enforce user to input max_num_batched_tokens, max_num_seqs, max_model_len to reduce chance of perf degradation

Signed-off-by: Chenyaaang <[email protected]>

Author: Chenyaaang

examples/online_serving/chart-helm/values.yaml

@@ -8,7 +8,7 @@ image:
   # -- Image tag
   tag: "latest"
   # -- Container launch command
-  command: ["vllm", "serve", "/data/", "--served-model-name", "opt-125m", "--dtype", "bfloat16", "--host", "0.0.0.0", "--port", "8000"]
+  command: ["vllm", "serve", "/data/", "--served-model-name", "opt-125m", "--dtype", "bfloat16", "--host", "0.0.0.0", "--port", "8000", "--max-num-batched-tokens", "2048", "--max-num-seqs", "16", "--max-model-len", "2048"]
 
 # -- Container port
 containerPort: 8000

vllm/entrypoints/openai/cli_args.py

@@ -289,6 +289,19 @@ def validate_parsed_serve_args(args: argparse.Namespace):
         raise TypeError("Error: --enable-reasoning requires "
                         "--reasoning-parser")
 
+    # Ensure that --max-num-batched-tokens, --max-num-seqs, --max-model-len
+    # are passed within command on TPU.
+    from vllm.platforms import current_platform
+    if current_platform.is_tpu():
+        if args.max_num_batched_tokens is None:
+            raise ValueError("Requires --max-num-batched-tokens")
+
+        if args.max_num_seqs is None:
+            raise ValueError("Requires --max-num-seqs")
+
+        if args.max_model_len is None:
+            raise ValueError("Requires --max-model-len")
+
 
 def create_parser_for_docs() -> FlexibleArgumentParser:
     parser_for_docs = FlexibleArgumentParser(