compressed-tensors 0.7.0__py3-none-any.whl → 0.8.0__py3-none-any.whl

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/version.py

@@ -17,7 +17,7 @@ Functionality for storing and setting the version info for SparseML
 """
 
 
-version_base = "0.7.0"
+version_base = "0.8.0"
 is_release = True  # change to True to set the generated version as a release version
 
 

{compressed_tensors-0.7.0.dist-info → compressed_tensors-0.8.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors
-Version: 0.7.0
+Version: 0.8.0
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.7.0.dist-info → compressed_tensors-0.8.0.dist-info}/RECORD

@@ -1,6 +1,6 @@
 compressed_tensors/__init__.py,sha256=UtKmifNeBCSE2TZSAfduVNNzHY-3V7bLjZ7n7RuXLOE,812
 compressed_tensors/base.py,sha256=73HYH7HY7O2roC89yG_piPFnZwrBfn_i7HmKl90SKc0,875
-compressed_tensors/version.py,sha256=RTYptXdV8f4QbYCRQ13eGeEsq4grNJs6EXgejoZl9EE,1585
+compressed_tensors/version.py,sha256=Z9w80ldLHldBZrnrRolznhe-AZsAg5ftvHw17kgPs10,1585
 compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1tWZewqVnOb4lM,825
 compressed_tensors/compressors/base.py,sha256=D9TNwQcjanDiAHODPbg8JUqc66e3j50rctY7A708NEs,6743
 compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
@@ -15,36 +15,27 @@ compressed_tensors/compressors/sparse_compressors/base.py,sha256=Ua4rUSGyucEs-YJ
 compressed_tensors/compressors/sparse_compressors/dense.py,sha256=lSKNWRx6H7aUqaJj1j4qbXk8Gkm1UohbnvW1Rvq6Ra4,1284
 compressed_tensors/compressors/sparse_compressors/sparse_bitmask.py,sha256=4fKwCG7ZM8mUtSnjPvubzEHl-mTnxMzwjmcs7L43WLY,6622
 compressed_tensors/compressors/sparse_quantized_compressors/__init__.py,sha256=4f_cwcKXB1nVVMoiKgTFAc8jAPjPLElo-Df_EDm1_xw,675
-compressed_tensors/compressors/sparse_quantized_compressors/marlin_24.py,sha256=akqE7eW8CLTslpWRxERaZ8R0TSm1lS7D1bgZXKL0xi8,9427
+compressed_tensors/compressors/sparse_quantized_compressors/marlin_24.py,sha256=BMIQWTLlnUvxy14iEJegtiP75WHJeOVojey9mKOK1hE,9427
 compressed_tensors/config/__init__.py,sha256=ZBqWn3r6ku1qfmlHHYp0mQueY0i7Pwhr9rbQk9dDlMc,704
-compressed_tensors/config/base.py,sha256=BNTFKy12isY7qblwxdi_R1f00EzgrNOXLrfxqLCPT8w,1903
+compressed_tensors/config/base.py,sha256=3bFAdwDZjOt-U3fneOeL8dRci-PS8DqstnXuQVtkfiQ,3435
 compressed_tensors/config/dense.py,sha256=NgSxnFCnckU9-iunxEaqiFwqgdO7YYxlWKR74jNbjks,1317
