compressed-tensors 0.12.3a20251003__py3-none-any.whl → 0.12.3a20251007__py3-none-any.whl

compressed_tensors/quantization/lifecycle/initialize.py

@@ -14,14 +14,13 @@
 
 
 import logging
-import math
-import warnings
-from typing import Optional
+from typing import Optional, Tuple
 
 import torch
 from compressed_tensors.quantization import (
     FP8_E4M3_DATA,
     ActivationOrdering,
+    DynamicType,
     KVCacheScaleType,
     QuantizationArgs,
     QuantizationMetadata,
@@ -32,7 +31,11 @@ from compressed_tensors.quantization import (
 from compressed_tensors.quantization.lifecycle.forward import (
     wrap_module_forward_quantized,
 )
-from compressed_tensors.quantization.utils import is_fp4, is_kv_cache_quant_scheme
+from compressed_tensors.quantization.utils import (
+    is_fp4,
+    is_kv_cache_quant_scheme,
+    strategy_cdiv,
+)
 from compressed_tensors.utils import (
     disable_hf_hook,
     get_execution_device,
@@ -44,6 +47,7 @@ from torch.nn import Module, Parameter
 __all__ = [
     "initialize_module_for_quantization",
     "is_attention_module",
+    "initialize_qparams",
 ]
 
 
@@ -69,10 +73,8 @@ def initialize_module_for_quantization(
     :param force_zero_point: whether to force initialization of a zero point for
         symmetric quantization
     """
-    # TODO: don't initialize parameters when running decompression
     scheme = scheme or getattr(module, "quantization_scheme", None)
     if scheme is None:
-        # no scheme passed and layer not targeted for quantization - skip
         return
 
     QuantizationMetadata.clear_all_qparams(module)
@@ -82,38 +84,52 @@ def initialize_module_for_quantization(
         _initialize_attn_scales(module)
 
     else:
+        if not isinstance(module, torch.nn.Linear):
+            _LOGGER.warning(f"Attempting to quantize module of type {type(module)}")
+
+        # use weight to determine observed shapes and dtype
+        if hasattr(module, "weight"):
+            weight = module.weight
+            assert isinstance(weight, torch.Tensor)
+        else:
+            # Note that a weight is required for both weight and activation
+            # quantization in order to know the dtype of activation scales
+            _LOGGER.warning(
+                f"module type {type(module)} targeted for quantization but "
+                f"has no attribute weight, skipping quantization for {type(module)}"
+            )
+            return
+
         if scheme.input_activations is not None:
-            _initialize_scale_zero_point(
+            initialize_qparams(
                 module,
                 "input",
                 scheme.input_activations,
+                observed_shape=weight.shape[-1:],
+                observed_dtype=weight.dtype,
                 force_zero_point=force_zero_point,
             )
 
         if scheme.weights is not None:
-            if hasattr(module, "weight"):
-                weight_shape = None
-                if isinstance(module, torch.nn.Linear):
-                    weight_shape = module.weight.shape
-                _initialize_scale_zero_point(
-                    module,
-                    "weight",
-                    scheme.weights,
-                    weight_shape=weight_shape,
-                    force_zero_point=force_zero_point,
-                )
-            else:
-                _LOGGER.warning(
-                    f"module type {type(module)} targeted for weight quantization but "
-                    "has no attribute weight, skipping weight quantization "
-                    f"for {type(module)}"
-                )
-
-        if scheme.output_activations is not None:
-            if not is_kv_cache_quant_scheme(scheme):
-                _initialize_scale_zero_point(
-                    module, "output", scheme.output_activations
-                )
+            initialize_qparams(
+                module,
+                "weight",
+                scheme.weights,
+                observed_shape=weight.shape,
+                observed_dtype=weight.dtype,
+                force_zero_point=force_zero_point,
+            )
+
+        output_is_kv_cache = is_kv_cache_quant_scheme(scheme)
+        if scheme.output_activations is not None and not output_is_kv_cache:
+            initialize_qparams(
+                module,
+                "output",
+                scheme.output_activations,
+                observed_shape=weight.shape[:-1],
+                observed_dtype=weight.dtype,
+                force_zero_point=force_zero_point,
+            )
 
         module.quantization_scheme = scheme
         module.quantization_status = QuantizationStatus.INITIALIZED
@@ -132,22 +148,40 @@ def is_attention_module(module: Module):
     )
 
