use helper in initialize

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>

Author: shanjiaz

src/compressed_tensors/quantization/lifecycle/initialize.py

@@ -198,27 +198,27 @@ def initialize_qparams(
         return
 
     # 1. Infer expected scale/zp shape
-    if strategy == QuantizationStrategy.TENSOR:
-        expected_shape = (1,)
-
-    elif strategy == QuantizationStrategy.TOKEN:
+    if strategy == QuantizationStrategy.TOKEN:
         raise ValueError("Cannot perform static token quantization")
 
-    elif strategy == QuantizationStrategy.CHANNEL:
-        if len(observed_shape) < 2:
+    elif strategy in (
+        QuantizationStrategy.TENSOR,
+        QuantizationStrategy.CHANNEL,
+        QuantizationStrategy.GROUP,
+        QuantizationStrategy.TENSOR_GROUP,
+    ):
+        # Validate shape requirements
+        if strategy == QuantizationStrategy.CHANNEL and 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")
+        if 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")
 
         # Use shared calculation to avoid floor division bugs
-        _, expected_shape = calculate_qparam_shape(
-            torch.Size(observed_shape), quantization_args
-        )
+        # Note: observed_shape may contain None for dynamic dimensions (e.g., sequence length)
+        # but calculate_qparam_shape only accesses specific indices that are concrete
+        _, expected_shape = calculate_qparam_shape(observed_shape, quantization_args)
 
         # initialize activation ordering if applicable
         if actorder == ActivationOrdering.GROUP:

src/compressed_tensors/quantization/utils/helpers.py

@@ -127,11 +127,9 @@ def calculate_qparams(
     # 5. 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
-        )
+        dtype=quantization_args.scale_dtype
+        if quantization_args.scale_dtype is not None
+        else scales.dtype
     )
     scales = torch.where(
         scales == 0,
@@ -483,20 +481,19 @@ def calculate_qparam_shape(
 
     if strategy == QuantizationStrategy.TENSOR:
         num_groups = 1
-        expected_shape = torch.Size((1,))
+        expected_shape = (1,)
 
     elif strategy == QuantizationStrategy.CHANNEL:
         num_groups = 1
-        expected_shape = torch.Size((weight_shape[0], 1))
+        expected_shape = (weight_shape[0], 1)
 
     elif strategy in (QuantizationStrategy.GROUP, QuantizationStrategy.TENSOR_GROUP):
         group_size = quantization_args.group_size
         if group_size is None:
             raise ValueError(f"{strategy} quantization requires group_size to be set")
 
-        # Use strategy_cdiv for proper ceiling division and validation
         num_groups = strategy_cdiv(weight_shape[-1], group_size, strategy)
-        expected_shape = torch.Size((weight_shape[0], num_groups))
+        expected_shape = (weight_shape[0], num_groups)
 
     else:
         raise ValueError(