v0.78.0-dev20260820

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/32316703403

LLK (low-level kernels)

• feat: extract list of tensix functional workers from exalens PR 53642
• SFPU edge cases testing phase 3 PR 52938
• Retire the Blackhole INT32_MIN reduce xfail; the kernel is fixed PR 53658
• Improve Quasar LLK perf coverage and stability PR 53128
• Unpack Tilize fp8 perf improvement PR 51680
• Add Quasar parallel FPU and SFPU perf coverage PR 53072
• Quasar: migrate tt-llk tests to the BFD allocator (and remove construct_tdma_desc) PR 53598
• Use 32-bit BH packer dest-offset writes PR 53653
• Migrate sub_bcast_col unpack onto the BFD allocator PR 53694
• Reconcile topk_xl with tt-blaze: local_sort_generic wrappers and post-copy reinit PR 53350
• Implement atan2 for Quasar PR 52449
• ci: add quasar to llk ttsim weekly workflow PR 53674

Metalium (tt-metal core)

• ttnn: remove non-relocatable build-dir path from INSTALL_RPATH PR 53605
• Fix all_gather to derive axis topology and link count from real fabric wiring PR 53596
• Zero-pad out-of-range runtime-arg reads (RTA bounds-check) PR 52241
• Demote expected num_routing_planes override warning to debug PR 52705
• Add multi-node tests for Quasar PR 53539
• DEVICE PRINT: helper to print type name PR 53374
• Feature: Trace allocation safety tracker PR 50156
• Refactor DEVICE_PRINT tests to utilize Metal 2.0 API PR 53407
• Suppress -Wdeprecated-declarations on Tracy's own build targets PR 53606
• Quasar: migrate tt-llk tests to the BFD allocator (and remove construct_tdma_desc) PR 53598
• Remove is_noc_mapping_enabled() from Cluster PR 53349
• unify internal API folder PR 51065
• Give Quasar dispatch engines their own L1 memory layout PR 53054
• Migrate sub_bcast_col unpack onto the BFD allocator PR 53694
• emule tests modernization PR 53452
• Feature: Support multichip DRAM harvesting in tensor prefetcher PR 50481
• Reconcile topk_xl with tt-blaze: local_sort_generic wrappers and post-copy reinit PR 53350
• Imporving MGD Pinning algoirthm to allow for more fine tuned control PR 53066
• Implement atan2 for Quasar PR 52449

TT-NN

• Fix incorrect int32 results for ttnn.remainder and ttnn.fmod with a scalar operand PR 53476
• Prep moreh_nll_loss_unreduced_backward for Metal 2.0 Port PR 53534
• Fix fused ring indexer runtime-prefix all-gather scaling PR 53445
• ttnn: remove non-relocatable build-dir path from INSTALL_RPATH PR 53605
• Fix all_gather to derive axis topology and link count from real fabric wiring PR 53596
• Make ttnn.prod accuracy observable in test_prod_all PR 52452
• Test Only: reduction op merge gate tests, one test per program factory PR 53547
• Update topk large indices perf baselines PR 53659
• Trim redundant test_split/test_repeat_interleave/test_slice_for_conv parametrization to fix nightly data_movement timeout PR 53109
• Migrate moreh_layer_norm_backward, moreh_group_norm_backward to Metal 2.0 PR 53040
• Migrate tile-metadata free calls to Device 2.0 accessors (matmul/reduction/normalization) PR 52826
• Remove DSV3.2 tests PR 52881
• Metal 2.0 port: moreh/moreh_nll_loss_backward PR 53096
• #53646: Preserve 32-bit TopK indices in multi-core sorting PR 53647
• kernels: forbid get_read_ptr on cb_reserve_back'd CBs (fixes #39432) PR 45454
• Move Galaxy Llama 3.3 70B references to Hugging Face and complete the e2e/demo cutover PR 50717
• pad: replace get_dynamic_runtime_args with a factory override_runtime_arguments PR 52556
• #53648: move slowest sim dram group_norm test shapes to nightly PR 53650
• experimental/dropout: replace get_dynamic_runtime_args with override_runtime_arguments PR 50341
• Fix ttnn.where and ttnn.where_bw docs and examples PR 53571
• Feature: Trace allocation safety tracker PR 50156
• Remove unused variables in permute writer kernel causing compiler warnings PR 53431
• rng (uniform/bernoulli/dropout): get_dynamic_runtime_args -> override_runtime_arguments (factory) PR 50734
• ttnn.topk: route large-k, wide/non-pow2 small-k, and MoE-gate shapes onto topk_large_indices (BH) PR 53464
• Port pool/upsample Interleaved/Sharded/NearestFloat factories (3 of 4) PR 52503
• Wire SiTU-GLU (Kimi K3) into the routed expert kernel PR 53602
• cover all sanity-tests legs with per-job and per-run summaries PR 53524
• unify internal API folder PR 51065
• Metal 2.0 port: embedding PR 53425
• Remove unused variable causing repeated compiler warning in groupnorm kernel PR 53024
• Feature: Support multichip DRAM harvesting in tensor prefetcher PR 50481
• Revert "[Cleanup] Replace LLK workaround with (now working) mul_reuse_dest_tiles (#53419)" PR 53689

