consolidate accuracy check APIs

[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

---

There are 2 APIs to check accuracy

1. the one used in BaseSearch would handle str inputs (like PROCESS_GROUP_NAME)
2. the one added for custom check function can allow small percent of mismatch. Will feng added it. I found it useful for fp8 kernels.

Consolidate these 2 APIs so we only need maintain one.

[jansel]

I do wonder if we even need the alternate strategy. Where is it used?

[jansel]

How is this different from atol/rtol?

[shunting314]

I didn't use these 2 arguments and just copied over from the existing code. @yf225 is there any usage of them?

[jansel]

Isn't it semantically different to do this in a chunked way? You are measuring percentage different on a chunk not on the entire tensor.

[shunting314]

Chunk is handled inside _chunked_assert_close. This part of code operates on the entire tensor though

[shunting314]

I do wonder if we even need the alternate strategy. Where is it used?

The alternative strategy allowing small percent of mismatch was added here #1733 by @yf225 . I used it in a fp8 kernel. There is a small percent of mismatch. If we just raise tolerance, we need a quite large tolerance.

[jansel]

@yf225 can you comment on if we could remove this?