fix(proxy): treat malformed cost-map token limits as absent on /v1/models

[yuneng-berri]

Relevant issues

Backport of #33903 onto patch-1.93.0rc2

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

patch-1.93.0rc2 carries the regression, so the 1.93.0 stable cut ships a 500 on GET /v1/models without this pick

$ git grep -n "cost_map_input" origin/patch-1.93.0rc2 -- litellm/proxy/utils.py
origin/patch-1.93.0rc2:litellm/proxy/utils.py:6133:        cost_map_input = model_cost_info.get("max_input_tokens")
origin/patch-1.93.0rc2:litellm/proxy/utils.py:6134:        if cost_map_input is not None:
origin/patch-1.93.0rc2:litellm/proxy/utils.py:6135:            max_input_tokens = int(cost_map_input)

The line also already carries the router-side guard from #33864, so it is in the same state staging was in: one path guarded, the cost-map path not

$ git grep -n "_as_int" origin/patch-1.93.0rc2 -- litellm/router.py
origin/patch-1.93.0rc2:litellm/router.py:8374:        def _as_int(value: object) -> "int | None":

Runtime proof was captured on the staging PR (#33903) against a live proxy: 500 before, 200 with the malformed deployment's limits omitted after, plus a real OpenAI call through the same proxy. On this line, the picked commit resolves and degrades gracefully rather than raising, at db190ba

$ python -c "
import litellm
from litellm import Router
from litellm.proxy.utils import create_model_info_response
r = Router(model_list=[{'model_name':'openai/probe','litellm_params':{'model':'openai/probe'},'model_info':{'max_input_tokens':'128,000'}}])
print('RESULT:', create_model_info_response(model_id='openai/probe', provider='openai', llm_router=r))
"
RESULT: {'id': 'openai/probe', 'object': 'model', 'created': 1677610602, 'owned_by': 'openai'}

$ python -m pytest tests/test_litellm/proxy/test_proxy_utils.py tests/test_litellm/test_router.py -k "create_model_info_response or token_limits"
23 passed, 158 deselected

Type

🐛 Bug Fix

Changes

Clean cherry-pick of ab02127 from #33903, no conflicts and no adaptation needed for this line

create_model_info_response cast the cost-map max_input_tokens and max_output_tokens with a bare int(), outside the try/except that guards the get_model_info lookup. A deployment whose model_info carries a non-numeric limit therefore raised inside the per-model listing loop and failed the whole GET /v1/models and /models response with a 500, taking every healthy deployment down with it

Both call sites now share coerce_token_limit in litellm_core_utils/core_helpers.py, which returns None for anything that is not a usable number so the listing omits that one limit and keeps serving

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[greptile-apps]

Greptile Summary

This PR backports a targeted fix for a 500 error on GET /v1/models caused by a bare int() cast on cost-map token limits that sat outside the existing try/except guard, so any deployment with a non-numeric limit (e.g. "128,000") would crash the entire listing.

• Introduces coerce_token_limit() in core_helpers.py as a shared utility that returns None for booleans, non-numeric strings, unsupported collection types, and infinity/NaN floats, replacing the bare int() in proxy/utils.py and the local _as_int helper in router.py.
• Adds three new mock-only tests covering parametrized bad values, mixed valid/malformed limits, and the end-to-end path where a Router registers a malformed limit into litellm.model_cost; global state is saved and restored correctly in the integration-style test.

Confidence Score: 5/5

Safe to merge — the change is a one-directional tightening that degrades gracefully instead of raising, with no impact on well-formed deployments.

The fix is minimal and surgical: coerce_token_limit is a pure function with no side effects, and its behavior is identical to the old _as_int for all values that previously worked correctly. The only observable difference is that booleans and other non-numeric types now return None instead of raising, which is the intended fix. Tests cover the newly guarded path and the pre-existing router path, and global state mutation in the integration test is properly contained in a try/finally block.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/core_helpers.py	Adds coerce_token_limit() shared utility that safely converts token-limit values to int, treating booleans, non-numeric strings, and unexpected types as absent (None).
litellm/proxy/utils.py	Replaces bare int() calls on cost-map token limits with coerce_token_limit(), preventing a 500 when a deployment carries a non-numeric limit string like "128,000".
litellm/router.py	Removes the local _as_int helper from get_configured_token_limits and replaces it with the shared coerce_token_limit; functionally equivalent (both return None for booleans, non-numeric strings, and unsupported types).
tests/test_litellm/proxy/test_proxy_utils.py	Adds three new tests covering malformed cost-map limits (parametrized), mixed valid/malformed limits, and the end-to-end path where a Router registers a bad limit into litellm.model_cost; global state is correctly saved and restored.

_{Reviews (1): Last reviewed commit: "fix(proxy): treat malformed cost-map tok..." | Re-trigger Greptile}