compressed-tensors-nightly 0.7.1.20241031__py3-none-any.whl → 0.7.1.20241102__py3-none-any.whl

compressed_tensors/quantization/__init__.py

@@ -19,4 +19,3 @@ from .quant_args import *
 from .quant_config import *
 from .quant_scheme import *
 from .lifecycle import *
-from .cache import QuantizedKVParameterCache

compressed_tensors/quantization/lifecycle/__init__.py

@@ -15,9 +15,7 @@
 # flake8: noqa
 # isort: skip_file
 
-from .calibration import *
 from .forward import *
-from .frozen import *
 from .initialize import *
 from .compressed import *
 from .apply import *

compressed_tensors/quantization/lifecycle/apply.py

@@ -22,13 +22,9 @@ from typing import Union
 
 import torch
 from compressed_tensors.config import CompressionFormat
-from compressed_tensors.quantization.lifecycle.calibration import (
-    set_module_for_calibration,
-)
 from compressed_tensors.quantization.lifecycle.compressed import (
     compress_quantized_weights,
 )
-from compressed_tensors.quantization.lifecycle.frozen import freeze_module_quantization
 from compressed_tensors.quantization.lifecycle.initialize import (
     initialize_module_for_quantization,
 )
@@ -233,6 +229,7 @@ def apply_quantization_status(model: Module, status: QuantizationStatus):
     :param model: model to apply quantization to
     :param status: status to update the module to
     """
+
     current_status = infer_quantization_status(model)
 
     if status >= QuantizationStatus.INITIALIZED > current_status:
@@ -243,18 +240,6 @@ def apply_quantization_status(model: Module, status: QuantizationStatus):
             )
         )
 
-    if current_status < status >= QuantizationStatus.CALIBRATION > current_status:
-        # only quantize weights up front when our end goal state is calibration,
-        # weight quantization parameters are already loaded for frozen/compressed
-        quantize_weights_upfront = status == QuantizationStatus.CALIBRATION
-        model.apply(
-            lambda module: set_module_for_calibration(
-                module, quantize_weights_upfront=quantize_weights_upfront
-            )
-        )
-    if current_status < status >= QuantizationStatus.FROZEN > current_status:
-        model.apply(freeze_module_quantization)
-
     if current_status < status >= QuantizationStatus.COMPRESSED > current_status:
         model.apply(compress_quantized_weights)
 

compressed_tensors/quantization/lifecycle/forward.py

@@ -14,14 +14,9 @@
 
 from functools import wraps
 from math import ceil
-from typing import Callable, Optional
+from typing import Optional
 
 import torch
-from compressed_tensors.quantization.cache import QuantizedKVParameterCache
-from compressed_tensors.quantization.observers.helpers import (
-    calculate_range,
-    compute_dynamic_scales_and_zp,
-)
 from compressed_tensors.quantization.quant_args import (
     QuantizationArgs,
     QuantizationStrategy,
@@ -29,7 +24,11 @@ from compressed_tensors.quantization.quant_args import (
 )
 from compressed_tensors.quantization.quant_config import QuantizationStatus
 from compressed_tensors.quantization.quant_scheme import QuantizationScheme
-from compressed_tensors.utils import safe_permute, update_parameter_data
+from compressed_tensors.quantization.utils import (
+    calculate_range,
+    compute_dynamic_scales_and_zp,
+)
+from compressed_tensors.utils import safe_permute
 from torch.nn import Module
 
 
@@ -39,7 +38,6 @@ __all__ = [
     "fake_quantize",
     "wrap_module_forward_quantized",
     "forward_quantize",
-    "calibrate_activations",
 ]
 
 
@@ -276,19 +274,7 @@ def wrap_module_forward_quantized(module: Module, scheme: QuantizationScheme):
         compressed = module.quantization_status == QuantizationStatus.COMPRESSED
 
         if scheme.input_activations is not None:
-            # calibrate and (fake) quantize input activations when applicable
-            # NOTE: will be moved out of compressed-tensors
-            if (
-                module.quantization_status == QuantizationStatus.CALIBRATION
-                and not scheme.input_activations.dynamic
-            ):
-                calibrate_activations(
-                    module=module,
-                    value=input_,
-                    base_name="input",
-                    quantization_args=scheme.input_activations,
-                )
-
+            # prehook should calibrate activations before forward call
             input_ = forward_quantize(module, input_, "input", scheme.input_activations)
 
         if scheme.weights is not None and not compressed:
@@ -302,31 +288,22 @@ def wrap_module_forward_quantized(module: Module, scheme: QuantizationScheme):
         output = forward_func_orig.__get__(module, module.__class__)(
             input_, *args[1:], **kwargs
         )
-        if scheme.output_activations is not None:
 
-            # calibrate and (fake) quantize output activations when applicable
-            # kv_cache scales updated on model self_attn forward call in
-            # wrap_module_forward_quantized_attn
+        # restore back to unquantized_value
+        if scheme.weights is not None and not compressed:
+            self.weight.data = unquantized_weight
 
+        if scheme.output_activations is not None:
+            # forward-hook should calibrate/forward_quantize
             if (
                 module.quantization_status == QuantizationStatus.CALIBRATION
                 and not scheme.output_activations.dynamic
             ):
-                calibrate_activations(
-                    module=module,
-                    value=output,
-                    base_name="output",
-                    quantization_args=scheme.ouput_activations,
-                )
+                return output
 
             output = forward_quantize(
                 module, output, "output", scheme.output_activations
             )
-
-        # restore back to unquantized_value
-        if scheme.weights is not None and not compressed:
-            self.weight.data = unquantized_weight
-
         return output
 
     # bind wrapped forward to module class so reference to `self` is correct
@@ -335,77 +312,6 @@ def wrap_module_forward_quantized(module: Module, scheme: QuantizationScheme):
     setattr(module, "forward", bound_wrapped_forward)
 
 
-def wrap_module_forward_quantized_attn(module: Module, scheme: QuantizationScheme):
-    # expects a module already initialized and injected with the parameters in
-    # initialize_module_for_quantization
-    if hasattr(module.forward, "__func__"):
-        forward_func_orig = module.forward.__func__
-    else:
-        forward_func_orig = module.forward.func
-
-    @wraps(forward_func_orig)  # ensures docstring, names, etc are propagated
-    def wrapped_forward(self, *args, **kwargs):
-
-        # kv cache stored under weights
-        if module.quantization_status == QuantizationStatus.CALIBRATION:
-            quantization_args: QuantizationArgs = scheme.output_activations
-            past_key_value: QuantizedKVParameterCache = quantization_args.get_kv_cache()
-            kwargs["past_key_value"] = past_key_value
-
-            # QuantizedKVParameterCache used for obtaining k_scale, v_scale only,
-            # does not store quantized_key_states and quantized_value_state
-            kwargs["use_cache"] = False
-
-            attn_forward: Callable = forward_func_orig.__get__(module, module.__class__)
-
-            past_key_value.reset_states()
-
-            rtn = attn_forward(*args, **kwargs)
-
-            update_parameter_data(
-                module, past_key_value.k_scales[module.layer_idx], "k_scale"
-            )
-            update_parameter_data(
-                module, past_key_value.v_scales[module.layer_idx], "v_scale"
-            )
-
-            return rtn
-
-        return forward_func_orig.__get__(module, module.__class__)(*args, **kwargs)
-
-    # bind wrapped forward to module class so reference to `self` is correct
-    bound_wrapped_forward = wrapped_forward.__get__(module, module.__class__)
-    # set forward to wrapped forward
-    setattr(module, "forward", bound_wrapped_forward)
-
-
-def calibrate_activations(
-    module: Module,
-    value: torch.Tensor,
-    base_name: str,
-    quantization_args: QuantizationArgs,
-):
-    # If empty tensor, can't update zp/scale
-    # Case for MoEs
-    if value.numel() == 0:
-        return
-    # calibration mode - get new quant params from observer
-    if not hasattr(module, f"{base_name}_observer"):
-        from compressed_tensors.quantization.lifecycle import initialize_observers
-
-        initialize_observers(
-            module=module, base_name=base_name, quantization_args=quantization_args
-        )
-
-    observer = getattr(module, f"{base_name}_observer")
-
-    updated_scale, updated_zero_point = observer(value)
-
-    # update scale and zero point
-    update_parameter_data(module, updated_scale, f"{base_name}_scale")
-    update_parameter_data(module, updated_zero_point, f"{base_name}_zero_point")
-
-
 def forward_quantize(
     module: Module, value: torch.Tensor, base_name: str, args: "QuantizationArgs"
 ) -> torch.Tensor:
@@ -426,10 +332,10 @@ def forward_quantize(
     g_idx = getattr(module, "weight_g_idx", None)
 
