v0.77.0

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/32177505475

LLK (low-level kernels)

• Add back the sfpu_math value so the perf report headers match again PR 52048
• fix(perf): regenerate catalog for formats.sfpu_math — unbreak main gate PR 52058
• Exempt reduce_block_max_row negative controls from the bit-exact check PR 51677
• fast_tilize: fix PERF_RUN_TYPE being shadowed by a macro, and the run-type paths that unblocks PR 51918
• Fix ttnn.sort returning incorrect values for UInt16 values + UInt32 indices (WH) PR 50709
• Fix LLK xdist artifact cleanup race PR 52223
• Remove fixed sleeps from LLK stream polling PR 52221
• Correct fast-math options PR 52236
• Fix duplicate perf-CSV headers and enforce uniqueness PR 51485
• guard REDUCE_OP parameter use in Quasar tilizeA_B api PR 52510
• llk tests: fix the undefined name in counter_report PR 52439
• perf tests: fix four measurement bugs across seven perf kernels PR 52440
• Fix perf reorg leftover: _write_run_parquet imports helpers.perf.parquet PR 53132
• LLK e2e: skip SyncFull dest handoff for tiny-tile matmul occupancy > 8 PR 53502
• Add golden perf-CSV header catalog and drift gate PR 51484
• prevent parametrize expansion for skipped tests PR 52059
• matmul_custom THROTTLE levels + unpack_A_custom standalone PR 50655
• Cover experimental generalized_moe_gate LLKs (Blackhole + Wormhole B0) PR 50566
• generalized_moe_gate: sanitize scratch so GATE stays bit-exact PR 52699
• LLK Test Infra transpose dest tests with Dest bank switching PR 51673
• SFPU edge cases initial pass ( expand ranges on ops) PR 52172
• deepseek_moe_gate: sanitize scratch so GATE stays bit-exact PR 52949
• Add RoPE SFPU test PR 53337
• Add tests for blaze-authored SFPU activation/filter LLKs PR 53123
• Limit LLK tilize polluter negative controls to same-format cases PR 52226
• add unpack tilize to fuser PR 51978
• Enable Quasar selection in LLK ttsim regression script PR 51649
• cleanup(llk): remove dead llk_pack_untilize_hw_configure PR 51838
• unpack_tilize tiny tiles compute api bringup PR 51405
• Add blaze-authored Blackhole LLKs to experimental PR 51361
• add pack untilize kernel to fuser PR 52041
• add unary broadcast kernel to fuser PR 52061
• add transpose dest operation to fuser PR 51610
• add UnpackReduceTilize to fuser PR 52079
• SFPU accuracy plotting tool PR 47249
• Fix ThreadId enum, mailbox_read/write TDMA path, and matmul … PR 51508
• Add topk_xl tests PR 51777
• Add llk-wave-debug skill for Quasar FSDB waveform diagnosis PR 51977
• Drive Int32 min through the LLK MIN reduce instead of -max(-x) PR 49589
• Add shared wide nullable perf-report schema (v1) PR 51594
• tilizeA_B tiny tile port PR 51418
• Implement CSV -> Parquet conversion PR 51737
• Add QuaSAR per-DFB TDMA guard to detect wait/pop and reser… PR 52369
• DeepSeek-V3 face-granular (16×16) compressed matmul: LLK kernels + metal API stack PR 50929
• golden, codegen and validation cleanup PR 52205
• add tests for uplifted blaze experimental LLKs PR 52163
• Add tests for experimental SFPU SDPA kernels PR 52710
• Fix SFPU RNG correlation and improve FP32 uniform random quality PR 52024
• Disarm TDMA guard on single-tile llk_pack PR 52748
• fix: improve tanh bf16 performance PR 52732
• Add static asserts to dest dvalid sync functions on llk api level PR 52843
• binary op sfpu (add) in parallel with matmul PR 47469
• #49944: adding RNE rounding to SFPU ADD, SUB, RSUB llk API PR 51060
• Add tests for hadamard and deepseek_moe_gate kernels PR 52827
• sfpu edge case testing phase 2 PR 52416
• Implement dev tool for easier regression check PR 52850
• refactor(llk-perf): group perf helpers under helpers/perf/ package PR 52706
• enable layernorm test for quasar PR 52950
• Add blaze-authored SFPU activation/filter LLKs to experimental (batch 3) PR 53025
• Remove ai_gen folder from llk tests PR 53049
• Pipeline reorg: Migrate Blackhole Sanity into sanity-tests PR 48943
• Sfploadmacro exp PR 52522
• Implement expiremental eltwise_binary_sub_bcast_cols_custom PR 51593
• Historical CSV→Parquet migration PR 51951
• Add publish_run: one run's CSVs -> one typed run.parquet PR 53019
• Add functional and performance coverage comparison script PR 51692
• Promote rmsnorm, add_rsqrt, eltwise_mul_scalar LLKs to experimental PR 52709
• Promote the SFPU RoPE LLK to experimental PR 53119
• Update ttsim-version PR 53184
• Add support to generic moe gate for k=9..15 PR 52867
• #47382: reconfig_data_format — replace is_tile_dim_reconfig_en with intent-named APIs PR 52253
• Fix Blackhole destination-reuse synchronization PR 52256
• #40182: Replace int32_to_float with convert PR 53214
• Add softcap unary op and situ_glu binary activation (Kimi K3 SiTU-GLU) PR 52448
• llk-audit: deterministic recall augmentor for the race audits PR 49657
• Document FP16B unity constant PR 53503
• trigger quasar build on fuser changes PR 52069
• ci: update tt-exalens version to 0.3.31 in LLK test harness PR 53514
• Promote the hw_cleanup LLK family to experimental PR 53296
• Implement cumsum for Quasar PR 52935

Metalium (tt-metal core)

