Use regular division inside Scale Estimation (#3210)

### Changes

Compute division inside SE algorithm always as `a/b` instead of
`a*(1/b)` in some cases.

### Reason for changes

During implementation #2727 some choices were made regarding how
division operation is computed in order for the changes to be completely
aligned with the previous implementation. Namely, before #2727 some
divisions were computed as `a*(1/b)`, and this is currently still the
case.

The way these divisions are computed originally was not intended. Now,
all divisions are aligned to the `a/b` form.

Compression time and memory are roughly the same.

| Model | Compression | Compression Time Develop (sec.) | Compression
Time Branch (sec.) | Peak Memory Develop (MiB) | Peak Memory Branch
(MiB) |

|--------------|--------------|---------------------------------|--------------------------------|---------------------------|--------------------------|
| tiny-llama | int4, SE | 222* | 228* | 3030 | 3032 |
| phi4-mini | in4, SE | 789* | 790* | 10817 | 10768 |
| llama-3.1-8b | int4, SE | 1776* | 1801* | 17756 | 18224 |

*time column includes PT -> OV conversion time.

### Related tickets

163286

### Tests

- https://github.com/openvinotoolkit/nncf/actions/runs/13368886294
- NNCF/job/manual/job/post_training_weight_compression/324/
- OVVP validation ✅

Author: nikita-savelyevv

nncf/openvino/optimized_functions/functions.py

@@ -33,7 +33,6 @@ def do_int_quantization(
     reduction_axes: Optional[ReductionAxes] = None,
     precomputed_scale: Tensor = None,
     precomputed_zero_point: Tensor = None,
-    **kwargs,
 ) -> Tuple[Tensor, Tensor, Tensor]:
     """
     Quantizes the given weight tensor.
@@ -50,10 +49,7 @@ def do_int_quantization(
     scale_shape = None if precomputed_scale is None else precomputed_scale.shape
     zero_point_shape = None if precomputed_zero_point is None else precomputed_zero_point.shape
 
-    ov_model_params = OVModelParameters(
-        dynamic_shapes=kwargs.get("dynamic_shapes") is True,
-        convertable_division=kwargs.get("convertable_division") is True,
-    )
+    ov_model_params = OVModelParameters()
     ov_model_params.input_dtypes["weight"] = weight.dtype
     if precomputed_scale is not None:
         ov_model_params.input_dtypes["scale"] = precomputed_scale.dtype
@@ -108,7 +104,6 @@ def quantize_dequantize_weight(
     precomputed_scale: Optional[Tensor] = None,
     precomputed_zero_point: Optional[Tensor] = None,
     return_compressed_weight: Optional[bool] = False,
-    **kwargs,
 ) -> Union[Tensor, Tuple[Tensor, Tensor, Tensor, Tensor]]:
     """
     Quantizes the given weight tensor and then dequantizes it back to obtain float32 values.
@@ -133,10 +128,7 @@ def quantize_dequantize_weight(
     scale_shape = precomputed_scale.shape if precomputed_scale is not None else None
     zero_point_shape = precomputed_zero_point.shape if precomputed_zero_point is not None else None
 
-    ov_model_params = OVModelParameters(
-        dynamic_shapes=kwargs.get("dynamic_shapes") is True,
-        convertable_division=kwargs.get("convertable_division") is True,
-    )
+    ov_model_params = OVModelParameters()
     ov_model_params.input_dtypes["weight"] = weight.dtype
     if precomputed_scale is not None:
         ov_model_params.input_dtypes["scale"] = precomputed_scale.dtype

nncf/quantization/algorithms/weight_compression/scale_estimation.py

@@ -246,10 +246,6 @@ def calculate_quantization_params(
         zero_scale = 0.001
         zero_mask = zero_scale * zero_mask.astype(original_weight.dtype)
 
