Disable Sage Attention sm90 backend due to confetti/noisy output

[arrdel]

What does this PR do?

Fixes #12783

This PR temporarily disables the Sage Attention sm90 backend which is causing confetti/noisy output on SM 9.0+ (Hopper) GPUs.

The Problem

The _SAGE_QK_INT8_PV_FP8_CUDA_SM90 backend was automatically being selected on SM 9.0+ GPUs (Hopper architecture) due to the constraint:

constraints=[_check_device_cuda_atleast_smXY(9, 0), _check_shape]

However, this backend is producing incorrect output (described as "confetti" or "noisy" output), indicating a bug in the underlying sageattention library's sm90 implementation.

The Solution

Temporarily disabled the sm90 backend by commenting out its registration:

• Users on SM 9.0+ GPUs will now fall back to the standard Sage Attention backends
• Added a comment referencing issue Sage Attention sm90 causes confetti/noisy output #12783 for future reference
• This is a temporary workaround until the upstream sageattention library fixes the sm90 implementation

Impact

• ✅ Fixes the confetti/noisy output issue on Hopper GPUs
• ✅ Users can still use other Sage Attention backends
• ✅ No breaking changes for users not on SM 9.0+ devices
• ⚠️ SM 9.0+ users won't get sm90-specific optimizations until upstream fixes the bug

Future Work

This backend can be re-enabled once the sageattention library fixes the sm90 implementation bug.

[Copilot]

Pull request overview

This PR temporarily disables the Sage Attention sm90 backend to fix confetti/noisy output issues on SM 9.0+ (Hopper) GPUs by commenting out the backend registration and implementation.

Key Changes:

• Commented out the _sage_qk_int8_pv_fp8_cuda_sm90_attention function registration to prevent automatic selection on Hopper GPUs
• Added a comment referencing issue #12783 to track re-enabling when upstream fixes the bug

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Commenting out code is not the recommended approach for temporarily disabling functionality. This creates several issues:

1. The AttentionBackendName._SAGE_QK_INT8_PV_FP8_CUDA_SM90 enum member still exists (line 195) but now has no registered implementation
2. The backend is still referenced in _check_attention_backend_requirements() (line 459) which could cause confusion
3. Commented code creates maintenance burden and can become stale

Recommended approach:
Instead of commenting out the function, keep it registered but add a runtime check that raises a clear error or logs a warning and falls back to another backend. For example:

@_AttentionBackendRegistry.register(
    AttentionBackendName._SAGE_QK_INT8_PV_FP8_CUDA_SM90,
    constraints=[_check_device_cuda_atleast_smXY(9, 0), _check_shape],
)
def _sage_qk_int8_pv_fp8_cuda_sm90_attention(
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    is_causal: bool = False,
    scale: Optional[float] = None,
    return_lse: bool = False,
    _parallel_config: Optional["ParallelConfig"] = None,
) -> torch.Tensor:
    # Temporarily disabled due to issue #12783 - sm90 backend causes confetti/noisy output
    raise NotImplementedError(
        "The sm90 backend for Sage Attention is temporarily disabled due to producing "
        "incorrect output (see issue #12783). Please use a different attention backend."
    )

Alternatively, if you want automatic fallback, modify the constraints to never match, or add a feature flag to control this behavior.

Suggested change

	# AttentionBackendName._SAGE_QK_INT8_PV_FP8_CUDA_SM90,
	# constraints=[_check_device_cuda_atleast_smXY(9, 0), _check_shape],
	# )
	# def _sage_qk_int8_pv_fp8_cuda_sm90_attention(
	# query: torch.Tensor,
	# key: torch.Tensor,
	# value: torch.Tensor,
	# is_causal: bool = False,
	# scale: Optional[float] = None,
	# return_lse: bool = False,
	# _parallel_config: Optional["ParallelConfig"] = None,
	# ) -> torch.Tensor:
	# return sageattn_qk_int8_pv_fp8_cuda_sm90(
	# q=query,
	# k=key,
	# v=value,
	# tensor_layout="NHD",
	# is_causal=is_causal,
	# sm_scale=scale,
	# return_lse=return_lse,
	# )
	@_AttentionBackendRegistry.register(
	AttentionBackendName._SAGE_QK_INT8_PV_FP8_CUDA_SM90,
	constraints=[_check_device_cuda_atleast_smXY(9, 0), _check_shape],
	)
	def _sage_qk_int8_pv_fp8_cuda_sm90_attention(
	query: torch.Tensor,
	key: torch.Tensor,
	value: torch.Tensor,
	is_causal: bool = False,
	scale: Optional[float] = None,
	return_lse: bool = False,
	_parallel_config: Optional["ParallelConfig"] = None,
	) -> torch.Tensor:
	# Temporarily disabled due to issue #12783 - sm90 backend causes confetti/noisy output
	raise NotImplementedError(
	"The sm90 backend for Sage Attention is temporarily disabled due to producing "
	"incorrect output (see issue #12783). Please use a different attention backend."
	)