frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Quant][CPU] fix fake_quantize_per_tensor_affine of inf values (#155109)

Browse Source 
Fixes #154328

**Summary**
Fail reason:
The input value is infinity in float and it has undefined behavior to convert it to int64_t. On X86, it will be converted to the min value of int64_t, which is not expected.

Fix:
Clamping `(input * inv_scale + zero_point)` to `[quant_min, quant_max]` before converting it to int64_t.

**Test plan**
```
pytest test/quantization/core/test_workflow_ops.py -k test_fake_quantize_per_tensor_affine_inf
```

Pull Request resolved: https://github.com/pytorch/pytorch/pull/155109
Approved by: https://github.com/leslie-fang-intel, https://github.com/jerryzh168
This commit is contained in:
Xia, Weiwen
2025-06-14 14:12:38 +00:00
committed by PyTorch MergeBot
parent 1a568f4e5d
commit e375d21bb9
2 changed files with 23 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
aten/src/ATen/native/quantized/cpu/kernels/QuantizedOpKernels.cpp
View File

@ -2699,10 +2699,11 @@ void _fake_quantize_tensor_helper(
bool* mask_val = (bool*)(data[1] + i * strides[1]);
scalar_t* input_val = (scalar_t*)(data[2] + i * strides[2]);
const auto qval = static_cast<int64_t>(z_point + std::nearbyint(*input_val * inv_scale));
if (fake_quant_on) {
*output_val = (std::fmin(std::fmax(qval, quant_min), quant_max) - z_point) * sc;
*mask_val = ((quant_min <= qval) && (qval <= quant_max));
auto qval_f = z_point + std::nearbyint(*input_val * inv_scale);
const auto qval = static_cast<int64_t>(std::fmin(std::fmax(qval_f, quant_min), quant_max));
*output_val = (qval - z_point) * sc;
*mask_val = ((quant_min <= qval_f) && (qval_f <= quant_max));
} else {
*output_val = *input_val;
*mask_val = 1;
@ -2718,10 +2719,11 @@ void _fake_quantize_tensor_helper(
bool* mask_val = (bool*)(data[1] + i * strides[1]);
scalar_t* input_val = (scalar_t*)(data[2] + i * strides[2]);
const auto qval = static_cast<int64_t>(z_point + std::nearbyint(*input_val * inv_scale));
if (fake_quant_on) {
*output_val = (std::fmin(std::fmax(qval, quant_min), quant_max) - z_point) * sc;
*mask_val = ((quant_min <= qval) && (qval <= quant_max));
auto qval_f = z_point + std::nearbyint(*input_val * inv_scale);
const auto qval = static_cast<int64_t>(std::fmin(std::fmax(qval_f, quant_min), quant_max));
*output_val = (qval - z_point) * sc;
*mask_val = ((quant_min <= qval_f) && (qval_f <= quant_max));
} else {
*output_val = *input_val;
*mask_val = 1;

15
test/quantization/core/test_workflow_ops.py
View File

@ -1038,6 +1038,21 @@ class TestFakeQuantizeOps(TestCase):
input, scale, zero_point, axis, quant_min, quant_max
)
@skipIfTorchDynamo("Not a suitable test for TorchDynamo")
@given(dtype=st.sampled_from([torch.float, torch.float64, torch.half, torch.bfloat16]),
device=st.sampled_from(['cpu', 'cuda'] if torch.cuda.is_available() else ['cpu']))
def test_fake_quantize_per_tensor_affine_inf(self, dtype, device) -> None:
# https://github.com/pytorch/pytorch/issues/154328
input_tensor = torch.tensor([torch.inf], dtype=dtype).to(device)
scale = 0.01
zero_point = 0
quant_min = 0
quant_max = 255
result = torch.fake_quantize_per_tensor_affine(input_tensor, scale, zero_point, quant_min, quant_max)
ref_result = (min(quant_max, max(quant_min, torch.round(input_tensor / scale) + zero_point)) - zero_point) * scale
ref_result = torch.Tensor([ref_result]).to(dtype).to(device)
self.assertEqual(result, ref_result)
class TestFusedObsFakeQuant(TestCase):
@given(device=st.sampled_from(['cpu', 'cuda'] if torch.cuda.is_available() else ['cpu']),