Add sub-byte data types: float4_e2m1fn, float6_e2m3fn, float6_e3m2fn

CHANGELOG.md

@@ -25,6 +25,8 @@ To release a new version (e.g. from `1.0.0` -> `2.0.0`):
 
 * Added new 8-bit float types following IEEE 754 convention:
   `ml_dtypes.float8_e4m3` and `ml_dtypes.float8_e3m4`.
+* Added new 4-bit and 6-bit float types:
+  `ml_dtypes.float4_e2m1fn`, `ml_dtypes.float6_e2m3fn` and `ml_dtypes.float6_e3m2fn`.
 * Fix outputs of float `divmod` and `floor_divide` when denominator is zero.
 
 ## [0.4.0] - 2024-04-1

README.md

@@ -17,6 +17,10 @@
   * `float8_e4m3fnuz`
   * `float8_e5m2`
   * `float8_e5m2fnuz`
+- Microscaling (MX) sub-byte floating point representations including:
+  * `float4_e2m1fn`
+  * `float6_e2m3fn`
+  * `float6_e3m2fn`
 - `int2`, `int4`, `uint2` and `uint4`: low precision integer types.
 
 See below for specifications of these number formats.
@@ -66,6 +70,39 @@ A `bfloat16` number is a single-precision float truncated at 16 bits.
 
 Exponent: 8, Mantissa: 7, exponent bias: 127. IEEE 754, with NaN and inf.
 
+### `float4_e2m1`
+
+Exponent: 2, Mantissa: 1, bias: 1.
+
+Extended range: no inf, no NaN.
+
+Microscaling format, 4 bits (encoding: `0bSEEM`) using byte storage (higher 4
+bits are unused). NaN representation is undefined.
+
+Possible values: [0, 0.5, 1, 1.5, 2, 3, 4, 6]
+
+### `float6_e2m3`
+
+Exponent: 2, Mantissa: 3, bias: 1.
+
+Extended range: no inf, no NaN.
+
+Microscaling format, 6 bits (encoding: `0bSEEMMM`) using byte storage (higher 2
+bits are unused). NaN representation is undefined.
+
+Possible values range: [-7.5; 7.5]
+
+### `float6_e3m2`
+
+Exponent: 3, Mantissa: 2, bias: 3.
+
+Extended range: no inf, no NaN.
+
+Microscaling format, 4 bits (encoding: `0bSEEEMM`) using byte storage (higher 2
+bits are unused). NaN representation is undefined.
+
+Possible values range: [-28; 28]
+
 ### `float8_e3m4`
 
 Exponent: 3, Mantissa: 4, bias: 3. IEEE 754, with NaN and inf.

ml_dtypes/__init__.py

@@ -17,6 +17,9 @@
     "__version__",
     "bfloat16",
     "finfo",
+    "float4_e2m1fn",
+    "float6_e2m3fn",
+    "float6_e3m2fn",
     "float8_e3m4",
     "float8_e4m3",
     "float8_e4m3b11fnuz",
@@ -36,6 +39,9 @@
 from ml_dtypes._finfo import finfo
 from ml_dtypes._iinfo import iinfo
 from ml_dtypes._ml_dtypes_ext import bfloat16
+from ml_dtypes._ml_dtypes_ext import float4_e2m1fn
+from ml_dtypes._ml_dtypes_ext import float6_e2m3fn
+from ml_dtypes._ml_dtypes_ext import float6_e3m2fn
 from ml_dtypes._ml_dtypes_ext import float8_e3m4
 from ml_dtypes._ml_dtypes_ext import float8_e4m3
 from ml_dtypes._ml_dtypes_ext import float8_e4m3b11fnuz
@@ -50,6 +56,9 @@
 import numpy as np
 
 bfloat16: Type[np.generic]
+float4_e2m1fn: Type[np.generic]
+float6_e2m3fn: Type[np.generic]
+float6_e3m2fn: Type[np.generic]
 float8_e3m4: Type[np.generic]
 float8_e4m3: Type[np.generic]
 float8_e4m3b11fnuz: Type[np.generic]

ml_dtypes/_finfo.py

@@ -17,6 +17,9 @@
 from typing import Dict
 
 from ml_dtypes._ml_dtypes_ext import bfloat16
+from ml_dtypes._ml_dtypes_ext import float4_e2m1fn
+from ml_dtypes._ml_dtypes_ext import float6_e2m3fn
+from ml_dtypes._ml_dtypes_ext import float6_e3m2fn
 from ml_dtypes._ml_dtypes_ext import float8_e3m4
 from ml_dtypes._ml_dtypes_ext import float8_e4m3
 from ml_dtypes._ml_dtypes_ext import float8_e4m3b11fnuz
@@ -27,6 +30,9 @@
 import numpy as np
 
