Skip to content

Add test config for coqui/XTTS-v2 model #5376

Draft
ashokkumarkannan1 wants to merge 1 commit into
mainfrom
akannan/xtts_v2-bringup
Draft

Add test config for coqui/XTTS-v2 model #5376
ashokkumarkannan1 wants to merge 1 commit into
mainfrom
akannan/xtts_v2-bringup

Conversation

@ashokkumarkannan1

@ashokkumarkannan1 ashokkumarkannan1 commented Jun 25, 2026

Copy link
Copy Markdown
Contributor

Ticket

Problem description

  • XTTS-v2 needs a runner test_config entry and tt-xla-side coverage. Bringing it up surfaced a model-independent functionalization bug: the HiFiGAN decoder's no-op .squeeze(1) lowers to prims::view_of, which the tt functionalization pass cannot handle, so the model fails FE compilation.

What's changed

Results

Check Result
CPU equivalence test (custom forward vs native inference) PCC 1.0
Runner test on current stack (no workaround) xfail — FE compilation prims::view_of (#5375)
With #5388 fix compiles & runs on TT, PCC 0.82 vs CPU

Checklist

  • New/Existing tests provide coverage for changes

Logs

@codecov-commenter

Copy link
Copy Markdown

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 33.85%. Comparing base (f4c45f5) to head (3603d3c).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files
@@            Coverage Diff             @@
##             main    #5376      +/-   ##
==========================================
+ Coverage   33.81%   33.85%   +0.04%     
==========================================
  Files          37       37              
  Lines        4992     4992              
==========================================
+ Hits         1688     1690       +2     
+ Misses       3304     3302       -2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

@ashokkumarkannan1 ashokkumarkannan1 force-pushed the akannan/xtts_v2-bringup branch from 3603d3c to abecb37 Compare June 26, 2026 06:38
Adds the XTTS-v2 (coqui/XTTS-v2) runner test_config entry and a CPU
logic-equivalence test, and bumps tt_forge_models to the XTTS-v2 loader.

The model compiles and runs end-to-end on TT (PCC 0.82 vs CPU) once the
HiFiGAN decoder's no-op squeezes are rewritten as reshape. Without that
workaround it fails FE compilation on prims::view_of, a tt functionalization
limitation tracked in #5375. This branch intentionally omits the decoder
workaround so the failure stays visible; config is marked KNOWN_FAILURE_XFAIL
(FAILED_FE_COMPILATION) linking #5375.

The CPU equivalence test (tests/torch/models/xtts_v2) confirms the custom
forward matches native Xtts.inference at PCC 1.0, isolating the issue to the
compiler.
@ashokkumarkannan1 ashokkumarkannan1 force-pushed the akannan/xtts_v2-bringup branch from abecb37 to 128ff82 Compare June 26, 2026 06:47
@ashokkumarkannan1 ashokkumarkannan1 changed the title Bringup XTTS-v2 model + track prims::view_of functionalization bug Add test config for XTTS-v2 model Jun 26, 2026
@ashokkumarkannan1 ashokkumarkannan1 changed the title Add test config for XTTS-v2 model Add test config for coqui/XTTS-v2 model Jun 26, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants