[Bugfix] Better error message for MLPSpeculator when `num_speculative…

…_tokens` is set too high (vllm-project#5894) Signed-off-by: Thomas Parnell <[email protected]> Signed-off-by: Alvant <[email protected]>
compressa-ai · Oct 26, 2024 · 6f23fa1 · 6f23fa1
1 parent dd01281
commit 6f23fa1
Showing 1 changed file with 3 additions and 3 deletions.
diff --git a/vllm/config.py b/vllm/config.py
@@ -956,9 +956,9 @@ def maybe_create_spec_config(
                     # Verify provided value doesn't exceed the maximum
                     # supported by the draft model.
                     raise ValueError(
-                        "Expected both speculative_model and "
-                        "num_speculative_tokens to be provided, but found "
-                        f"{speculative_model=} and {num_speculative_tokens=}.")
+                        "This speculative model supports a maximum of "
+                        f"num_speculative_tokens={n_predict}, but "
+                        f"{num_speculative_tokens=} was provided.")
 
             draft_model_config.max_model_len = (
                 SpeculativeConfig._maybe_override_draft_max_model_len(