[release/2.6][SWDEV-523736] Skip&Fix some testcases for archs without SDPA or Navi4x #2213
Conversation
|
pruthvistony please have a look. |
k-artem
changed the title
test/inductor/test_torchinductor.py
Outdated
| @@ -10611,6 +10611,7 @@ def fn(q, k, v): | |||
| ) | |||
|
|
|||
| @expectedFailureXPU | |||
| @unittest.skipIf(not PLATFORM_SUPPORTS_FLASH_ATTENTION, "Some archs don't support SDPA") | |||
There was a problem hiding this comment.
SDPA is support for Navi4 through AOTriton, we shouldnt skip the whole test
There was a problem hiding this comment.
It's true for Navi48, but not for Navi44, currently no support it but it's planned as experimental in AOTriton 0.10.
|
Jenkins build for 5d697ad2d52d7d659dc4cc2fb76f7a224dc7b2a3 commit finished as FAILURE |
k-artem
changed the title
|
The commit updated, pruthvistony please have a look. |
|
Jenkins build for 03c4453a242a92f069a50552f0592ba3a38f46ee commit finished as FAILURE |
* Most part of testcases work properly on Navi48(gfx1201) with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1, in this commit enable it for this arch. No support of AOTriton currently for Navci44(gfx1200), so these testcases just skipped. * test_qconv2d_int8_mixed_bf16 skipped because it was originally skipped in pytorch#112550 but later lost. * test_sac_ilp_case1 skipped as per SWDEV-509011 * test_distributed_checkpoint_state_dict_type[0-1]_cuda fixed bug with arguments.
|
@xinyazhang , |
test/dynamo/test_repros.py
Outdated
| @@ -6408,6 +6410,7 @@ def fn(x): | |||
| self.assertEqual(fn(inp), opt_fn(inp)) | |||
|
|
|||
| @requires_cuda | |||
| @skipIfRocmArch(NAVI44_ARCH) | |||
There was a problem hiding this comment.
Forgot to change to PLATFORM_SUPPORTS_FLASH_ATTENTION or is this intentional?
|
|
||
| def CDNA2OrLater(): | ||
| return evaluate_gfx_arch_within(["gfx90a", "gfx942"]) | ||
|
|
||
| def evaluate_platform_supports_flash_attention(): | ||
| if TEST_WITH_ROCM: | ||
| arch_list = ["gfx90a", "gfx942", "gfx1100"] | ||
| arch_list = ["gfx90a", "gfx942", "gfx1100", "gfx1201"] |
There was a problem hiding this comment.
Should take TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL into considration. Check the upstream's pattern as a reference.
There was a problem hiding this comment.
It takes in line 48, but this patch is implemented another logic because target version is ROCm 6.4.2/pytorch/2.6, I enabled TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL in evaluate_gfx_arch_within
+ if result and gcn_arch_name == "gfx1201":
+ os.environ['TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL'] = '1'
There was a problem hiding this comment.
Don't do this.
UT should follow the guide of TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL rather than changing its value
|
@xinyazhang updated according comments, please review |
|
Jenkins build for adbbad0e7b7c4cc21c1841fa68e7fe93a789ddbd commit finished as FAILURE |
|
@pruthvistony please review. |
|
Jenkins build for 53a544d3b4f846ba5bb5934e1734ad3f01167ee7 commit finished as FAILURE |
[release/2.6][SWDEV-523736] Skip some testcases for archs without SDPA or Navi4x
[SWDEV-523736] Fix some unittests for Navi4x
TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1, in this commit enable it
for this arch. No support of AOTriton currently for Navci44(gfx1200),
so these testcases just skipped.
[Inductor] [Quant] Enable QConv2d Unary int8-mixed-bf16 Lowering pytorch/pytorch#112550 but later lost.
with arguments.
Fixes #SWDEV-523736