• jit_build: fix kernel-cache temp-path collision across forked processes PR 51570
• Fix Tensix hang in unary broadcast compute API PR 51772
• Quasar - Fixing watcher to always analyze unmapped TCs PR 52087
• Updating experimental::quasar DM CreateKernel to skip DM0 & DM1 PR 52095
• Per-fiber my_logical_x_/y_ globals + build_core_map stale-SWEmuleChip cache invalidation PR 52071
• #51409 on splitting dfb id to be program unique and adding a device facing id that is unique within core group PR 51499
• Dedupe repeated "Unknown motherboard" warning in physical system discovery PR 51643
• Fix TC capacity to be 16bits PR 52117
• Stride the dispatcher's paged write bank walk by the aligned page size PR 52227
• Return Metal logical DRAM coords from get_metal_dram_cores, unbreaking watcher on Blackhole PR 52031
• Pass tile size to matmul test-kernel TensorAccessors PR 52209
• Fix ttnn.sort returning incorrect values for UInt16 values + UInt32 indices (WH) PR 50709
• Use per-core shard address for host data movement on per-core-allocated L1 buffers PR 52080
• Correct fast-math options PR 52236
• Minor Scratchpad fixes for Metal 2.0 PR 52381
• Fix Quasar unicast and iDMA VC assignments PR 52240
• Compile only the general relay_paged read loop on an erisc, unbreaking the eth prefetcher PR 52501
• Fix protobufs visibility issues by avoiding exporting scaleout_tools symbols PR 52259
• Draw the random packed write payload in whole words, and stop the clamp converging on none PR 52237
• emule: bind per-core CB config by core_ranges membership (tt-emule-blaze#153) PR 52726
• Use monotonic semaphores in Quasar pipeline test kernels PR 51601
• Synchronize device tests before host readback PR 52466
• Fix tensor prefetcher stop() deadlocking under the Tracy device profiler PR 52529
• Make runtime_noc_debugging run and pass in CI PR 52672
• Metal 2.0: record dataflow buffers in graph-capture CB accounting PR 52757
• Fixing Producer Injection Counter Overflow PR 52396
• emule: resolve a fabric connection's direction per worker, not per chip PR 52888
• tests: fix FP32 SFPU binary broadcast unpack path PR 52758
• tests: skip nightly fabric UDM suite before initialization PR 52740
• Resolve Quasar dispatch core type before choosing core descriptor PR 53085
• Fix matmul integration tests for DRAM accessors and async completion PR 52425
• emule: converge the runner with the emule-blaze fork PR 53162
• NOC debug tool fixes PR 50115
• Always validate Metal 2.0 program args on cache miss, gate only hit-path re-validation PR 53084
• Preserve descriptor-backed simulator topology classification PR 52785
• Reset the shared GO mailboxes from the last CQ, once the others have drained PR 52777
• Set KERNEL_COMPILE_TIME_ARG_MAP via a generated header instead of -D define PR 53404
• Update per-core allocation tensor spec test expectation PR 52210
• Add Quasar fast dispatch stress tests PR 52104
• tests: handle missing mock cluster setup for PSD tests PR 52479
• Compile mux-v2 kernels at O3 and refresh WH/BH goldens PR 53023
• Changing Quasar DPRINT tests to use Metal 2.0 API PR 53005
• Llama 8 New MGD PR 53048
• asan: robustify Illegal-Semaphore tests for the provenance exemption PR 53052
• Tensor prefetcher: take a MeshCommandQueue, not a cq_id PR 51635
• Automated UMD Bump 03.08.2026 PR 51947
• Automated UMD Bump 04.08.2026 PR 52055
• Collapse fabric stability intensity forks into one YAML (gtest-style name filters) PR 50270
• Collapse BH-glx torus/superpod YAMLs into test_fabric_2d_torus_stability.yaml PR 50282
• Preallocate vector capacity and move vectors into structured return values PR 51988
• unpack_tilize tiny tiles compute api bringup PR 51405
• Performance: deepen the fast dispatch prefetcher's relay_paged scratch to a 3-buffer ring PR 51272
• Add blaze-authored Blackhole LLKs to experimental PR 51361
• Collapse per-core-type command queue dispatch layouts into a single layout PR 52105
• Break DispatchCoreManager/ServiceCoreManager MetalContext::instance() cycle PR 52023
• Bump ttsim version to v1.9.9 PR 52130
• Add INT8 Support in tt-metal PR 50391
• Automated UMD Bump 05.08.2026 PR 52132
• Preallocate vector capacity and move vectors into structured return values PR 51989
• Demote real-time profiler sync-complete log to debug PR 52013
• Demote 'Using pre-compiled firmware from' log to debug level PR 51794
• Fix ThreadId enum, mailbox_read/write TDMA path, and matmul … PR 51508
• Add missing include sysmem_buffer.hpp to tt_cluster.hpp PR 52199
• Reduce dispatch engine prefetch ringbuffer size for two command queues PR 52123
• Adding UnitMeshFixture to modernize single-device tests PR 52222
• Add topk_xl tests PR 51777
• Performance: cut the per-page cost of the fast dispatch prefetcher's relay_paged read loop PR 51793
• Mesh coord offset for mesh shape override mapping/composing PR 51639
• Remove dependency to umd::Firmware PR 52200
• Parallelize loading bfloat in host PR 51469
• Correct ISA selection PR 52239
• Pass MetalEnvImpl into DataCollector and ProfilerStateManager constructors PR 52245
• Drive Int32 min through the LLK MIN reduce instead of -max(-x) PR 49589
• Add common-field accessors to ComputeHardwareConfig in Metal 2.0 PR 52350
• Fix #22943: Eltwise binary + broadcast init cleanup — migrate all call sites PR 50745
• Declare watcher ASSERT before including risc_common.h PR 52349
• SFPI 7.68.0 803 PR 52476
• ci: skip Tracy/emscripten build overhead where the profiler isn't needed PR 52552
• DeepSeek-V3 face-granular (16×16) compressed matmul: LLK kernels + metal API stack PR 50929
• Port sum_reduce_scalar from blaze to metal PR 52076
• emule: fabric route ordering race, $TMPDIR JIT scratch, and two triage diagnostics PR 52355
• Fix SFPU RNG correlation and improve FP32 uniform random quality PR 52024
• emule: resolve mcast semaphore increments only to worker cores PR 51504
• MeshDevice wrappers for slow dispatch PR 52478
• populate per-CB face geometry at CB/DFB setup (tt-emule-blaze#191) PR 52742
• emule: enable multi-dispatch socket pipelines under host-interleaved dispatch PR 52334
• Auto-mapper support for Disagg Prefill Decode PR 43175
• Lowering log level for logging in dataflow buffer file PR 52338
• Implement vararg CTA for metal 2.0 PR 52391
• Automated UMD Bump 07.08.2026 PR 52447
• Consolidate mailbox tests; cover DM->TRISC for all compute threads PR 52791
• Weight-load + JIT-compile observability in logs PR 52258
• Graduate runtime tensor out of experimental PR 51640
• #49944: adding RNE rounding to SFPU ADD, SUB, RSUB llk API PR 51060
• lock/unlock support for DFB PR 47975
• Add tests for hadamard and deepseek_moe_gate kernels PR 52827
• Demote per-device "Fabric initialized on Device" log to debug level PR 51849
• Demote per-device Watcher attach/detach logs to debug severity PR 52122
• docs: fix Doxygen argument tables to match compute API function signatures PR 52833
• Automated UMD Bump 11.08.2026 PR 52815
• Automated UMD Bump 12.08.2026 PR 52926
• Add support for per core allocator on socket buffers PR 52551
• DFB tests modernization PR 52965
• Surface Area Reduction profiler headers PR 52766
• Shrink the DRISC mailbox to fit its own mailboxes_t PR 52980
• Automated UMD Bump 13.08.2026 PR 53008
• SFPI 7.69.0 822 PR 53044
• Prepare for upcoming get_heartbeat() UMD API change PR 53032
• Deleting old unused tests PR 53013
• Add x280 support to HD sockets PR 52017
• Thread context_id into ProgramImpl and its downstream PR 52771
• Uniform passing convention for MeshCommandQueue in Runtime API PR 51559
• Remove data_types.hpp from the tt-metalium API surface PR 52977
• Remove vector_aligned.hpp from tt_metal api PR 52129
• MeshDeviceSingleCardFixture-based tests modernization PR 52998
• Feature: Adding CrossNodeDFB support for WH/BH PR 52776
• Implement expiremental eltwise_binary_sub_bcast_cols_custom PR 51593
• Intra-galaxy pipeline-parallel prefill PR 51874
• Promote rmsnorm, add_rsqrt, eltwise_mul_scalar LLKs to experimental PR 52709
• Automated UMD Bump 14.08.2026 PR 53121
• Refactoring device ID access in dispatch tests PR 53007
• Promote the SFPU RoPE LLK to experimental PR 53119
• Bump ttsim version to v1.10.1 PR 53180
• Add support to generic moe gate for k=9..15 PR 52867
• #47382: reconfig_data_format — replace is_tile_dim_reconfig_en with intent-named APIs PR 52253
• Fix Blackhole destination-reuse synchronization PR 52256
• Demote per-iteration Rx launch/status logs to debug in 1D fabric test PR 52807
• Metal tests update to Mesh APIs PR 53163
• Promote pack_block_contiguous_uninit to the experimental compute API PR 53117
• Add softcap unary op and situ_glu binary activation (Kimi K3 SiTU-GLU) PR 52448
• Fabric tests update to Mesh APIs PR 53169
• #52756: Performance: cut host-side EnqueueMeshWorkload cost on multi-device meshes PR 52775
• Per-core H2D socket FIFO in hybrid mode PR 53322
• Automated UMD Bump 18.08.2026 PR 53479
• Promote the hw_cleanup LLK family to experimental PR 53296

