v0.78.0-dev20260826

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/32914502401

LLK (low-level kernels)

• Don't clear the LLK artifact tree on a collect-only pytest pass PR 54186
• SrcS format fix PR 53592
• Fix #52930: Guard the legacy-compat reciprocal's pole so 1/0 returns inf PR 53758
• Return 1 for SFPU pow(x, 0) PR 53759
• TDMA guard: per-TRISC thread_local armed_mask (TEN-4746) PR 54201
• add convolutions tests to fuser PR 53480
• Add tests for newly promoted kernels from blaze PR 53130
• feat: import out-of-tree LLK test via suite python_tests PR 54286
• Fix #46325: Remove default CB id values from unpack and pack thread LLK APIs PR 49233
• Use idiomatic form of SFPLOAD whose only side effect is AddrMod PR 54231
• Advance-test scaffolds for 8 demo-fork experimental LLKs (Blackhole) PR 52720
• fix(sfpu): atan2(inf, 0) returns 0/pi instead of pi/2 (IEEE 754 special case) PR 54240
• add support for perf testing on Quasar PR 53910
• Allow runtime dest switching in compute APIs PR 52887
• Align WH/BH LLK functional and perf test naming PR 53934

Metalium (tt-metal core)

• TDMA guard: per-TRISC thread_local armed_mask (TEN-4746) PR 54201
• Fix routing plane calculation and re-golden t3k PR 54339
• Fix #46325: Remove default CB id values from unpack and pack thread LLK APIs PR 49233
• Poll for RT profiler drain on close instead of sleeping 100ms PR 51596
• Fall back to achievable torus in auto-discovery topology mapper PR 48169
• Allow runtime dest switching in compute APIs PR 52887
• Run host-only (no-silicon) validation-smoke suites on GitHub-hosted CPU runners PR 54024

TT-NN

• reshape_on_device tile-reshape deadlock PR 52828
• ttnn.topk: honour optional indices_tensor and its index width PR 54168
• Cap matmul per-core block to output dimensions PR 53662
• untilize: budget codegen CB plan against live L1 occupancy PR 53399
• Fix incorrect output tensor accessor in all_broadcast PR 54229
• Test Only: cut nightly moreh test runtime via module-scoped device PR 53991
• Fix failing CCL WH Galaxy test in L2 nightly PR 54316
• Forward output_mem_config on the outer multiply in addcdiv_bw PR 53899
• Port moreh_nll_loss_unreduced_backward to Metal 2.0 PR 54147
• repeat: drop the implementation selector from the public op PR 53323
• fix(sfpu): atan2(inf, 0) returns 0/pi instead of pi/2 (IEEE 754 special case) PR 54240
• Port kv_cache PR 54193
• Allow runtime dest switching in compute APIs PR 52887
• Port moreh_nll_loss_step1 to Metal 2.0 PR 54272
• Port S2I to Metal 2.0 PR 52207

tt-train

• grpo_remote_rollout: enable prefill warmup in TttGenerationWorker PR 54250
• Allow runtime dest switching in compute APIs PR 52887

Models

• #54220: Re-center Kimi-K2.6-1T traced chunked prefill perf baseline PR 54222
• Uncollecting deepseek prefill combine test cases instead of skipping them PR 53059
• #54280: Re-center Kimi-K3-2.8T MoE perf baseline PR 54282
• Drop deprecated num_links/topology args from ttnn.all_gather in sampling fallback PR 52657
• Auto-skip HF weight load on warm ttnn cache for all e2e tests (generalize #48531) PR 50550
• AttnRes 1/4 — torch references for attention residuals PR 53114
• Migrate sibling models to common runtime and vLLM-ready executors PR 52724
• Allow runtime dest switching in compute APIs PR 52887

TT-STL

• Run host-only (no-silicon) validation-smoke suites on GitHub-hosted CPU runners PR 54024

Infrastructure & CI

• Bump tenstorrent-skills-reviewer pins to pick up rule corrections PR 54253
• Fix Matt Pocock reviewer never firing on PR open PR 54247
• Add tenstorrent-skills-reviewer workflow; remove daily-repo-status PR 54236
• Bump skills reviewers from Sonnet 4.6 to Sonnet 5 PR 54263
• repeat: drop the implementation selector from the public op PR 53323
• Auto-skip HF weight load on warm ttnn cache for all e2e tests (generalize #48531) PR 50550
• Migrate DeepSeek-V3 unit and module tests to tier 1 Models unit CI PR 54184
• Fall back to achievable torus in auto-discovery topology mapper PR 48169
• Align WH/BH LLK functional and perf test naming PR 53934
• CI: rename the Jira scripts and correct the AI/IP requirement inventory PR 53853
• CI: only file release-board tickets for real release failures PR 54205
• Update the code owners of tt_metal/hw/inc/api/tensor/ PR 54312
• Run host-only (no-silicon) validation-smoke suites on GitHub-hosted CPU runners PR 54024
• Feature: add /test agentic workflow for launching optional CI pipelines from a PR comment PR 54367

Tooling

• #0 - Link node tt-telemetry Grafana dashboard from health check tickets PR 53794
• Add portable cluster health reporter for Exabox validation PR 53836

Other

• Fix incorrect output tensor accessor in all_broadcast PR 54229