
+
+
+def assert_loop_stayed_free(took: float, lags: tuple[float, ...]) -> None:
+    assert max(lags) < took / 4, f"the event loop stalled {max(lags):.3f}s during a {took:.3f}s count"

tests/test_litellm/litellm_core_utils/test_token_counter.py

@@ -1,10 +1,15 @@
 #### What this tests ####
 #    This tests litellm.token_counter.token_counter() function
+import asyncio
 import importlib
+import threading
 import time
 import traceback
+from concurrent.futures import Future, wait
+from typing import Final
 from unittest.mock import MagicMock
 
+import anyio.to_thread
 import pytest
 import tiktoken
 
@@ -14,9 +19,21 @@ import litellm
 from litellm import create_pretrained_tokenizer, decode, encode, get_modified_max_tokens
 from litellm import token_counter as token_counter_old
 import litellm.constants
-from litellm.litellm_core_utils.token_counter import _get_tiktoken_count_function
+from litellm.constants import TOKEN_COUNTER_MAX_CONCURRENT_COUNTS
+from litellm.litellm_core_utils.asyncify import asyncify
+from litellm.litellm_core_utils.token_counter import (
+    _get_exact_count_function,
+    _get_extrapolating_count_function,
+    _get_tiktoken_count_function,
+    offload_token_count,
+)
 from litellm.litellm_core_utils.token_counter import token_counter as token_counter_new
 from tests.large_text import text
+from tests.test_litellm.litellm_core_utils.event_loop_lag import (
+    assert_loop_stayed_free,
+    timed_with_loop_lags,
+    warm_tokenizer,
+)
 from tests.test_litellm.litellm_core_utils.messages_with_counts import (
     MESSAGES_TEXT,
     MESSAGES_WITH_IMAGES,
@@ -120,6 +137,135 @@ def test_valid_chunk_size_config_is_honoured(monkeypatch):
         importlib.reload(litellm.constants)
 
 
+async def test_huggingface_count_in_a_worker_thread_leaves_the_event_loop_free():
+    warm_tokenizer("claude-fable-5")
+
+    tokens, took, lags = await timed_with_loop_lags(
+        lambda: asyncify(token_counter_new)(model="claude-fable-5", text=text * 100)
+    )
+
+    assert tokens > 0
+    assert_loop_stayed_free(took, lags)
+
+
+@pytest.mark.parametrize("max_exact_chars", [64, 1_000, 2_500])
+def test_count_above_the_cap_samples_the_whole_string_and_scales(max_exact_chars: int):
+    count_exactly: Final = MagicMock(side_effect=lambda chunk: chunk.count("a") + len(chunk))
+    front_heavy: Final = "a" * 1_000 + "b" * 4_000
+    exact: Final = 1_000 + len(front_heavy)
+
+    estimate: Final = _get_extrapolating_count_function(count_exactly, max_exact_chars=max_exact_chars)(front_heavy)
+
+    assert abs(estimate - exact) <= exact // 100
+    assert sum(len(call.args[0]) for call in count_exactly.call_args_list) <= max_exact_chars
+
+
+def test_count_at_or_below_the_cap_is_exact():
+    count_exactly: Final = MagicMock(side_effect=len)
+
+    assert _get_extrapolating_count_function(count_exactly, max_exact_chars=5_000)("a" * 5_000) == 5_000
+    assert count_exactly.call_args_list == [(("a" * 5_000,),)]
+
+
+class _SlowEncoder:
+    def __init__(self) -> None:
+        self._lock: Final = threading.Lock()
+        self.in_flight = 0
+        self.peak_in_flight = 0
+
+    def encode_batch_fast(self, texts: list[str]) -> list[list[int]]:
+        with self._lock:
+            self.in_flight += 1
+            self.peak_in_flight = max(self.peak_in_flight, self.in_flight)
+        time.sleep(0.1)
+        with self._lock:
+            self.in_flight -= 1
+        return [[0] * len(text) for text in texts]
+
+
+@pytest.mark.asyncio
+async def test_offloaded_counts_do_not_borrow_from_the_shared_thread_pool():
+    encoder: Final = _SlowEncoder()
+    count: Final = _get_exact_count_function(None, {"type": "huggingface_tokenizer", "tokenizer": encoder})
+    shared_pool: Final = anyio.to_thread.current_default_thread_limiter()
+    burst: Final = 2 * TOKEN_COUNTER_MAX_CONCURRENT_COUNTS
+
+    async def shared_pool_borrowed_until_done(counting: asyncio.Future[list[int]]) -> tuple[int, ...]:
+        if counting.done():
+            return ()
+        await asyncio.sleep(0.01)
+        return (shared_pool.borrowed_tokens, *await shared_pool_borrowed_until_done(counting))
+
+    counting: Final = asyncio.ensure_future(asyncio.gather(*(offload_token_count(count)("abc") for _ in range(burst))))
+    borrowed: Final = await shared_pool_borrowed_until_done(counting)
+
+    assert await counting == [3] * burst
+    assert len(borrowed) > 1 and max(borrowed) == 0
+    assert 1 < encoder.peak_in_flight <= TOKEN_COUNTER_MAX_CONCURRENT_COUNTS
+
+
+def _count_in_a_fresh_event_loop(text: str, result: Future[int]) -> None:
+    def slow_count(counted: str) -> int:
+        time.sleep(0.1)
+        return len(counted)
+
+    result.set_result(asyncio.run(offload_token_count(slow_count)(text)))
+
+
+def test_offloaded_counts_finish_in_every_event_loop_that_shares_the_process():
+    loops: Final = 2 * TOKEN_COUNTER_MAX_CONCURRENT_COUNTS
+    results: Final = tuple(Future[int]() for _ in range(loops))
+    threads: Final = tuple(
+        threading.Thread(target=_count_in_a_fresh_event_loop, args=("a" * size, result), daemon=True)
+        for size, result in enumerate(results, start=1)
+    )
+    for thread in threads:
+        thread.start()
+
+    _, pending = wait(results, timeout=5)
+
+    assert not pending
+    assert tuple(result.result() for result in results) == tuple(range(1, loops + 1))
+
+
+@pytest.mark.parametrize(
+    ("configured", "expected"),
+    [("8", 8), ("0", 4), ("not-an-int", 4)],
+)
+def test_max_concurrent_counts_config_is_honoured(monkeypatch: pytest.MonkeyPatch, configured: str, expected: int):
+    monkeypatch.setenv("TOKEN_COUNTER_MAX_CONCURRENT_COUNTS", configured)
+    try:
+        assert importlib.reload(litellm.constants).TOKEN_COUNTER_MAX_CONCURRENT_COUNTS == expected
+    finally:
+        monkeypatch.delenv("TOKEN_COUNTER_MAX_CONCURRENT_COUNTS")
+        importlib.reload(litellm.constants)
+
+
+def test_token_counter_applies_the_default_cap():
+    max_exact_chars: Final = litellm.constants.TOKEN_COUNTER_MAX_EXACT_CHARS
+    prose: Final = ("The quick brown fox jumps over the lazy dog. " * (max_exact_chars // 45 + 1))[:max_exact_chars]
+    over_the_cap: Final = prose + "a" * 200_000
+    exact: Final = _get_exact_count_function("gpt-5.6")(over_the_cap)
+
+    estimate: Final = token_counter_new(model="gpt-5.6", text=over_the_cap)
+
+    assert estimate != exact
+    assert abs(estimate - exact) <= exact // 100
+
+
+@pytest.mark.parametrize(
+    ("configured", "expected"),
+    [("2048", 2048), ("0", 4_000_000), ("not-an-int", 4_000_000)],
+)
+def test_max_exact_chars_config_is_honoured(monkeypatch: pytest.MonkeyPatch, configured: str, expected: int):
+    monkeypatch.setenv("TOKEN_COUNTER_MAX_EXACT_CHARS", configured)
+    try:
+        assert importlib.reload(litellm.constants).TOKEN_COUNTER_MAX_EXACT_CHARS == expected
+    finally:
+        monkeypatch.delenv("TOKEN_COUNTER_MAX_EXACT_CHARS")
+        importlib.reload(litellm.constants)
+
+
 def test_token_counter_with_prefix():
     messages = [
         {"role": "user", "content": "Who won the world cup in 2022?"},

tests/test_litellm/proxy/hooks/test_tpm_concurrent.py

@@ -3670,5 +3670,42 @@ async def test_post_call_success_hook_contains_header_merge_failures(
     )
 
 
+@pytest.mark.asyncio
+async def test_the_project_itpm_reservation_counts_the_request_off_the_event_loop(rate_limiter):
+    from tests.large_text import text
+    from tests.test_litellm.litellm_core_utils.event_loop_lag import (
+        assert_loop_stayed_free,
+        timed_with_loop_lags,
+        warm_tokenizer,
+    )
+
+    handler, _cache = rate_limiter
+    stash = get_or_create_request_stash()
+    warm_tokenizer("claude-fable-5")
+    data: dict[str, object] = {
+        "model": "claude-fable-5",
+        "messages": [{"role": "user", "content": text * 100}],
+    }
+    itpm_descriptor = {
+        "key": PROJECT_ITPM_DESCRIPTOR_KEY,
+        "value": "proj-loop:claude-fable-5",
+        "rate_limit": {"tokens_per_unit": 10_000_000, "window_size": 60},
+    }
+
+    _, took, lags = await timed_with_loop_lags(
+        lambda: handler._reserve_project_io_tokens_or_raise(
+            descriptors=[itpm_descriptor],
+            data=data,
+            requested_model="claude-fable-5",
+            user_api_key_dict=UserAPIKeyAuth(api_key=hash_token("sk-itpm-loop"), project_id="proj-loop"),
+            tpm_reservation_scopes=[],
+            tpm_reservation_amount=0,
+        )
+    )
+
+    assert stash.rate_limit_response is not None
+    assert_loop_stayed_free(took, lags)
+
+
 if __name__ == "__main__":
     pytest.main([__file__, "-v", "-s"])

tests/test_litellm/proxy/test_proxy_server.py

@@ -6,6 +6,7 @@ import os
 import re
 import socket
 import subprocess
+import time
 import types
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
@@ -28,7 +29,7 @@ from litellm.caching.caching import RedisCache
 from litellm.caching.redis_cluster_cache import RedisClusterCache
 from litellm.litellm_core_utils.get_model_cost_map import ModelCostMapReloaded
 from litellm.caching.dual_cache import DualCache
-from litellm.proxy._types import LitellmUserRoles, UserAPIKeyAuth
+from litellm.proxy._types import LitellmUserRoles, TokenCountRequest, UserAPIKeyAuth
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 from litellm.proxy.proxy_server import app, initialize
 from litellm.utils import _invalidate_model_cost_lowercase_map
@@ -12954,3 +12955,59 @@ async def test_update_general_settings_keeps_yaml_openai_websocket_passthrough()
         import litellm.proxy.proxy_server as ps
 
