Upload export_apochat_litert.py with huggingface_hub

export_apochat_litert.py

@@ -221,12 +221,13 @@ def dequantize_mlx_to_pytorch(
         if name.endswith(".scales") or name.endswith(".biases"):
             continue
 
-        base = name
-        is_quantized = base in quantized
+        # The quantized group is the key prefix without the final `.weight`.
+        group_base = name[: -len(".weight")] if name.endswith(".weight") else name
+        is_quantized = group_base in quantized
 
         if is_quantized:
-            scales = weights[f"{base}.scales"]
-            biases = weights[f"{base}.biases"]
+            scales = weights[f"{group_base}.scales"]
+            biases = weights[f"{group_base}.biases"]
             # Dequantize to bfloat16 on the MLX device.
             arr = mx.dequantize(arr, scales, biases, group_size=64, bits=4).astype(mx.bfloat16)
         elif arr.dtype != mx.bfloat16: