[None][chore] Prevent negative max_tokens passed into tllm request (#9037)

Signed-off-by: Junyi Xu <[email protected]>

Author: JunyiXu-nv

tensorrt_llm/executor/base_worker.py

@@ -486,15 +486,13 @@ def _deduce_max_tokens(request: GenerationRequest,
             splited_prompt_len = int(len(prompt_token_ids) / cp_size)
             default_max_tokens = max_seq_len - splited_prompt_len - query_token_len
             if default_max_tokens <= 0:
-                logger.warning(
-                    f"`default_max_tokens` ({default_max_tokens}) should be greater than 0, "
+                # Raise error on `default_max_tokens` not enough, since max_tokens should be less than `default_max_tokens``
+                raise ValueError(
+                    f"`default_max_tokens` ({default_max_tokens}) must be greater than 0, "
                     f"`default_max_tokens` ({default_max_tokens}) = max_seq_len ({max_seq_len})"
                     f" - `splited_prompt_len` ({splited_prompt_len}) - `query_token_len` ({query_token_len})"
                 )
-                if max_tokens is None:
-                    raise ValueError(
-                        "`max_tokens` must be set when `default_max_tokens` is illegal"
-                    )
+
             # default_max_tokens is the biggest available value
             if max_tokens is None:
                 return default_max_tokens
@@ -504,7 +502,11 @@ def _deduce_max_tokens(request: GenerationRequest,
                     f"`default_max_tokens` ({default_max_tokens}), using default_max_tokens instead."
                 )
                 return default_max_tokens
-            return max_tokens
+            elif max_tokens <= 0:
+                raise ValueError(
+                    f"`max_tokens` ({max_tokens}) must be greater than 0")
+            else:
+                return max_tokens
 
         try:
             executor_request = tllm.Request(