TT-NN

• Avoid optimized sharded tilize for row-major tiny-tile inputs PR 51517
• Fix moe_routing_remap expert_parallel_size vs mesh-axis validation PR 51686
• RingJointSDPA: restore non-sliding multi-batch support PR 52073
• max_pool2d - flush L2 after halo pad fill PR 52187
• Support ND-sharding in ttnn.all_gather PR 51515
• Reland "[Bug fix] Make sampling and top-k deterministic on sub-core grids." with int32 tie-break index reduce PR 52177
• Fix ttnn.sort returning incorrect values for UInt16 values + UInt32 indices (WH) PR 50709
• Use per-core shard address for host data movement on per-core-allocated L1 buffers PR 52080
• Fix ttnn.group_norm fused kernel for non-tile-aligned H*W (#50682) PR 51159
• #50684: Shrink CoreRangeSet in generate_transpose_shard_spec to populated shards PR 51156
• #50714: Fix SliceRmShardedProgramFactory silent corruption on width-slice PR 51565
• Fix one-tile all-gather scatter header PR 52431
• Fix dense token-map stride in selective_reduce_combine PR 52377
• Remove expert-count-dependent TRISC2 stack allocation PR 52511
• Drain pending NoC atomics before ComputeOnly exit PR 52520
• fix flux clip regression by reversing ring attention accumulation order PR 52547
• fix(generic-op): update all-gather writer ABI PR 52291
• Clear DeepSeek dispatch writer NOC transaction tags PR 52432
• Reset DeepSeek combine writer transaction tag per row PR 52433
• Fix ttnn.sort wrong indices for bf16 PR 51562
• Fix high_bw_all_gather cached-launch synchronization PR 52633
• Extend matmul K-tile padding to shared-exponent dtypes PR 52302
• Fix Blackhole hang on RM reshape to [N,1] PR 50967
• Avoid use of hard-coded bfloat16 intermediate CBs in Layernorm PR 52081
• Bug #48267 Pair printed shards with their own coordinates in to_string PR 52708
• fix device_id remap skips buffer_pages_by_address snapshot format and buffer_pages PR 51678
• Metal 2.0: record dataflow buffers in graph-capture CB accounting PR 52757
• Fix AllGather BH CI machine requirement PR 52862
• Fix ttnn.split padded-shape leak and non-aligned TILE fast-path PR 51768
• Softmax: Fixed block_size (4/8) in attention-optimized interleaved softmax to maximize dest usage PR 51145
• #52572: Fix conv2d PACK format state before tilize PR 52906
• Fix sort compute kernel API misuses causing silent wrong data PR 52271
• Make the MMRS L1 MM-output hand-off opt-in PR 52863
• Matmul program config for single-core sharded grids PR 52039
• Fix MoE compute stride and no-bias K-padding for small WH sweeps PR 51847
• Fix slice tensor-args reader kernel upper-dim offset (Horner loop order) PR 52902
• Fix tilize kernel API misuses causing silent wrong data PR 51903
• Mask the tile-padding rows instead of back-correcting for them PR 52924
• Remove DeepSeek combine JIT warnings PR 53145
• Fix GPT-OSS first-group chunked sliding RingJointSDPA PR 53153
• Fix prefill dispatch on routed groups PR 53171
• Fix prefill combine on routed groups PR 53172
• Always validate Metal 2.0 program args on cache miss, gate only hit-path re-validation PR 53084
• ttnn.topk multi-core: fix silent value corruption for >32 flattened rows PR 53453
• #52680: Fix noc_async_{read,write}_sharded per-chunk size underflow PR 52154
• #52681: Clamp tilize retile shrink-case output tile-rows to buffer capacity PR 52180
• Add accurate fp32 SFPU path for ttnn.max via fast_and_approximate_mode PR 52718
• #51235: Fix reduce_to_root intermediate tensor spec and c_18 circular buffer overflow PR 52852
• Fix Ring SDPA LLK assert program size and FP32 hangs PR 51963
• Fix pad failure in Nightly tests PR 52392
• #52414: update reduction test accuracy thresholds PR 52415
• xfail moe nontile sweep on WH 7x8 grid (tt-metal#52246) PR 52248
• Gate high_bw_all_gather tests to Blackhole PR 52454
• Drop unreachable reshard assertion from legacy DRAM block-sharded reject test PR 52869
• prefill(unified expert perf tests CI enablement) PR 52918
• Fix sparse matmul tile mismatch test expectation PR 53186
• #52660: create prototype copy of llama quasar ops tests PR 52663
• Add fp32 sort index test PR 53327
• prune the reduce test grid and re-enable test_reduction.py on ttsim PR 53360
• test(graph_query): assert softmax DFB peak instead of CB peak PR 53446
• Tensor prefetcher: take a MeshCommandQueue, not a cq_id PR 51635
• #47500: Traceable chunked prefill — trace layer + on-device MoE padding config PR 51624
• Fix nonzero row-major block-sharded L1 reads PR 47148
• Port experimental/transformer/rotary_embedding_llama factories PR 51490
• Fix full ND sharded TILE test cases to use tile-aligned shard shapes PR 49576
• Preallocate vector capacity and move vectors into structured return values PR 51988
• Revert "[Bug fix] Make sampling and top-k deterministic on sub-core g… PR 52088
• Add support for sharded joint on the sequence dimension to ring_joint_scaled_dot_product_sdpa PR 48677
• DS Prefill :: Improve fabric links usage for dispatch operation PR 51019
• Prepare untilize_with_unpadding for Device 2.0 Port PR 51695
• Add INT8 Support in tt-metal PR 50391
• Migrate remote semaphore increments to Semaphore<>::up() PR 49605
• sparse_matmul: reject block/subblock configs that deadlock mcast_in0 PR 51514
• repeat: codegen port, device-2.0 re-verification PR 49888
• Metal 2.0 port: normalization/layernorm_distributed PR 52006
• Universal I/O support for repeat_interleave PR 51307
• Migrate Llama standalone ops to Device 2.0 API PR 49129
• ttnn: add experimental RGB→YUV 4:2:0 conversion op PR 50733
• Device 2.0 Migration for Experimental SSM Ops PR 52078
• Support 1D Reduce Scatter on reshaped mesh device PR 50777
• unary + moreh_adam/adamw: patch cache-hit args in place, no descriptor rebuild PR 51785
• Preallocate vector capacity and move vectors into structured return values PR 51989
• Fix ThreadId enum, mailbox_read/write TDMA path, and matmul … PR 51508
• Metal 2.0 port: reduction/manual_seed PR 52007
• Remove unused out_block_hw_actual variable in groupnorm kernels PR 52108
• Fix offset pointers in I2S and S2I PR 51747
• Harden flaky batch_norm PCC test PR 49784
• Metal 2.0 port: reduction/moe PR 52225
• Add exact float32 support to ttnn.topk PR 50521
• Port Reshard to Metal 2.0 PR 51495
• rotary_embedding_indexed: port to Metal 2.0 named-arg program factory PR 52191
• Layer Acknowledgements via D2H Service PR 49247
• Port Gelu Backwards to Metal 2.0 PR 51771
• Mesh coord offset for mesh shape override mapping/composing PR 51639
