v0.78.0-dev20260825

Note

If you are installing from a release, please refer to the README, INSTALLATION instructions, and any other documentation packaged with the release, not on the main branch. There may be differences between the latest main and the previous release.

The changelog will now follow, showing the changes from last release.

This release was generated by the CI workflow https://github.com/tenstorrent/tt-metal/actions/runs/32793015476

LLK (low-level kernels)

• Add Int8 support for TTNN quantize/dequantize/requantize PR 49907
• Add MOP-free Matmul LLK and Compute Kernel Support for Quasar PR 53518
• Refactor LLK test harness to support out-of-tree kernel testing PR 54182
• Fix incorrect comment to match code PR 54210

Metalium (tt-metal core)

• ASan: add teardown script and fix out-of-bounds check in buffer implementation PR 53819
• Increase compute MOP timeout in Quasar TRISC firmware PR 53617
• Fix SDPA forward mask=None hang, Wormhole scale rounding, and backward validation gaps PR 53217
• Add Int8 support for TTNN quantize/dequantize/requantize PR 49907
• Automated UMD Bump 21.08.2026 PR 53955
• Revert "Fix active-ERISC kernel link: define __global_pointer$" PR 54148
• Add MOP-free Matmul LLK and Compute Kernel Support for Quasar PR 53518
• Update SFPI to version 7.71.0 PR 54153
• jit_build: opt-in Zve32f (RVV) plumbing for TRISC2 kernel compiles (enable_trisc2_rvv) PR 54121
• Add all-solutions enumeration and sweep runner to scale-out rank binding tools PR 49860
• Replace single-device Buffer with MeshBuffer in tests PR 53995

TT-NN

• Fix invalid 1D DRAM shard grid in CopyToDeviceFiltered_EmptyFilter_NoChange test PR 53876
• Fix off-by-one error in DRAM shard grid size for Llama MLP width-sharded-to-interleaved test PR 54091
• Fix incorrect scalar_face value in experimental pool op on Quasar PR 54194
• Test Only: remove all eltwise tests from tt_dnn sweep_tests PR 53894
• Add merge-gate tests for normalization ops and fix kernel packaging PR 53677
• Test Only: reduction test cleanup: repeated tests, stale skips, masked negative tests PR 53573
• Add Int8 support for TTNN quantize/dequantize/requantize PR 49907
• Add ordered affine-transform reduction PR 52786
• Add accurate fp32 SFPU reduction path for ttnn.min and ttnn.sum using fast_and_approximate_mode PR 53636
• Fix performance regression in width-reduce (dim=-1) for height-sharded tensors PR 51158
• Port moreh_linear_backward bias-gradient op to Metal 2.0 program factory PR 53664
• Test Only: remove a stale trailing comment in nightly reduction tests PR 53864
• Port data_movement/copy DefaultRowMajor/DefaultTilized factories (2 of 3) PR 53570
• Port fill_pad data movement operation to Metal 2.0 PR 52545
• Migrate 6 of 7 untilize program factories to the Metal 2.0 program spec API PR 53586
• Remove inert RuntimeTensorShape accessor args from transpose op program factories PR 53580
• Migrate row-major reshape kernel to CircularBuffer API PR 54188
• jit_build: opt-in Zve32f (RVV) plumbing for TRISC2 kernel compiles (enable_trisc2_rvv) PR 54121
• ttnn.topk: use model-based multi-core split selection and tile-row-aware eligibility for up to 10x speedup on narrow-row shapes PR 53793

tt-train

• Fix cross-entropy target page overrun, bounds check, cluster_axis cache hash, and backward padding gradients PR 53219
• Fix moe_group NOC staging race condition, Blackhole page stride, grouped tail initialization, and async-write-zeros barrier mismatches PR 53220
• Fix SDPA forward mask=None hang, Wormhole scale rounding, and backward validation gaps PR 53217
• Migrate tt-train tests from galaxy-perf workflow to end-to-end models T3 workflow PR 52553

Models

• Restore batch-1 Galaxy inference on main PR 53513
• Correct DeepSeek post-FF2 reduce-scatter reference PR 53825
• Split the TT-Transformers module tests between tier 1 Models unit and sweep CI PR 53911
• Expose performance metrics for Whisper PR 53717
• Support Linear topology for SP prefill (unblocks non-torus galaxies) PR 54164
• Add all-solutions enumeration and sweep runner to scale-out rank binding tools PR 49860

Infrastructure & CI

• Blaze prefill: add missing TorusXY fabric profile to three Kimi-K3 CI matrix rows PR 54174
• Fix job log downloads broken by gh 2.97 escape-sequence guard PR 54185
• Don't report infra faults as test failures or pass-rate drops PR 54027
• Add Matt Pocock Skills automated PR reviewer workflow PR 54129
• Split the TT-Transformers module tests between tier 1 Models unit and sweep CI PR 53911
• ci: give the vLLM structured output benchmark its own timeout PR 53628
• Migrate tt-train tests from galaxy-perf workflow to end-to-end models T3 workflow PR 52553
• chore: Update ccache to 4.14 PR 54200
• Runtime: Add randomized dispatch sanity test to the package-and-release pipeline PR 53696
• ci: skip the AI run summary on cancelled runs PR 54230
• Increase wh_n150_civ2 data-movement timeout to 50 minutes PR 54003
• Add setup runner failure signature PR 53377

Tooling

• Health check: upload pre-reboot failure as discarded so self-heal reboots are visible without skewing stats PR 53916
• Add all-solutions enumeration and sweep runner to scale-out rank binding tools PR 49860

Other

• Don't report infra faults as test failures or pass-rate drops PR 54027