Fix wrong scale eps applied

Author: alexsamardzic

test/quantization/test_quant_primitives.py

@@ -961,6 +961,69 @@ def test_float8_quant_primitives(self, hp_dtype, float8_dtype):
         torch.testing.assert_close(expected_quantized, quantized)
         torch.testing.assert_close(expected_dequantized, dequantized)
 
+    @parameterized.expand(
+        [
+            torch.float64,
+            torch.float32,
+            torch.bfloat16,
+            torch.float16,
+        ]
+    )
+    def test_choose_qparams_affine_for_inf_scale_reciprocal(self, hp_dtype):
+        # Fixed by #1770, the test will fail for all the variants
+        # before that fix, and will pass afterwards.
+        #
+        # The scale value must be checked, within
+        # _choose_qparams_affine() function, (that
+        # choose_qparams_affine() and others call into) to a large
+        # enough number so that its reciprocal does not become Inf.
+        # Otherwise during the quantization, by multiplying with scale
+        # reciprocal, all the values will be quantized to Inf value,
+        # except from zero value that would produce NaN (0*Inf) as
+        # quantized value.
+        #
+        # The minimal normalized value for given floating point data
+        # type is given by torch.finfo(hp_dtype).tiny - let's call
+        # this value "tiny".  It could be checked, that for
+        # all of torch.float64, torch.float32, torch.float16 and
+        # torch.floatb16, denormalized number that is equal to tiny/4
+        # will produce Inf as its reciprocal.
+        #
+        # Thus, to reproduce the problem, one would create a tensor
+        # with such values that their absolute maximum, after being
+        # divided with the range of quantized data (that is 57344 for
+        # torch.float8_e5m2), would produce scale smaller than tiny/4.
+        # Also, eps parameter should be set to value no greater than
+        # tiny/4, as scale is clamped from below to that value.  With
+        # such inputs, choose_qparams_affine() would produce Inf as
+        # scale value, if not checked for it.
+        #
+        # Note that this may seem as contrieved reproducer.  However,
+        # there are cases with existing code that would pass
+        # torch.finfo(torch.float32).eps as eps value, no matters of
+        # scale_dtype.  The float16 has rather small range, so this
+        # value is well bellow torch.finfo(torch.float32).eps, and for
+        # such eps value, the code bellow would produce Inf scale even
+        # for float16 tensor that has 0.5 as its maximum value.
+        float8_dtype = torch.float8_e5m2
+        tiny = torch.finfo(hp_dtype).tiny
+        x = torch.tensor([[0, 100 * tiny]], dtype=hp_dtype)
+
+        import pytest
+
+        with pytest.raises(AssertionError) as exc_info:
+            scale, _ = choose_qparams_affine(
+                input=x,
+                mapping_type=MappingType.SYMMETRIC,
+                block_size=[1, 2],
+                target_dtype=float8_dtype,
+                eps=tiny / 4,
+                scale_dtype=hp_dtype,
+                preserve_zero=True,
+                zero_point_domain=ZeroPointDomain.NONE,
+            )
+        assert str(exc_info.value) == "Invalid scale value"
+
 
 if __name__ == "__main__":
     unittest.main()

torchao/quantization/quant_primitives.py

@@ -862,6 +862,7 @@ def _choose_qparams_affine(
     3. calculate quantization parameters based on min_val/max_val based on args like `preserve_zero`
        and `zero_point_domain`
     """
+
     quant_min, quant_max = _get_and_check_qmin_qmax(target_dtype, quant_min, quant_max)
     assert mapping_type in [
         MappingType.SYMMETRIC.name,
@@ -877,7 +878,7 @@ def _choose_qparams_affine(
         if scale_dtype is None:
             scale_dtype = input.dtype
         if eps is None:
-            eps = torch.finfo(input.dtype).eps
+            eps = torch.finfo(input.dtype).tiny
 
         assert (
             len(block_size) == input.dim()
@@ -900,7 +901,7 @@ def _choose_qparams_affine(
         if scale_dtype is None:
             scale_dtype = min_val.dtype
         if eps is None:
-            eps = torch.finfo(min_val.dtype).eps
+            eps = torch.finfo(min_val.dtype).tiny
 
     if preserve_zero:
         min_val_neg = torch.min(min_val, torch.zeros_like(min_val))
@@ -950,6 +951,13 @@ def _choose_qparams_affine(
         assert mapping_type == MappingType.ASYMMETRIC.name
         scale = (max_val_pos - min_val_neg) / float(quant_max - quant_min)
         scale = torch.clamp(scale, min=eps)
+        # Prevent scale reciprocal to become Inf.
+        assert torch.all(
+            scale
+            >= torch.finfo(
+                scale.dtype if torch.is_floating_point(scale) else torch.float32
+            ).tiny
+        ), "Invalid scale value"
         if zero_point_domain == ZeroPointDomain.NONE.name:
             zero_point = None
         elif zero_point_domain == ZeroPointDomain.INT.name:
@@ -969,7 +977,11 @@ def _choose_qparams_affine(
 
     if zero_point is not None:
         zero_point = zero_point.to(dtype=zero_point_dtype)
-    return scale.to(dtype=scale_dtype), zero_point
+    scale = scale.to(dtype=scale_dtype)
+    # Prevent scale reciprocal to become Inf.
+    if torch.is_floating_point(scale):
+        assert torch.all(scale >= torch.finfo(scale_dtype).tiny), "Invalid scale value"
+    return scale, zero_point
 
 
 def choose_qparams_and_quantize_affine_qqq(