fix(sm100): make arch check robust against CUTLASS DSL Arch enum changes

[lingolin128]

Summary

The current SM100 forward kernel uses a range-based comparison for the architecture
check:

assert self.arch >= Arch.sm_100 and self.arch <= Arch.sm_110f, \
    "Only SM 10.x and 11.x are supported"

This relies on the ordering of Arch enum values, which is fragile. Specifically, in
nvidia-cutlass-dsl 4.5.0, certain Arch enum entries have unexpected .value tuples,
which can cause this assertion to misbehave on valid SM 10.x architectures (e.g.
SM 10.3a / B300).

Fix

fix: #25564
Replace the range comparison with a major-version check, which is independent of how
the suffix variants (a, f, etc.) are ordered in the enum:

  arch_major = self.arch.value[0]
  assert arch_major in [10, 11], "Only SM 10.x and 11.x are supported"

This continues to gate the kernel to SM 10.x and 11.x as intended, but no longer
depends on the relative ordering of Arch.sm_10* / Arch.sm_11* variants in the
upstream CUTLASS DSL package.

Test

Verified on B300 (SM 10.3a) with nvidia-cutlass-dsl==4.5.0:

  import torch
  from flash_attn.cute import flash_attn_func

  q = torch.randn(1, 128, 4, 128, dtype=torch.bfloat16, device='cuda')
  k = torch.randn(1, 128, 4, 128, dtype=torch.bfloat16, device='cuda')
  v = torch.randn(1, 128, 4, 128, dtype=torch.bfloat16, device='cuda')
  out, lse = flash_attn_func(q, k, v, causal=False)
  # torch.Size([1, 128, 4, 128])

[lingolin128]

@Johnsonms @jayhshah Hi there, please review this PR, really appreciate it!

[janbernloehr]

Seems to be a duplicate of #2572

[lingolin128]

Seems to be a duplicate of #2572

@janbernloehr You're right, I realize this is duplicated with PR #2572. I actually encountered and verified this issue during my practical usage last week, but I didn’t sort it out and submit the PR earlier. I still hope this change can be merged. It’s totally fine if it cannot be merged eventually. Thanks a lot for your review!

[lingolin128]

Regardless, my pull request was submitted afterward. Junrong Lin is an exceptional engineer whom I look up to. Kindly merge his PR #2572 . I will keep contributing to open source and strive for continuous improvement. ^^

[Johnsonms]

Thanks @lingolin128! You’re very welcome to continue contributing to the community.