-        # This is required for alignment with a previous OpenVINO models implementation
-        # TODO(Nikita Savelyev): remove this
-        opt_fns_kwargs = dict(dynamic_shapes=False, convertable_division=True)
-
         # iterative rectification of initial scale
         for i in range(initial_steps):
             near_to_ideal_scale = estimate_scales(original_weight, target, zero_mask, importance)
@@ -264,7 +260,6 @@ def calculate_quantization_params(
                     config,
                     precomputed_scale=near_to_ideal_scale,
                     precomputed_zero_point=zp,
-                    **opt_fns_kwargs,
                 )
 
             q_weights_ = fns.zeros_like(original_weight) + out
@@ -299,7 +294,6 @@ def calculate_quantization_params(
                         config,
                         precomputed_scale=near_to_ideal_scale,
                         precomputed_zero_point=zp,
-                        **opt_fns_kwargs,
                     )
                 compressed_weights = fns.zeros_like(original_weight) + out
                 target, zero_mask = get_target_zero_mask(compressed_weights, zp)
@@ -318,7 +312,6 @@ def calculate_quantization_params(
                     config,
                     precomputed_scale=scaled_scale,
                     precomputed_zero_point=zp,
-                    **opt_fns_kwargs,
                 )
             compressed_weights = fns.zeros_like(original_weight) + out
 
@@ -336,7 +329,6 @@ def calculate_quantization_params(
                     config,
                     precomputed_scale=near_to_ideal_scale,
                     precomputed_zero_point=zp,
-                    **opt_fns_kwargs,
                 )
             q_weights_ = fns.zeros_like(original_weight) + out
 

nncf/quantization/algorithms/weight_compression/weight_lowering.py

@@ -443,7 +443,6 @@ def do_int_quantization(
     reduction_axes: Optional[ReductionAxes] = None,
     precomputed_scale: Tensor = None,
     precomputed_zero_point: Tensor = None,
-    **kwargs,
 ) -> Tuple[Tensor, Tensor, Tensor]:
     """
     Performs integer quantization on the given weight tensor.
@@ -475,9 +474,7 @@ def do_int_quantization(
     if _can_run_optimized(weight.backend):
         from nncf.openvino.optimized_functions import do_int_quantization as do_int_quantization_ov
 
-        return do_int_quantization_ov(
-            weight, config, reduction_axes, precomputed_scale, precomputed_zero_point, **kwargs
-        )
+        return do_int_quantization_ov(weight, config, reduction_axes, precomputed_scale, precomputed_zero_point)
 
     # Reference implementation
     if weight.backend == TensorBackend.ov:
@@ -507,7 +504,6 @@ def quantize_dequantize_weight(
     precomputed_scale: Optional[Tensor] = None,
     precomputed_zero_point: Optional[Tensor] = None,
     return_compressed_weight: Optional[bool] = False,
-    **kwargs,
 ) -> Union[Tensor, Tuple[Tensor, Tensor, Tensor, Tensor]]:
     """
     First quantizes the given weight tensor and then dequantizes it back to obtain float32 values.
@@ -533,7 +529,6 @@ def quantize_dequantize_weight(
             precomputed_scale,
             precomputed_zero_point,
             return_compressed_weight,
-            **kwargs,
         )
 
     # Reference implementation

tests/post_training/data/wc_reference_data.yaml

@@ -36,11 +36,11 @@ tinyllama_data_aware_gptq_scale_estimation_stateful_backend_OV:
   num_int8: 124
   metrics_xfail_reason: "Issue-148819"
 tinyllama_scale_estimation_per_channel_backend_OV:
-  metric_value: 0.81389
+  metric_value: 0.80873
   num_int4: 188
   num_int8: 124
 tinyllama_scale_estimation_per_channel_backend_TORCH:
-  metric_value: 0.81389
+  metric_value: 0.80873
   num_int4: 188
   num_int8: 124
   atol: 0.006 # difference across devices: 0.80873 vs 0.81389

tests/post_training/data/wc_reference_data_2025.0.yaml

@@ -14,7 +14,7 @@ tinyllama_data_aware_backend_TORCH:
   num_int8: 124
   num_compressed_xfail_reason: "Issue-160006"
 tinyllama_scale_estimation_per_channel_backend_TORCH:
-  metric_value: 0.81389
+  metric_value: 0.80873
   num_int4: 188
   num_int8: 124
   atol: 0.006 # difference across devices: 0.80873 vs 0.81389