only check accuracy of mutated args

[shunting314]

Stacked PRs:

• fix misleading benchmarking for fp8 gemm #1980
• warning for process group name not found conditionally #1973
• fix _clone_args when there are non-tensor input #1963
• autotune fp8 matmul reduce scatter in unit test #1962
• ->only check accuracy of mutated args #1961
• consolidate accuracy check APIs #1910

---

only check accuracy of mutated args

Previously if there are mutated args, we'll check every argument no matter if it's actually mutated. It would cause issue for SymmMemTensor signal_pad ptr. When we clone a symm memory tensor, we replace the signal_pad ptr as well. It does not make sense to compare the value of signal_pad ptr. Sometimes they pass the check since pointer addresses can be close. But it's also common that they fail the check.

[jansel]

Are the indices still correct after you flatten things? If you have f([a,b], c) is arg 2 b or c? It seems like this is changing that from c to b.

[shunting314]

arg 2 (count starts from 1) would be b here.

The self._mutated_arg_indices is constructed like this:

1. flat the args
2. filter away non-tensor args
3. track the index of mutated tensors in the tensor list

I was thinking about simplifying things by letting self._mutated_arg_indices track the index in the non-filtered flattened args, but the following comment makes me hesitate:

helion/helion/autotuner/base_search.py

Line 462 in 8c6f87c

# we should only count tensors, since they won't be bound or removed

[jansel]

Let's try to clean this up. I think we need:

1. A better name that reflects the filtered+flattened nature.
2. Some helpers methods or data structures to make working this easier. I don't want to see the same flattening and filtering code duplicated in multiple places.