 
-def _initialize_scale_zero_point(
+def initialize_qparams(
     module: Module,
     base_name: str,
     quantization_args: QuantizationArgs,
-    weight_shape: Optional[torch.Size] = None,
+    observed_shape: Tuple[int],
+    observed_dtype: torch.dtype,
     force_zero_point: bool = True,
 ):
-    if quantization_args.dynamic is True:
-        return
+    """
+    Initialize quantization parameters for a given basename according to the passed
+    quantization args. The shape and dtype of the observed weight/activation must also
+    be provided.
+
+    Scales will always be initialized. Global scales are initialized depending on args.
+    Zero points will be initialized if not symmetric or if `force_zero_point` is True.
+
+    :param module: module to register qparams to
+    :param base_name: base name of qparams, for example "input", "weight", "k", "v"
+    :param quantization_args: arguments for quantization
+    :param observed_shape: last (right-most) known dimensions of the observed weight/act
+    :param observed_dtype: dtype of the observed weight/actt
+    :param force_zero_point: force the zero_point parameter to be initialized
+    """
+    strategy = quantization_args.strategy
+    dynamic = quantization_args.dynamic
+    actorder = quantization_args.actorder
+    device = get_execution_device(module)  # avoid performing intialization ops on cpu
 
-    # initialize on execution device to avoid performing quantized ops on cpu
-    device = get_execution_device(module)
+    # Skip all intialization for fully dynamic quantization
+    if dynamic is True:
+        return
 
-    # 1. Create global_scales for tensor_group - generates
-    # a per tensor scale
-    if quantization_args.strategy == QuantizationStrategy.TENSOR_GROUP:
+    # 0. Create global scale for tensor-group quantization
+    if strategy == QuantizationStrategy.TENSOR_GROUP:
         init_global_scale = Parameter(
             torch.empty(1, dtype=torch.float32, device=device),
             requires_grad=False,
@@ -156,56 +190,55 @@ def _initialize_scale_zero_point(
             module, f"{base_name}_global_scale", init_global_scale
         )
 
-    # 2. Infer expected scale/zero point shape
-    if quantization_args.strategy == QuantizationStrategy.TOKEN:
+    # Skip scale/zp initialization for locally dynamic quantization
+    if dynamic == DynamicType.LOCAL:
+        return
+
+    # 1. Infer expected scale/zp shape
+    if strategy == QuantizationStrategy.TENSOR:
+        expected_shape = (1,)
+
+    elif strategy == QuantizationStrategy.TOKEN:
         expected_shape = (1, 1)
+
+    elif strategy == QuantizationStrategy.CHANNEL:
+        if len(observed_shape) < 2:
+            raise ValueError("Channel quant requires at least 2 observed dimensions")
+
+        expected_shape = (observed_shape[-2], 1)
+
+    elif strategy in (QuantizationStrategy.GROUP, QuantizationStrategy.TENSOR_GROUP):
+        assert quantization_args.group_size is not None
+        if len(observed_shape) < 1:
+            raise ValueError("Group quant requires at least 1 observed dimension")
+
+        group_size = quantization_args.group_size
+        num_groups = strategy_cdiv(observed_shape[-1], group_size, strategy)
+        expected_shape = (*observed_shape[:-1], num_groups)
+
+        # initialize activation ordering if applicable
+        if actorder == ActivationOrdering.GROUP:
+            init_g_idx = Parameter(
+                torch.full((observed_shape[-1],), -1, device=device, dtype=torch.int),
+                requires_grad=False,
+            )
+            register_offload_parameter(module, f"{base_name}_g_idx", init_g_idx)
+
+    elif strategy == QuantizationStrategy.BLOCK:
+        assert quantization_args.block_structure is not None
+        if len(observed_shape) < 2:
+            raise ValueError("Block quant requires at least 2 observed dimensions")
+
+        block_structure = quantization_args.block_structure
+        num_rows = strategy_cdiv(observed_shape[-2], block_structure[-2], strategy)
+        num_cols = strategy_cdiv(observed_shape[-1], block_structure[-1], strategy)
+        expected_shape = (num_rows, num_cols)
+
     else:
-        expected_shape = 1
-
-    if base_name == "weight" and weight_shape is not None:
-        if quantization_args.strategy == QuantizationStrategy.CHANNEL:
-            # (output_channels, 1) - only for weights
-            expected_shape = (weight_shape[0], 1)
-        elif quantization_args.strategy in (
-            QuantizationStrategy.TENSOR_GROUP,
-            QuantizationStrategy.GROUP,
-        ):
-            # GROUP/TENSOR_GROUP for both weights and activations
-            num_groups = math.ceil(weight_shape[1] / quantization_args.group_size)
-            expected_shape = (weight_shape[0], max(num_groups, 1))
-        elif quantization_args.strategy == QuantizationStrategy.BLOCK:
-            # For block quantization, scale shape should match number of blocks - only
-            # for weights
-            if quantization_args.block_structure is None:
-                raise ValueError(
-                    "Block quantization requires block_structure to be specified"
-                )
-            block_height, block_width = quantization_args.block_structure
-            rows, cols = weight_shape[-2], weight_shape[-1]
-            num_rows_blocks = math.ceil(rows / block_height)
-            num_cols_blocks = math.ceil(cols / block_width)
-
-            # Warn if dimensions don't divide evenly
-            if rows % block_height != 0 or cols % block_width != 0:
-                warnings.warn(
-                    f"Block quantization: tensor shape {weight_shape} does not divide"
-                    f"evenly by block structure {quantization_args.block_structure}. "
-                    f"Some blocks will be incomplete which may affect quantization"
-                    "quality.",
-                    UserWarning,
-                )
-
-            expected_shape = (num_rows_blocks, num_cols_blocks)
-    elif quantization_args.strategy == QuantizationStrategy.BLOCK:
-        warnings.warn(
-            f"BLOCK quantization not supported for {base_name} activations. "
-            f"Falling back to tensor-level quantization.",
-            UserWarning,
-        )
-        expected_shape = 1
+        assert False, f"Unknown strategy {strategy}"
 
