refactor: regex-based auto-resolution for tool call and reasoning parsers

src/prime_rl/configs/inference.py

@@ -6,6 +6,7 @@
 from pydantic_config import BaseConfig
 
 from prime_rl.configs.shared import BaseModelConfig, SlurmConfig
+from prime_rl.utils.parsers import resolve_reasoning_parser, resolve_tool_call_parser
 from prime_rl.utils.utils import rgetattr, rsetattr
 
 # TODO: Set thinking/ solution budget
@@ -41,7 +42,12 @@ def __str__(self) -> str:
 
 
 class ModelConfig(BaseModelConfig):
-    """Configures the inference model. Most arguments are passed directly to the vLLM LLM class (https://docs.vllm.ai/en/latest/api/vllm.LLM.html)."""
+    """Configures the inference model. Most arguments are passed directly to the vLLM LLM class (https://docs.vllm.ai/en/latest/api/vllm.LLM.html).
+
+    Parser fields (tool_call_parser, reasoning_parser) default to "auto",
+    which resolves to a concrete parser name at validation time based on
+    the model name.  Set to None to disable.
+    """
 
     dtype: Annotated[
         Literal["auto", "float16", "bfloat16", "float32"],
@@ -82,17 +88,16 @@ class ModelConfig(BaseModelConfig):
     tool_call_parser: Annotated[
         str | None,
         Field(
-            description="The tool call parser to use. Passed to vLLM as `--tool-call-parser`. "
-            'Set to "auto" to infer from the model name.',
+            description='The tool call parser to use. Set to "auto" (default) to detect from the model name, or None to disable.',
         ),
     ] = "auto"
 
     reasoning_parser: Annotated[
         str | None,
         Field(
-            description="Parser for extracting reasoning content from model outputs. Passed to vLLM as `--reasoning-parser`. Setting this enables reasoning mode.",
+            description='Parser for extracting reasoning content from model outputs. Set to "auto" (default) to detect from the model name, or None to disable.',
         ),
-    ] = None
+    ] = "auto"
 
     rope_scaling: Annotated[
         dict[str, Any] | str | None,
@@ -101,6 +106,15 @@ class ModelConfig(BaseModelConfig):
         ),
     ] = None
 
+    @model_validator(mode="after")
+    def auto_resolve_parsers(self):
+        """Resolve "auto" parser values to concrete parser names from model name."""
+        if self.tool_call_parser == "auto":
+            self.tool_call_parser = resolve_tool_call_parser(self.name)
+        if self.reasoning_parser == "auto":
+            self.reasoning_parser = resolve_reasoning_parser(self.name)
+        return self
+
 
 class WeightBroadcastConfig(BaseConfig):
     """Configures weight broadcast settings."""
@@ -545,14 +559,19 @@ def to_vllm(self) -> Namespace:
         # Set `logprobs_mode` to `processed_logprobs` by default
         rsetattr(namespace, "logprobs_mode", "processed_logprobs")
 
-        # Remove chat_template if not set (vLLM doesn't accept None)
-        if namespace.chat_template is None:
-            delattr(namespace, "chat_template")
+        # Remove tool_call_parser if not set (vLLM doesn't accept None)
+        if namespace.tool_call_parser is None:
+            delattr(namespace, "tool_call_parser")
+        namespace.enable_auto_tool_choice = hasattr(namespace, "tool_call_parser")
 
         # Remove reasoning_parser if not set (vLLM doesn't accept None)
         if namespace.reasoning_parser is None:
             delattr(namespace, "reasoning_parser")
 
+        # Remove chat_template if not set (vLLM doesn't accept None)
+        if namespace.chat_template is None:
+            delattr(namespace, "chat_template")
+
         # Remove lora_target_modules if not set (vLLM doesn't accept None)
         if hasattr(namespace, "lora_target_modules") and namespace.lora_target_modules is None:
             delattr(namespace, "lora_target_modules")

src/prime_rl/inference/vllm/server.py

@@ -22,117 +22,6 @@
 from prime_rl.configs.inference import InferenceConfig
 from prime_rl.utils.logger import get_logger
 
