compressed-tensors 0.12.3a20251023__tar.gz → 0.12.3a20251028__tar.gz

{compressed_tensors-0.12.3a20251023/src/compressed_tensors.egg-info → compressed_tensors-0.12.3a20251028}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.3a20251023
+Version: 0.12.3a20251028
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/vllm-project/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.12.3a20251023 → compressed_tensors-0.12.3a20251028}/src/compressed_tensors/quantization/quant_args.py

@@ -25,6 +25,7 @@ from pydantic import BaseModel, ConfigDict, Field, field_validator, model_valida
 __all__ = [
     "FP8_E4M3_DATA",
     "FP4_E2M1_DATA",
+    "BFLOAT16_DATA",
     "FloatArgs",
     "QuantizationType",
     "QuantizationStrategy",
@@ -38,9 +39,9 @@ __all__ = [
 class FloatArgs:
     exponent: int
     mantissa: int
-    bits: int
-    max: float
-    min: float
+    bits: Optional[int] = None
+    max: Optional[float] = None
+    min: Optional[float] = None
     dtype: Optional[torch.dtype] = None
 
 
@@ -76,6 +77,11 @@ class FP8_E4M3_DATA(FloatArgs):
     dtype = torch.float8_e4m3fn
 
 
+class BFLOAT16_DATA(FloatArgs):
+    exponent = 8
+    mantissa = 7
+
+
 class QuantizationType(str, Enum):
     """
     Enum storing quantization type options

{compressed_tensors-0.12.3a20251023 → compressed_tensors-0.12.3a20251028}/src/compressed_tensors/quantization/utils/__init__.py

@@ -14,3 +14,4 @@
 
 # flake8: noqa
 from .helpers import *
+from .mxfp4_utils import *

compressed_tensors-0.12.3a20251028/src/compressed_tensors/quantization/utils/mxfp4_utils.py

@@ -0,0 +1,97 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from compressed_tensors.quantization.quant_args import BFLOAT16_DATA, FP4_E2M1_DATA
+
+
+__all__ = ["convert_mxfp4_exp_scale", "generate_mxfp4_scales", "round_to_power_2"]
+
+# Reference: https://github.com/vllm-project/vllm/blob/main/tests/quantization/reference_mxfp4.py # noqa: E501
+
+
+def convert_mxfp4_exp_scale(
+    scale: torch.Tensor, dtype: torch.dtype = torch.bfloat16
+) -> torch.Tensor:
+    """
+    Converts mxfp4 scales. Scales are powers of 2, with the
+    exponents stored in uint8. Converts to dense dtype so that
+    they can be applied to the weights and activations during QDQ
+
+    :param scale: uint8 exponent scale
+    :param dtype: dense dtype
+    """
+    assert scale.dtype == torch.uint8
+    scale_exp = scale.to(torch.int32) - 127
+    scale = 2.00 ** (scale_exp.to(torch.float))
+    return scale.to(dtype)
+
+
+def round_to_power_2(x: torch.Tensor) -> torch.Tensor:
+    """
+    Round values to the closest power of 2.
+    This is done by masking the values with BFLOAT16_SIGN_EXPONENT_MASK
+    which essentially removes the mantissa and keeps the exponent.
+    i.e the closest power of 2 for the input_value.
+
+    E.g:
+        0.0825 = 1.32 (mantissa) x 2**-4 (exponent)
+        0.0825 ==> -4 (exponent) + 127 = 123 = 01111011 (8 bits for bfloat16)
+        0.0825 ==> 0.32 (mantissa) = 0101001 (7 bits for bfloat16)
+        0.0825 == 0b01111011_0101001 (bfloat16)
+        0b01111011_0101001 & 111111111_0000000 == 0b01111011_0000000
+        Keep the exponent + sign bit to give you the closest power of 2, 0.0625
+
+    :param x: tensor to round to closest power of 2
+    """
+    assert x.dtype == torch.bfloat16
+    x = x.view(torch.uint16).to(torch.int32)
+
+    # Find closest power of 2
+    BFLOAT16_VAL_TO_ADD = 1 << (BFLOAT16_DATA.mantissa - FP4_E2M1_DATA.mantissa - 1)
+    # Add value to push the value to the next exponent
+    BFLOAT16_SIGN_EXPONENT_MASK = (
+        (1 << (BFLOAT16_DATA.exponent + 1)) - 1
+    ) << BFLOAT16_DATA.mantissa
+    # mask to only keep exponent - we conservatively round down
+    # to better represent smaller numbers / prevent overflow
+    block_max_uint = torch.bitwise_and(
+        x + BFLOAT16_VAL_TO_ADD, BFLOAT16_SIGN_EXPONENT_MASK
+    )
+    return block_max_uint.to(torch.uint16).view(torch.bfloat16)
+
+
+def generate_mxfp4_scales(x: torch.Tensor) -> torch.Tensor:
+    """
+    Generate mxfp4 scales. The scales require the following steps
+    1. Round to the closest power of 2
+    2. Convert to exponent
+    3. Store in uint8
+
+    Called when calculating qparams using observers.
+
+    :param x: tensor to round to closest power of 2
+    :returns uint8 scales as exponents
+    """
+    # Round to closest power of 2
+    scale_power_2 = round_to_power_2(x)
+    # Convert to exponent
+    scale_exp = 127 + torch.floor(torch.log2(scale_power_2)).to(torch.int32) - 2
+    # Clamp and store in uint8, as expected by mxfp4
+    scale_exp = torch.clamp(
+        scale_exp,
+        max=torch.iinfo(torch.uint8).max,
+        min=torch.iinfo(torch.uint8).min,
+    )
+    return scale_exp.to(torch.uint8)

