docs: add flash-attn import troubleshooting

[zcxGGmu]

Summary

• Add troubleshooting for NameError: name 'flash_attn_func' is not defined in the install guide.
• Show a direct import check for the active environment and clarify that the flash-attn wheel must match Python / PyTorch / CUDA / C++ ABI.
• Point balance_serve users to the patched FlashInfer note instead of replacing that package.

Refs #944

Test Plan

• python documentation guard check for the new flash-attn troubleshooting text and link target
• git diff --check HEAD~1..HEAD

[greptile-apps]

Greptile Summary

This PR adds a flash-attn import troubleshooting block to the installation guide, addressing issue #944. The addition covers the NameError: name 'flash_attn_func' is not defined failure, provides a one-liner import check, and points balance_serve users to the correct FlashInfer note.

• Adds a diagnostic python -c command to verify the active environment has flash-attn installed correctly.
• Clarifies wheel compatibility requirements (Python, CUDA, PyTorch, C++ ABI) and links to the balance-serve installation guide for users of that backend.

Confidence Score: 5/5

Documentation-only change with no runtime impact; the link target and diagnostic command are both correct.

The change adds a self-contained troubleshooting block to a Markdown file. The ./balance-serve.md#installation-guide anchor resolves to the correct heading in that file, and the python -c import check is syntactically valid. No code is modified.

Files Needing Attention: No files require special attention.

Important Files Changed

Filename	Overview
doc/en/install.md	Adds 12-line troubleshooting block for flash-attn import errors; link target and diagnostic command are both valid.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[flash_attn_func NameError at runtime] --> B{Run import check}
    B --> |"python -c 'from flash_attn import flash_attn_func'"| C{Import succeeds?}
    C --> |Yes| D[Reinstall KTransformers / check sys.path]
    C --> |No| E{Using balance_serve backend?}
    E --> |Yes| F[Follow balance-serve FlashInfer note\nDo NOT replace patched FlashInfer]
    E --> |No| G[Download matching flash-attn wheel\nPython + CUDA + PyTorch + C++ ABI]
    G --> H[Reinstall wheel & retry]

Loading

_{Reviews (2): Last reviewed commit: "[docs]: refine flash-attn troubleshootin..." | Re-trigger Greptile}

[zcxGGmu]

Thanks for the review — I pushed 39aef7d0130c addressing the feedback.

Changes made:

• Removed duplicate PyTorch wording from the flash-attn troubleshooting guidance.

Validation:

• git diff --check — passed.
• Content guard — passed: flash-attn troubleshooting now focuses on wheel compatibility.
• Diff scan — passed: no secret/debug-marker patterns found.