• Add high-bandwidth all-gather for sparse MLA prefill PR 51134
• Migrate precomposed NoC addresses to UnicastEndpoints (CCL) PR 49577
• Port normalization/batch_norm (2/2 device-operations, 8 kernels) PR 51974
• Add token count aware decompression path to fp8 decompression operator PR 50853
• Perf improvements for fp8 compression and decompression PR 50639
• Remove dead enough_space_width heuristic from tilize and untilize PR 52092
• Remove unused reset_index and output_tile_index in groupnorm kernels PR 52111
• Port data_movement/permute (all 5 factories) PR 51095
• Port data_movement/fold factories PR 51042
• Drive Int32 min through the LLK MIN reduce instead of -max(-x) PR 49589
• Fuse sparse MLA ring indexer over ND-sharded caches PR 51419
• Fix ttnn.all_gather topology creation with invalid shard dims PR 52333
• data_movement: patch cache-hit args in place instead of rebuilding the descriptor PR 51784
• Add support for non_rectangular corerangeset inputs to layernorm PR 52277
• Fix #22943: Eltwise binary + broadcast init cleanup — migrate all call sites PR 50745
• Migrate all_gather_minimal_matmul_async set_remote to Semaphore::relay_unicast PR 49706
• Performance: Fuse per-channel quantize/dequantize with a scalar zero-point PR 51138
• Metal 2.0 port: reduction/sampling PR 52322
• Metal 2.0 port: reduction/topk PR 52364
• Port moreh/moreh_getitem (2/2 factories) PR 52160
• #52509: ifdef out function calls not defined on quasar in kernel lib helper PR 52512
• Port normalization/softmax + coupled moreh/moreh_softmax (all 7 + 5 factories) PR 51753
• mesh_partition: share slice's cache-hit patch instead of rebuilding PR 51786
• Port data_movement/bcast to Metal 2.0 (all 5 factories) PR 50901
• Add validate_program_args config (skip Metal 2.0 host-side arg validation off-CI) PR 50945
• GroupNorm: Added FP32 support to GroupNorm across interleaved and sharded (Welford and legacy) configurations PR 49219
• #52085: update quasar resnet model and ops PR 52535
• exp_ring_joint_sdpa: replace get_dynamic_runtime_args with a program-factory override PR 52554
• Port transpose off get_dynamic_runtime_args onto override_runtime_arguments PR 52566
• sparse_sdpa_msa: replace get_dynamic_runtime_args with override_runtime_arguments PR 50352
• Port moreh/moreh_sum (6/6 factories) PR 51995
• Support Row Major interleaved input/output on the legacy GroupNorm path PR 49501
• Generalize multi_scale_deformable_attn to support D values that are multiples of 16 PR 52380
• Fix SFPU RNG correlation and improve FP32 uniform random quality PR 52024
• Layer Ack Service Multi host compatible PR 49644
• Migrate multicast semaphores to Semaphore::set_multicast / inc_multicast PR 52361
• Auto-mapper support for Disagg Prefill Decode PR 43175
• Fused CCL + matmul: fabric-bound minimal-matmul, strided AGMM and MMRS PR 52513
• Expose forwarding link indices to Python PR 52642
• Add KDA reference semantics and test utilities PR 52781
• Sparse MLA: move gathers to high-bandwidth all-gather PR 52606
• #47644: Extend universal input/output support for ttnn::fold PR 50385
• reduce_scatter "direct" algorithm for small shapes PR 51741
• Graduate runtime tensor out of experimental PR 51640
• #49944: adding RNE rounding to SFPU ADD, SUB, RSUB llk API PR 51060
• Neighbor-pad halo exchange and halo-mode conv3d PR 52514
• Ring joint SDPA and the LTX-2.3 distilled AV pipeline PR 52730
• Decouple the socket page size from the tensor page size PR 52457
• Preserve ND provenance in permute fallback PR 46594
• Bug fix: honor rank>4 untilize_with_unpadding output_tensor_end PR 46939
• Add support for per core allocator on socket buffers PR 52551
• Surface Area Reduction profiler headers PR 52766
• data_movement: move shared codegen kernels out of repeat/ PR 52806
• Mean & Sum: Added H-axis split to the ROW_MAJOR H-reduce for tall-H shapes PR 50393
• Fix quantize composite path ignoring memory_config PR 52992
• Add x280 support to HD sockets PR 52017
• add support fused/virtual concat to matmul variants (except agmm) PR 51802
• Add Minimax H3 support PR 52874
• third_party: upgrade nanobind pin 2.12.0 -> 2.14.0 PR 53057
• sdpa: revert even-ring split forwarding from ring-joint SDPA PR 53076
• Uniform passing convention for MeshCommandQueue in Runtime API PR 51559
• Remove data_types.hpp from the tt-metalium API surface PR 52977
• Port untilize (build_untilize_tile) — validated on wormhole_b0 + blackhole PR 50178
• Port moreh/moreh_softmax_backward (5/5 factories) PR 52986
• Optimize generic all-to-all packet scheduling PR 52674
• ttnn.sparse_matmul: compact sparse output PR 52954
• Metal 2.0 port: data_movement/sort (SortProgramFactorySingleRowMultiCore) PR 52528
• #47382: reconfig_data_format — replace is_tile_dim_reconfig_en with intent-named APIs PR 52253
• Fix Blackhole destination-reuse synchronization PR 52256
• sampling: TTSampling hands ttnn.topk an indices_tensor half the width of the logits PR 53001
• Add softcap unary op and situ_glu binary activation (Kimi K3 SiTU-GLU) PR 52448
• Migrate dit_fused_distributed_rmsnorm kernels to the Device 2.0 kernel API PR 52755
• Remove references to grayskull board types PR 53125
• add fused swiglu to strided agmm PR 53203
• Add indexed/gather mode to ttnn.sparse_matmul PR 50654
• port(metal_2.0): moreh/moreh_matmul → MetalV2FactoryConcept (verified passing) PR 51093
• Command-lifetime semaphore bank and demo numbers PR 53050
• descriptors: override_runtime_arguments belongs on the program factory, not the operation PR 51637
• Port reduction/prod ProdAll/ProdNc factories PR 51096
• port(metal_2.0): moreh/moreh_norm_backward → MetalV2FactoryConcept (verified passing) PR 51103
• port(metal_2.0): moreh/moreh_mean_backward → MetalV2FactoryConcept (verified passing) PR 51102
• Derive cb_scaler tile count from a shared constant (#48487) PR 48650
• #48140: Add INT32 index support to topk PR 48843
• Migrate copy row-major kernels to Device 2.0 data-movement API PR 52232
• MultiDeviceGlobalSemaphore removal PR 53213
• group_norm: synthesize per-group masks in the kernel PR 48640
• Remove TT_METAL_QSR_TILIZE_UNPACK_TO_DEST PR 53339
• #9147: Add optional output tensor support for ttnn::repeat PR 52686
• Replace LLK workaround with (now working) mul_reuse_dest_tiles PR 53419
• Relax TTNN op numeric tolerances PR 45240
• Add sigmoid-gated RMSNorm PR 52783
• high_bw_all_gather: restore bandwidth for batch-indexed gathers PR 53368