 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=vnBB5zasO_XpHomZvzUPVVbzyCz2VgebsHePm0kANzY,6831
-compressed_tensors/quantization/quant_args.py,sha256=73KevZXHyrkMCT_3CxbYHz70fI3i-wcF8NvN0wsBPK4,8271
+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=uFgp6ECU6ZkHWkeKlAVAzZTLDbrTrzPSPrY23eJluaw,5931
-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=IuLeRkVQPrMxkMcIjr4OMFlIUMHkqjH4qAxC2KiUBGw,2673
+compressed_tensors/quantization/quant_scheme.py,sha256=5ggPz5sqEfTUgvJJeiPIINA74QtO-08hb3szsm7UHGE,6000
+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=eLup6QDRUUp_Ozcas7RDRLIXBWjFbxn5gWbcAIJEGlw,15715
-compressed_tensors/quantization/lifecycle/frozen.py,sha256=NiJw7NP7pcT6idWFa8vksgiLoT8oQ975e57S4QfD2QQ,1874
+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=4_YG7jKl7d2-Cy58pOkMtInFRhvYahxYchesWMPxPVM,8862
-compressed_tensors/quantization/observers/__init__.py,sha256=4Sa7rqi5RB_S5bPO8KmncETiqDsoMBhwP37arlQym8s,764
-compressed_tensors/quantization/observers/base.py,sha256=5ovQicWPYHjIxr6-EkQ4lgOX0PpI9g23iSzKpxjM1Zg,8420
-compressed_tensors/quantization/observers/helpers.py,sha256=s_A23Qa_BLfOdHJCN5bm-qPWkhjjj_RIVrhSp1Y9Dtk,4211
-compressed_tensors/quantization/observers/memoryless.py,sha256=jH_c6K3gxf4W3VNXQ7tbnP-J_86QTrEfjBn6Kh1C-H8,2165
-compressed_tensors/quantization/observers/min_max.py,sha256=sQXqU3z-voxIDfR_9mQzwQUflZj2sASm_G8CYaXntFw,3865
-compressed_tensors/quantization/observers/mse.py,sha256=Aeh-253Vbab1F8cYuBiGNn4OXWJ67wXQ_JVfl3mu2a8,6034
+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
@@ -52,8 +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-0.7.0.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.7.0.dist-info/METADATA,sha256=Lgcl4rU8ifo0PY-FrurFApAkuTD9HBeJohuULjVqebs,6782
-compressed_tensors-0.7.0.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-compressed_tensors-0.7.0.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.7.0.dist-info/RECORD,,
+compressed_tensors-0.8.0.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.8.0.dist-info/METADATA,sha256=lRjH5wempREQ2lTFNqzMusIW95YHN4rF8yd73MVvOe0,6782
+compressed_tensors-0.8.0.dist-info/WHEEL,sha256=bFJAMchF8aTQGUgMZzHJyDDMPTO3ToJ7x23SLJa1SVo,92
+compressed_tensors-0.8.0.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.8.0.dist-info/RECORD,,