         assert ps.general_settings["enable_openai_websocket_passthrough"] is False
+
+
+async def test_token_counter_keeps_the_event_loop_free_during_a_huggingface_count(monkeypatch):
+    from tests.large_text import text
+    from tests.test_litellm.litellm_core_utils.event_loop_lag import (
+        assert_loop_stayed_free,
+        timed_with_loop_lags,
+        warm_tokenizer,
+    )
+
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None)
+    warm_tokenizer("claude-fable-5")
+
+    response, took, lags = await timed_with_loop_lags(
+        lambda: proxy_server_module.token_counter(TokenCountRequest(model="claude-fable-5", prompt=text * 100))
+    )
+
+    assert response.total_tokens > 0
+    assert_loop_stayed_free(took, lags)
+
+
+async def test_token_counter_loads_a_custom_tokenizer_off_the_event_loop(monkeypatch):
+    from tokenizers import Tokenizer
+
+    from litellm import Router
+    from tests.test_litellm.litellm_core_utils.event_loop_lag import assert_loop_stayed_free, timed_with_loop_lags
+
+    claude_tokenizer: Final = litellm.utils._select_tokenizer("claude-fable-5")["tokenizer"]
+
+    class SlowHubTokenizer:
+        @staticmethod
+        def from_pretrained(identifier: str, revision: str = "main", token: str | None = None) -> Tokenizer:
+            time.sleep(0.3)
+            return claude_tokenizer
+
+    monkeypatch.setattr(litellm.utils, "Tokenizer", SlowHubTokenizer)
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.llm_router",
+        Router(
+            model_list=[
+                {
+                    "model_name": "self-hosted",
+                    "litellm_params": {"model": "openai/self-hosted-model", "api_base": "http://localhost:8080/v1"},
+                    "model_info": {"custom_tokenizer": {"identifier": "my-org/tokenizer", "revision": "main", "auth_token": None}},
+                }
+            ]
+        ),
+    )
+
+    response, took, lags = await timed_with_loop_lags(
+        lambda: proxy_server_module.token_counter(TokenCountRequest(model="self-hosted", prompt="count me off the loop"))
+    )
+
+    assert response.tokenizer_type == "huggingface_tokenizer"
+    assert response.total_tokens > 0
+    assert_loop_stayed_free(took, lags)

tests/test_litellm/proxy/test_proxy_utils.py

@@ -1823,6 +1823,44 @@ def test_a_dispatched_failure_lifts_the_four_fields_the_spend_log_needs():
     assert lifted["standard_logging_object"] == {"id": "log-1"}
 
 
+@pytest.mark.asyncio
+async def test_a_dispatched_failure_is_counted_off_the_event_loop():
+    from unittest.mock import AsyncMock, patch
+
+    from tests.large_text import text
+    from tests.test_litellm.litellm_core_utils.event_loop_lag import (
+        assert_loop_stayed_free,
+        timed_with_loop_lags,
+        warm_tokenizer,
+    )
+
+    warm_tokenizer("claude-fable-5")
+    request_data = {
+        "litellm_logging_obj": _LoggingObj(
+            {
+                "first_api_call_start_time": 1700000000.0,
+                "call_type": "acompletion",
+                "model": "claude-fable-5",
+                "messages": [{"role": "user", "content": text * 100}],
+            }
+        ),
+        "metadata": {},
+    }
+    proxy_logging_obj = ProxyLogging(user_api_key_cache=DualCache())
+    proxy_logging_obj.alert_types = []
+    with patch.object(proxy_logging_obj, "update_request_status", new=AsyncMock()):
+        _, took, lags = await timed_with_loop_lags(
+            lambda: proxy_logging_obj.post_call_failure_hook(
+                request_data=request_data,
+                original_exception=Exception("boom"),
+                user_api_key_dict=UserAPIKeyAuth(),
+            )
+        )
+
+    assert request_data["combined_usage_object"].prompt_tokens > 0
+    assert_loop_stayed_free(took, lags)
+
+
 @pytest.mark.asyncio
 async def test_proxy_only_error_expected_4xx_skips_traceback_for_both_handlers(monkeypatch):
     """Regression for LIT-6043: an expected 4xx must not format a traceback for

tests/test_litellm/router_utils/pre_call_checks/test_prompt_caching_deployment_check.py

