v0.75.0-dev20260717

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/29545485095

LLK (low-level kernels)

• Enable true INT32 for relu_min/relu_max/relu6 PR 49721
• Add --op flag to filter tests by SFPU op PR 49832
• Blackhole SFPU MEDIUM-bucket sfpi conversions (net wins) PR 49826
• LLK] perf: add cross-thread zone-alignment barrier (sync_point) to the profiler PR 49741
• 16x32 tiny-tile support for SDPA reduce_block_max_row + sub_bcast_col_custom PR 49270

Metalium (tt-metal core)

• Quickfix remove problematic warning message from telemetry PR 50133
• Enable true INT32 for relu_min/relu_max/relu6 PR 49721
• disable device profiler for mock device PR 49596
• 16x32 tiny-tile support for SDPA reduce_block_max_row + sub_bcast_col_custom PR 49270

TT-NN

• #49987: Skip test for alignment mismatch in index_fill op PR 49990
• fix hangs in Wan on quad and realign ccl multihost ring geometry introduced from PR#46138 PR 49856
• #46565: Extend universal input/output support for ttnn::gather PR 47361
• Fix HEIGHT_SHARDED width front-pad offset PR 49493
• Enable true INT32 for relu_min/relu_max/relu6 PR 49721
• SP×TP seq-sharded DSA indexer PR 49496
• pack_scaled_fp8_kv_cache: Add mixed-format row packing PR 49812
• Use f32 intermediate values in SDPA PR 48753
• matmul: use framework default program hash (decouple descriptor helper) PR 49869
• generalized_moe_gate: drop redundant custom compute_program_hash PR 49870
• Adds fused full-tile bias support to MatmulMultiCoreReuse PR 49825
• Document I/O specs and fix examples for CCL ops PR 49861
• batch_norm RunningStatistics: drop redundant custom compute_program_hash PR 49871
• Expose ccl get_usable_topology to Python PR 49838
• #50157: static cast size_t to uint32_t in hetergeneous_data_structs.hpp PR 50159
• Remove WH-only LLK workarounds PR 49859

Models

• fix(mla): guard dense k_chunk=640 against DeepSeek-V3.1 L1 OOM PR 50005
• test(deepseek_v3_d_p): fix stale GLM52Adapter import in conftest PR 50001
• test(sparse_mla): size standalone GLM-5.2 index cache to num_full_indexer_layers PR 50003
• Bump Mochi VAE threshold PR 50160
• Small accuracy improvements for some DiT models PR 49328
• fix hangs in Wan on quad and realign ccl multihost ring geometry introduced from PR#46138 PR 49856
• SP×TP seq-sharded DSA indexer PR 49496
• Add Janus Pro Conv2d patch embedding with parity test PR 47333
• Use f32 intermediate values in SDPA PR 48753
• Added zero_padded_kv_cache in bh-e2e CI PR 49954
• Remove tile config from row-major tensors in deepseek_v3_b1 tests PR 49961
• #50137: add channel parameter to pad_and_fold_conv_filters_for_unity_stride PR 50153

Infrastructure & CI

• ci: docker SBOM and bake optimizations PR 49882
• refactor(pr-gate): remove duplicate wheel build from the PR gate [Performance] PR 49897
• Add Janus Pro Conv2d patch embedding with parity test PR 47333
• ci: remove nightly/comprehensive sweep + validation cron schedules PR 49891
• Add minimax m3 prefill owners PR 49906
• Organize merge gate by team PR 49978
• Added zero_padded_kv_cache in bh-e2e CI PR 49954
• Modular Docker build with LLM: deepseek-pd-worker image PR 49709

Documentation

• Document sfpi API additions PR 49915

Other

• ci: remove nightly/comprehensive sweep + validation cron schedules PR 49891
• Use f32 intermediate values in SDPA PR 48753
• Fix concat_fuse mesh mapper to shard along the concat dim PR 49865
• Expose ccl get_usable_topology to Python PR 49838