run on windows

[ziyanxzy]

compiled_fn = self.call_user_compiler(gm)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^

File "C:\Users\SAS\miniforge3\envs\inf\Lib\site-packages\torch_dynamo\output_graph.py", line 1416, in call_user_compiler
return self._call_user_compiler(gm)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\SAS\miniforge3\envs\inf\Lib\site-packages\torch_dynamo\output_graph.py", line 1465, in _call_user_compiler
raise BackendCompilerFailed(self.compiler_fn, e) from e
torch._dynamo.exc.BackendCompilerFailed: backend='inductor' raised:
ImportError: DLL load failed while importing cuda_utils: The specified module could not be found.

You can suppress this exception and fall back to eager by setting:
import torch._dynamo
torch._dynamo.config.suppress_errors = True

[JeyesHan]

@ziyanxzy Thanks for your attention to Infinity. It seems that the error comes from Flex_Attn. Maybe you could run the code with --use_flex_attn=False
Besides, note that this codebase runs on Linux machines. We don't know whether it could be run on Windows. We are glad that you could share your experiences here, regardless of whether they were successful or unsuccessful.

[ziyanxzy]

Thanks han. i have use_flex_attn=0,but still have triton issue. can i do not use triton kernel in infinity?

[ziyanxzy]

and if i donot use flash_attn, it still has flash_attn_varlen_kvpacked_func function in crossattention, can i replace it?