From e426575e23f11abc4f6b1e214fc9dc2503026769 Mon Sep 17 00:00:00 2001
From: Tai An <antai12232931@outlook.com>
Date: Fri, 24 Jul 2026 21:08:56 -0700
Subject: [PATCH] fix(model): match inline-load requests by resolved path so
 symlink aliases don't reload

An inline model request identifies the loaded model by comparing the requested name against container.model_dir.name. Because the container is created from model_path.resolve(), model_dir.name is always the real directory name. When a model is addressed through a symlink alias (e.g. a 'coder' symlink to 'ArtusDev_Qwen_...-EXL3'), the names never match, so the already-loaded model is needlessly unloaded and reloaded on every request.

Compare resolved paths instead of bare directory names in both the inline fast-path guard and load_model_gen, so a symlink that points at the loaded model is recognized as the same model.

Fixes #379
---
 common/model.py                |  9 +++++++--
 endpoints/OAI/utils/common_.py | 12 ++++++++----
 2 files changed, 15 insertions(+), 6 deletions(-)

diff --git a/common/model.py b/common/model.py
index 871a52a7..91082b00 100644
--- a/common/model.py
+++ b/common/model.py
@@ -145,7 +145,12 @@ async def load_model_gen(model_path: pathlib.Path, **kwargs):
         if container and container.model:
             loaded_model_name = container.model_dir.name
 
-            if loaded_model_name == model_path.name and container.loaded:
+            # Compare resolved paths, not just directory names, so a symlink
+            # alias pointing at the already-loaded model is recognized as the
+            # same model instead of triggering a spurious unload/reload (#379).
+            already_loaded = container.model_dir.resolve() == model_path.resolve()
+
+            if already_loaded and container.loaded:
                 xlogger.info(f'Model "{loaded_model_name}" is already loaded')
 
                 # Emit a terminal progress event so API clients always
@@ -320,4 +325,4 @@ def check_context_length(
             container.validate_context_length(prompt, params, mm_embeddings)
     except ContextLengthExceededError as exc:
         error_message = handle_request_error(str(exc), exc_info=False).error.message
-        raise ContextLengthHTTPException(error_message) from exc
+        raise ContextLengthHTTPException(error_message) from exc
\ No newline at end of file
diff --git a/endpoints/OAI/utils/common_.py b/endpoints/OAI/utils/common_.py
index 3a68ea8c..46710099 100644
--- a/endpoints/OAI/utils/common_.py
+++ b/endpoints/OAI/utils/common_.py
@@ -62,9 +62,13 @@ def aggregate_usage_stats(usage_stats_list: list[UsageStats]) -> UsageStats:
 async def load_inline_model(model_name: str, request: Request):
     """Load a model from the data.model parameter"""
 
-    # Return if the model container already exists and the model is fully loaded
-    if model.container and model.container.model_dir.name == model_name and model.container.loaded:
-        return
+    # Return if the model container already exists and the model is fully loaded.
+    # Compare resolved paths so a symlink alias for the loaded model is treated
+    # as already loaded instead of triggering a spurious reload (#379).
+    if model.container and model.container.loaded:
+        requested_dir = pathlib.Path(config.model.model_dir) / model_name
+        if model.container.model_dir.resolve() == requested_dir.resolve():
+            return
 
     # Return if inline loading is disabled
     # Also warn if an admin key is used
@@ -113,4 +117,4 @@ async def load_inline_model(model_name: str, request: Request):
     await model.load_model(
         model_path,
         draft_model=config.draft_model.model_dump(include={"draft_model_dir"}),
-    )
+    )
\ No newline at end of file