Update docs (#1839)

python/sglang/srt/mem_cache/flush_cache.py

@@ -29,5 +29,5 @@ if __name__ == "__main__":
     parser.add_argument("--url", type=str, default="http://localhost:30000")
     args = parser.parse_args()
 
-    response = requests.get(args.url + "/flush_cache")
+    response = requests.post(args.url + "/flush_cache")
     assert response.status_code == 200

python/sglang/srt/model_executor/model_runner.py

@@ -124,7 +124,7 @@ class ModelRunner:
                 "Automatically turn off --chunked-prefill-size and adjust --mem-fraction-static for multimodal models."
             )
             server_args.chunked_prefill_size = None
-            server_args.mem_fraction_static *= 0.95
+            self.mem_fraction_static *= 0.95
             # TODO: qwen2-vl does not support radix cache now, set disable_radix_cache=True automatically
             if self.model_config.hf_config.architectures == [
                 "Qwen2VLForConditionalGeneration"

python/sglang/srt/server.py

@@ -139,7 +139,7 @@ async def get_server_args():
     return dataclasses.asdict(tokenizer_manager.server_args)
 
 
-@app.get("/flush_cache")
+@app.post("/flush_cache")
 async def flush_cache():
     """Flush the radix cache."""
     tokenizer_manager.flush_cache()
@@ -180,7 +180,7 @@ async def get_memory_pool_size():
 
         return ret
     except Exception as e:
-        return JSONResponse(
+        return ORJSONResponse(
             {"error": {"message": str(e)}}, status_code=HTTPStatus.BAD_REQUEST
         )