Allow 1 manitissa bit diff in TestFused8BitRowwiseQuantizationConversion (#2015)

Summary:
Pull Request resolved: #2015

The reference implementation of FP8 quantization is in Python, but the
actual implementation is in C++/CUDA.  Upon summerdengfb's investigation,
Python has a known floating point representation issue
(https://www.geeksforgeeks.org/floating-point-error-in-python/). This
could cause quantization result discrepancy.  To workaround this
issue, we allow 1 bit difference in the FP8 quantization result (LSB
of mantissa) in `TestFused8BitRowwiseQuantizationConversion`.

Reviewed By: q10, shintaro-iwasaki

Differential Revision: D49255499

fbshipit-source-id: b28294f8076bda61589e10699119375f03b091a8

Author: sryap

fbgemm_gpu/test/quantize_ops_test.py

@@ -118,7 +118,10 @@ def test_quantize_op(
             ncols_aligned = (ncols + 4 - 1) // 4 * 4
             # compare quantized data
             np.testing.assert_allclose(
-                quantized_data_numpy[:, :ncols], reference[:, :ncols]
+                quantized_data_numpy[:, :ncols],
+                reference[:, :ncols],
+                # Allow 1 mantissa bit difference (LSB)
+                atol=1,
             )
             # compare scales
             np.testing.assert_array_almost_equal(