tt-train

• Fix memory usage underreporting with gradient accumulation PR 50909
• Adding Vocab Parallel Embedding for Qwen3 TP Tied Weights PR 51428
• Test only - Add INT8 as an unsupported type PR 52761
• Preallocate vector capacity and move vectors into structured return values PR 51988
• Sparse moe train PR 43508
• Bug fix - Rebind library paths for TTML in CI PR 51993
• Fix #22943: Eltwise binary + broadcast init cleanup — migrate all call sites PR 50745
• Re-enable performance-inefficient-vector-operation clang-tidy check in tt-train PR 52285
• Fix SFPU RNG correlation and improve FP32 uniform random quality PR 52024
• Improve training stability and introduce new training (scheduler/optimizer) knobs PR 48716
• Generate and upload loss plot summaries to CI PR 51051
• Surface Area Reduction profiler headers PR 52766
• Uniform passing convention for MeshCommandQueue in Runtime API PR 51559
• GRPO example: reverse text (single p150, Qwen3-0.6B) PR 53003

Models

• Fix moe_routing_remap expert_parallel_size vs mesh-axis validation PR 51686
• Reland "[Bug fix] Make sampling and top-k deterministic on sub-core grids." with int32 tie-break index reduce PR 52177
• Fixup Wan LoRa I2V PR 52257
• relax 12B 1x4 full-model PCC gate 0.94 -> 0.935 PR 52411
• Raise ValueError for invalid sampling_splits PR 52444
• cast supplied env variable I2V_ENCODE_FRAMES to int PR 52568
• Reserve tile-aligned routed-expert dispatch capacity PR 52434
• #52630: Don't dequantize an already-dequantized Kimi checkpoint PR 52743
• Fix MoE reduce score-channel shape inference PR 52435
• generator: fix batched prefill's slot vs prefill-order index spaces PR 52808
• Bump up l1_small_size for failing tests PR 52934
• LTX: capture the encode trace last, keep the encoder measured PR 52968
• ds_prefill(k3_mla_gate) - route K3 output gate through high_bw_all_gather PR 52975
• Combine tests fixes and pruning PR 52084
• Qwen3.6-27B: Added bucketing fix and TP=8 support PR 53062
• Fix GPT-OSS first-group chunked sliding RingJointSDPA PR 53153
• Forward decode kwargs through Qwen VL wrappers PR 53131
• Let a line-cabled mesh run the H3 FFN PR 53179
• Restrict Blackhole DeepSeek HCA end-to-end tests PR 53511
• Add LTX 2.3 Blackhole T2V and I2V tests to CI PR 51767
• add WAN rope/attention to models CI, remove all wan tests from t3k. PR 52262
• Replace repeat2 with ci-eval-1 in Llama 3.3-70B galaxy e2e tests PR 52474
• SDXL VAE pcc relaxed PR 50930
• Add Wan2.2 Models to QB2 PR 52772
• Gate Chunked Kimi perf (code_debug 55k, trace) at +/-3% per chunk PR 53101
• Gate Chunked Kimi perf (code_debug 55k, no trace) at +/-5% per chunk PR 53288
• #52660: create prototype copy of llama quasar ops tests PR 52663
• Set the LayerAck env in Prefill Runner test PR 53495
• #47500: Traceable chunked prefill — trace layer + on-device MoE padding config PR 51624
• Fix merged KV chunk table stage layout PR 51996
• SP-only merged KVPE+indexer KV chunk address table PR 51917
• multirank external runner pcc test PR 51269
• MiniMax-M3 prefill zone profiler PR 51776
• remove _trace_prefill_supported_seq_lens PR 52032
• tt_transformers: remove unreachable per_core_M arm in prefill QKV config PR 52044
• Warm the SP ring cache-read path in prefill compile() PR 52022
• Migrate vLLM nightly to "vLLM Model Tests" (tiered-format, single cadence) PR 50910
• Revert "[Bug fix] Make sampling and top-k deterministic on sub-core g… PR 52088
• Add support for sharded joint on the sequence dimension to ring_joint_scaled_dot_product_sdpa PR 48677
• DS Prefill :: Improve fabric links usage for dispatch operation PR 51019
• Replace transformers fx-availability shim with direct torch.fx import PR 52067
• sparse_matmul: reject block/subblock configs that deadlock mcast_in0 PR 51514
• prefill: add 8-galaxy connected MGD + 8-rank D2D rank binding PR 52189
• Add MiniMax-M3 Blaze chunked-prefill throughput regression PR 51574
• Qwen3-32B Blackhole Galaxy: prefill unit tests + full-model e2e accuracy PR 50985
• Feature: Test owner reassignment and querying PR 49945
• Layer Acknowledgements via D2H Service PR 49247
• Mesh coord offset for mesh shape override mapping/composing PR 51639
• Enable force_argmax fast-path on non-Galaxy (single-chip P150/N150) PR 48886
• Migrate model traces dataset from Neon to Snowflake PR 48687
• Fuse sparse MLA ring indexer over ND-sharded caches PR 51419
• Update llama3.1-8b bh_quietbox_2 b4 perf targets PR 52409
• Fix #22943: Eltwise binary + broadcast init cleanup — migrate all call sites PR 50745
• Bring up Gemma-2 text models (2B / 9B) on tt_transformers PR 52471
• GroupNorm: Added FP32 support to GroupNorm across interleaved and sharded (Welford and legacy) configurations PR 49219
• Test Prefill Transformer Cleanup PR 52352
• #52085: update quasar resnet model and ops PR 52535
• Migrate Mistral-Small-3.1-24B and ViT to the tier 2 Models CI pipelines PR 52299
• Remove 11 small vision/CNN models from CI, retire single-card perf-models pipeline PR 52318
• prefill runner: add connected sub-torus (Y/XY) mesh-graph descriptors PR 50403
• Fix SFPU RNG correlation and improve FP32 uniform random quality PR 52024
• Layer Ack Service Multi host compatible PR 49644
• Verify migrated KV from the migration driver PR 52455
• GDN decode: fuse exp into the decay multiply PR 50089
• Add bf8 support PR 50361
• Auto-mapper support for Disagg Prefill Decode PR 43175
• Fix single-rank prefill KV migration for non-MLA models (MiniMax-M3) PR 52113
• Fused CCL + matmul: fabric-bound minimal-matmul, strided AGMM and MMRS PR 52513
• Add KDA reference semantics and test utilities PR 52781
• Sparse MLA: move gathers to high-bandwidth all-gather PR 52606
• Kimi-K3 MLA: NoPE, output gate, 96 heads (MLA layer only, random weights) PR 52068
• prefill: remove standalone serving mode (default 11-chunk KV, fix rank claims) PR 52213
• Weight-load + JIT-compile observability in logs PR 52258
• Clean up/document TT-DiT test organization PR 52504
• #47644: Extend universal input/output support for ttnn::fold PR 50385
• Qwen3.6-27B: Added bucketing fix and TP=8 support PR 51556
• Ring joint SDPA and the LTX-2.3 distilled AV pipeline PR 52730
• Decouple the socket page size from the tensor page size PR 52457
• #52838: Migrate Stable Diffusion 3.5 Large to the tiered Models CI pipelines PR 52892
• #0: GLM 5.2 prefill migration changes PR 52920
• Migrate 7 models to the tiered Models CI — 6 as tier 3 (N150), ResNet-50 as tier 1 off T3K / Galaxy PR 52829
• Mean & Sum: Added H-axis split to the ROW_MAJOR H-reduce for tall-H shapes PR 50393
• Enable Qwen2.5-Coder-32B on Blackhole QuietBox 2 PR 52880
• add support fused/virtual concat to matmul variants (except agmm) PR 51802
• Add Minimax H3 support PR 52874
• Retire the demo deepseek_moe_gate fork onto canonical generalized_moe_gate PR 52747
• Remove unmaintained OFT model PR 52919
• Intra-galaxy pipeline-parallel prefill PR 51874
• Add opt-in file-export transport for prefill migration device map PR 52872
• Prefill migration with KV chunk table and bring-up hooks PR 52292
• Promote rmsnorm, add_rsqrt, eltwise_mul_scalar LLKs to experimental PR 52709
• Migrate mochi model to tiered ci PR 53002
• MoE prefill: op fusion, TP-sharded residual, and a gate-bias bug fix PR 52698
• #52838: Migrate Z-Image-Turbo to the tiered Models CI pipelines PR 53104
• #47382: reconfig_data_format — replace is_tile_dim_reconfig_en with intent-named APIs PR 52253
• MiniMax-H3: run on a 4x32 Blackhole quad galaxy PR 53086
• sampling: TTSampling hands ttnn.topk an indices_tensor half the width of the logits PR 53001
• RTDETRv1/v2 Bounty PR 50680
• DiffusionGemma block-diffusion serving runtime PR 52955
• Kimi-K3 prefill: LatentMoE bring-up (896 experts / top-16, 3584 latent) PR 52453
• #52838: Migrate Motif-Image-6B to the tiered Models CI pipelines PR 53173
• Command-lifetime semaphore bank and demo numbers PR 53050
• kimi27 quad runner/producer addition in blaze prefill tests PR 52192
• DeepSeek V4 HCA: functional prefill PR 52008
• group_norm: synthesize per-group masks in the kernel PR 48640
• prefill: drop the runner self-test / KV validation PR 52214
• Remove TT_METAL_QSR_TILIZE_UNPACK_TO_DEST PR 53339
• Migrate BGE-M3 to the tiered Models CI as tier 2, and retire (Single-card) Demo tests PR 53042

TT-STL

• Preallocate vector capacity and move vectors into structured return values PR 51988

