fix test case

neuralmagic · Oct 13, 2024 · 2567660 · 2567660
1 parent 34fecdb
commit 2567660
Show file tree

Hide file tree

Showing 2 changed files with 9 additions and 3 deletions.
diff --git a/src/compressed_tensors/quantization/lifecycle/forward.py b/src/compressed_tensors/quantization/lifecycle/forward.py
@@ -22,7 +22,6 @@
     calculate_range,
     compute_dynamic_scales_and_zp,
 )
-from compressed_tensors.quantization.lifecycle.initialize import initialize_observers
 from compressed_tensors.quantization.quant_args import (
     QuantizationArgs,
     QuantizationStrategy,
@@ -279,7 +278,10 @@ def wrapped_forward(self, *args, **kwargs):
         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:
+            if (
+                module.quantization_status == QuantizationStatus.CALIBRATION
+                and not scheme.input_activations.dynamic
+            ):
                 calibrate_activations(
                     module=module,
                     value=input_,
@@ -307,7 +309,10 @@ def wrapped_forward(self, *args, **kwargs):
             # wrap_module_forward_quantized_attn
 
             # NOTE: will be removed from compressed-tensors
-            if module.quantization_status == QuantizationStatus.CALIBRATION:
+            if (
+                module.quantization_status == QuantizationStatus.CALIBRATION
+                and not scheme.output_activations.dynamic
+            ):
                 calibrate_activations(
                     module=module,
                     value=output,

diff --git a/tests/test_quantization/lifecycle/test_dynamic_lifecycle.py b/tests/test_quantization/lifecycle/test_dynamic_lifecycle.py
@@ -23,6 +23,7 @@
 
 
 def test_apply_tinyllama_dynamic_activations():
+    # NOTE: should not calibrate dynamic quant
     quant_config = get_sample_dynamic_tinyllama_quant_config()
     model = get_tinyllama_model()