compressed-tensors 0.12.3a20251030__tar.gz → 0.12.3a20251110__tar.gz

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/.github/workflows/test-check.yaml

@@ -12,7 +12,7 @@ on:
 
 jobs:
   python-tests:
-    runs-on: ubuntu-22.04
+    runs-on: ibm-wdc-k8s-vllm-h100-solo
     env:
         HF_TOKEN: ${{ secrets.HF_RED_HAT_READ_ONLY }}
     steps:

{compressed_tensors-0.12.3a20251030/src/compressed_tensors.egg-info → compressed_tensors-0.12.3a20251110}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.3a20251030
+Version: 0.12.3a20251110
 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-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/src/compressed_tensors/compressors/quantized_compressors/base.py

@@ -90,7 +90,6 @@ class BaseQuantizationCompressor(BaseCompressor):
         desc = "Compressing with quantization"
         for name in tqdm(uncompressed_names, desc=desc, disable=(not show_progress)):
             value = model_state[name]
-
             # compress weights
             if name.endswith("weight"):
                 prefix = name.removesuffix("weight")
@@ -129,10 +128,18 @@ class BaseQuantizationCompressor(BaseCompressor):
                 if name.endswith("zero_point") and self._skip_zp(name, names_to_scheme):
                     continue
 
+                if name.endswith("weight_scale") and self._skip_scale():
+                    continue
+
                 compressed_dict[name] = value.to(compression_device)
 
         return compressed_dict
 
+    def _skip_scale(self):
+        from compressed_tensors.compressors import NVFP4PackedCompressor
+
+        return isinstance(self, NVFP4PackedCompressor)
+
     def _skip_zp(
         self, name: str, names_to_scheme: Dict[str, QuantizationScheme]
     ) -> bool:

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/src/compressed_tensors/compressors/quantized_compressors/fp4_quantized.py

@@ -26,7 +26,7 @@ from compressed_tensors.quantization.lifecycle.forward import dequantize, quanti
 from torch import Tensor
 
 
-__all__ = ["pack_fp4_to_uint8", "unpack_fp4_from_uint8"]
+__all__ = ["pack_fp4_to_uint8", "unpack_fp4_from_uint8", "NVFP4PackedCompressor"]
 
 FLOAT_TO_E2M1 = [
     0.0,
@@ -103,6 +103,7 @@ class NVFP4PackedCompressor(BaseQuantizationCompressor):
         if device is not None:
             weight_packed = weight_packed.to(device)
         compressed_dict["weight_packed"] = weight_packed
+        compressed_dict["weight_scale"] = scale.to(quantization_args.scale_dtype)
         return compressed_dict
 
     def decompress_weight(
@@ -111,8 +112,8 @@ class NVFP4PackedCompressor(BaseQuantizationCompressor):
         quantization_args: Optional[QuantizationArgs] = None,
     ) -> torch.Tensor:
         weight = compressed_data["weight_packed"]
-        scale = compressed_data["weight_scale"]
         global_scale = compressed_data["weight_global_scale"]
+        scale = compressed_data["weight_scale"]
         m, n = weight.shape
         # TODO: use a user provided dequant dtype
         unpacked = unpack_fp4_from_uint8(weight, m, n * 2)

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/src/compressed_tensors/quantization/lifecycle/forward.py

@@ -21,7 +21,7 @@ from compressed_tensors.quantization.quant_args import (
     DynamicType,
     QuantizationArgs,
     QuantizationStrategy,
-    round_to_quantized_type,
+    round_to_quantized_type_args,
 )
 from compressed_tensors.quantization.quant_config import QuantizationStatus
 from compressed_tensors.quantization.quant_scheme import QuantizationScheme
@@ -466,20 +466,17 @@ def _quantize(
     # if a global scale is optionally provided, use it
     # to further scale the local `scale` parameter
     if global_scale is not None:
-        scale = scale.to(global_scale.dtype) / global_scale
+        scale = scale / global_scale
 
     scaled = x / scale
 
     if zero_point is not None:
         scaled += zero_point.to(x.dtype)
 
