Re-organize CI tests (#1052)

python/sglang/srt/constrained/base_tool_cache.py

@@ -54,7 +54,7 @@ class BaseToolCache:
         return val
 
     def init_value(self, key):
-        raise NotImplementedError
+        raise NotImplementedError()
 
     def get_cache_hit_rate(self):
         if self.metrics["total"] == 0:

python/sglang/srt/managers/tp_worker.py

@@ -410,13 +410,16 @@ class ModelTpServer:
 
         # Print stats
         if self.tp_rank == 0:
-            self.tree_cache_metrics["total"] += (
-                adder.log_input_tokens + adder.log_hit_tokens
-            ) / 10**9
-            self.tree_cache_metrics["hit"] += (adder.log_hit_tokens) / 10**9
-            tree_cache_hit_rate = (
-                self.tree_cache_metrics["hit"] / self.tree_cache_metrics["total"]
-            )
+            if isinstance(self.tree_cache, RadixCache):
+                self.tree_cache_metrics["total"] += (
+                    adder.log_input_tokens + adder.log_hit_tokens
+                ) / 10**9
+                self.tree_cache_metrics["hit"] += (adder.log_hit_tokens) / 10**9
+                tree_cache_hit_rate = (
+                    self.tree_cache_metrics["hit"] / self.tree_cache_metrics["total"]
+                )
+            else:
+                tree_cache_hit_rate = 0.0
             logger.info(
                 f"[gpu={self.gpu_id}] Prefill batch. "
                 f"#new-seq: {len(can_run_list)}, "

python/sglang/srt/mem_cache/chunk_cache.py

@@ -68,7 +68,7 @@ class ChunkCache(BasePrefixCache):
         req.last_node = entry
 
     def insert(self):
-        raise NotImplementedError
+        raise NotImplementedError()
 
     def evict(self, num_tokens: int, evict_callback: Callable):
         pass

python/sglang/srt/server.py

@@ -447,6 +447,15 @@ def _wait_and_warmup(server_args, pipe_finish_writer):
         print(f"Initialization failed. warmup error: {last_traceback}", flush=True)
         sys.exit(1)
 
+    # Print warnings here
+    if server_args.disable_radix_cache and server_args.chunked_prefill_size is not None:
+        logger.warning(
+            "You set both `--disable-radix-cache` and `--chunked-prefill-size`. "
+            "This combination is an experimental feature and we noticed it can lead to "
+            "wrong generation results. If you want to use chunked prefill, it is recommended "
+            "not using `--disable-radix-cache`."
+        )
+
     logger.info("The server is fired up and ready to roll!")
     if pipe_finish_writer is not None:
         pipe_finish_writer.send("init ok")