{compressed_tensors-0.7.0.dist-info → compressed_tensors-0.8.0.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: bdist_wheel (0.44.0)
+Generator: bdist_wheel (0.45.0)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

compressed_tensors/quantization/cache.py

@@ -1,201 +0,0 @@
-# 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.
-
-
-from enum import Enum
-from typing import Any, Dict, List, Optional, Tuple
-
-from compressed_tensors.quantization.observers import Observer
-from compressed_tensors.quantization.quant_args import QuantizationArgs
-from torch import Tensor
-from transformers import DynamicCache as HFDyanmicCache
-
-
-class KVCacheScaleType(Enum):
-    KEY = "k_scale"
-    VALUE = "v_scale"
-
-
-class QuantizedKVParameterCache(HFDyanmicCache):
-
-    """
-    Quantized KV cache used in the forward call based on HF's dynamic cache.
-    Quantization strategy (tensor, group, channel) set from Quantization arg's strategy
-    Singleton, so that the same cache gets reused in all forward call of self_attn.
-    Each time forward is called, .update() is called, and ._quantize(), ._dequantize()
-     gets called appropriately.
-    The size of tensor is
-     `[batch_size, num_heads, seq_len - residual_length, head_dim]`.
-
-
-    Triggered by adding kv_cache_scheme in the recipe.
-
-    Example:
-
-    ```python3
-    recipe = '''
-    quant_stage:
-        quant_modifiers:
-            QuantizationModifier:
-                kv_cache_scheme:
-                    num_bits: 8
-                    type: float
-                    strategy: tensor
-                    dynamic: false
-                    symmetric: true
-    '''
-
-    """
-
-    _instance = None
-    _initialized = False
-
-    def __new__(cls, *args, **kwargs):
-        """Singleton"""
-        if cls._instance is None:
-            cls._instance = super(QuantizedKVParameterCache, cls).__new__(cls)
-        return cls._instance
-
-    def __init__(self, quantization_args: QuantizationArgs):
-        if not self._initialized:
-            super().__init__()
-
-            self.quantization_args = quantization_args
-
-            self.k_observers: List[Observer] = []
-            self.v_observers: List[Observer] = []
-
-            # each index corresponds to layer_idx of the attention layer
-            self.k_scales: List[Tensor] = []
-            self.v_scales: List[Tensor] = []
-
-            self.k_zps: List[Tensor] = []
-            self.v_zps: List[Tensor] = []
-
-            self._initialized = True
-
-    def update(
-        self,
-        key_states: Tensor,
-        value_states: Tensor,
-        layer_idx: int,
-        cache_kwargs: Optional[Dict[str, Any]] = None,
-    ) -> Tuple[Tensor, Tensor]:
-        """
-        Get the k_scale and v_scale and output the
-         fakequant-ed key_states and value_states
-        """
-
-        if len(self.k_observers) <= layer_idx:
-            k_observer = self.quantization_args.get_observer()
-            v_observer = self.quantization_args.get_observer()
-
-            self.k_observers.append(k_observer)
-            self.v_observers.append(v_observer)
-
-        q_key_states = self._quantize(
-            key_states.contiguous(), KVCacheScaleType.KEY, layer_idx
-        )
-        q_value_states = self._quantize(
-            value_states.contiguous(), KVCacheScaleType.VALUE, layer_idx
-        )
-
-        qdq_key_states = self._dequantize(q_key_states, KVCacheScaleType.KEY, layer_idx)
-        qdq_value_states = self._dequantize(
-            q_value_states, KVCacheScaleType.VALUE, layer_idx
-        )
-
-        keys_to_return, values_to_return = qdq_key_states, qdq_value_states
-
-        return keys_to_return, values_to_return
-
-    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
-        """
-        Returns the sequence length of the cached states.
-        A layer index can be optionally passed.
-        """
-        if len(self.key_cache) <= layer_idx:
-            return 0
-        # since we cannot get the seq_length of each layer directly and
-        # rely on `_seen_tokens` which is updated every "layer_idx" == 0,
-        # this is a hack to get the actual seq_length for the given layer_idx
-        # this part of code otherwise fails when used to
-        # verify attn_weight shape in some models
-        return self._seen_tokens if layer_idx == 0 else self._seen_tokens - 1
-
-    def reset_states(self):
-        """reset the kv states (used in calibration)"""
-        self.key_cache: List[Tensor] = []
-        self.value_cache: List[Tensor] = []
-        # Used in `generate` to keep tally of how many tokens the cache has seen
-        self._seen_tokens = 0
-        self._quantized_key_cache: List[Tensor] = []
-        self._quantized_value_cache: List[Tensor] = []
-
-    def reset(self):
-        """
-        Reset the instantiation, create new instance on init
-        """
-        QuantizedKVParameterCache._instance = None
-        QuantizedKVParameterCache._initialized = False
-
-    def _quantize(self, tensor, kv_type, layer_idx):
-        """Quantizes a key/value using a defined quantization method."""
-        from compressed_tensors.quantization.lifecycle.forward import quantize
-
-        if kv_type == KVCacheScaleType.KEY:  # key type
-            observer = self.k_observers[layer_idx]
