From 05a9d313ac6c7556071de726f3a12ecafa0b4b35 Mon Sep 17 00:00:00 2001
From: songkuan-zheng <252822057+songkuan-zheng@users.noreply.github.com>
Date: Wed, 17 Jun 2026 11:24:19 +0000
Subject: [PATCH] perf(proxy): drop dead proxy_model_list fetch on the
 /v1/models user-filter hot path

Greptile P2 thread on #29748 flagged that `proxy_model_list` flows into
`_apply_user_models_filter` only to be passed through to `get_user_models`,
where it is consumed exclusively on the `all-proxy-models` branch.
`_apply_user_models_filter` already short-circuits and returns
`all_models` unchanged when `all-proxy-models` is in the user's models
list, so the parameter is never read on any reachable code path.

Concrete cost: `apply_user_models_filter_to_deployments` (called on
every `/v2/model/info` and the by-id branch of `/v1/model/info`)
materializes the full router model-name list via
`llm_router.get_model_names()` on each call, then threads it through
three layers down to a function that always ignores it.

Fix:
- Drop `proxy_model_list` from `_apply_user_models_filter` signature.
- Update both callers (`get_available_models_for_user`,
  `apply_user_models_filter_to_deployments`) to stop passing it. The
  former still computes `proxy_model_list` for legitimate downstream
  consumers (`get_complete_model_list`, `get_key_models`, etc.); the
  latter no longer needs `llm_router.get_model_names()` at all and
  drops the call.
- Inside `_apply_user_models_filter`, pass `[]` to the surviving
  `get_user_models(...)` call with an inline comment naming the dead
  branch, so the structural contract of `get_user_models` is preserved
  for hypothetical future callers that do hit the `all-proxy-models`
  path.

Local verification:
- 59/59 tests/test_litellm/proxy/discovery_endpoints/
- 87/87 across discovery_endpoints + proxy_server/test_routes_model_info +
  proxy_server/test_team_model_name_translation
---
 litellm/proxy/utils.py | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index c8ffb869c8f..beddf498dbd 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -6444,7 +6444,6 @@ async def get_available_models_for_user(
     all_models = await _apply_user_models_filter(
         all_models=all_models,
         user_api_key_dict=user_api_key_dict,
-        proxy_model_list=proxy_model_list,
         model_access_groups=model_access_groups,
         prisma_client=prisma_client,
         proxy_logging_obj=proxy_logging_obj,
@@ -6457,7 +6456,6 @@ async def get_available_models_for_user(
 async def _apply_user_models_filter(
     all_models: List[str],
     user_api_key_dict: "UserAPIKeyAuth",
-    proxy_model_list: List[str],
     model_access_groups: Dict[str, List[str]],
     prisma_client: Optional["PrismaClient"],
     proxy_logging_obj: Optional["ProxyLogging"],
@@ -6534,9 +6532,14 @@ async def _apply_user_models_filter(
     if SpecialModelNames.all_proxy_models.value in user_models:
         return all_models
 
+    # The `all-proxy-models` short-circuit above returned already, so
+    # `user_models` here never contains it; `get_user_models` reads
+    # `proxy_model_list` only on that branch, so passing [] is a no-op
+    # here and lets the caller skip a `llm_router.get_model_names()` call
+    # on every /v1/models and /v2/model/info request.
     user_allowed = get_user_models(
         user_models=user_models,
-        proxy_model_list=proxy_model_list,
+        proxy_model_list=[],
         model_access_groups=model_access_groups,
     )
     return filter_models_by_user_access(
@@ -6579,10 +6582,8 @@ async def apply_user_models_filter_to_deployments(
         return deployments
 
     if llm_router is None:
-        proxy_model_list: List[str] = []
         model_access_groups: Dict[str, List[str]] = {}
     else:
-        proxy_model_list = llm_router.get_model_names()
         model_access_groups = llm_router.get_model_access_groups()
 
     distinct_model_names = list(
@@ -6591,7 +6592,6 @@ async def apply_user_models_filter_to_deployments(
     allowed_model_names = await _apply_user_models_filter(
         all_models=distinct_model_names,
         user_api_key_dict=user_api_key_dict,
-        proxy_model_list=proxy_model_list,
         model_access_groups=model_access_groups,
         prisma_client=prisma_client,
         proxy_logging_obj=proxy_logging_obj,