-    # 3. Identify quantization scale and zp dtype
-    scale_dtype = module.weight.dtype
+    # 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
@@ -221,14 +254,12 @@ def _initialize_scale_zero_point(
             scale_dtype = torch.bfloat16
         zp_dtype = quantization_args.pytorch_dtype()
 
-    # 4. Initializes empty scale, zero point, and g_idx parameters for the module
-    # do not init scales for quantzation_args.dynamic == DynamicType.local
-    if not quantization_args.dynamic:
-        init_scale = Parameter(
-            torch.empty(expected_shape, dtype=scale_dtype, device=device),
-            requires_grad=False,
-        )
-        register_offload_parameter(module, f"{base_name}_scale", init_scale)
+    # 3. Initializes scale/zp for the module
+    init_scale = Parameter(
+        torch.empty(expected_shape, dtype=scale_dtype, device=device),
+        requires_grad=False,
+    )
+    register_offload_parameter(module, f"{base_name}_scale", init_scale)
 
     if force_zero_point or not quantization_args.symmetric:
         init_zero_point = Parameter(
@@ -237,16 +268,6 @@ def _initialize_scale_zero_point(
         )
         register_offload_parameter(module, f"{base_name}_zero_point", init_zero_point)
 
-    # only grouped activation ordering has g_idx
-    if quantization_args.actorder == ActivationOrdering.GROUP:
-        g_idx_shape = (weight_shape[1],)
-        g_idx_dtype = torch.int
-        init_g_idx = Parameter(
-            torch.full(g_idx_shape, -1, device=device, dtype=g_idx_dtype),
-            requires_grad=False,
-        )
-        register_offload_parameter(module, f"{base_name}_g_idx", init_g_idx)
-
 
 def _initialize_attn_scales(module: Module) -> None:
     """Initlaize k_scale, v_scale for  self_attn"""

compressed_tensors/quantization/utils/helpers.py

@@ -27,6 +27,7 @@ from compressed_tensors.quantization.quant_args import (
 )
 from compressed_tensors.quantization.quant_scheme import QuantizationScheme
 from compressed_tensors.utils import deprecated
+from loguru import logger
 from torch import FloatTensor, IntTensor, Tensor
 from torch.nn import Module
 
@@ -47,6 +48,7 @@ __all__ = [
     "calculate_qparams",
     "generate_gparam",
     "is_fp4",
+    "strategy_cdiv",
 ]
 
 # target the self_attn layer
@@ -461,3 +463,26 @@ def generate_gparam(
     max_val_pos = torch.max(torch.abs(min_vals), torch.abs(max_vals))
     global_scale = scale_data.max * quant_data.max / max_val_pos
     return global_scale.to(dtype).reshape([1])
+
+
+def strategy_cdiv(
+    value: int,
+    divisor: int,
+    strategy: Optional[QuantizationStrategy],
+    strict: bool = False,
+) -> int:
+    dividend = math.ceil(value / divisor)
+    if dividend * divisor != value:
+        message = (
+            f"{strategy} quantization strategy requires strict division of "
+            f"weight/activation size {value} and group/block size {divisor}. "
+            "consider reducing the group/block size or ignoring modules with "
+            f"weights not divisible by {divisor}"
+        )
+        if strict:
+            raise ValueError(message)
+
+        else:
+            logger.bind(log_once=True).warning(message)
+
+    return dividend

compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.12.3.a20251003'
+__version__ = version = '0.12.3.a20251007'
 __version_tuple__ = version_tuple = (0, 12, 3)

