[MXFP4] Add MXFP4 Compressor (#502)

* add mxfp4_formaat

* fix name

Author: dsikka

src/compressed_tensors/compressors/quantized_compressors/__init__.py

@@ -14,6 +14,6 @@
 # flake8: noqa
 
 from .base import *
+from .fp4_quantized import *
 from .naive_quantized import *
-from .nvfp4_quantized import *
 from .pack_quantized import *

src/compressed_tensors/compressors/quantized_compressors/nvfp4_quantized.py renamed to src/compressed_tensors/compressors/quantized_compressors/fp4_quantized.py

@@ -123,6 +123,15 @@ def decompress_weight(
         return decompressed_weight
 
 
+@BaseCompressor.register(name=CompressionFormat.mxfp4_pack_quantized.value)
+class MXFP4PackedCompressor(NVFP4PackedCompressor):
+    """
+    Alias for mxfp4 quantized models
+    """
+
+    pass
+
+
 @torch.compile(fullgraph=True, dynamic=True)
 def pack_fp4_to_uint8(x: torch.Tensor) -> torch.Tensor:
     """

src/compressed_tensors/config/base.py

@@ -34,6 +34,7 @@ class CompressionFormat(Enum):
     marlin_24 = "marlin-24"
     mixed_precision = "mixed-precision"
     nvfp4_pack_quantized = "nvfp4-pack-quantized"
+    mxfp4_pack_quantized = "mxfp4-pack-quantized"
 
 
 @unique

tests/test_compressors/quantized_compressors/test_nvfp4_quant.py renamed to tests/test_compressors/quantized_compressors/test_fp4_quant.py

@@ -14,7 +14,7 @@
 
 import pytest
 import torch
-from compressed_tensors.compressors.quantized_compressors.nvfp4_quantized import (
+from compressed_tensors.compressors.quantized_compressors.fp4_quantized import (
     pack_fp4_to_uint8,
     unpack_fp4_from_uint8,
 )