torch._standard_gamma() has accuracy gap compared to scipy and torch.cpu

[daisyden]

🐛 Describe the bug

test_distributions_xpu.py::TestDistributionsXPU::test_gamma_gpu_sample_xpu FAILED [100%]

============================================================================== FAILURES ===============================================================================
___________________________________________________________ TestDistributionsXPU.test_gamma_gpu_sample_xpu ____________________________________________________________
Traceback (most recent call last):
File "/home/gta/daisyden/rebase/n1_ci/third_party/torch-xpu-ops/test/xpu/test_distributions_xpu.py", line 119, in _test_gamma_gpu_sample
self._check_sampler_sampler(
File "/home/gta/daisyden/rebase/n1_ci/third_party/torch-xpu-ops/test/xpu/../../../../test/distributions/test_distributions.py", line 1289, in _check_sampler_sampler
self.assertLess(-threshold, bias, message)
File "/home/gta/miniforge3/envs/n2/lib/python3.10/unittest/case.py", line 1232, in assertLess
self.fail(self._formatMessage(msg, standardMsg))
File "/home/gta/miniforge3/envs/n2/lib/python3.10/unittest/case.py", line 675, in fail
raise self.failureException(msg)
AssertionError: -0.06743384440339613 not less than -0.259 : Gamma(alpha=1.0, beta=0.1).sample() is biased:
[-0.259 -0.264 -0.243 -0.059 1. 0.951 -0.279 -0.274 -0.296 -0.277]

To execute this test, run the following from the base repo dir:
python test/distributions/test_distributions.py TestDistributionsXPU.test_gamma_gpu_sample_xpu

Versions

with 2024/12/09 nightly wheel, no such issue on preci.

[daisyden]

Tried cpu _standard_gamma tensor, no such issue.

[PenghuiCheng]

fixed：#1161

[daisyden]

Seems this issue exists or not on different driver versions, investigation is WIP.