-            scales = self.k_scales
-            zps = self.k_zps
-        else:
-            assert kv_type == KVCacheScaleType.VALUE
-            observer = self.v_observers[layer_idx]
-            scales = self.v_scales
-            zps = self.v_zps
-
-        scale, zp = observer(tensor)
-        if len(scales) <= layer_idx:
-            scales.append(scale)
-            zps.append(zp)
-        else:
-            scales[layer_idx] = scale
-            zps[layer_idx] = scale
-
-        q_tensor = quantize(
-            x=tensor,
-            scale=scale,
-            zero_point=zp,
-            args=self.quantization_args,
-        )
-        return q_tensor
-
-    def _dequantize(self, qtensor, kv_type, layer_idx):
-        """Dequantizes back the tensor that was quantized by `self._quantize()`"""
-        from compressed_tensors.quantization.lifecycle.forward import dequantize
-
-        if kv_type == KVCacheScaleType.KEY:
-            scale = self.k_scales[layer_idx]
-            zp = self.k_zps[layer_idx]
-        else:
-            assert kv_type == KVCacheScaleType.VALUE
-            scale = self.v_scales[layer_idx]
-            zp = self.v_zps[layer_idx]
-
-        qdq_tensor = dequantize(
-            x_q=qtensor,
-            scale=scale,
-            zero_point=zp,
-            args=self.quantization_args,
-        )
-        return qdq_tensor

compressed_tensors/quantization/lifecycle/calibration.py

@@ -1,70 +0,0 @@
-# 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 logging
-
-from compressed_tensors.quantization.quant_config import QuantizationStatus
-from compressed_tensors.utils import is_module_offloaded, update_parameter_data
-from torch.nn import Module
-
-
-__all__ = [
-    "set_module_for_calibration",
-]
-
-
-_LOGGER = logging.getLogger(__name__)
-
-
-def set_module_for_calibration(module: Module, quantize_weights_upfront: bool = True):
-    """
-    marks a layer as ready for calibration which activates observers
-    to update scales and zero points on each forward pass
-
-    apply to full model with `model.apply(set_module_for_calibration)`
-
-    :param module: module to set for calibration
-    :param quantize_weights_upfront: whether to automatically
-       run weight quantization at the start of calibration
-    """
-    if not getattr(module, "quantization_scheme", None):
-        # no quantization scheme nothing to do
-        return
-    status = getattr(module, "quantization_status", None)
-    if not status or status != QuantizationStatus.INITIALIZED:
-        _LOGGER.warning(
-            f"Attempting set module with status {status} to calibration mode. "
-            f"but status is not {QuantizationStatus.INITIALIZED} - you may "
-            "be calibrating an uninitialized module which may fail or attempting "
-            "to re-calibrate a frozen module"
-        )
-
-    if quantize_weights_upfront and module.quantization_scheme.weights is not None:
-        # set weight scale and zero_point up front, calibration data doesn't affect it
-        observer = module.weight_observer
-        g_idx = getattr(module, "weight_g_idx", None)
-
-        offloaded = is_module_offloaded(module)
-        if offloaded:
-            module._hf_hook.pre_forward(module)
-
-        scale, zero_point = observer(module.weight, g_idx=g_idx)
-        update_parameter_data(module, scale, "weight_scale")
-        update_parameter_data(module, zero_point, "weight_zero_point")
-
-        if offloaded:
-            module._hf_hook.post_forward(module, None)
-
-    module.quantization_status = QuantizationStatus.CALIBRATION

compressed_tensors/quantization/lifecycle/frozen.py

@@ -1,55 +0,0 @@
-# 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.
-
-
-from compressed_tensors.quantization.quant_config import QuantizationStatus
-from compressed_tensors.quantization.utils import is_kv_cache_quant_scheme
-from torch.nn import Module
-
-
-__all__ = [
-    "freeze_module_quantization",
-]
-
-
-def freeze_module_quantization(module: Module):
-    """
-    deletes observers so static quantization is completed.
-
-    apply to full model with `model.apply(freeze_module_quantization)`
-
-    :param module: module to freeze quantization for
-    """
-    scheme = getattr(module, "quantization_scheme", None)
-    if not scheme:
-        # no quantization scheme nothing to do
-        return
-
-    if module.quantization_status == QuantizationStatus.FROZEN:
-        # nothing to do, already frozen
-        return
-
-    # delete observers from module if not dynamic
-    if scheme.input_activations and not scheme.input_activations.dynamic:
-        delattr(module, "input_observer")
-    if scheme.weights and not scheme.weights.dynamic:
-        delattr(module, "weight_observer")
-    if (
-        scheme.output_activations
-        and not is_kv_cache_quant_scheme(scheme)
-        and not scheme.output_activations.dynamic
-    ):
-        delattr(module, "output_observer")
-
-    module.quantization_status = QuantizationStatus.FROZEN

compressed_tensors/quantization/observers/__init__.py

@@ -1,22 +0,0 @@
-# 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.
-
-# flake8: noqa
-# isort: skip_file
-
-from .helpers import *
-from .base import *
-from .memoryless import *
-from .min_max import *
-from .mse import *