Beta 0.4.0

Major Improvements

• Fully transitioned to torch-xla/PJRT backend for PyTorch models
  • This allows us to merch tt-torch and tt-xla
  • tt-torch will be deprecated after this release
• Added tensor-parallel support and brought up Llama3 8B
• Achieved parity for blackhole model support

📗 Getting started

Install via pip

pip install tt-torch==0.4.0 --extra-index-url https://pypi.eng.aws.tenstorrent.com/

Docker container

docker pull ghcr.io/tenstorrent/tt-torch-slim:0.4.0

More detailed instructions can be found in the Getting Started docker section

Tests:

• Workflow:
  • Run link: https://github.com/tenstorrent/tt-torch/actions/runs/17862937518
  • Status: ✅
• Docker:
  • Basic: https://github.com/tenstorrent/tt-forge/actions/runs/18112249621
  • Demo: https://github.com/tenstorrent/tt-forge/actions/runs/18112253251

🎯 Other Changes

• Uplift third_party/tt-mlir to b758a62c96ebd0ca9b17796f8279ca8fadfb5308 2025-08-15 (#1170) by @vmilosevic
• Uplift LLVM, stableHLO and shardy versions. (#1172) by @tapspatel
• llama3.2 generative tests simplified (#1171) by @gengelageTT
• Uplift to latest tt-forge-models Aug16 + support to install per-model requirements.txt (#1169) by @kmabeeTT
• Switch tt-forge-models run-tests.yml job to run single test and set HF_TOKEN etc (#1178) by @kmabeeTT
• Switch CIv2 runner label to new format (#1167) by @jmcgrathTT
• Uplift third_party/tt-mlir to 94af4a5f6f9eaeda05a82b68245bc5e2e4434f17 2025-08-17 (#1177) by @vmilosevic
• Uplift third_party/tt-mlir to 84b226f2298c285b31fb0792f2cc4685e0065876 2025-08-19 (#1179) by @vmilosevic
• Add support for TT_TORCH_IR_LOG_LEVEL for experimental backend. (#1181) by @mmanzoorTT
• Uplift third_party/tt-mlir to a3ea3703fb0db66c76a92173d4ba1af94ca27f67 2025-08-21 (#1180) by @vmilosevic
• Uplift third_party/tt-mlir to b03d526f953d654ef167faf534a20fd2b357ba76 2025-08-22 (#1187) by @vmilosevic
• Use tt-mlir from tt-xla by default, and uplift tt-xla only (#1174) by @jameszianxuTT
• Various nightly fixes/improvements (BH tests enable, re-enable PCC check for many BH tests, remove xfail, high-memory machines) (#1190) by @kmabeeTT
• Uplift third_party/tt-xla to 4ecd0b45c5bbb9dd26a6e5b7553595d58272b6de 2025-08-23 (#1191) by @vmilosevic
• Few Nightly PCC data_parallel relax + tt-forge-models tighten based on branch results (#1196) by @kmabeeTT
• use ModelLoader for sd unet n300 test (#1200) by @gengelageTT
• Add global sharding registry to allow mark_sharding() to work through torch.compile path (#1189) by @jameszianxuTT
• Add Eager Tensor Parallel tests to CI (#1202) by @hshahTT
• Uplift third_party/tt-xla to c3ea12ca0414ff574a9dc3da09ce1f2e1538ba55 2025-08-26 (#1197) by @vmilosevic
• Reschedule the nightly tests for 2 hours earlier (#1198) by @vvukomanTT
• Downgrade gliner (#1204) by @ddilbazTT
• Uplift third_party/tt-xla to a7c11621c20f6adefca77150a5220354a2c43f3d 2025-08-27 (#1206) by @vmilosevic
• Uplift third_party/tt-xla to 876c5772c2e63b018d1a494d33e54b661a400111 2025-08-29 (#1208) by @vmilosevic
• Uplift third_party/tt-xla to c3568c51fa5e93d246115b77c9c8f304e910a609 2025-08-29 (#1210) by @vmilosevic
• Add bge-m3 model tests (#1205) by @gengelageTT
• Uplift tt-xla to 9dc6e4c9de8db8d67a58f9b256c316d64442b8ea (#1212) by @jameszianxuTT
• Uplift tt-xla/ python 3.11 (#1213) by @ddilbazTT
• Add permission needed for tt-forge's perf benchmark workflow (#1214) by @vvukomanTT
• uplift torch-xla wheel 1adbe97 (python3.11 jax==0.7.1) (#1217) by @jazpurTT
• Uplift tt-xla to 69dcd00a87954f2353577597d857623eaf9d7f18 (#1219) by @jameszianxuTT
• Uplift third_party/tt-xla to 0ac1278c577f501900c8be25ff55d0be2e024ae9 2025-09-13 (#1211) by @vmilosevic
• Uplift third_party/tt-xla to a8b5af11a1d9bbb73fa062dc65ec632e876ed367 2025-09-14 (#1221) by @vmilosevic
• Fix nightly due to failures on Sept 12 (#1226) by @jameszianxuTT