     if args.dynamic:
-        # dynamic quantization - no need to invoke observer
+        # dynamic quantization - determine the scale/zp on the fly
         scale, zero_point = compute_dynamic_scales_and_zp(value=value, args=args)
     else:
-        # static quantization - get previous scale and zero point from layer
+        # static quantization - get scale and zero point from layer
         scale = getattr(module, f"{base_name}_scale")
         zero_point = getattr(module, f"{base_name}_zero_point", None)
 

compressed_tensors/quantization/lifecycle/initialize.py

@@ -14,13 +14,12 @@
 
 
 import logging
+from enum import Enum
 from typing import Optional
 
 import torch
-from compressed_tensors.quantization.cache import KVCacheScaleType
 from compressed_tensors.quantization.lifecycle.forward import (
     wrap_module_forward_quantized,
-    wrap_module_forward_quantized_attn,
 )
 from compressed_tensors.quantization.quant_args import (
     ActivationOrdering,
@@ -34,12 +33,21 @@ from compressed_tensors.utils import get_execution_device, is_module_offloaded
 from torch.nn import Module, Parameter
 
 
-__all__ = ["initialize_module_for_quantization", "initialize_observers"]
+__all__ = [
+    "initialize_module_for_quantization",
+    "is_attention_module",
+    "KVCacheScaleType",
+]
 
 
 _LOGGER = logging.getLogger(__name__)
 
 
+class KVCacheScaleType(Enum):
+    KEY = "k_scale"
+    VALUE = "v_scale"
+
+
 def initialize_module_for_quantization(
     module: Module,
     scheme: Optional[QuantizationScheme] = None,
@@ -64,9 +72,7 @@ def initialize_module_for_quantization(
         return
 
     if is_attention_module(module):
-        # wrap forward call of module to perform
         # quantized actions based on calltime status
-        wrap_module_forward_quantized_attn(module, scheme)
         _initialize_attn_scales(module)
 
     else:
@@ -107,6 +113,7 @@ def initialize_module_for_quantization(
         module.quantization_status = QuantizationStatus.INITIALIZED
 
         offloaded = False
+        # What is this doing/why isn't this in the attn case?
         if is_module_offloaded(module):
             try:
                 from accelerate.hooks import add_hook_to_module, remove_hook_from_module
@@ -144,14 +151,12 @@ def initialize_module_for_quantization(
                 module._hf_hook.weights_map = new_prefix_dict
 
 
-def initialize_observers(
-    module: Module,
-    base_name: str,
-    quantization_args: QuantizationArgs,
-):
-    # initialize observer module and attach as submodule
-    observer = quantization_args.get_observer()
-    module.register_module(f"{base_name}_observer", observer)
+def is_attention_module(module: Module):
+    return "attention" in module.__class__.__name__.lower() and (
+        hasattr(module, "k_proj")
+        or hasattr(module, "v_proj")
+        or hasattr(module, "qkv_proj")
+    )
 
 
 def _initialize_scale_zero_point(
@@ -209,14 +214,6 @@ def _initialize_scale_zero_point(
         module.register_parameter(f"{base_name}_g_idx", init_g_idx)
 
 
-def is_attention_module(module: Module):
-    return "attention" in module.__class__.__name__.lower() and (
-        hasattr(module, "k_proj")
-        or hasattr(module, "v_proj")
-        or hasattr(module, "qkv_proj")
-    )
-
-
 def _initialize_attn_scales(module: Module) -> None:
     """Initlaize k_scale, v_scale for  self_attn"""
 

compressed_tensors/quantization/quant_args.py

@@ -114,20 +114,7 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
         """
         :return: torch quantization FakeQuantize built based on these QuantizationArgs
         """
-        from compressed_tensors.quantization.observers.base import Observer
-
-        # No observer required for the dynamic case
-        if self.dynamic:
-            self.observer = None
-            return self.observer
-
-        return Observer.load_from_registry(self.observer, quantization_args=self)
-
-    def get_kv_cache(self):
-        """Get the singleton KV Cache"""
-        from compressed_tensors.quantization.cache import QuantizedKVParameterCache
-
-        return QuantizedKVParameterCache(self)
+        return self.observer
 
     @field_validator("type", mode="before")
     def validate_type(cls, value) -> QuantizationType:
@@ -210,6 +197,7 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
                 "activation ordering"
             )
 
+        # infer observer w.r.t. dynamic
         if dynamic:
             if strategy not in (
                 QuantizationStrategy.TOKEN,
@@ -221,18 +209,19 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
                     "quantization",
                 )
             if observer is not None:
-                warnings.warn(
-                    "No observer is used for dynamic quantization, setting to None"
-                )
-                model.observer = None
+                if observer != "memoryless":  # avoid annoying users with old configs
+                    warnings.warn(
+                        "No observer is used for dynamic quantization, setting to None"
+                    )
+                observer = None
 
-        # if we have not set an observer and we
-        # are running static quantization, use minmax
-        if not observer and not dynamic:
-            model.observer = "minmax"
+        elif observer is None:
+            # default to minmax for non-dynamic cases
+            observer = "minmax"
 
         # write back modified values
         model.strategy = strategy
+        model.observer = observer
         return model
 
     def pytorch_dtype(self) -> torch.dtype:

compressed_tensors/quantization/utils/helpers.py

@@ -16,9 +16,14 @@ import logging
 from typing import Generator, List, Optional, Tuple
 
 import torch
-from compressed_tensors.quantization.observers.base import Observer
-from compressed_tensors.quantization.quant_args import QuantizationArgs
+from compressed_tensors.quantization.quant_args import (
+    FP8_DTYPE,
+    QuantizationArgs,
+    QuantizationStrategy,
+    QuantizationType,
+)
 from compressed_tensors.quantization.quant_scheme import QuantizationScheme
+from torch import FloatTensor, IntTensor, Tensor
 from torch.nn import Module
 from tqdm import tqdm
 
@@ -36,6 +41,9 @@ __all__ = [
     "is_kv_cache_quant_scheme",
     "iter_named_leaf_modules",
     "iter_named_quantizable_modules",
+    "compute_dynamic_scales_and_zp",
+    "calculate_range",
+    "calculate_qparams",
 ]
 
 # target the self_attn layer
@@ -45,6 +53,105 @@ KV_CACHE_TARGETS = ["re:.*self_attn$"]
 _LOGGER: logging.Logger = logging.getLogger(__name__)
 
 
+def calculate_qparams(
+    min_vals: Tensor, max_vals: Tensor, quantization_args: QuantizationArgs
+) -> Tuple[FloatTensor, IntTensor]:
+    """
+    :param min_vals: tensor of min value(s) to calculate scale(s) and zero point(s)
+        from
+    :param max_vals: tensor of max value(s) to calculate scale(s) and zero point(s)
+        from
+    :param quantization_args: settings to quantization
+    :return: tuple of the calculated scale(s) and zero point(s)
+    """
+    min_vals = torch.min(min_vals, torch.zeros_like(min_vals))
+    max_vals = torch.max(max_vals, torch.zeros_like(max_vals))
+    device = min_vals.device
+
+    bit_min, bit_max = calculate_range(quantization_args, device)
+    bit_range = bit_max - bit_min
+    zp_dtype = quantization_args.pytorch_dtype()
+
+    if quantization_args.symmetric:
+        max_val_pos = torch.max(torch.abs(min_vals), torch.abs(max_vals))
+        scales = max_val_pos / (float(bit_range) / 2)
+        scales = torch.clamp(scales, min=torch.finfo(torch.float32).eps)
+        zero_points = torch.zeros(scales.shape, device=device, dtype=min_vals.dtype)
+    else:
+        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
+    zero_points = zero_points.to(zp_dtype)
+
+    if scales.ndim == 0:
+        scales = scales.reshape(1)
+        zero_points = zero_points.reshape(1)
+
+    return scales, zero_points
+
+
+def compute_dynamic_scales_and_zp(value: Tensor, args: QuantizationArgs):
+    """
+    Returns the computed scales and zero points for dynamic activation
+    qunatization.
+
+    :param value: tensor to calculate quantization parameters for
+    :param args: quantization args
+    :param reduce_dims: optional tuple of dimensions to reduce along,
+        returned scale and zero point will be shaped (1,) along the
+        reduced dimensions
+    :return: tuple of scale and zero point derived from the observed tensor
+    """
+    if args.strategy == QuantizationStrategy.TOKEN:
+        dim = {1, 2}
+        reduce_dims = tuple(idx for idx in range(value.ndim) if idx not in dim)
+    elif args.strategy == QuantizationStrategy.TENSOR:
+        reduce_dims = None
+    else:
+        raise ValueError(
+            f"One of {QuantizationStrategy.TOKEN} or {QuantizationStrategy.TENSOR} ",
+            "must be used for dynamic quantization",
+        )
+
+    if not reduce_dims:
+        min_val, max_val = torch.aminmax(value)
+    else:
+        min_val = torch.amin(value, dim=reduce_dims, keepdims=True)
+        max_val = torch.amax(value, dim=reduce_dims, keepdims=True)
+
+    return calculate_qparams(min_val, max_val, args)
+
+
+def calculate_range(quantization_args: QuantizationArgs, device: str) -> Tuple:
+    """
+    Calculated the effective quantization range for the given Quantization Args
+
+    :param quantization_args: quantization args to get range of
+    :param device: device to store the range to
+    :return: tuple endpoints for the given quantization range
+    """
+    if quantization_args.type == QuantizationType.INT:
+        bit_range = 2**quantization_args.num_bits
+        q_max = torch.tensor(bit_range / 2 - 1, device=device)
+        q_min = torch.tensor(-bit_range / 2, device=device)
+    elif quantization_args.type == QuantizationType.FLOAT:
+        if quantization_args.num_bits != 8:
+            raise ValueError(
+                "Floating point quantization is only supported for 8 bits,"
+                f"got {quantization_args.num_bits}"
+            )
+        fp_range_info = torch.finfo(FP8_DTYPE)
+        q_max = torch.tensor(fp_range_info.max, device=device)
+        q_min = torch.tensor(fp_range_info.min, device=device)
+    else:
+        raise ValueError(f"Invalid quantization type {quantization_args.type}")
+
+    return q_min, q_max
+
+
 def infer_quantization_status(model: Module) -> Optional["QuantizationStatus"]:  # noqa
     """
     Checks the quantization status of a model. Assumes all modules in the model have
@@ -118,12 +225,17 @@ def iter_named_leaf_modules(model: Module) -> Generator[Tuple[str, Module], None
     """
     for name, submodule in model.named_modules():
         children = list(submodule.children())
-        if len(children) == 0 and not isinstance(submodule, Observer):
+        # TODO: verify if an observer would ever be attached in this case/remove check
+        if len(children) == 0 and "observer" in name:
             yield name, submodule
         else:
+            if len(children) > 0:
+                named_children, children = zip(*list(submodule.named_children()))
             has_non_observer_children = False
-            for child in children:
-                if not isinstance(child, Observer):
+            for i in range(len(children)):
+                child_name = named_children[i]
+
+                if "observer" not in child_name:
                     has_non_observer_children = True
 