 _bfloat16_dtype = np.dtype(bfloat16)
+_float4_e2m1fn_dtype = np.dtype(float4_e2m1fn)
+_float6_e2m3fn_dtype = np.dtype(float6_e2m3fn)
+_float6_e3m2fn_dtype = np.dtype(float6_e3m2fn)
 _float8_e3m4_dtype = np.dtype(float8_e3m4)
 _float8_e4m3_dtype = np.dtype(float8_e4m3)
 _float8_e4m3b11fnuz_dtype = np.dtype(float8_e4m3b11fnuz)
@@ -45,6 +51,33 @@ def __init__(self):
     self.smallest_subnormal = bfloat16(smallest_subnormal)
 
 
+class _Float4E2m1fnMachArLike:
+
+  def __init__(self):
+    smallest_normal = float.fromhex("0x1p0")
+    self.smallest_normal = float4_e2m1fn(smallest_normal)
+    smallest_subnormal = float.fromhex("0x0.8p0")
+    self.smallest_subnormal = float4_e2m1fn(smallest_subnormal)
+
+
+class _Float6E2m3fnMachArLike:
+
+  def __init__(self):
+    smallest_normal = float.fromhex("0x1p0")
+    self.smallest_normal = float6_e2m3fn(smallest_normal)
+    smallest_subnormal = float.fromhex("0x0.2p0")
+    self.smallest_subnormal = float6_e2m3fn(smallest_subnormal)
+
+
+class _Float6E3m2fnMachArLike:
+
+  def __init__(self):
+    smallest_normal = float.fromhex("0x1p-2")
+    self.smallest_normal = float6_e3m2fn(smallest_normal)
+    smallest_subnormal = float.fromhex("0x0.4p-2")
+    self.smallest_subnormal = float6_e3m2fn(smallest_subnormal)
+
+
 class _Float8E3m4MachArLike:
 
   def __init__(self):
@@ -110,7 +143,7 @@ def __init__(self):
 
 class finfo(np.finfo):  # pylint: disable=invalid-name,missing-class-docstring
   __doc__ = np.finfo.__doc__
-  _finfo_cache: Dict[np.dtype, np.finfo] = {}
+  _finfo_cache: Dict[type, np.finfo] = {}
 
   @staticmethod
   def _bfloat16_finfo():
@@ -157,6 +190,120 @@ def float_to_str(f):
     # pylint: enable=protected-access
     return obj
 
