[Test] Mark kokoro/Kokoro-82M single-device inference KNOWN_FAILURE_X…

[saiarthiraguram]

Ticket

Github Issue

Problem description

Bring up hexgrad/Kokoro-82M (StyleTTS2 TTS / iSTFTNet vocoder) on the tt-xla
PyTorch runner. The model now has a tt-forge-models loader (companion PR:
tenstorrent/tt-forge-models#); this PR wires it into the tt-xla single-device
inference suite and records its status.

The model compiles and runs end-to-end on n150 with trained weights and produces
a finite, sane-magnitude waveform, but waveform PCC tops out at ~0.14 (need
0.99): the iSTFTNet vocoder (sine-phase cumsum + STFT/iSTFT over a long 1-D
waveform) is sensitive to bf16 tensor storage. Decoder output PCC is already
0.948 and the vocoder collapses it to 0.138; fp32_dest_acc_en+hifi4 is
bit-identical (storage-bound, not accumulation-bound). Root cause and the full
device-vs-CPU bisect are in #5332.

What's changed

• third_party/tt_forge_models — submodule pointer bumped to include the
  Kokoro loader (companion tt-forge-models PR). (Apply after that PR merges.)
• tests/runner/test_config/torch/test_config_inference_single_device.yaml —
  new kokoro/pytorch-hexgrad/Kokoro-82M-single_device-inference entry,
  KNOWN_FAILURE_XFAIL, reason documenting the bf16 waveform-PCC wall and
  linking Kokoro-82M: whole-model PCC collapses (-0.0019, best 0.24) — bf16 tile storage quantizes large iSTFTNet sine-phase accumulation, causing catastrophic cancellation #5332.
• python_package/tt_torch/utils.py — Dynamo guard-repr patch fix:
  self.get(guard) → self.get(guard.name) (a prerequisite that unblocked
  compilation during bringup). (Drop this hunk if it has already landed on
  main from another change.)
• tests/torch/ops/kokoro/ — two self-contained single-device op sanities
  reproducing the underlying device numerical-robustness gap (bf16
  InstanceNorm1d variance catastrophic-cancellation on a near-constant input
  → rsqrt → inf/FLT_MAX/nan): test_adain_sanity.py (minimal bare op) and
  test_adain_chain_sanity.py (Conv1d → AdaIN1d, the Kokoro structure). Both
  xfail(strict=False), #5332. *(Currently staged on branch
  sai_arthi_raguram/kokoro-instancenorm-bf16-repro)

Checklist

• New/Existing tests provide coverage for changes — model node runs E2E on
  n150 (xfail on PCC); op sanities reproduce the root-cause gap in isolation.

Logs

debug_report.md
iter_n150_13_fp32acc_realw_run.log