[Flash Attention] Disable packed sequences with pos ids only during torch compile

[vasqu]

Draft, only as a reference as what could be done. It would allow for full graph compile when using no attention mask.

Supported compile:

• Bsz 1
  • No mask
    • Before: No full graph, recompilations
    • After: Full graph
  • Attn mask
    • Before: No full graph, recompilations
    • After: No full graph, recompilations
  • Pos ids, no mask
    • Before: No full graph, recompilations
    • After: Not supported, silent wrong computations (if packed)
  • Fa kwargs, no mask
    • Before: Full graph
    • After: Full graph
• Bsz > 1
  • No mask
    • Before: Full graph
    • After: Full graph
  • Attn mask
    • Before: Same as bsz 1
    • After: Same as bsz 1

Tl;dr: core changes are

• No attn mask: Full graph support vs recompilations and no full graph (bsz == 1)
• Position ids but no attn mask: Not supported for compile vs recompilations and no full graph

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.