{compressed_tensors-0.12.3a20251023 → compressed_tensors-0.12.3a20251028}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.12.3.a20251023'
+__version__ = version = '0.12.3.a20251028'
 __version_tuple__ = version_tuple = (0, 12, 3)

{compressed_tensors-0.12.3a20251023 → compressed_tensors-0.12.3a20251028/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.3a20251023
+Version: 0.12.3a20251028
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/vllm-project/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.12.3a20251023 → compressed_tensors-0.12.3a20251028}/src/compressed_tensors.egg-info/SOURCES.txt

@@ -75,6 +75,7 @@ src/compressed_tensors/quantization/lifecycle/helpers.py
 src/compressed_tensors/quantization/lifecycle/initialize.py
 src/compressed_tensors/quantization/utils/__init__.py
 src/compressed_tensors/quantization/utils/helpers.py
+src/compressed_tensors/quantization/utils/mxfp4_utils.py
 src/compressed_tensors/registry/__init__.py
 src/compressed_tensors/registry/registry.py
 src/compressed_tensors/transform/__init__.py
@@ -142,6 +143,7 @@ tests/test_quantization/test_configs/__init__.py
 tests/test_quantization/test_configs/test_bit_depths.py
 tests/test_quantization/test_configs/test_strategies.py
 tests/test_quantization/test_utils/test_helpers.py
+tests/test_quantization/test_utils/test_mxfp4_utils.py
 tests/test_transform/conftest.py
 tests/test_transform/test_transform_args.py
 tests/test_transform/test_transform_config.py

compressed_tensors-0.12.3a20251028/tests/test_quantization/test_utils/test_mxfp4_utils.py

@@ -0,0 +1,79 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from compressed_tensors.quantization.utils import (
+    convert_mxfp4_exp_scale,
+    generate_mxfp4_scales,
+    round_to_power_2,
+)
+
+
+def test_round_power_2_noise():
+    powers = torch.Tensor(
+        [
+            [2**-10, 2**-9, 2**-8, 2**-7, 2**-6],
+            [2**-5, 2**-4, 2**-3, 2**-2, 2**-1],
+            [2**0, 2**1, 2**-10, 2**-9, 2**-8],
+            [2**-7, 2**-6, 2**-5, 2**-4, 2**-3],
+            [2**-2, 2**-1, 2**0, 2**1, 2**-10],
+        ]
+    ).to(torch.bfloat16)
+
+    noise = torch.rand_like(powers) * 0.2
+    powers_noisy = powers * (1 + noise)
+    rounded = round_to_power_2(powers_noisy)
+    assert torch.equal(rounded, powers)
+
+
+def test_round_power_2():
+    x = torch.Tensor(
+        (
+            [5.687891, -8.291567, -1.540329, -0.315635, 0.965272],
+            [-6.944130, 0.073246, -0.451778, 8.571118, -9.856593],
+            [-0.040571, -0.708509, 2.485657, -4.003352, -0.995600],
+            [0.224199, 5.032586, -1.309816, -0.621958, 7.290238],
+            [-9.848001, -0.290731, 1.501562, 0.379829, -5.312081],
+        )
+    ).to(torch.bfloat16)
+    x_rounded = torch.Tensor(
+        (
+            [4.000000, -8.000000, -1.000000, -0.250000, 1.000000],
+            [-4.000000, 0.062500, -0.500000, 8.000000, -8.000000],
+            [-0.0312, -0.500000, 2.000000, -4.000000, -1.000000],
+            [0.250000, 4.000000, -1.000000, -0.500000, 8.000000],
+            [-8.000000, -0.250000, 1.000000, 0.250000, -4.000000],
+        )
+    ).to(torch.bfloat16)
+    rounded = round_to_power_2(x)
+    assert torch.equal(rounded, x_rounded)
+
+
+def test_mxfp4_scales_e2e():
+    mock_weight = torch.normal(mean=0.0002, std=0.0576, size=(2880, 2880))
+
+    x = mock_weight.reshape(*mock_weight.shape[:-1], -1, 32).to(torch.bfloat16)
+    min_vals = torch.amin(x, dim=-1)
+    max_vals = torch.amax(x, dim=-1)
+
+    min_vals = torch.min(min_vals, torch.zeros_like(min_vals))
+    max_vals = torch.max(max_vals, torch.zeros_like(max_vals))
+    block_max = torch.max(torch.abs(min_vals), torch.abs(max_vals))
+
+    scales_generated = generate_mxfp4_scales(block_max)
+    converted_ct = convert_mxfp4_exp_scale(scales_generated)
+
+    scales_exp = torch.log2(converted_ct)
+    block_max_exp = torch.floor(torch.log2(round_to_power_2(block_max))) - 2
+    assert torch.equal(scales_exp, block_max_exp)