-MODEL_TOOL_CALL_PARSER: dict[str, str] = {
-    # GLM-4.5
-    "zai-org/GLM-4.5": "glm45",
-    "zai-org/GLM-4.5-FP8": "glm45",
-    "zai-org/GLM-4.5-Base": "glm45",
-    "zai-org/GLM-4.5-Air": "glm45",
-    "zai-org/GLM-4.5-Air-FP8": "glm45",
-    "zai-org/GLM-4.5-Air-Base": "glm45",
-    "zai-org/GLM-4.5V": "glm45",
-    "zai-org/GLM-4.5V-FP8": "glm45",
-    # GLM-4.7
-    "zai-org/GLM-4.7": "glm47",
-    "zai-org/GLM-4.7-FP8": "glm47",
-    "zai-org/GLM-4.7-Flash": "glm47",
-    # GLM-5
-    "zai-org/GLM-5": "glm47",
-    "zai-org/GLM-5-FP8": "glm47",
-    # GLM-5.1
-    "zai-org/GLM-5.1": "glm47",
-    "zai-org/GLM-5.1-FP8": "glm47",
-    # MiniMax M2
-    "MiniMaxAI/MiniMax-M2": "minimax_m2",
-    "MiniMaxAI/MiniMax-M2.1": "minimax_m2",
-    "MiniMaxAI/MiniMax-M2.5": "minimax_m2",
-    # INTELLECT-3
-    "PrimeIntellect/INTELLECT-3": "hermes",
-    "PrimeIntellect/INTELLECT-3-FP8": "hermes",
-    "PrimeIntellect/INTELLECT-3.1": "hermes",
-    # Qwen3 dense
-    "Qwen/Qwen3-0.6B": "hermes",
-    "Qwen/Qwen3-0.6B-Base": "hermes",
-    "Qwen/Qwen3-0.6B-FP8": "hermes",
-    "Qwen/Qwen3-1.7B": "hermes",
-    "Qwen/Qwen3-1.7B-Base": "hermes",
-    "Qwen/Qwen3-1.7B-FP8": "hermes",
-    "Qwen/Qwen3-4B": "hermes",
-    "Qwen/Qwen3-4B-Base": "hermes",
-    "Qwen/Qwen3-4B-FP8": "hermes",
-    "Qwen/Qwen3-8B": "hermes",
-    "Qwen/Qwen3-8B-Base": "hermes",
-    "Qwen/Qwen3-8B-FP8": "hermes",
-    "Qwen/Qwen3-14B": "hermes",
-    "Qwen/Qwen3-14B-Base": "hermes",
-    "Qwen/Qwen3-14B-FP8": "hermes",
-    "Qwen/Qwen3-32B": "hermes",
-    "Qwen/Qwen3-32B-FP8": "hermes",
-    # Qwen3 MoE
-    "Qwen/Qwen3-30B-A3B": "hermes",
-    "Qwen/Qwen3-30B-A3B-Base": "hermes",
-    "Qwen/Qwen3-30B-A3B-FP8": "hermes",
-    "Qwen/Qwen3-235B-A22B": "hermes",
-    "Qwen/Qwen3-235B-A22B-FP8": "hermes",
-    # Qwen3 2507
-    "Qwen/Qwen3-4B-Instruct-2507": "hermes",
-    "Qwen/Qwen3-4B-Thinking-2507": "hermes",
-    "Qwen/Qwen3-4B-Instruct-2507-FP8": "hermes",
-    "Qwen/Qwen3-4B-Thinking-2507-FP8": "hermes",
-    "Qwen/Qwen3-30B-A3B-Instruct-2507": "hermes",
-    "Qwen/Qwen3-30B-A3B-Thinking-2507": "hermes",
-    "Qwen/Qwen3-30B-A3B-Instruct-2507-FP8": "hermes",
-    "Qwen/Qwen3-30B-A3B-Thinking-2507-FP8": "hermes",
-    "Qwen/Qwen3-235B-A22B-Instruct-2507": "hermes",
-    "Qwen/Qwen3-235B-A22B-Thinking-2507": "hermes",
-    "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8": "hermes",
