Preserve histogram counts during entropy calibration

CHANGELOG.rst

@@ -101,6 +101,8 @@ Changelog
 
 **Bug Fixes**
 
+- Preserve accumulated histogram counts when computing an entropy calibration threshold with NumPy histograms.
+
 - Fix Megatron unified HF export of MoE models with grouped-GEMM experts when only the experts are quantized (e.g. ``nvfp4_experts_only-*`` recipes): ``hf_quant_config.json`` and the ``quantization_config`` in ``config.json`` were not written, so the quantized experts were served as unquantized weights. Re-export such checkpoints.
 - Fix Megatron-Core checkpoint saving for quantized grouped MoE experts when tensor and expert parallelism are both enabled.
 - Fix unified HuggingFace export of RADIO-based VLMs retaining post-conversion vision and

modelopt/torch/quantization/calib/histogram.py

@@ -218,7 +218,7 @@ def _normalize_distr(distr):
         if summ != 0:
             distr = distr / summ
 
-    bins = calib_hist[:]
+    bins = calib_hist.copy()
     bins[0] = bins[1]
 
     total_data = np.sum(bins)

tests/unit/torch/quantization/test_calibrator.py

@@ -89,6 +89,32 @@ def test_track_amax_raises(self):
 
 
 class TestHistogramCalibrator:
+    @pytest.mark.parametrize("torch_hist", [False, True])
+    def test_entropy_preserves_collected_histogram(self, torch_hist):
+        calibrator = calib.HistogramCalibrator(4, None, False, num_bins=32, torch_hist=torch_hist)
+        reference = calib.HistogramCalibrator(4, None, False, num_bins=32, torch_hist=torch_hist)
+        first_batch = torch.cat((torch.zeros(100), torch.arange(1, 33).float()))
+        calibrator.collect(first_batch)
+        reference.collect(first_batch)
+        expected_hist = np.asarray(calibrator._calib_hist).copy()
+        expected_edges = np.asarray(calibrator._calib_bin_edges).copy()
+        percentile = calibrator.compute_amax("percentile", percentile=50)
+
+        first_entropy = calibrator.compute_amax("entropy", start_bin=16)
+        second_entropy = calibrator.compute_amax("entropy", start_bin=16)
+
+        np.testing.assert_array_equal(calibrator._calib_hist, expected_hist)
+        np.testing.assert_array_equal(calibrator._calib_bin_edges, expected_edges)
+        assert calibrator.compute_amax("percentile", percentile=50) == percentile
+        assert first_entropy == second_entropy
+        next_batch = torch.tensor([0.0, 2.0, 4.0, 64.0])
+        calibrator.collect(next_batch)
+        reference.collect(next_batch)
+        np.testing.assert_array_equal(calibrator._calib_hist, reference._calib_hist)
+        assert calibrator.compute_amax("percentile", percentile=50) == reference.compute_amax(
+            "percentile", percentile=50
+        )
+
     @pytest.mark.skip(reason="TODO: Fix assertions in test_grow")
     def test_grow(self, verbose):
         x_1 = torch.tensor([0, 255, 255, 255, 255, 255])