[bug fix] fix: detect FP8 weights from safetensors header instead of ass… (#23414)

python/sglang/srt/server_args.py

@@ -43,6 +43,7 @@ from sglang.srt.utils.common import (
     get_device_sm,
     get_nvidia_driver_version,
     get_quantization_config,
+    has_fp8_weights_in_checkpoint,
     human_readable_int,
     is_blackwell_supported,
     is_cpu,
@@ -1772,13 +1773,22 @@ class ServerArgs:
                     self.quantization is None
                     and not self._quantization_explicitly_unset
                 ):
-                    # Default DeepSeek V3/R1 native FP8 when not explicitly set,
+                    # DeepSeek V3/R1 uses native FP8 MoE experts without
-                    # Because we need this condition for an assertion in
+                    # declaring it in quantization_config.  However, other
-                    # flashinfer_trtllm MoE runner backend.
+                    # models that share the same architecture class (e.g.
+                    # Moonlight-16B-A3B) are purely BF16.  Check the actual
+                    # safetensors header instead of assuming FP8 by arch name.
                     if quant_method is None and model_arch in ["DeepseekV3ForCausalLM"]:
+                        if has_fp8_weights_in_checkpoint(self.model_path):
                             self.quantization = "fp8"
                             logger.info(
-                            "Quantization not specified, default to fp8 for DeepSeek on sm100"
+                                "Detected FP8 expert weights in checkpoint, "
+                                "default to fp8 for DeepSeek on sm100"
+                            )
+                        else:
+                            logger.info(
+                                "No FP8 expert weights found in checkpoint, "
+                                "keeping bf16 for DeepSeek-arch model on sm100"
                             )
                     else:
                         self.quantization = quant_method

python/sglang/srt/utils/common.py

@@ -2704,6 +2704,54 @@ def get_quantization_config(hf_config) -> str | None:
     return None
 
 
+def has_fp8_weights_in_checkpoint(model_path: str) -> bool:
+    """Check if a model checkpoint actually contains FP8 (float8_e4m3fn) expert
+    weight tensors by reading safetensors metadata headers.
+
+    This is needed because some models (e.g. DeepSeek V3/R1) use native FP8 MoE
+    experts without declaring it in quantization_config, while other models
+    sharing the same architecture (e.g. Moonlight) are purely BF16.
+
+    Only reads the safetensors header (a few KB of JSON), not the actual weights.
+    """
+    import json
+    import struct
+
+    try:
+        index_path = os.path.join(model_path, "model.safetensors.index.json")
+        if os.path.exists(index_path):
+            with open(index_path) as f:
+                index = json.load(f)
+            weight_map = index.get("weight_map", {})
+            expert_files = {
+                v for k, v in weight_map.items() if "experts" in k and "weight" in k
+            }
+            shard_file = next(iter(expert_files), None) or next(
+                iter(set(weight_map.values())), None
+            )
+            if shard_file is None:
+                return False
+            shard_path = os.path.join(model_path, shard_file)
+        else:
+            shard_path = os.path.join(model_path, "model.safetensors")
+
+        if not os.path.exists(shard_path):
+            return False
+
+        with open(shard_path, "rb") as f:
+            header_len = struct.unpack("<Q", f.read(8))[0]
+            header = json.loads(f.read(header_len))
+
+        for key, meta in header.items():
+            if key == "__metadata__":
+                continue
+            if "experts" in key and "weight" in key:
+                return meta.get("dtype") == "F8_E4M3"
+        return False
+    except Exception:
+        return False
+
+
 def flatten_nested_list(nested_list):
     if isinstance(nested_list, list):
         return [