+  @staticmethod
+  def _float4_e2m1fn_finfo():
+    obj = object.__new__(np.finfo)
+    obj.dtype = _float4_e2m1fn_dtype
+    obj.bits = 4
+    obj.eps = 0.5
+    obj.epsneg = 0.5
+    obj.machep = -1
+    obj.negep = -1
+    obj.max = float4_e2m1fn(6.0)
+    obj.min = float4_e2m1fn(-6.0)
+    obj.nexp = 2
+    obj.nmant = 1
+    obj.iexp = obj.nexp
+    obj.maxexp = 3
+    obj.minexp = 0
+    obj.precision = 0
+    obj.resolution = float4_e2m1fn(1.0)
+    # pylint: disable=protected-access
+    obj._machar = _Float4E2m1fnMachArLike()
+    tiny = obj._machar.smallest_normal
+    if not hasattr(obj, "tiny"):
+      obj.tiny = tiny
+    if not hasattr(obj, "smallest_normal"):
+      obj.smallest_normal = tiny
+    obj.smallest_subnormal = obj._machar.smallest_subnormal
+
+    float_to_str = str
+    obj._str_tiny = float_to_str(tiny)
+    obj._str_smallest_normal = float_to_str(tiny)
+    obj._str_smallest_subnormal = float_to_str(obj.smallest_subnormal)
+    obj._str_max = float_to_str(obj.max)
+    obj._str_epsneg = float_to_str(obj.epsneg)
+    obj._str_eps = float_to_str(obj.eps)
+    obj._str_resolution = float_to_str(obj.resolution)
+    # pylint: enable=protected-access
+    return obj
+
+  @staticmethod
+  def _float6_e2m3fn_finfo():
+    obj = object.__new__(np.finfo)
+    obj.dtype = _float6_e2m3fn_dtype
+    obj.bits = 6
+    obj.eps = 0.125
+    obj.epsneg = 0.125
+    obj.machep = -3
+    obj.negep = -3
+    obj.max = float6_e2m3fn(7.5)
+    obj.min = float6_e2m3fn(-7.5)
+    obj.nexp = 2
+    obj.nmant = 3
+    obj.iexp = obj.nexp
+    obj.maxexp = 3
+    obj.minexp = 0
+    obj.precision = 0
+    obj.resolution = float4_e2m1fn(1.0)
+    # pylint: disable=protected-access
+    obj._machar = _Float6E2m3fnMachArLike()
+    tiny = obj._machar.smallest_normal
+    if not hasattr(obj, "tiny"):
+      obj.tiny = tiny
+    if not hasattr(obj, "smallest_normal"):
+      obj.smallest_normal = tiny
+    obj.smallest_subnormal = obj._machar.smallest_subnormal
+
+    float_to_str = str
+    obj._str_tiny = float_to_str(tiny)
+    obj._str_smallest_normal = float_to_str(tiny)
+    obj._str_smallest_subnormal = float_to_str(obj.smallest_subnormal)
+    obj._str_max = float_to_str(obj.max)
+    obj._str_epsneg = float_to_str(obj.epsneg)
+    obj._str_eps = float_to_str(obj.eps)
+    obj._str_resolution = float_to_str(obj.resolution)
+    # pylint: enable=protected-access
+    return obj
+
+  @staticmethod
+  def _float6_e3m2fn_finfo():
+    obj = object.__new__(np.finfo)
+    obj.dtype = _float6_e3m2fn_dtype
+    obj.bits = 6
+    obj.eps = 0.25
+    obj.epsneg = 0.125
+    obj.machep = -2
+    obj.negep = -3
+    obj.max = float6_e3m2fn(28.0)
+    obj.min = float6_e3m2fn(-28.0)
+    obj.nexp = 3
+    obj.nmant = 2
+    obj.iexp = obj.nexp
+    obj.maxexp = 5
+    obj.minexp = -2
+    obj.precision = 0
+    obj.resolution = float6_e3m2fn(1.0)
+    # pylint: disable=protected-access
+    obj._machar = _Float6E3m2fnMachArLike()
+    tiny = obj._machar.smallest_normal
+    if not hasattr(obj, "tiny"):
+      obj.tiny = tiny
+    if not hasattr(obj, "smallest_normal"):
+      obj.smallest_normal = tiny
+    obj.smallest_subnormal = obj._machar.smallest_subnormal
+
+    float_to_str = str
+    obj._str_tiny = float_to_str(tiny)
+    obj._str_smallest_normal = float_to_str(tiny)
+    obj._str_smallest_subnormal = float_to_str(obj.smallest_subnormal)
+    obj._str_max = float_to_str(obj.max)
+    obj._str_epsneg = float_to_str(obj.epsneg)
+    obj._str_eps = float_to_str(obj.eps)
+    obj._str_resolution = float_to_str(obj.resolution)
+    # pylint: enable=protected-access
+    return obj
+
   @staticmethod
   def _float8_e3m4_finfo():
     def float_to_str(f):
@@ -472,71 +619,30 @@ def float_to_str(f):
     # pylint: enable=protected-access
     return obj
 