Models

• MiniMax-H3 audio: fix conv_pre and tap_matmul under T-sharding (rebased onto main) PR 53073
• sampling: adaptive single-device vocab splits and per-request seed salting PR 53167
• Flux.1 attention + transformer block tests: enumerate mesh sizes directly, require exact device count PR 53588
• Sampling module tests (handover of #38752) PR 51576
• Remove DSV3.2 tests PR 52881
• Feature: Add TTTv2 Llama 3.1 8B runtime for demo and vLLM integration PR 51184
• Move Galaxy Llama 3.3 70B references to Hugging Face and complete the e2e/demo cutover PR 50717
• DFlash prefill: multi-turn + multi-user drafter context-KV PR 52413
• Wire SiTU-GLU (Kimi K3) into the routed expert kernel PR 53602
• Add flux.2 to tt_dit PR 53608
• K3 MLA chunked test against its golden trace PR 53477
• Add LiquidAI LFM2.5-VL-1.6B multimodal support PR 52523
• #52838: Migrate Qwen-Image to the tiered Models CI pipelines (BH QuietBox 2) remove stale pipelines and references PR 53191

Infrastructure & CI

• Fix fused ring indexer runtime-prefix all-gather scaling PR 53445
• sampling: adaptive single-device vocab splits and per-request seed salting PR 53167
• ci-digest: resolve the run summary by attempt and scope PR 53644
• Add multi-node tests for Quasar PR 53539
• Sampling module tests (handover of #38752) PR 51576
• Remove DSV3.2 tests PR 52881
• #53563: Rename jobs in blaze_models_prefill_tests.yaml to a consistent convention PR 53565
• kernels: forbid get_read_ptr on cb_reserve_back'd CBs (fixes #39432) PR 45454
• Feature: Add TTTv2 Llama 3.1 8B runtime for demo and vLLM integration PR 51184
• Move Galaxy Llama 3.3 70B references to Hugging Face and complete the e2e/demo cutover PR 50717
• Feature: Trace allocation safety tracker PR 50156
• Remove @kevinmi-TT from CODEOWNERS PR 53671
• DFlash prefill: multi-turn + multi-user drafter context-KV PR 52413
• Skip code-analysis for unrelated changes PR 53401
• Add Blaze Models Prefill tests to CICD data production PR 53680
• cover all sanity-tests legs with per-job and per-run summaries PR 53524
• ci: anchor the wheel-artifact match when reusing a run's build PR 53682
• unify internal API folder PR 51065
• K3 MLA chunked test against its golden trace PR 53477
• Collapse build-artifact summary blocks into
  sections PR 51510
• Remove codeowner-bypass group from CODEOWNERS PR 53712
• Imporving MGD Pinning algoirthm to allow for more fine tuned control PR 53066
• #52838: Migrate Qwen-Image to the tiered Models CI pipelines (BH QuietBox 2) remove stale pipelines and references PR 53191
• ci: add quasar to llk ttsim weekly workflow PR 53674

Documentation

• DEVICE PRINT: helper to print type name PR 53374
• Feature: Trace allocation safety tracker PR 50156

Tooling

• Rediscover ethernet links after reset so retrain success is detected PR 50620
• #0 - Sync health check infra versions PR 53641
• Imporving MGD Pinning algoirthm to allow for more fine tuned control PR 53066

Other

• sampling: adaptive single-device vocab splits and per-request seed salting PR 53167
• Fix all_gather to derive axis topology and link count from real fabric wiring PR 53596
• kernels: forbid get_read_ptr on cb_reserve_back'd CBs (fixes #39432) PR 45454
• Move Galaxy Llama 3.3 70B references to Hugging Face and complete the e2e/demo cutover PR 50717
• ttnn.topk: route large-k, wide/non-pow2 small-k, and MoE-gate shapes onto topk_large_indices (BH) PR 53464
• #52838: Migrate Qwen-Image to the tiered Models CI pipelines (BH QuietBox 2) remove stale pipelines and references PR 53191