Infrastructure & CI

• Batch lead-model sweep jobs by device key: 71 -> 0 test failures on Galaxy PR 51817
• fix(silencer): report-then-scan every run, not report-or-scan [skip ci] PR 52140
• Use path relative to GITHUB_ACTION_PATH when run scripts PR 52139
• #1358: Use dedicated write-only PAT for triage issue creation PR 52268
• Kimi prefill runner CI: opt out of the dispatch-hang triage PR 52070
• Update Blaze pipeline DSv3 weights paths on Exabox PR 52075
• Fix silent clang-tidy PR-comment drop for violations with no auto-fix PR 52550
• Remove old wan2.2 integration test component PR 52739
• Configure the fabric on the generic device path PR 52320
• Bug #48267 Pair printed shards with their own coordinates in to_string PR 52708
• Make runtime_noc_debugging run and pass in CI PR 52672
• Disambiguate in-place wheel artifact name by Python version PR 52971
• Qwen3.6-27B: Added bucketing fix and TP=8 support PR 53062
• Restrict Blackhole DeepSeek HCA end-to-end tests PR 53511
• Add LTX 2.3 Blackhole T2V and I2V tests to CI PR 51767
• add WAN rope/attention to models CI, remove all wan tests from t3k. PR 52262
• Replace repeat2 with ci-eval-1 in Llama 3.3-70B galaxy e2e tests PR 52474
• Adding support for ops pytests PR 52760
• Retune MiniMax-M3 prefill perf gate to 4500 +/- 7% PR 52722
• Remove passing tests from ttsim skip list PR 52871
• Add Wan2.2 Models to QB2 PR 52772
• Llama 8 New MGD PR 53048
• prune the reduce test grid and re-enable test_reduction.py on ttsim PR 53360
• repo-assist: gate confident bug root-cause claims behind verification PR 51791
• #47500: Traceable chunked prefill — trace layer + on-device MoE padding config PR 51624
• SP-only merged KVPE+indexer KV chunk address table PR 51917
• Collapse fabric stability intensity forks into one YAML (gtest-style name filters) PR 50270
• Collapse BH-glx torus/superpod YAMLs into test_fabric_2d_torus_stability.yaml PR 50282
• Enable sc16 deepseek blitz tests PR 51980
• ttop: call the await loop via $GITHUB_ACTION_PATH, not a nested uses PR 52028
• Migrate vLLM nightly to "vLLM Model Tests" (tiered-format, single cadence) PR 50910
• chore: update gh-aw agentic workflows to v0.84.4 [skip ci] PR 51888
• Switch nightly to run from standalone plugin PR 48919
• repeat: codegen port, device-2.0 re-verification PR 49888
• ttnn: add experimental RGB→YUV 4:2:0 conversion op PR 50733
• Carry CI failure grouping state in an artifact instead of GitHub issues PR 51841
• Add MiniMax-M3 Blaze chunked-prefill throughput regression PR 51574
• Qwen3-32B Blackhole Galaxy: prefill unit tests + full-model e2e accuracy PR 50985
• Fix publish-release-image: dynamic tool list + stop bypassing the proxy for Harbor PR 52204
• Feature: Test owner reassignment and querying PR 49945
• Add high-bandwidth all-gather for sparse MLA prefill PR 51134
• Fix Silencer docs: JIT/device-kernel warning evidence lives in test-execution logs, not the host build PR 52216
• Cap OpenMP host threads and use passive wait in ttnn-sanity-tests PR 50525
• Move CCL tests from end-to-end pipelines to L2 nightly PR 52238
• Migrate model traces dataset from Neon to Snowflake PR 48687
• Increase timeout on kimi_chunked_padded traced 15 accuracy test PR 52294
• remove t3k wan performance tests PR 52254
• Reassign test owner_id per 2026-08-06 escalation (#52275) PR 52281
• Strip cmd comments in prepare_test_matrix.py to fix apostrophe/quote-splice bug PR 52341
• Bug fix - Rebind library paths for TTML in CI PR 51993
• Phase 1: Set release testing to use models e2e test configuration PR 52164
• Extend release-models image to include migration-layer files PR 52303
• Update CI ownership and clarify CCL pipelines PR 52353
• Use new exabox runners for sc4 and sc16 deepseek tests PR 51666
• Fix #22943: Eltwise binary + broadcast init cleanup — migrate all call sites PR 50745
• Add metalium-developers-infra as codeowners of METALIUM_GUIDE.md PR 52465
• Fix build-wrapper.yaml fetch-depth type error on workflow_dispatch PR 52386
• Bring up Gemma-2 text models (2B / 9B) on tt_transformers PR 52471
• ci: route T3K sanity fast tests to CIv2 (wh_llmbox_civ2) PR 52313
• Remove Wormhole cards from tt-train CI runs PR 52357
• Enable export TT_METAL_SHM_TRACKING_DISABLED flag to speed up audio decode in the ci PR 52263
• Fix CI path to model weights so they don't get downloaded from HF PR 52438
• ci: surface clang-tidy violations, not just a count, in the failing step PR 52498
• Add nightly workflow to close inactive PRs after 32 days PR 52343
• Restrict stale-PR workflow to Monday-Thursday PR 52546
• ci: skip Tracy/emscripten build overhead where the profiler isn't needed PR 52552
• Add validate_program_args config (skip Metal 2.0 host-side arg validation off-CI) PR 50945