-    # clamp first because cast isn't guaranteed to be saturated (ie for fp8)
-    clamped_value = torch.clamp(
-        scaled,
-        q_min,
-        q_max,
+    # clamp and round
+    quantized_value = round_to_quantized_type_args(
+        tensor=scaled, args=args, min=q_min, max=q_max
     )
-    quantized_value = round_to_quantized_type(clamped_value, args)
 
     if dtype is not None:
         quantized_value = quantized_value.to(dtype)
@@ -499,7 +496,7 @@ def _dequantize(
     # if a global scale is optionally provided, use it
     # to further scale the local `scale` parameter
     if global_scale is not None:
-        scale = scale.to(global_scale.dtype) / global_scale
+        scale = scale / global_scale
 
     dequant_value = x_q.to(scale.dtype)
 

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/src/compressed_tensors/quantization/lifecycle/initialize.py

@@ -24,7 +24,6 @@ from compressed_tensors.modeling import (
     QuantizedKVCache,
 )
 from compressed_tensors.quantization import (
-    FP8_E4M3_DATA,
     ActivationOrdering,
     DynamicType,
     QuantizationArgs,
@@ -36,7 +35,7 @@ from compressed_tensors.quantization import (
 from compressed_tensors.quantization.lifecycle.forward import (
     wrap_module_forward_quantized,
 )
-from compressed_tensors.quantization.utils import is_fp4, strategy_cdiv
+from compressed_tensors.quantization.utils import strategy_cdiv
 from compressed_tensors.utils import (
     disable_hf_hook,
     get_execution_device,
@@ -250,20 +249,13 @@ def initialize_qparams(
 
     # 2. Identify quantization scale and zp dtype
     scale_dtype = observed_dtype
-
-    if is_fp4(quantization_args=quantization_args):
-        scale_dtype = zp_dtype = FP8_E4M3_DATA.dtype
-    else:
-        # TODO: consider erroring out in the future as if the dtype if not one of these,
-        # there is likely bug
-        if scale_dtype not in [
-            torch.float16,
-            torch.bfloat16,
-            torch.float32,
-            torch.float64,
-        ]:
-            scale_dtype = torch.bfloat16
-        zp_dtype = quantization_args.pytorch_dtype()
+    if scale_dtype not in [
+        torch.float16,
+        torch.bfloat16,
+        torch.float32,
+        torch.float64,
+    ]:
+        scale_dtype = torch.float16
 
     # 3. Initializes scale/zp for the module
     init_scale = Parameter(
@@ -274,7 +266,9 @@ def initialize_qparams(
 
     if force_zero_point or not quantization_args.symmetric:
         init_zero_point = Parameter(
-            torch.zeros(expected_shape, device=device, dtype=zp_dtype),
+            torch.zeros(
+                expected_shape, device=device, dtype=quantization_args.zp_dtype
+            ),
             requires_grad=False,
         )
         register_offload_parameter(module, f"{base_name}_zero_point", init_zero_point)

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/src/compressed_tensors/quantization/quant_args.py

@@ -19,7 +19,15 @@ from typing import Any, Dict, List, Optional, Union
 import torch
 from compressed_tensors.utils import Aliasable
 from compressed_tensors.utils.helpers import deprecated
-from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
+from compressed_tensors.utils.type import TorchDtype
+from pydantic import (
+    BaseModel,
+    ConfigDict,
+    Field,
+    field_serializer,
+    field_validator,
+    model_validator,
+)
 
 
 __all__ = [
@@ -30,7 +38,8 @@ __all__ = [
     "QuantizationType",
     "QuantizationStrategy",
     "QuantizationArgs",
-    "round_to_quantized_type",
+    "round_to_quantized_type_args",
+    "round_to_quantized_type_dtype",
     "ActivationOrdering",
     "DynamicType",
 ]
@@ -174,6 +183,8 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
     block_structure: Optional[List[int]] = None
     dynamic: Union[DynamicType, bool] = False
     actorder: Union[ActivationOrdering, bool, None] = None
+    scale_dtype: Optional[TorchDtype] = None
+    zp_dtype: Optional[TorchDtype] = None
     observer: Optional[str] = Field(
         default=None,
         description=(
@@ -189,6 +200,12 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
         ),
     )
 