             if not has_non_observer_children:
@@ -144,14 +256,19 @@ def iter_named_quantizable_modules(
     :returns: generator tuple of (name, submodule)
     """
     for name, submodule in model.named_modules():
+        # TODO: verify if an observer would ever be attached in this case/remove check
         if include_children:
             children = list(submodule.children())
-            if len(children) == 0 and not isinstance(submodule, Observer):
+            if len(children) == 0 and "observer" not in name:
                 yield name, submodule
             else:
+                if len(children) > 0:
+                    named_children, children = zip(*list(submodule.named_children()))
                 has_non_observer_children = False
-                for child in children:
-                    if not isinstance(child, Observer):
+                for i in range(len(children)):
+                    child_name = named_children[i]
+
+                    if "observer" not in child_name:
                         has_non_observer_children = True
 
                 if not has_non_observer_children:

compressed_tensors/registry/registry.py

@@ -258,7 +258,7 @@ def get_from_registry(
         retrieved_value = _import_and_get_value_from_module(module_path, value_name)
     else:
         # look up name in alias registry
-        name = _ALIAS_REGISTRY[parent_class].get(name)
+        name = _ALIAS_REGISTRY[parent_class].get(name, name)
         # look up name in registry
         retrieved_value = _REGISTRY[parent_class].get(name)
         if retrieved_value is None:

{compressed_tensors_nightly-0.7.1.20241031.dist-info → compressed_tensors_nightly-0.7.1.20241102.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.7.1.20241031
+Version: 0.7.1.20241102
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors_nightly-0.7.1.20241031.dist-info → compressed_tensors_nightly-0.7.1.20241102.dist-info}/RECORD

