From 7973b8cfe8fd51a3e330bcdfb617cdef3672fa72 Mon Sep 17 00:00:00 2001
From: Awni Hannun <awni@apple.com>
Date: Tue, 30 Dec 2025 09:19:36 -0800
Subject: [PATCH] allow mxfp8 and nvfp4 (#709)

---
 mlx_lm/convert.py | 12 +++++++++---
 mlx_lm/utils.py   | 20 ++++++++++++++++----
 2 files changed, 25 insertions(+), 7 deletions(-)

diff --git a/mlx_lm/convert.py b/mlx_lm/convert.py
index 09408f1..6fb2304 100644
--- a/mlx_lm/convert.py
+++ b/mlx_lm/convert.py
@@ -187,17 +187,23 @@ def configure_parser() -> argparse.ArgumentParser:
         "-q", "--quantize", help="Generate a quantized model.", action="store_true"
     )
     parser.add_argument(
-        "--q-group-size", help="Group size for quantization.", type=int, default=64
+        "--q-group-size",
+        help="Group size for quantization.",
+        type=int,
+        default=None,
     )
     parser.add_argument(
-        "--q-bits", help="Bits per weight for quantization.", type=int, default=4
+        "--q-bits",
+        help="Bits per weight for quantization.",
+        type=int,
+        default=None,
     )
     parser.add_argument(
         "--q-mode",
         help="The quantization mode.",
         type=str,
         default="affine",
-        choices=["affine", "mxfp4"],
+        choices=["affine", "mxfp4", "nvfp4", "mxfp8"],
     )
     parser.add_argument(
         "--quant-predicate",
diff --git a/mlx_lm/utils.py b/mlx_lm/utils.py
index e798816..ada948d 100644
--- a/mlx_lm/utils.py
+++ b/mlx_lm/utils.py
@@ -611,8 +611,8 @@ def save_model(
 def quantize_model(
     model: nn.Module,
     config: dict,
-    group_size: int,
-    bits: int,
+    group_size: Optional[int],
+    bits: Optional[int],
     mode: str = "affine",
     quant_predicate: Optional[Callable[[str, nn.Module], Union[bool, dict]]] = None,
 ) -> Tuple[nn.Module, dict]:
@@ -622,8 +622,8 @@ def quantize_model(
     Args:
         model (nn.Module): The model to be quantized.
         config (dict): Model configuration.
-        group_size (int): Group size for quantization.
-        bits (int): Bits per weight for quantization.
+        group_size (Optional[int]): Group size for quantization.
+        bits (Optional[int]): Bits per weight for quantization.
         mode (str): The quantization mode.
         quant_predicate (Callable): A callable that decides how to quantize
           each layer based on the path. Accepts the layer `path` and the
@@ -633,9 +633,21 @@ def quantize_model(
     Returns:
         Tuple: Tuple containing quantized model and config.
     """
+
+    def defaults_for_mode(mode, group_size, bits):
+        mode_defaults = {
+            "affine": (64, 4),
+            "mxfp4": (32, 4),
+            "nvfp4": (16, 4),
+            "mxfp8": (32, 8),
+        }
+        default_group_size, default_bits = mode_defaults[mode]
+        return group_size or default_group_size, bits or default_bits
+
     quantized_config = copy.deepcopy(config)
 
     quant_predicate = quant_predicate or getattr(model, "quant_predicate", None)
+    group_size, bits = defaults_for_mode(mode, group_size, bits)
     quant_params = {"group_size": group_size, "bits": bits, "mode": mode}
     if "quantization" in quantized_config:
         # If the model is already partially quantized, return params so that