+    @field_serializer("zp_dtype")
+    def serialize_dtype(self, dtype: torch.dtype):
+        if self.symmetric:
+            return None
+        return str(dtype)
+
     @field_validator("type", mode="before")
     def validate_type(cls, value) -> QuantizationType:
         if isinstance(value, str):
@@ -266,6 +283,7 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
         dynamic = model.dynamic
         observer = model.observer
         dynamic = model.dynamic
+        zp_dtype = model.zp_dtype
 
         # infer strategy
         if strategy is None:
@@ -353,9 +371,16 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
             # default to minmax for non-dynamic cases
             observer = "minmax"
 
+        if zp_dtype is None:
+            if model.num_bits == 4 and model.type == QuantizationType.FLOAT:
+                zp_dtype = FP8_E4M3_DATA.dtype
+            else:
+                zp_dtype = model.pytorch_dtype()
+
         # write back modified values
         model.strategy = strategy
         model.observer = observer
+        model.zp_dtype = zp_dtype
         return model
 
     def pytorch_dtype(self) -> torch.dtype:
@@ -381,18 +406,56 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
     model_config = ConfigDict(extra="forbid")
 
 
-def round_to_quantized_type(
-    tensor: torch.Tensor, args: QuantizationArgs
+def round_to_quantized_type_dtype(
+    tensor: torch.Tensor,
+    dtype: torch.dtype,
+    cast_to_original_dtype: Optional[bool] = True,
 ) -> torch.Tensor:
     """
-    Rounds each element of the input tensor to the nearest quantized representation,
-    keeping to original dtype
+    Rounds an input tensor to the nearest quantized representation given a dtype.
+    The original dtype is kept post-rounding.
 
     :param tensor: tensor to round
-    :param args: QuantizationArgs to pull appropriate dtype from
+    :param dtype: dtype to use for rounding
+    :param cast_to_original_dtype: whether or not we cast the rounded tensor to
+        the original dtype
     :return: rounded tensor
     """
     original_dtype = tensor.dtype
+    if torch.is_floating_point(torch.tensor([], dtype=dtype)):
+        finfo = torch.finfo(dtype)
+        rounded = torch.clamp(tensor, finfo.min, finfo.max).to(dtype)
+    else:
+        iinfo = torch.iinfo(dtype)
+        rounded = torch.round(torch.clamp(tensor, iinfo.min, iinfo.max)).to(dtype)
+
+    if cast_to_original_dtype:
+        return rounded.to(original_dtype)
+    return rounded
+
+
+def round_to_quantized_type_args(
+    tensor: torch.Tensor,
+    args: QuantizationArgs,
+    min: torch.Tensor,
+    max: torch.Tensor,
+    cast_to_original_dtype: Optional[bool] = True,
+) -> torch.Tensor:
+    """
+    Rounds an input tensor to the nearest quantized representation given
+    qunatization args. The original dtype is kept post-rounding.
+
+    :param tensor: tensor to round
+    :param args: quantization args to use for rounding
+    :param min: min value to use for clamping
+    :param max: max value to use for clamping
+    :param cast_to_original_dtype: whether or not we cast the rounded tensor to
+        the original dtype
+    :return: rounded tensor
+    """
+
+    original_dtype = tensor.dtype
+    tensor = torch.clamp(tensor, min, max)
     if args.type == QuantizationType.FLOAT:
         if args.num_bits == 8:
             rounded = tensor.to(FP8_E4M3_DATA.dtype)
@@ -405,4 +468,6 @@ def round_to_quantized_type(
     else:
         raise ValueError(f"Invalid quantization type {args.type}")
 
-    return rounded.to(original_dtype)
+    if cast_to_original_dtype:
+        return rounded.to(original_dtype)
+    return rounded

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/src/compressed_tensors/quantization/quant_config.py

@@ -11,7 +11,6 @@
 # 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.
-
 from collections import defaultdict
 from enum import Enum
 from typing import Annotated, Any, Dict, List, Optional, Set, Union

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/src/compressed_tensors/quantization/quant_scheme.py