{compressed_tensors-0.12.3a20251003.dist-info → compressed_tensors-0.12.3a20251007.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.3a20251003
+Version: 0.12.3a20251007
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.12.3a20251003.dist-info → compressed_tensors-0.12.3a20251007.dist-info}/RECORD

@@ -1,7 +1,7 @@
 compressed_tensors/__init__.py,sha256=SRqNYFVvxAaLa4SImhoiIBKfoOSj7EUdx0CxXjGC2PA,884
 compressed_tensors/base.py,sha256=-gxWvDF4LCkyeDP8YlGzvBBKxo4Dk9h4NINPD61drFU,921
 compressed_tensors/logger.py,sha256=sTm1Od1cV0aDxBm3YN-PPvsOATxY_2tBV62TQE4HiPw,4032
-compressed_tensors/version.py,sha256=8cvrsHpab_kfxFsOLRqkafZiztV6QoX8qSdtWAslAE8,523
+compressed_tensors/version.py,sha256=wUAf4k-SKBfmo_lva-t_YeLxw6mVgmYfFIXdh6YQLP4,523
 compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1tWZewqVnOb4lM,825
 compressed_tensors/compressors/base.py,sha256=nvWsv4xEw1Tkxkxth6TmHplDYXfBeP22xWxOsZERyDY,7204
 compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
@@ -37,9 +37,9 @@ compressed_tensors/quantization/lifecycle/apply.py,sha256=1zRc7tQbE5OAVJ5VRgU9FZ
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=_gTH0CnLe8MxkTY1hrCCeSYAMzuvIwoCTT4hxW1TPk4,2354
 compressed_tensors/quantization/lifecycle/forward.py,sha256=MAw049L4a9ha4P5D4MjOMoIcSwv9_ZXizahYzHJaaQI,17550
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
-compressed_tensors/quantization/lifecycle/initialize.py,sha256=1vCHN1LLSIoh6IgGAyE-Jan64HvuiChwdGhslbYkXhk,9832
+compressed_tensors/quantization/lifecycle/initialize.py,sha256=xebqRiQz3hiSTYwCQQsovg-IKJtHkAbuj6eWygf5yKY,10259
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
-compressed_tensors/quantization/utils/helpers.py,sha256=fyAjq0YbrZG6_UbEU7n6Z1i-Bg4PLA-mQiu7mHS9fs0,16420
+compressed_tensors/quantization/utils/helpers.py,sha256=BA-twfAKk-HMBr_OZHZnSQN7F1a0l5zB1kJhml6j-cI,17146
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
 compressed_tensors/registry/registry.py,sha256=cWnlwZ66lgG0w9OAUEAgq5XVxqsgFm1o8ZYdNhkNvJY,11957
 compressed_tensors/transform/__init__.py,sha256=v2wfl4CMfA6KbD7Hxx_MbRev63y_6QLDlccZq-WTtdw,907
@@ -65,8 +65,8 @@ compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVy
 compressed_tensors/utils/safetensors_load.py,sha256=Vql34aCTDHwmTZXJHzCyBISJo7iA7EQ78LdTlMjdpZo,12023
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
 compressed_tensors/utils/type.py,sha256=bNwoo_FWlvLuDpYAGGzZJITRg0JA_Ngk9LGPo-kvjeU,2554
-compressed_tensors-0.12.3a20251003.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.12.3a20251003.dist-info/METADATA,sha256=8jYvFdUw29ME2nRPHzr3TcPKJ6wtt_iP2pfjIF6WzY0,7027
-compressed_tensors-0.12.3a20251003.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-compressed_tensors-0.12.3a20251003.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.12.3a20251003.dist-info/RECORD,,
+compressed_tensors-0.12.3a20251007.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.12.3a20251007.dist-info/METADATA,sha256=8uUWt8bF7sZhbMQd2Llj4PDLC7I4ALKJE_eFNa8DBWI,7027
+compressed_tensors-0.12.3a20251007.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+compressed_tensors-0.12.3a20251007.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.12.3a20251007.dist-info/RECORD,,