• Migrate Mistral-Small-3.1-24B and ViT to the tier 2 Models CI pipelines PR 52299
• Remove 11 small vision/CNN models from CI, retire single-card perf-models pipeline PR 52318
• Copilot review: comments must explain why, not what changed PR 52347
• Scope ttnn-core CODEOWNERS to core tests, not op tests PR 52736
• Fix curl URL-globbing failure on bot PR authors in codeowners ping PR 52746
• MINFRA-1082: Add distinct InfraError buckets for common merge-gate failure signatures [skip ci] PR 51370
• Auto-mapper support for Disagg Prefill Decode PR 43175
• Close stale PRs: upgrade actions/stale to v11, scan oldest-first, raise operations budget PR 52729
• Add KDA reference semantics and test utilities PR 52781
• Kimi-K3 MLA: NoPE, output gate, 96 heads (MLA layer only, random weights) PR 52068
• Grant actions: write so the stale-PR state cache can refresh PR 52765
• Use new exabox runners and remove old ones PR 52803
• Remove ownership requirement for .github/deprecations.json PR 52848
• Clean up/document TT-DiT test organization PR 52504
• Stop ttnn-core owning ttnn/cpp/ttnn/kernel/ PR 52846
• Graduate runtime tensor out of experimental PR 51640
• Generate and upload loss plot summaries to CI PR 51051
• Qwen3.6-27B: Added bucketing fix and TP=8 support PR 51556
• Vm/trigger prefill worker build on dispatch PR 52463
• Make setup multihost action PR 52907
• Add explicit CODEOWNERS for the LLK perf infrastructure PR 52821
• Widen MiniMax-M3 prefill perf gate margin to 15% PR 52937
• Fix unpack and install for multihost envs PR 52945
• Bug fix: honor rank>4 untilize_with_unpadding output_tensor_end PR 46939
• #52838: Migrate Stable Diffusion 3.5 Large to the tiered Models CI pipelines PR 52892
• Disable per-op hang detection for fabric stability tests PR 52368
• Run test owner escalation weekly instead of daily PR 52967
• Remove DeepSeek prefill tests from T3K e2e pipeline PR 53000
• ci: use system SFPI toolchain in CI-internal builds to avoid redundant network downloads PR 52999
• Upgrade mold to v2.42.0 PR 53014
• Migrate 7 models to the tiered Models CI — 6 as tier 3 (N150), ResNet-50 as tier 1 off T3K / Galaxy PR 52829
• Let wget show status when getting sfpi PR 53051
• Enable Qwen2.5-Coder-32B on Blackhole QuietBox 2 PR 52880
• Thread context_id into ProgramImpl and its downstream PR 52771
• ci: fix LLK nightly pipelines PR 53045
• Pipeline reorg: Migrate Blackhole Sanity into sanity-tests PR 48943
• naming fix: Make Sanity workflow run names consistent PR 53074
• Require PR justification for CI time-budget changes and remove stale pool data PR 53081
• Uniform passing convention for MeshCommandQueue in Runtime API PR 51559
• Remove unmaintained OFT model PR 52919
• Emit Parquet alongside CSVs in perf artifacts PR 53020
• Stop convolutions owning experimental/CMakeLists.txt PR 53136
• Docs: add AI-IP 0.76.0 release documentation PR 53102
• Port untilize (build_untilize_tile) — validated on wormhole_b0 + blackhole PR 50178
• CODEOWNERS: add @mtairum to models/common/sampling/ PR 53165
• Revise release notes for v0.76.0 PR 53168
• Migrate mochi model to tiered ci PR 53002
• Update AI IP v0.76.0 release notes PR 53177
• Revise release notes for v0.75.0 and v0.76.0 PR 53178
• #52838: Migrate Z-Image-Turbo to the tiered Models CI pipelines PR 53104
• GRPO example: reverse text (single p150, Qwen3-0.6B) PR 53003
• #47382: reconfig_data_format — replace is_tile_dim_reconfig_en with intent-named APIs PR 52253
• Update Quasar program versions in documentation PR 53195
• Add vllm-tt-plugin host tests to vLLM models CI. PR 53155
• Kimi-K3 prefill: LatentMoE bring-up (896 experts / top-16, 3584 latent) PR 52453
• Move all gh-aw agentic workflows to v0.86.2 PR 53316
• Remove prefill-team tests from demo SP release pipeline PR 53265
• llk-audit: deterministic recall augmentor for the race audits PR 49657
• fix(ci): check-harbor fallback doesn't fire when caller is schedule-triggered PR 53286
• #52838: Migrate Motif-Image-6B to the tiered Models CI pipelines PR 53173
• kimi27 quad runner/producer addition in blaze prefill tests PR 52192
• ci: fix compile-time-tracker push auth broken by actions/checkout v6 PR 53385
• DeepSeek V4 HCA: functional prefill PR 52008
• Split codeowners-group-analysis: extract owner-selection into a Python script PR 52410
• trigger quasar build on fuser changes PR 52069
• Migrate BGE-M3 to the tiered Models CI as tier 2, and retire (Single-card) Demo tests PR 53042
• prefill: move sc4 KV-PCC leg budget to shield team PR 53491
• Update owner_id in models_e2e_tests.yaml PR 53536
• emule: add wormhole+blackhole smoke-test script PR 49636
• Add sigmoid-gated RMSNorm PR 52783
• add jonathansuTT to /models/ codeowners PR 53449

