Update CuTe namespace and enhance dependencies

[LoserCheems]

Summary

• This update addresses issues related to the namespace of the CuTe library and enhances the Python version requirement and optional dependencies.

Root Cause

• The previous implementation used an outdated namespace for CuTe, which caused import errors and compatibility issues with newer Python versions.

Changes

• Refactored the CuTe namespace to flash_sparse_attn.ops.cute.
• Updated the Python version requirement to 3.10.
• Enhanced the sync_cute_subtree script for better error handling and temporary worktree support.

Reproduction

• Ensure the environment meets the new Python version requirement and attempt to import the updated CuTe modules.

Tests

• Validated changes by running existing tests and confirming that imports function correctly under the new namespace.

Compatibility

• Migration to the new namespace may require updates to existing code that relies on the old flash_attn.cute imports.

Checklist

• Linked issue provided
• Adds or updates tests
• Updates docs if needed
• No perf regressions

[Copilot]

Pull request overview

This PR updates the vendored CuTe/FlashAttention-4 integration to live under the flash_sparse_attn.ops.cute namespace, updates packaging metadata (Python >= 3.10 and new optional deps), and improves the subtree sync scripts by adding namespace-rewrite and a temporary-worktree flow for dirty repos.

Changes:

• Refactor CuTe Python sources to import via flash_sparse_attn.ops.cute instead of flash_attn.cute.
• Add a rewrite_cute_namespace.py script and integrate it into the subtree sync scripts (bash + PowerShell), including a temporary worktree mode.
• Update pyproject.toml to require Python >= 3.10 and add a cute optional-dependency extra.

Reviewed changes

Copilot reviewed 27 out of 27 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
scripts/sync_cute_subtree.sh	Adds temporary-worktree sync path, CuTe import rewrite step, and improved reporting.
scripts/sync_cute_subtree.ps1	PowerShell equivalent of the enhanced sync workflow with rewrite + temporary worktree support.
scripts/rewrite_cute_namespace.py	New helper to rewrite vendored CuTe Python imports to the local namespace.
pyproject.toml	Bumps minimum Python version and adds cute + enhanced dev optional dependencies.
flash_sparse_attn/ops/cute/init.py	Updates distribution version lookup and patches compile helper import path.
flash_sparse_attn/ops/cute/tile_scheduler.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/softmax.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/paged_kv.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/pack_gqa.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/mask.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/interface.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/flash_fwd.py	Rewrites internal imports and updates SM90 lazy import to new namespace.
flash_sparse_attn/ops/cute/flash_fwd_sm90.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/flash_fwd_sm100.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/flash_fwd_sm120.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/flash_fwd_combine.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/flash_bwd.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/flash_bwd_sm90.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/flash_bwd_sm100.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/flash_bwd_sm120.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/flash_bwd_preprocess.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/flash_bwd_postprocess.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/compute_block_sparsity.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/block_sparsity.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/block_sparse_utils.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/block_info.py	Rewrites internal imports to the new CuTe namespace.
flash_sparse_attn/ops/cute/blackwell_helpers.py	Rewrites internal imports to the new CuTe namespace.

---

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

[Copilot]

The reported synced commit range can be inaccurate when syncing via a temporary worktree. In particular, if the rewrite commit is skipped (because there are existing local changes under $PREFIX) then SYNC_END_HEAD is still taken from the temporary worktree’s HEAD, even though that commit wasn’t cherry-picked back. Consider recomputing SYNC_START_HEAD/SYNC_END_HEAD from the current worktree after cherry-picking (or have invoke_temporary_worktree_sync update these vars based on what was actually applied).

[Copilot]

REWRITE_COMMIT_MESSAGE is defined and used to identify the rewrite commit later, but the commit here hard-codes the message string. If the message ever changes, the filter logic can silently break. Use the shared REWRITE_COMMIT_MESSAGE value when creating the commit to keep these paths consistent.

Suggested change

	invoke_git -C "$work_repo_root" commit -m "Rewrite vendored CuTe namespace to flash_sparse_attn.ops.cute"
	invoke_git -C "$work_repo_root" commit -m "$REWRITE_COMMIT_MESSAGE"

[Copilot]

The new --no-temporary-worktree option line in the usage output is mis-indented compared to the other options, which makes the help text harder to read. Align the indentation with the rest of the option list.

Suggested change

	--no-temporary-worktree Fail instead of using a temporary worktree when the current tree is dirty.
	--no-temporary-worktree Fail instead of using a temporary worktree when the current tree is dirty.

[Copilot]

The printed synced commit range can be misleading when using the temporary-worktree path. Invoke-CoreSync returns StartHead/EndHead from the temporary worktree, but the current worktree may not end up at EndHead if you skip cherry-picking the rewrite commit (and instead apply the rewrite without committing). Consider recomputing EndHead from the current repo after cherry-picking (or have Invoke-TemporaryWorktreeSync return the actual applied range).

Suggested change

	Write-Host "Done."
	Write-Host "Upstream source: $UpstreamRepo"
	Write-Host "Upstream cache used for subtree split: $upstreamRepoForSplit"
	Write-Host "Synced commit range: $($syncResult.StartHead) -> $($syncResult.EndHead)"
	# Recompute the effective end commit from the current repository after syncing,
	# since the temporary worktree's EndHead may not match the final HEAD.
	$printedEndHead = Get-GitOutput -Repo $repoRoot -Arguments @("rev-parse", "HEAD")
	Write-Host "Done."
	Write-Host "Upstream source: $UpstreamRepo"
	Write-Host "Upstream cache used for subtree split: $upstreamRepoForSplit"
	Write-Host "Synced commit range: $($syncResult.StartHead) -> $printedEndHead"

[Copilot]

$RewriteCommitMessage is used to detect the rewrite commit later, but the commit message here is hard-coded. If either string changes, the skip logic in Invoke-TemporaryWorktreeSync can stop matching. Use $RewriteCommitMessage when creating the commit to keep the behavior consistent.

Suggested change

	Invoke-Git -Repo $WorkRepoRoot -Arguments @("commit", "-m", "Rewrite vendored CuTe namespace to flash_sparse_attn.ops.cute")
	Invoke-Git -Repo $WorkRepoRoot -Arguments @("commit", "-m", $RewriteCommitMessage)