-    "Qwen/Qwen3-235B-A22B-Thinking-2507-FP8": "hermes",
-    # Qwen3-Next
-    "Qwen/Qwen3-Next-80B-A3B-Instruct": "hermes",
-    "Qwen/Qwen3-Next-80B-A3B-Thinking": "hermes",
-    "Qwen/Qwen3-Next-80B-A3B-Instruct-FP8": "hermes",
-    "Qwen/Qwen3-Next-80B-A3B-Thinking-FP8": "hermes",
-    # Qwen3-Coder
-    "Qwen/Qwen3-Coder-480B-A35B-Instruct": "hermes",
-    "Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8": "hermes",
-    "Qwen/Qwen3-Coder-30B-A3B-Instruct": "hermes",
-    "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8": "hermes",
-    # Qwen3-Coder-Next
-    "Qwen/Qwen3-Coder-Next": "hermes",
-    "Qwen/Qwen3-Coder-Next-Base": "hermes",
-    "Qwen/Qwen3-Coder-Next-FP8": "hermes",
-    # Qwen3.5 dense (uses qwen3_coder tool format, not hermes)
-    "Qwen/Qwen3.5-0.8B": "qwen3_coder",
-    "Qwen/Qwen3.5-0.8B-Base": "qwen3_coder",
-    "Qwen/Qwen3.5-2B": "qwen3_coder",
-    "Qwen/Qwen3.5-2B-Base": "qwen3_coder",
-    "Qwen/Qwen3.5-4B": "qwen3_coder",
-    "Qwen/Qwen3.5-4B-Base": "qwen3_coder",
-    "Qwen/Qwen3.5-9B": "qwen3_coder",
-    "Qwen/Qwen3.5-9B-Base": "qwen3_coder",
-    "Qwen/Qwen3.5-27B": "qwen3_coder",
-    "Qwen/Qwen3.5-27B-FP8": "qwen3_coder",
-    # Qwen3.5 MoE (uses qwen3_coder tool format, not hermes)
-    "Qwen/Qwen3.5-35B-A3B": "qwen3_coder",
-    "Qwen/Qwen3.5-35B-A3B-Base": "qwen3_coder",
-    "Qwen/Qwen3.5-35B-A3B-FP8": "qwen3_coder",
-    "Qwen/Qwen3.5-122B-A10B": "qwen3_coder",
-    "Qwen/Qwen3.5-122B-A10B-FP8": "qwen3_coder",
-    "Qwen/Qwen3.5-397B-A17B": "qwen3_coder",
-    "Qwen/Qwen3.5-397B-A17B-FP8": "qwen3_coder",
-    # NemotronH
-    "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16": "qwen3_coder",
-    "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16": "qwen3_coder",
-}
-
-
-def resolve_tool_call_parser(model_name: str, tool_call_parser: str | None) -> str | None:
-    """Resolve tool_call_parser from model name if set to "auto"."""
-    if tool_call_parser == "auto":
-        return MODEL_TOOL_CALL_PARSER.get(model_name)
-    return tool_call_parser
-
-
 logger = get_logger()
 from prime_rl.inference.patches import (
     monkey_patch_harmony_stop_token_propagation,
@@ -340,11 +229,7 @@ def server(config: InferenceConfig, vllm_extra: dict[str, Any] | None = None):
     assert args is not None
     validate_parsed_serve_args(args)
 
