[Fix] Fix FP8 Online Quantization (#26415)

Co-authored-by: yichiche@amd.com <jacky.cheng>

python/sglang/multimodal_gen/runtime/layers/quantization/fp8.py

@@ -75,7 +75,11 @@ logger = logging.getLogger(__name__)
 
 
 class Fp8Config(QuantizationConfig):
-    """Config class for FP8."""
+    """Config class for FP8.
+
+    No-arg ``Fp8Config()`` selects online (post-load) weight quantization:
+    ``is_checkpoint_fp8_serialized=False`` with ``activation_scheme="dynamic"``.
+    """
 
     def __init__(
         self,

python/sglang/multimodal_gen/runtime/layers/quantization/mxfp4.py

@@ -44,7 +44,8 @@ class Mxfp4Config(QuantizationConfig):
     """
     MXFP4 quantization config for diffusion models.
 
-    Supports online quantization from unquantized BF16/FP16 checkpoints.
+    Supports online quantization from unquantized BF16/FP16 checkpoints;
+    no-arg ``Mxfp4Config()`` selects that online (post-load) path.
     Note: MXFP4 requires ROCm and MI350+ (gfx95x).
     """
 

python/sglang/multimodal_gen/runtime/loader/transformer_load_utils.py

@@ -503,12 +503,16 @@ def _resolve_quant_config(
         )
 
         # modelslim requires a per-layer quant description file; load it from
-        # the component directory rather than returning an empty config.
+        # the component directory rather than constructing an empty config.
         if server_args.quantization == "modelslim":
             return get_quant_config(hf_config, component_model_path)
 
+        # Online-quant convention: for `fp8` and `mxfp4`, a no-arg
+        # QuantizationConfig() selects the post-load path -- weights load
+        # in source dtype and are quantized in
+        # process_weights_after_loading.
         quant_cls = get_quantization_config(server_args.quantization)
-        return quant_cls.from_config({})
+        return quant_cls()
 
     quant_config = get_quant_config(hf_config, component_model_path)
     if quant_config is None and server_args.transformer_weights_path:

python/sglang/multimodal_gen/runtime/server_args.py

@@ -1331,9 +1331,11 @@ class ServerArgs(DisaggArgsMixin):
             help=(
                 "Quantization method for the transformer. If omitted, the method is "
                 "auto-detected from the checkpoint config or safetensors metadata when "
-                "possible. Applies to both pre-quantized checkpoints and online "
+                "possible. Use this flag to override auto-detection. "
-                "quantization. Use this flag to override auto-detection. "
+                "Online (post-load) quantization from a BF16/FP16 checkpoint "
-                "Options: 'fp8', 'mxfp8', 'mxfp4', 'mxfp4_npu', 'modelslim'. "
+                "is supported for 'fp8' and 'mxfp4'. Other methods "
+                "('modelopt', 'modelopt_fp8', 'modelopt_fp4', 'mxfp8', "
+                "'mxfp4_npu', 'modelslim') require a pre-quantized checkpoint. "
                 "Note: 'mxfp4' targets ROCm + MI350+ (gfx95x); "
                 "'mxfp4_npu' / 'mxfp8' target Ascend NPU (A5 series for mxfp4_npu)."
             ),