+  _finfo_type_map = {
+    bfloat16: _bfloat16_finfo,
+    float4_e2m1fn: _float4_e2m1fn_finfo,
+    float6_e2m3fn: _float6_e2m3fn_finfo,
+    float6_e3m2fn: _float6_e3m2fn_finfo,
+    float8_e3m4: _float8_e3m4_finfo,
+    float8_e4m3: _float8_e4m3_finfo,
+    float8_e4m3fn: _float8_e4m3fn_finfo,
+    float8_e4m3fnuz: _float8_e4m3fnuz_finfo,
+    float8_e4m3b11fnuz: _float8_e4m3b11fnuz_finfo,
+    float8_e5m2: _float8_e5m2_finfo,
+    float8_e5m2fnuz: _float8_e5m2fnuz_finfo,
+  }
+  _finfo_name_map = {t.__name__: t for t in _finfo_type_map}
+
   def __new__(cls, dtype):
-    if (
-        isinstance(dtype, str)
-        and dtype == "bfloat16"
-        or dtype == _bfloat16_dtype
-    ):
-      if _bfloat16_dtype not in cls._finfo_cache:
-        cls._finfo_cache[_bfloat16_dtype] = cls._bfloat16_finfo()
-      return cls._finfo_cache[_bfloat16_dtype]
-    if (
-        isinstance(dtype, str)
-        and dtype == "float8_e3m4"
-        or dtype == _float8_e3m4_dtype
-    ):
-      if _float8_e3m4_dtype not in cls._finfo_cache:
-        cls._finfo_cache[_float8_e3m4_dtype] = cls._float8_e3m4_finfo()
-      return cls._finfo_cache[_float8_e3m4_dtype]
-    if (
-        isinstance(dtype, str)
-        and dtype == "float8_e4m3"
-        or dtype == _float8_e4m3_dtype
-    ):
-      if _float8_e4m3_dtype not in cls._finfo_cache:
-        cls._finfo_cache[_float8_e4m3_dtype] = cls._float8_e4m3_finfo()
-      return cls._finfo_cache[_float8_e4m3_dtype]
-    if (
-        isinstance(dtype, str)
-        and dtype == "float8_e4m3b11fnuz"
-        or dtype == _float8_e4m3b11fnuz_dtype
-    ):
-      if _float8_e4m3b11fnuz_dtype not in cls._finfo_cache:
-        cls._finfo_cache[_float8_e4m3b11fnuz_dtype] = (
-            cls._float8_e4m3b11fnuz_finfo()
-        )
-      return cls._finfo_cache[_float8_e4m3b11fnuz_dtype]
-    if (
-        isinstance(dtype, str)
-        and dtype == "float8_e4m3fn"
-        or dtype == _float8_e4m3fn_dtype
-    ):
-      if _float8_e4m3fn_dtype not in cls._finfo_cache:
-        cls._finfo_cache[_float8_e4m3fn_dtype] = cls._float8_e4m3fn_finfo()
-      return cls._finfo_cache[_float8_e4m3fn_dtype]
-    if (
-        isinstance(dtype, str)
-        and dtype == "float8_e4m3fnuz"
-        or dtype == _float8_e4m3fnuz_dtype
-    ):
-      if _float8_e4m3fnuz_dtype not in cls._finfo_cache:
-        cls._finfo_cache[_float8_e4m3fnuz_dtype] = cls._float8_e4m3fnuz_finfo()
-      return cls._finfo_cache[_float8_e4m3fnuz_dtype]
-    if (
-        isinstance(dtype, str)
-        and dtype == "float8_e5m2"
-        or dtype == _float8_e5m2_dtype
-    ):
-      if _float8_e5m2_dtype not in cls._finfo_cache:
-        cls._finfo_cache[_float8_e5m2_dtype] = cls._float8_e5m2_finfo()
-      return cls._finfo_cache[_float8_e5m2_dtype]
-    if (
-        isinstance(dtype, str)
-        and dtype == "float8_e5m2fnuz"
-        or dtype == _float8_e5m2fnuz_dtype
-    ):
-      if _float8_e5m2fnuz_dtype not in cls._finfo_cache:
-        cls._finfo_cache[_float8_e5m2fnuz_dtype] = cls._float8_e5m2fnuz_finfo()
-      return cls._finfo_cache[_float8_e5m2fnuz_dtype]
+    key = cls._finfo_name_map.get(dtype) if isinstance(dtype, str) else \
+      dtype.type if isinstance(dtype, np.dtype) else dtype
+    finfo = cls._finfo_cache.get(key)
+    if finfo is not None:
+      return finfo
+
+    init = cls._finfo_type_map.get(key)
+    if init is not None:
+      cls._finfo_cache[dtype] = init()
+      return cls._finfo_cache[dtype]
     return super().__new__(cls, dtype)