-    args.tool_call_parser = resolve_tool_call_parser(args.model, args.tool_call_parser)
-    args.enable_auto_tool_choice = args.tool_call_parser is not None
     args.reset_prefix_cache_after_update = config.experimental.reset_prefix_cache_after_update
-    if args.tool_call_parser is not None:
-        logger.info(f"Using tool_call_parser='{args.tool_call_parser}' for model '{args.model}'")
 
     # Set the worker extension class based on the broadcast backend
     args.worker_extension_cls = WORKER_EXTENSION_CLS[config.weight_broadcast.type]

src/prime_rl/utils/parsers.py

@@ -0,0 +1,45 @@
+import re
+
+# (regex, parser_name) — first match wins.
+TOOL_CALL_PARSER_PATTERNS: list[tuple[re.Pattern[str], str]] = [
+    (re.compile(r"^deepseek-ai/DeepSeek-V3\.2"), "deepseek_v32"),
+    (re.compile(r"^deepseek-ai/DeepSeek-V3\.1"), "deepseek_v31"),
+    (re.compile(r"^zai-org/GLM-4\.5"), "glm45"),
+    (re.compile(r"^zai-org/GLM-4\.7"), "glm47"),
+    (re.compile(r"^zai-org/GLM-5"), "glm47"),
+    (re.compile(r"^MiniMaxAI/MiniMax-M2"), "minimax_m2"),
+    (re.compile(r"^PrimeIntellect/INTELLECT-3"), "qwen3_coder"),
+    (re.compile(r"^nvidia/NVIDIA-Nemotron-3"), "qwen3_coder"),
+    (re.compile(r"^stepfun-ai/Step-3\.5"), "step3p5"),
+    # Qwen3.5 and Qwen3-Coder use qwen3_coder — must be before the Qwen3 catch-all
+    (re.compile(r"^Qwen/Qwen3\.5-"), "qwen3_coder"),
+    (re.compile(r"^Qwen/Qwen3-Coder"), "qwen3_coder"),
+    (re.compile(r"^Qwen/Qwen3-"), "hermes"),
+]
+
+REASONING_PARSER_PATTERNS: list[tuple[re.Pattern[str], str]] = [
+    (re.compile(r"^deepseek-ai/DeepSeek-V3\.[12]"), "deepseek_r1"),
+    (re.compile(r"^zai-org/GLM-"), "glm45"),
+    (re.compile(r"^MiniMaxAI/MiniMax-M2"), "minimax_m2_append_think"),
+    (re.compile(r"^PrimeIntellect/INTELLECT-3"), "deepseek_r1"),
+    (re.compile(r"^stepfun-ai/Step-3\.5"), "step3p5"),
+    # Only Qwen3 Thinking models reason — Instruct models do not
+    (re.compile(r"^Qwen/Qwen3-.*Thinking"), "deepseek_r1"),
+    (re.compile(r"^Qwen/Qwen3\.5-"), "qwen3"),
+]
+
+
+def resolve_parser(model_name: str, patterns: list[tuple[re.Pattern[str], str]]) -> str | None:
+    """Auto-detect parser from model name. Returns the first matching pattern's parser."""
+    for pattern, parser_name in patterns:
+        if pattern.search(model_name):
+            return parser_name
+    return None
+
+
+def resolve_tool_call_parser(model_name: str) -> str | None:
+    return resolve_parser(model_name, TOOL_CALL_PARSER_PATTERNS)
+
+
+def resolve_reasoning_parser(model_name: str) -> str | None:
+    return resolve_parser(model_name, REASONING_PARSER_PATTERNS)

tests/unit/test_configs.py

@@ -253,3 +253,31 @@ def test_subconfig_seq_len_wins_over_shared():
     )
     assert config.trainer.model.seq_len == 8192
     assert config.orchestrator.seq_len == 4096
+
+
+def test_shared_model_name_resolves_inference_parser():
+    """Shared model.name propagates to inference before ModelConfig runs its parser auto-resolver."""
+    config = RLConfig.model_validate(
+        {
+            "model": {"name": "Qwen/Qwen3-Coder-30B-A3B-Instruct"},
+            "trainer": {},
+            "orchestrator": {},
+            "inference": {},
+        }
+    )
+    assert config.inference is not None
+    assert config.inference.model.name == "Qwen/Qwen3-Coder-30B-A3B-Instruct"
+    assert config.inference.model.tool_call_parser == "qwen3_coder"
+
+
+def test_explicit_inference_parser_wins_over_auto():
+    config = RLConfig.model_validate(
+        {
+            "model": {"name": "Qwen/Qwen3-Coder-30B-A3B-Instruct"},
+            "trainer": {},
+            "orchestrator": {},
+            "inference": {"model": {"tool_call_parser": "hermes"}},
+        }
+    )
+    assert config.inference is not None
+    assert config.inference.model.tool_call_parser == "hermes"