@@ -22,28 +22,20 @@ compressed_tensors/config/dense.py,sha256=NgSxnFCnckU9-iunxEaqiFwqgdO7YYxlWKR74j
 compressed_tensors/config/sparse_bitmask.py,sha256=pZUboRNZTu6NajGOQEFExoPknak5ynVAUeiiYpS1Gt8,1308
 compressed_tensors/linear/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
 compressed_tensors/linear/compressed_linear.py,sha256=0jTTf6XxOAjAYs3tvFtgiNMAO4W10sSeR-pdH2M413g,3218
-compressed_tensors/quantization/__init__.py,sha256=nWP_fsl6Nn0ksEgZPzerGiETdvF-ZfNwPnwGlRiR5pY,805
-compressed_tensors/quantization/cache.py,sha256=Sf_9Nfe3RpX04V84iUJMgLN9pWNMFIYvZW02LXcPUQw,6830
-compressed_tensors/quantization/quant_args.py,sha256=k7NuZn8OqjgzmAVaN2-jHPQ1bgDkMuUoLJtLnhkvIOI,9085
+compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
+compressed_tensors/quantization/quant_args.py,sha256=osjNwCSB6tcyH9Qeg5sHEiB-bHyi3XJ8TzkGVJuGTc4,8711
 compressed_tensors/quantization/quant_config.py,sha256=NCiMvUMnnz5kTyAkDylxjtEGQnjgsIYIeNR2zyHEdTQ,10371
 compressed_tensors/quantization/quant_scheme.py,sha256=5ggPz5sqEfTUgvJJeiPIINA74QtO-08hb3szsm7UHGE,6000
-compressed_tensors/quantization/lifecycle/__init__.py,sha256=MXE2E7GfIfRRfhrdGy2Og3AZOz5N59B0ZGFcsD89y6c,821
-compressed_tensors/quantization/lifecycle/apply.py,sha256=czaayvpeUYyWRJhO_klffw6esptOgA9sBKL5TWQcRdw,15805
-compressed_tensors/quantization/lifecycle/calibration.py,sha256=fJ2RDL3E4hmWR8v8nYhq_tv31K8WV00o_4Y3xr7c37Y,3041
+compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
+compressed_tensors/quantization/lifecycle/apply.py,sha256=pdCqxXnVw7HoDDanaOtek13g8x_nb54CBUlfuMdhFG4,14993
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
-compressed_tensors/quantization/lifecycle/forward.py,sha256=8GjOnx4rwOZZqSDTdnejNOY2DVTjNDzH0DfY_rQam6k,16575
-compressed_tensors/quantization/lifecycle/frozen.py,sha256=71TsgS0Uxku0NomdWOBJsVfXCGTne-Gx9zUEMsCmw5Q,1764
+compressed_tensors/quantization/lifecycle/forward.py,sha256=QPL6-vKOFuKdKIEsVqMhsw4x552Jpm2sqO0oeChbnrM,12941
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
-compressed_tensors/quantization/lifecycle/initialize.py,sha256=lKoFy18PjbSklyum7f4hoLuWtHShBKax7JDTBzPlCqM,8839
-compressed_tensors/quantization/observers/__init__.py,sha256=DYrttzq-8MHLZUzpX-xzzm4hrw6HcXkMkux82KBKb1M,738
-compressed_tensors/quantization/observers/base.py,sha256=5ovQicWPYHjIxr6-EkQ4lgOX0PpI9g23iSzKpxjM1Zg,8420
-compressed_tensors/quantization/observers/helpers.py,sha256=nUFdNEIACiPBfFwNYDGCXOvw6tf7j6jfTvDwImHKMPg,5506
-compressed_tensors/quantization/observers/min_max.py,sha256=sQXqU3z-voxIDfR_9mQzwQUflZj2sASm_G8CYaXntFw,3865
-compressed_tensors/quantization/observers/mse.py,sha256=G5Y9v4MqXUVcKxBSmCFFW3p_7rlu-6scqLIN88ng-sE,6080
+compressed_tensors/quantization/lifecycle/initialize.py,sha256=C41hKA5VANyEwkB5FxzEn3Z0Da5tfxF1I07P8rUcyS0,8537
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
-compressed_tensors/quantization/utils/helpers.py,sha256=y4LEyC2oUd876ZMdALWKGH3Ct5EgBJZV4id_NUjTGH8,9531
+compressed_tensors/quantization/utils/helpers.py,sha256=DBP-sGRpGAY01K0LFE7qqonNj4hkTYL_mXrMs2LtAD8,14100
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
-compressed_tensors/registry/registry.py,sha256=fxjOjh2wklCvJhQxwofdy-zV8q7MkQ85SLG77nml2iA,11890
+compressed_tensors/registry/registry.py,sha256=vRcjVB1ITfSbfYUaGndBBmqhip_5vsS62weorVg0iXo,11896
 compressed_tensors/utils/__init__.py,sha256=gS4gSU2pwcAbsKj-6YMaqhm25udFy6ISYaWBf-myRSM,808
 compressed_tensors/utils/helpers.py,sha256=hWGIR0W7ENHwdC7wW2SQJJiCF9-xOu_u3fY2RzLyYg4,4101
 compressed_tensors/utils/offload.py,sha256=d9q8LNe8HyF8tOjgjA7QGLD3HRysmNp0d8eBbdqBgIM,4089
@@ -51,8 +43,8 @@ compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVy
 compressed_tensors/utils/permute.py,sha256=V6tJLKo3Syccj-viv4F7ZKZgJeCB-hl-dK8RKI_kBwI,2355
 compressed_tensors/utils/safetensors_load.py,sha256=m08ANVuTBxQdoa6LufDgcNJ7wCLDJolyZljB8VEybAU,8578
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors_nightly-0.7.1.20241031.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors_nightly-0.7.1.20241031.dist-info/METADATA,sha256=5wd-PII0N7nJCHB0UGxJKh0J226nAyuDQTbM8Otp4FU,6799
-compressed_tensors_nightly-0.7.1.20241031.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-compressed_tensors_nightly-0.7.1.20241031.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors_nightly-0.7.1.20241031.dist-info/RECORD,,
+compressed_tensors_nightly-0.7.1.20241102.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors_nightly-0.7.1.20241102.dist-info/METADATA,sha256=pQ8FXKctjUHKkisrXYyeDUuunknVPkjHnHvS-uJ89oI,6799
+compressed_tensors_nightly-0.7.1.20241102.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+compressed_tensors_nightly-0.7.1.20241102.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors_nightly-0.7.1.20241102.dist-info/RECORD,,