@@ -477,3 +477,65 @@ async def test_wildcard_route_resolves_underlying_model_minimum(local_model_cost
 
     assert deployments[0]["litellm_params"]["model"] == "anthropic/claude-opus-4-6"
     assert _get_min_token_count_for_deployments(deployments) == 4096
+
+
+@pytest.mark.asyncio
+async def test_async_filter_deployments_counts_the_prompt_off_the_event_loop():
+    from tests.large_text import text
+    from tests.test_litellm.litellm_core_utils.event_loop_lag import (
+        assert_loop_stayed_free,
+        timed_with_loop_lags,
+        warm_tokenizer,
+    )
+
+    warm_tokenizer("anthropic/claude-fable-5")
+    check = PromptCachingDeploymentCheck(cache=DualCache())
+    deployments = _deployments("anthropic/claude-fable-5")
+    messages = cast(List[AllMessageValues], [{"role": "user", "content": text * 100}])
+
+    result, took, lags = await timed_with_loop_lags(
+        lambda: check.async_filter_deployments(
+            model=MODEL_GROUP_ALIAS, healthy_deployments=deployments, messages=messages
+        )
+    )
+
+    assert result == deployments
+    assert_loop_stayed_free(took, lags)
+
+
+@pytest.mark.asyncio
+async def test_async_log_success_event_counts_the_prompt_off_the_event_loop():
+    from tests.large_text import text
+    from tests.test_litellm.litellm_core_utils.event_loop_lag import (
+        assert_loop_stayed_free,
+        timed_with_loop_lags,
+        warm_tokenizer,
+    )
+
+    warm_tokenizer("anthropic/claude-fable-5")
+    cache = DualCache()
+    check = PromptCachingDeploymentCheck(cache=cache)
+    messages = cast(
+        List[AllMessageValues],
+        [{"role": "user", "content": [{"type": "text", "text": text * 100, "cache_control": {"type": "ephemeral"}}]}],
+    )
+    standard_logging_object = {
+        "call_type": "acompletion",
+        "model": "anthropic/claude-fable-5",
+        "messages": messages,
+        "model_id": "dep-1",
+    }
+
+    _, took, lags = await timed_with_loop_lags(
+        lambda: check.async_log_success_event(
+            kwargs={"standard_logging_object": standard_logging_object},
+            response_obj=None,
+            start_time=None,
+            end_time=None,
+        )
+    )
+
+    assert await PromptCachingCache(cache=cache).async_get_model_id(messages=messages, tools=None) == {
+        "model_id": "dep-1"
+    }
+    assert_loop_stayed_free(took, lags)

tests/test_litellm/test_router/test_io_token_rate_limits.py

@@ -1039,3 +1039,31 @@ class TestContextSlotRetention:
         assert deployment is not None
         router._update_kwargs_with_deployment(deployment=deployment.model_dump(), kwargs=kwargs)
         assert get_io_token_rate_limit_request_kwargs() is kwargs
+
+
+@pytest.mark.asyncio
+async def test_the_deployment_itpm_reservation_counts_the_request_off_the_event_loop():
+    from litellm.utils import get_utc_datetime
+    from tests.large_text import text
+    from tests.test_litellm.litellm_core_utils.event_loop_lag import (
+        assert_loop_stayed_free,
+        timed_with_loop_lags,
+        warm_tokenizer,
+    )
+
+    dual_cache = DualCache()
+    check = ModelRateLimitingCheck(dual_cache=dual_cache)
+    warm_tokenizer("anthropic/claude-fable-5")
+    deployment = {
+        "litellm_params": {"model": "anthropic/claude-fable-5", "itpm": 10_000_000},
+        "model_info": {"id": "io-loop-id"},
+        "model_name": "claude",
+    }
+    set_io_token_rate_limit_request_kwargs({"messages": [{"role": "user", "content": text * 100}], "metadata": {}})
+
+    _, took, lags = await timed_with_loop_lags(lambda: check.async_pre_call_check(deployment))
+
+    minute = get_utc_datetime().strftime("%H-%M")
+    reserved = await dual_cache.async_get_cache(key=f"global_router:io-loop-id:anthropic/claude-fable-5:itpm:{minute}")
+    assert reserved > 100_000
+    assert_loop_stayed_free(took, lags)