[Frontend] Add --spec-method/--spec-model/--spec-tokens CLI aliases (#42476)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>

vllm/engine/arg_utils.py

@@ -607,6 +607,9 @@ class EngineArgs:
     reasoning_parser_plugin: str | None = None
 
     speculative_config: dict[str, Any] | None = None
+    spec_method: str | None = None
+    spec_model: str | None = None
+    spec_tokens: int | None = None
 
     show_hidden_metrics_for_version: str | None = (
         ObservabilityConfig.show_hidden_metrics_for_version
@@ -1435,6 +1438,12 @@ class EngineArgs:
         vllm_group.add_argument(
             "--speculative-config", "-sc", **vllm_kwargs["speculative_config"]
         )
+        speculative_kwargs = get_kwargs(SpeculativeConfig)
+        vllm_group.add_argument("--spec-method", **speculative_kwargs["method"])
+        vllm_group.add_argument("--spec-model", **speculative_kwargs["model"])
+        vllm_group.add_argument(
+            "--spec-tokens", **speculative_kwargs["num_speculative_tokens"]
+        )
         vllm_group.add_argument(
             "--kv-transfer-config", **vllm_kwargs["kv_transfer_config"]
         )
@@ -1634,6 +1643,21 @@ class EngineArgs:
         """Initializes and returns a SpeculativeConfig object based on
         `speculative_config`.
         """
+        for flag, key, value in (
+            ("--spec-method", "method", self.spec_method),
+            ("--spec-model", "model", self.spec_model),
+            ("--spec-tokens", "num_speculative_tokens", self.spec_tokens),
+        ):
+            if value is None:
+                continue
+            if self.speculative_config is None:
+                self.speculative_config = {}
+            if key in self.speculative_config:
+                raise ValueError(
+                    f"{flag} and --speculative-config['{key}'] are mutually exclusive"
+                )
+            self.speculative_config[key] = value
+
         if self.speculative_config is None:
             return None
 

vllm/entrypoints/llm.py

@@ -188,6 +188,10 @@ class LLM(PoolingOfflineMixin):
             dictionary or an AttentionConfig instance. If a dictionary, it will
             be converted to an AttentionConfig. Allows specifying the attention
             backend and other attention-related settings.
+        spec_method: Top-level alias for `speculative_config["method"]`.
+        spec_model: Top-level alias for `speculative_config["model"]`.
+        spec_tokens: Top-level alias for
+            `speculative_config["num_speculative_tokens"]`.
         **kwargs: Arguments for [`EngineArgs`][vllm.EngineArgs].
 
     Note:
@@ -236,6 +240,9 @@ class LLM(PoolingOfflineMixin):
         compilation_config: int | dict[str, Any] | CompilationConfig | None = None,
         quantization_config: dict[str, Any] | QuantizationConfigArgs | None = None,
         logits_processors: list[str | type[LogitsProcessor]] | None = None,
+        spec_method: str | None = None,
+        spec_model: str | None = None,
+        spec_tokens: int | None = None,
         **kwargs: Any,
     ) -> None:
         """LLM constructor."""
@@ -357,6 +364,9 @@ class LLM(PoolingOfflineMixin):
             compilation_config=compilation_config_instance,
             quantization_config=quantization_config,
             logits_processors=logits_processors,
+            spec_method=spec_method,
+            spec_model=spec_model,
+            spec_tokens=spec_tokens,
             **kwargs,
         )