Skip to content

v0.78.0-dev20260827

Pre-release
Pre-release

Choose a tag to compare

View all tags
@github-actions github-actions released this 27 Aug 03:21
· 115 commits to main since this release
Immutable release. Only release title and notes can be modified.
v0.78.0-dev20260827
20d1eee
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

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/33031051524

LLK (low-level kernels)

  • Bind test harness silu/clamp/log/hardtanh/abs to production metal kernels PR 53918
  • Add tt-llk unit tests for blaze->experimental promoted LLKs PR 53361
  • Harden LLK PR review workflow reliability PR 54297
  • Feature | Separate test outputs per run PR 53660

Metalium (tt-metal core)

  • Add checks to make sure there are not any attempts to do partial entry size noc transactions for dfbs PR 54345
  • Fix SFPU binary broadcast runtime dest API compile failure PR 54427
  • Topology Mapper hard host-cap full-packing(fixes #50253 — SC16 ring on SC24) PR 54013
  • Route NoC address construction through an arch-selected backend PR 54199
  • Extend documentation: equal block division across dest banks for untilize PR 54307
  • Automated UMD Bump 26.08.2026 PR 54414
  • SFPI 7.72.0 873 PR 54476
  • Bump ttsim version to v1.10.3 PR 54496
  • Thread MetalContext& into DeviceCommand and alike PR 53713

TT-NN

  • GroupNorm: reconfigure the unpacker data format on the fp32 paths PR 52712
  • fix(ttnn): correct addcmul composite-fallback multiply association PR 53478
  • Eltwise Unary: Update output dtypes for preallocated output tensor PR 54364
  • Isolate Tracy profiling zone stacks per thread PR 53824
  • #52683: Fix slice RM chunked reader CB ring-wrap straddle for odd num_chunks PR 52389
  • #54138: Bug fix - binary_ng program-cache key omissions and DRAM-sharded routing PR 54233
  • Fix sharded_to_interleaved corrupting row-major output when the shard is wider than the row PR 54226
  • decross the group_norm DRAM dtype grid, move 512x512 splits to nightly PR 53549
  • test: shrink the two slow core ttnn tests so ttsim can run them again PR 53956
  • Prune tt-metal-l2-nightly PR 53691
  • tests: reduce KDA test CI time PR 54265
  • Skip situ_glu sub-device test under slow dispatch PR 54416
  • #52694: Skip empty input-tile inner scan in gather multi-core reader PR 53112
  • prefill - Enable SiTU for shared expert and FFN for Kimi K3 PR 54149
  • Route NoC address construction through an arch-selected backend PR 54199
  • #52682: Fix roll DRAM_RM 3-shard OOB and reshard diff-width dangling ref PR 52274
  • Add affine exclusive scan PR 52788
  • Metal 2.0 port: data_movement/reshape_view PR 54202
  • Metal 2.0 port: data_movement/pad PR 54214
  • Add fused distributed GroupNorm PR 52500
  • port(metal_2.0): reduction/argmax ArgMaxDeviceOperation → ProgramSpecFactoryConcept (6 gtests / 66 / 937 passed) PR 54212

tt-train

  • Fix layernorm_bw dx double-subtract and non-L1 nested DST acquire PR 53215
  • Use shared TTI exp in cross-entropy fw/bw kernels PR 54310
  • tt-train/grpo: example cleanup, move logging & wandb init into ttml PR 53433

Models

  • Dump the DSA indexer key cache for src-KV validation in GLM5.2 PR 53428
  • Fix GPT-OSS Galaxy experts validation PR 53615
  • #51100: Update DeepSeek Galaxy decode op gap target PR 53553
  • Avoid eager DeepSeek weights in random component tests PR 53841
  • Combine op tests for externally owned models PR 54273
  • Prune blackhole-e2e-tests via uncollect_if predicates PR 54314
  • #54426: add graph tracing based llama quasar ops pytests PR 54428
  • prefill - Enable SiTU for shared expert and FFN for Kimi K3 PR 54149
  • Pipeline-parallel KV-cache migration for prefill PR 54313
  • Add fused distributed GroupNorm PR 52500

Infrastructure & CI

  • Fix: teach /test to narrow suite toggles and report the inputs it supplied PR 54373
  • Fix: restore the /test badge column and pin the report heading PR 54380
  • test-command: deterministically enforce the PR head ref on every dispatched pipeline PR 54486
  • Topology Mapper hard host-cap full-packing(fixes #50253 — SC16 ring on SC24) PR 54013
  • test: shrink the two slow core ttnn tests so ttsim can run them again PR 53956
  • tests: reduce KDA test CI time PR 54265
  • Prune blackhole-e2e-tests via uncollect_if predicates PR 54314
  • #54332: add more tests to quasar_local_tests.yaml PR 54421
  • Track (Runtime) pipelines in CICD data ingestion PR 54328
  • Harden LLK PR review workflow reliability PR 54297
  • Add sc4 high power sku PR 54424
  • Feature | Separate test outputs per run PR 53660
  • Increase Blackhole experimental L2 timeout to 40 minutes PR 54396
  • rearrange fabric pipeline dispatch controls. PR 54482

Tooling

  • sweep_rank_binding_solutions: fix producer tcp-interface plumbing + NFS index race PR 54248

Other

  • Fix sharded_to_interleaved corrupting row-major output when the shard is wider than the row PR 54226
Assets 27
Loading