@@ -18,6 +18,7 @@ from typing import List, Optional
 
 from compressed_tensors.config import CompressionFormat
 from compressed_tensors.quantization.quant_args import (
+    FP8_E4M3_DATA,
     DynamicType,
     QuantizationArgs,
     QuantizationStrategy,
@@ -160,6 +161,8 @@ NVFP4A16 = dict(
         symmetric=True,
         dynamic=False,
         group_size=16,
+        scale_dtype=FP8_E4M3_DATA.dtype,
+        zp_dtype=FP8_E4M3_DATA.dtype,
     )
 )
 
@@ -173,6 +176,8 @@ NVFP4 = dict(
         dynamic=False,
         group_size=16,
         observer="static_minmax",
+        scale_dtype=FP8_E4M3_DATA.dtype,
+        zp_dtype=FP8_E4M3_DATA.dtype,
     ),
     input_activations=QuantizationArgs(
         num_bits=4,
@@ -182,6 +187,8 @@ NVFP4 = dict(
         dynamic=DynamicType.LOCAL,
         group_size=16,
         observer="static_minmax",
+        scale_dtype=FP8_E4M3_DATA.dtype,
+        zp_dtype=FP8_E4M3_DATA.dtype,
     ),
 )
 

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/src/compressed_tensors/quantization/utils/helpers.py

@@ -24,6 +24,7 @@ from compressed_tensors.quantization.quant_args import (
     QuantizationArgs,
     QuantizationStrategy,
     QuantizationType,
+    round_to_quantized_type_dtype,
 )
 from compressed_tensors.quantization.quant_scheme import QuantizationScheme
 from compressed_tensors.utils import deprecated
@@ -46,7 +47,6 @@ __all__ = [
     "calculate_range",
     "calculate_qparams",
     "generate_gparam",
-    "is_fp4",
     "strategy_cdiv",
 ]
 
@@ -57,13 +57,6 @@ KV_CACHE_TARGETS = ["re:.*self_attn$"]
 _LOGGER: logging.Logger = logging.getLogger(__name__)
 
 
-def is_fp4(quantization_args: QuantizationArgs):
-    return (
-        quantization_args.num_bits == 4
-        and quantization_args.type == QuantizationType.FLOAT
-    )
-
-
 def calculate_qparams(
     min_vals: Tensor,
     max_vals: Tensor,
@@ -92,52 +85,50 @@ def calculate_qparams(
     bit_min, bit_max = calculate_range(quantization_args, device)
     bit_range = bit_max - bit_min
 
-    if is_fp4(quantization_args=quantization_args):
-        zp_dtype = FP8_E4M3_DATA.dtype
-    else:
-        zp_dtype = quantization_args.pytorch_dtype()
-
+    # 1. Generate scale and zero-point
     if quantization_args.symmetric:
         max_val_pos = torch.max(torch.abs(min_vals), torch.abs(max_vals))
-
-        if is_fp4(quantization_args=quantization_args) and global_scale is not None:
-            # Conditionally scale the generated local scale by a global_scale
-            scales = global_scale * (max_val_pos / FP4_E2M1_DATA.max)
-            scales = torch.clamp(scales, max=FP8_E4M3_DATA.max, min=FP8_E4M3_DATA.min)
-            scales = scales.to(FP8_E4M3_DATA.dtype)
-
-        else:
-            scales = max_val_pos / (float(bit_range) / 2)
-
-        # TODO: in the case of MoEs, the global_scale may also be 0/need to be clamped
-        if scales.dtype == FP8_E4M3_DATA.dtype:
-            # torch.clamp not supported for FP8
-            # use the next largest fp8 value from 0
-            scales = torch.where(
-                scales == 0,
-                torch.tensor(0.125, dtype=FP8_E4M3_DATA.dtype, device=device),
-                scales,
-            )
-        else:
-            scales = torch.clamp(scales, min=torch.finfo(torch.float32).eps)
-
+        scales = max_val_pos / (float(bit_range) / 2)
         zero_points = torch.zeros(scales.shape, device=device, dtype=min_vals.dtype)
     else:
-        if is_fp4(quantization_args=quantization_args):
+        if (
+            quantization_args.num_bits == 4
+            and quantization_args.type == QuantizationType.FLOAT
+        ):
             raise NotImplementedError(
                 "Asymmetric Quantization is not supported for FP4"
             )
-
         scales = (max_vals - min_vals) / float(bit_range)
-        scales = torch.clamp(scales, min=torch.finfo(torch.float32).eps)
         zero_points = bit_min - (min_vals / scales)
         zero_points = torch.clamp(zero_points, bit_min, bit_max)
 
-    # match zero-points to quantized type
-    # if casting to int, use round instead of truncate
-    if quantization_args.type == QuantizationType.INT:
-        zero_points = torch.round(zero_points)
-    zero_points = zero_points.to(zp_dtype)
+    # 2. Conditionally scale the generated local scale by a global_scale
+    if global_scale is not None:
+        scales = global_scale * scales
+
+    # 3. Conditionally round the scale to the quantized dtype, if scale_dtype is set
+    if quantization_args.scale_dtype is not None:
+        scales = round_to_quantized_type_dtype(
+            scales, dtype=quantization_args.scale_dtype
+        )
+
+    # 4. Update any 0s with small values to
+    # prevent div by 0
+    eps = _get_dtype_eps(
+        dtype=quantization_args.scale_dtype
+        if quantization_args.scale_dtype is not None
+        else scales.dtype
+    )
+    scales = torch.where(
+        scales == 0,
+        torch.tensor(eps, dtype=scales.dtype, device=device),
+        scales,
+    )
+
+    # 5. Round the zp to zp_dtype
+    zero_points = round_to_quantized_type_dtype(
+        zero_points, dtype=quantization_args.zp_dtype, cast_to_original_dtype=False
+    )
 
     if scales.ndim == 0:
         scales = scales.reshape(1)
@@ -455,3 +446,14 @@ def strategy_cdiv(
             logger.bind(log_once=True).warning(message)
 
     return dividend
+
+
+def _get_dtype_eps(dtype: torch.dtype) -> float:
+    if dtype == FP8_E4M3_DATA.dtype:
+        return 0.125
+    elif dtype == FP4_E2M1_DATA.dtype:
+        return 0.25
+    elif torch.is_floating_point(torch.tensor([], dtype=dtype)):
+        return torch.finfo(dtype).eps
+    else:
+        return 1

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.12.3.a20251030'
+__version__ = version = '0.12.3.a20251110'
 __version_tuple__ = version_tuple = (0, 12, 3)

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.3a20251030
+Version: 0.12.3a20251110
 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-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/src/compressed_tensors.egg-info/SOURCES.txt

@@ -8,13 +8,8 @@ setup.py
 .github/.gitkeep
 .github/actions/test/action.yml
 .github/scripts/step-status
-.github/workflows/build-test.yml
-.github/workflows/build.yml
-.github/workflows/post-release-nightly-build.yml
 .github/workflows/quality-check.yaml
 .github/workflows/test-check.yaml
-.github/workflows/test.yml
-.github/workflows/trigger-all.yml
 examples/bitmask_compression.ipynb
 examples/quantize_and_pack_int4.ipynb
 examples/bit_packing/ex_quantize_and_pack.py

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/tests/test_compressors/model_compressors/test_model_compressor.py

@@ -22,6 +22,7 @@ import torch.nn as nn
 from compressed_tensors.compressors import ModelCompressor
 from compressed_tensors.config import CompressionFormat, SparsityCompressionConfig
 from compressed_tensors.quantization import (
+    FP8_E4M3_DATA,
     QuantizationArgs,
     QuantizationConfig,
     QuantizationScheme,
@@ -425,8 +426,18 @@ def test_multiple_quant_compressors():
         format=CompressionFormat.float_quantized.value,
     )
 
-    input_activations = QuantizationArgs(num_bits=4, type="float")
-    weights = QuantizationArgs(num_bits=4, type="float")
+    input_activations = QuantizationArgs(
+        num_bits=4,
+        type="float",
+        scale_dtype=FP8_E4M3_DATA.dtype,
+        zp_dtype=FP8_E4M3_DATA.dtype,
+    )
+    weights = QuantizationArgs(
+        num_bits=4,
+        type="float",
+        scale_dtype=FP8_E4M3_DATA.dtype,
+        zp_dtype=FP8_E4M3_DATA.dtype,
+    )
 
     scheme_nvfp4 = QuantizationScheme(
         targets=["Linear"],

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/tests/test_quantization/lifecycle/test_apply.py

