[refactor] Migrate the first override families: Mistral/Pixtral dtype, MiniMaxM2, MiMoV2 (stack 7/15) (#30069)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

python/sglang/srt/arg_groups/overrides.py

@@ -29,13 +29,21 @@ Two declaration forms, keyed on ``hf_config.architectures[0]``:
 from __future__ import annotations
 
 import dataclasses
+import logging
 from typing import Any, Callable, Dict, Iterable, List, Optional, Sequence, Tuple
 
 from sglang.srt.arg_groups.arg_utils import model_overridable_fields
 from sglang.srt.runtime_context import resolve_flag_leaf
 
+logger = logging.getLogger(__name__)
+
 # Constant per-architecture overrides (populated by the migration sweeps).
-MODEL_OVERRIDES: Dict[str, Dict[str, Any]] = {}
+MODEL_OVERRIDES: Dict[str, Dict[str, Any]] = {
+    # These models run in bfloat16 regardless of the requested dtype
+    # (faithful port of the legacy unconditional arch branch).
+    "MistralLarge3ForCausalLM": {"dtype": "bfloat16"},
+    "PixtralForConditionalGeneration": {"dtype": "bfloat16"},
+}
 
 # Derived per-architecture override providers, in registration order.
 _MODEL_OVERRIDE_FNS: Dict[str, List[Callable[..., dict]]] = {}
@@ -83,6 +91,41 @@ def collect_model_override_declarations(
     return declarations
 
 
+# ---------------------------------------------------------------------------
+# Derived per-family declarations (faithful ports of legacy arch branches).
+# Callables read the PRISTINE server_args, never write; logging is kept
+# verbatim from the legacy branch for operator-visible fidelity.
+# ---------------------------------------------------------------------------
+
+
+def _register_for(*architectures: str):
+    """Register one provider for several architectures (family lists)."""
+
+    def decorator(fn: Callable[..., dict]) -> Callable[..., dict]:
+        for architecture in architectures:
+            register_model_override(architecture)(fn)
+        return fn
+
+    return decorator
+
+
+# Keep in sync with MIMO_V2_MODEL_ARCHS (server_args.py / configs/hf_config.py).
+@_register_for("MiMoV2ForCausalLM", "MiMoV2FlashForCausalLM")
+def _mimo_v2_overrides(server_args: Any, hf_config: Any) -> dict:
+    if server_args.speculative_algorithm == "EAGLE":
+        logger.info("Enable multi-layer EAGLE speculative decoding for MiMoV2 model.")
+        return {"enable_multi_layer_eagle": True}
+    return {}
+
+
+@_register_for("MiniMaxM2ForCausalLM")
+def _minimax_m2_overrides(server_args: Any, hf_config: Any) -> dict:
+    logger.info(
+        "Enable TF32 matmul for MiniMaxM2ForCausalLM model to improve gate gemm performance."
+    )
+    return {"enable_tf32_matmul": True}
+
+
 @dataclasses.dataclass(frozen=True)
 class OverrideRecord:
     """Provenance of one resolved write: ``base`` is the value before this

python/sglang/srt/runtime_context.py

@@ -306,6 +306,13 @@ class Flags(_StaticFlags):
     moe: MoeFlags = dataclasses.field(default_factory=MoeFlags)
     capture: CaptureFlags = dataclasses.field(default_factory=CaptureFlags)
 
+    # -- resolved config leaves (flat; materialized at publish) --------------
+    # Pristine user requests stay on the matching server_args fields; these
+    # leaves carry the model-resolved values.
+    dtype: str = "auto"
+    enable_tf32_matmul: bool = False
+    enable_multi_layer_eagle: bool = False
+
     def freeze(self) -> None:
         for field in dataclasses.fields(self):
             value = getattr(self, field.name)

python/sglang/srt/server_args.py

@@ -572,6 +572,7 @@ class ServerArgs:
                 '* "float32" for FP32 precision.'
             ),
             choices=["auto", "half", "float16", "bfloat16", "float", "float32"],
+            model_overridable=True,
         ),
     ] = "auto"
     quantization: A[
@@ -653,7 +654,10 @@ class ServerArgs:
     ] = None  # For flash_rl load format
     enable_tf32_matmul: A[
         bool,
-        "Enable float32 matmuls to use TensorFloat32 precision for better performance (via torch.set_float32_matmul_precision). CUDA only.",
+        Arg(
+            help="Enable float32 matmuls to use TensorFloat32 precision for better performance (via torch.set_float32_matmul_precision). CUDA only.",
+            model_overridable=True,
+        ),
     ] = False
 
     # -------------------------------------------------------------------------
@@ -1595,7 +1599,10 @@ class ServerArgs:
     ] = False
     enable_multi_layer_eagle: A[
         bool,
-        "Enable multi-layer Eagle speculative decoding.",
+        Arg(
+            help="Enable multi-layer Eagle speculative decoding.",
+            model_overridable=True,
+        ),
     ] = False
     speculative_adaptive: A[
         bool,
@@ -3741,12 +3748,6 @@ class ServerArgs:
         )
         apply_declarations_to_server_args(self, self._resolved_overrides)
 
-        if model_arch in [
-            "MistralLarge3ForCausalLM",
-            "PixtralForConditionalGeneration",
-        ]:
-            self.dtype = "bfloat16"
-
         if model_arch in [
             "DeepseekV4ForCausalLM",
         ]:
@@ -4205,11 +4206,8 @@ class ServerArgs:
                         f"attention TP size is {expected_attn_tp_size}."
                     )
 
-            if self.speculative_algorithm == "EAGLE":
-                self.enable_multi_layer_eagle = True
-                logger.info(
-                    "Enable multi-layer EAGLE speculative decoding for MiMoV2 model."
-                )
+            # enable_multi_layer_eagle for EAGLE moved to the override registry
+            # (arg_groups/overrides.py: _mimo_v2_overrides).
 
             if self.enable_hierarchical_cache:
                 if not envs.SGLANG_ENABLE_UNIFIED_RADIX_TREE.get():
@@ -4518,11 +4516,8 @@ class ServerArgs:
         elif model_arch in ["ZayaForCausalLM"]:
             self._handle_mamba_radix_cache(model_arch=model_arch)
 
-        elif model_arch in ["MiniMaxM2ForCausalLM"]:
-            self.enable_tf32_matmul = True
-            logger.info(
-                "Enable TF32 matmul for MiniMaxM2ForCausalLM model to improve gate gemm performance."
-            )
+        # MiniMaxM2ForCausalLM (enable_tf32_matmul) moved to the override registry
+        # (arg_groups/overrides.py: _minimax_m2_overrides).
 
         if (
             model_arch in ["Qwen3VLForConditionalGeneration"]