Add support for upstream split reference in sync scripts

[LoserCheems]

Summary

• Introduces support for upstream split references in synchronization scripts and enhances merge handling.

Root Cause

• The previous implementation lacked proper handling for upstream split references during synchronization.

Changes

• Updated synchronization scripts to accommodate upstream split references and improved merge logic.

Reproduction

• Steps to reproduce the issue include attempting to synchronize with upstream split references without the new handling.

Tests

• Validated changes through existing test cases and added new tests for the updated synchronization functionality.

Compatibility

• No backward compatibility issues identified; changes integrate seamlessly with existing functionality.

Checklist

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

[Copilot]

Pull request overview

Adds support for syncing the vendored CuTe subtree from the upstream repo’s default branch/ref (including “split ref” handling), improves subtree merge/cherry-pick behavior, and updates the CuTe-based FlashAttention-4 implementation to support additional SM100/2-CTA configurations and CUDA 13 packaging.

Changes:

• Update sync scripts (bash + PowerShell) to resolve upstream default ref, disable merge auto-edit, and handle merge commits when replaying sync commits.
• Extend CuTe tile scheduling / SM100 forward kernel tuning (cluster indexing support, register/ex2 emulation tuning, and enable 2-CTA path for head_dim padded 192).
• Add CUDA 13 optional dependency + uv index/source wiring and document the cu13 install extra.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
scripts/sync_cute_subtree.sh	Resolve upstream default ref for remote caches; avoid merge editor; verify sync includes upstream split; handle merge commits during cherry-pick.
scripts/sync_cute_subtree.ps1	PowerShell parity for upstream default ref resolution, merge auto-edit disabling, and merge-aware cherry-picking.
flash_sparse_attn/ops/cute/tile_scheduler.py	Add use_cluster_idx to support cluster-based block coordinate selection and adjusted grid sizing.
flash_sparse_attn/ops/cute/interface.py	Broaden 2-CTA eligibility to include head_dim padded 192.
flash_sparse_attn/ops/cute/flash_fwd_sm100.py	Add tuning table + cluster-index scheduling hooks; adjust Q copy thread layout logic.
flash_sparse_attn/ops/cute/pyproject.toml	Add cu13 extra and uv index/source configuration for CUDA 13 torch wheels.
flash_sparse_attn/ops/cute/README.md	Document installing with flash-attn-4[cu13].

---

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

[Copilot]

copy_utils.tiled_copy_2d is defined as tiled_copy_2d(dtype, major_mode_size, num_threads, is_async=False) in flash_sparse_attn/ops/cute/copy_utils.py, but this call passes an extra positional argument (async_copy_elems) and uses threads_per_row as the second argument. As written, this will raise a TypeError when use_tma_Q is false (and even if it didn’t, it would be using the helper with the wrong semantics). Update the call to match the helper’s signature (i.e., pass the actual major-mode size in elements and use the keyword is_async=), or update tiled_copy_2d/call sites consistently if the intended API includes num_copy_elems.