@@ -22,6 +22,7 @@ import torch
 from compressed_tensors.config import CompressionFormat
 from compressed_tensors.quantization import (
     DEFAULT_QUANTIZATION_METHOD,
+    FP8_E4M3_DATA,
     QuantizationArgs,
     QuantizationConfig,
     QuantizationScheme,
@@ -153,7 +154,11 @@ def test_apply_quantization_config_tinyllama():
                 "linear": QuantizationScheme(
                     targets=["Linear"],
                     input_activations=QuantizationArgs(
-                        num_bits=8, type="float", strategy="tensor"
+                        num_bits=8,
+                        type="float",
+                        strategy="tensor",
+                        scale_dtype=FP8_E4M3_DATA.dtype,
+                        zp_dtype=torch.float,
                     ),
                 )
             }
@@ -163,7 +168,11 @@ def test_apply_quantization_config_tinyllama():
                 "linear": QuantizationScheme(
                     targets=["Linear"],
                     input_activations=QuantizationArgs(
-                        num_bits=8, type="float", strategy="tensor"
+                        num_bits=8,
+                        type="float",
+                        strategy="tensor",
+                        scale_dtype=FP8_E4M3_DATA.dtype,
+                        zp_dtype=torch.float,
                     ),
                 )
             },
@@ -176,7 +185,11 @@ def test_apply_quantization_config_tinyllama():
         QuantizationConfig(
             config_groups={},
             kv_cache_scheme=QuantizationArgs(
-                num_bits=8, type="float", strategy="tensor"
+                num_bits=8,
+                type="float",
+                strategy="tensor",
+                scale_dtype=FP8_E4M3_DATA.dtype,
+                zp_dtype=torch.float,
             ),
         ),
         QuantizationConfig(
@@ -184,12 +197,20 @@ def test_apply_quantization_config_tinyllama():
                 "attention": QuantizationScheme(
                     targets=["LlamaAttention"],
                     input_activations=QuantizationArgs(
-                        num_bits=8, type="float", strategy="tensor"
+                        num_bits=8,
+                        type="float",
+                        strategy="tensor",
+                        scale_dtype=FP8_E4M3_DATA.dtype,
+                        zp_dtype=torch.float,
                     ),
                 )
             },
             kv_cache_scheme=QuantizationArgs(
-                num_bits=8, type="float", strategy="tensor"
+                num_bits=8,
+                type="float",
+                strategy="tensor",
+                scale_dtype=FP8_E4M3_DATA.dtype,
+                zp_dtype=torch.float,
             ),
         ),
     ],
@@ -448,7 +469,13 @@ def test_apply_kv_cache():
     with init_empty_weights():
         model = AutoModelForCausalLM.from_pretrained("nm-testing/llama2.c-stories15M")
 
-    args = QuantizationArgs(num_bits=8, type="float", strategy="tensor")
+    args = QuantizationArgs(
+        num_bits=8,
+        type="float",
+        strategy="tensor",
+        scale_dtype=FP8_E4M3_DATA.dtype,
+        zp_dtype=torch.float,
+    )
     config = QuantizationConfig(config_groups={}, kv_cache_scheme=args)
 
     apply_quantization_config(model, config)
@@ -468,7 +495,13 @@ def test_apply_attention():
 
     scheme = QuantizationScheme(
         targets=["LlamaAttention"],
-        input_activations=QuantizationArgs(num_bits=8, type="float", strategy="tensor"),
+        input_activations=QuantizationArgs(
+            num_bits=8,
+            type="float",
+            strategy="tensor",
+            scale_dtype=FP8_E4M3_DATA.dtype,
+            zp_dtype=torch.float,
+        ),
     )
     config = QuantizationConfig(config_groups={"attention": scheme})
 

{compressed_tensors-0.12.3a20251030 → compressed_tensors-0.12.3a20251110}/tests/test_quantization/lifecycle/test_dynamic_lifecycle.py

@@ -79,7 +79,7 @@ def _test_layer_dynamic_quantization_status(
 def get_tinyllama_model():
     return AutoModelForCausalLM.from_pretrained(
         "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T",
-        torch_dtype="auto",
+        torch_dtype=torch.bfloat16,
     )