[PyTorch] Fix fake kernel metadata mismatch for empty-input MoE permute ops

[pggPL]

Description

For empty input, moe_permute (both mask and index map) returns row_id_map created with torch.tensor([]), i.e. float32 with shape (0,), while the registered fake (meta) kernels declare int32 with shape (0, num_experts * 2 + 1) (mask map) / (num_tokens * topK,) (index map).

PyTorch validates fallback output dtype metadata against the fake kernel at runtime since pytorch#183731 (nightlies after 2026-07-22, torch 2.14), so torch.compile'd empty-input permute now fails with:

RuntimeError: expected dtype torch.int32 but got torch.float32
Error in op: torch.ops.te_moe.permute_mask_map_fwd.default

Reproduced by test_permutation_mask_map_empty_input[True-*] and test_permutation_mask_map_alongside_probs_empty_input[True-*] on torch 2.14 nightly. On torch <= 2.13 the mismatch passes silently (no metadata check).

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• permute_mask_map_fwd empty-input fast path returns row_id_map as int32 (0, num_experts * 2 + 1) and permuted_probs with probs.dtype, matching the fake kernel.
• permute_index_map empty-input fast path returns row_id_map as int32 (0,) instead of float32.
• _moe_permute_index_map_fake output shape now matches the empty-input fast path (0 rows, num_out_tokens ignored), mirroring the existing mask-map fake handling.
• _moe_permute_mask_map_forward_fake uses probs.dtype for permuted_probs also when out_rows == 0, keeping eager and fake consistent.
• Added test_permutation_index_map_empty_input (eager + compile) — the index-map path had the same bug but no empty-input test coverage.

Verified on torch 2.14.0.dev20260807 nightly: all empty_input tests pass (24/24; 9 compile variants fail without the fix), full tests/pytorch/test_permutation.py green (248 passed, 154 skipped).

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

🤖 Generated with Claude Code