enable torch.compile tests on mx on Blackwell

[vkuzo]

These tests were disabled in #1624 due to lack of triton support at the time.

I tried it again today with a fresh PyTorch build, and got the following error: https://gist.github.com/vkuzo/7f345ac91f3d4a520a866cf2c727ec12

E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp91 = tmp89 | tmp90                                                                                                                        
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp92 = tl.full([1], 0, tl.int32)                                                                                                            
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp93 = tl.where(tmp81, tmp92, tmp91)                                                                                                        
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp94 = tmp65 > tmp80                                                                                                                        
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp95 = tmp83 == tmp80
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp96 = tmp94 & tmp95                                                                                                                        
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp97 = tl.where(tmp96, tmp79, tmp93)
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp98 = tmp64.to(tl.int32)                                                                                                                   
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp99 = tmp98 << tmp58      
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp100 = tmp97 | tmp99      
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp101 = tmp100.to(tl.float32, bitcast=True)                                                                                                 
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tmp103 = tmp101 * tmp102
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     tl.store(out_ptr4 + (x2), tmp103, xmask)                                                                                                     
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] 
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] metadata: {'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'out_ptr4': '*fp32', 'xnumel': 'i32'}, 'device': 0, 'constants': {'XBLOCK': 64},
 'configs': [AttrsDescriptor.from_dict({'arg_properties': {'tt.divisibility': (0, 1, 2, 3), 'tt.equal_to': ()}, 'cls': 'AttrsDescriptor'})], 'device_type': 'cuda', 'num_warps': 1, 'num_stages': 1, 'debug': True, 'cc': 100}   
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] Traceback (most recent call last):
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]   File "/home/vasiliy/.conda/envs/pytorch/lib/python3.11/site-packages/triton/backends/nvidia/compiler.py", line 356, in make_cubin              
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     subprocess.run(ptxas_cmd, check=True, close_fds=False, stderr=flog)                                       
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]   File "/home/vasiliy/.conda/envs/pytorch/lib/python3.11/subprocess.py", line 569, in run                     
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     raise CalledProcessError(retcode, process.args,                                                        
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] subprocess.CalledProcessError: Command '['/home/vasiliy/.conda/envs/pytorch/lib/python3.11/site-packages/triton/backends/nvidia/bin/ptxas', '-lin
einfo', '-v', '--gpu-name=sm_100', '/tmp/tmpodrv3kxq.ptx', '-o', '/tmp/tmpodrv3kxq.ptx.o']' returned non-zero exit status 255.                                                                                                   
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]                              
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] During handling of the above exception, another exception occurred:
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]                                                                                                                                                  
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] Traceback (most recent call last):
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]   File "/data/users/vasiliy/pytorch/torch/_inductor/runtime/triton_heuristics.py", line 530, in _precompile_config
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     binary = triton.compile(*compile_args, **compile_kwargs)                   
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^                                                                                     
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]   File "/home/vasiliy/.conda/envs/pytorch/lib/python3.11/site-packages/triton/compiler/compiler.py", line 279, in compile
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     next_module = compile_ir(module, metadata)
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]   File "/home/vasiliy/.conda/envs/pytorch/lib/python3.11/site-packages/triton/backends/nvidia/compiler.py", line 389, in <lambda>
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     stages["cubin"] = lambda src, metadata: self.make_cubin(src, metadata, options, self.capability)
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]                                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^                                             
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]   File "/home/vasiliy/.conda/envs/pytorch/lib/python3.11/site-packages/triton/backends/nvidia/compiler.py", line 374, in make_cubin              
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532]     raise RuntimeError(f'{error}\n'                                                                                                              
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] RuntimeError: Internal Triton PTX codegen error
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] `ptxas` stderr:
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] ptxas fatal   : Value 'sm_100' is not defined for option 'gpu-name'
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] 
E0219 13:51:06.311000 2616135 torch/_inductor/runtime/triton_heuristics.py:532] Repro command: /home/vasiliy/.conda/envs/pytorch/lib/python3.11/site-packages/triton/backends/nvidia/bin/ptxas -lineinfo -v --gpu-name=sm_100 /tm
p/tmpodrv3kxq.ptx -o /tmp/tmpodrv3kxq.ptx.o