Documentation

• Fix #22943: Eltwise binary + broadcast init cleanup — migrate all call sites PR 50745
• Fix SFPU RNG correlation and improve FP32 uniform random quality PR 52024
• Remove data_types.hpp from the tt-metalium API surface PR 52977
• #47382: reconfig_data_format — replace is_tile_dim_reconfig_en with intent-named APIs PR 52253
• Add softcap unary op and situ_glu binary activation (Kimi K3 SiTU-GLU) PR 52448
• docs: remove stale ttnn demos page PR 53450

Tooling

• #0 - Telemetry port selection added in health check PR 52283
• Improve callstack aggregation PR 51733
• Fix protobufs visibility issues by avoiding exporting scaleout_tools symbols PR 52259
• Collapse BH-glx torus/superpod YAMLs into test_fabric_2d_torus_stability.yaml PR 50282
• DCAMP-1176: unify test scripts between k8s and bare metal slurm PR 52062
• Fix #22943: Eltwise binary + broadcast init cleanup — migrate all call sites PR 50745
• Feat: Query Class For FSD Hierarchy Fields PR 50947
• tools/scaleout: split source-file lists into sources.cmake PR 52385
• Auto-mapper support for Disagg Prefill Decode PR 43175
• Add concise Missing Links summary to validation analysis. PR 52345
• Migrate 7 models to the tiered Models CI — 6 as tier 3 (N150), ResNet-50 as tier 1 off T3K / Galaxy PR 52829
• tools/tracy: --process-logs-only resolving the .logs folder twice PR 52159
• Pipeline reorg: Migrate Blackhole Sanity into sanity-tests PR 48943
• Update tt-exalens version to 0.3.30 PR 53505

Other

• Fix moe_routing_remap expert_parallel_size vs mesh-axis validation PR 51686
• Batch lead-model sweep jobs by device key: 71 -> 0 test failures on Galaxy PR 51817
• RingJointSDPA: restore non-sliding multi-batch support PR 52073
• sdpa perf checks: scope the is_high_power gate to the galaxy mesh PR 52190
• Support ND-sharding in ttnn.all_gather PR 51515
• Configure the fabric on the generic device path PR 52320
• Prevent Galaxy CCL test hangs by using the full mesh shape PR 52731
• split_model_traced: use the traced split_size, and compare every chunk PR 52792
• Fix GPT-OSS first-group chunked sliding RingJointSDPA PR 53153
• sweeps: replay stale traced topk vectors without their indices tensor PR 53161
• Fix Ring SDPA LLK assert program size and FP32 hangs PR 51963
• Skip Blackhole ring-joint SDPA matrix in CI PR 52658
• re-enable skipped CI test for reduce scatter async PR 52809
• Add support for sharded joint on the sequence dimension to ring_joint_scaled_dot_product_sdpa PR 48677
• Ignore every sweep export directory with one rule PR 52143
• Support 1D Reduce Scatter on reshaped mesh device PR 50777
• Migrate model traces dataset from Neon to Snowflake PR 48687
• Update CONTRIBUTING.md: contribution flow, reorg, remove stale process docs PR 52250
• Fix #22943: Eltwise binary + broadcast init cleanup — migrate all call sites PR 50745
• exp_ring_joint_sdpa: replace get_dynamic_runtime_args with a program-factory override PR 52554
• Migrate Mistral-Small-3.1-24B and ViT to the tier 2 Models CI pipelines PR 52299
• Remove 11 small vision/CNN models from CI, retire single-card perf-models pipeline PR 52318
• Fused CCL + matmul: fabric-bound minimal-matmul, strided AGMM and MMRS PR 52513
• Kimi-K3 MLA: NoPE, output gate, 96 heads (MLA layer only, random weights) PR 52068
• #47644: Extend universal input/output support for ttnn::fold PR 50385
• reduce_scatter "direct" algorithm for small shapes PR 51741
• Graduate runtime tensor out of experimental PR 51640
• Neighbor-pad halo exchange and halo-mode conv3d PR 52514
• Ring joint SDPA and the LTX-2.3 distilled AV pipeline PR 52730
• ci: use system SFPI toolchain in CI-internal builds to avoid redundant network downloads PR 52999
• Migrate 7 models to the tiered Models CI — 6 as tier 3 (N150), ResNet-50 as tier 1 off T3K / Galaxy PR 52829
• add support fused/virtual concat to matmul variants (except agmm) PR 51802
• Pipeline reorg: Migrate Blackhole Sanity into sanity-tests PR 48943
• third_party: upgrade nanobind pin 2.12.0 -> 2.14.0 PR 53057
• Optimize generic all-to-all packet scheduling PR 52674
• Move all gh-aw agentic workflows to v0.86.2 PR 53316
• add fused swiglu to strided agmm PR 53203
• llk-audit: deterministic recall augmentor for the race audits PR 49657
• Allow overriding x86 -march for Ivy Bridge hosts PR 49041
• Migrate BGE-M3 to the tiered Models CI as tier 2, and retire (Single-card) Demo tests PR 53042