v0.44.0

📦 Uncategorized

• Update CreateBuffer to return shared_ptr, and Enqueue R/W buffer to accept std::shared_ptr
  • PR: #5154
• #4794: Implement DownBlock2D using ttnn for stable_diffusion model
  • PR: #5091
• #4797: Implement BasicTransformerBlock sub-module using ttnn for stab…
  • PR: #5081
• #0: write cluster config for FD mode, non tunneling cores as well
  • PR: #5161
• Update bw test, change mulsi calls to use *
  • PR: #5149
• #3003: updated tt-lib documentation
  • PR: #5179
• #0: Update to v0.44.0
  • PR: #5188
• #4003: added ability to trace ttnn operations using torchtrail library
  • PR: #5135
• Support moreh logsoftmax
  • PR: #4961
• #4614: gitmodules: Use https URLs for submodules
  • PR: #5183
• #0: add reviewers to frequently touched ops docs file
  • PR: #5190
• backward ops - hypot and atan2
  • PR: #5045
• #4885: Move program device map to program
  • PR: #5193
• #4858: Add support for float to int typecast
  • PR: #5058
• Matmul_block on a smaller grid size
  • PR: #5170
• Revert "#0: Add support for typecast float to int"
  • PR: #5199
• Add dst ethernet router support and remote command processor to accept FD packets on remote chip
  • PR: #5102
• Falcon40B TT Implementation
  • PR: #5046
• #5198: Fix moreh softmax related bug
  • PR: #5200
• #0: skip MOREH Softmax tests from main
  • PR: #5202
• #3122: Use device grid size in falcon_attention to be genereric...
  • PR: #5207
• #0: Add assertions for interleaved tensors for ops that don't support sharding
  • PR: #5195
• #5169: Add activation ops to ttnn
  • PR: #5217
• #3003: add duration to the ttnn operation nodes when TTNN_ENABLE_LOGGING=1 is used to compile the code
  • PR: #5201
• #5027: Optimize group attn matmul for Falcon40B decode
  • PR: #5127
• #0: add documentation about managing documentation
  • PR: #5227
• Adding docs for maxpool, avg pool and upsample
  • PR: #5223
• Revert "#0: skip MOREH Softmax tests from d5811b7…
  • PR: #5228
• #5165: Add hyperbolic ops to ttnn
  • PR: #5166
• #4866: Add grayskull open source llk-library
  • PR: #5136
• #5002: simplified preprocessing of CNNs using preprocess_model
  • PR: #5181
• Create GroupNorm sharded in TTNN
  • PR: #5221
• #5097: Support for dedicated completion queue thread
  • PR: #5098
• upsample test calculate grid
  • PR: #5238
• fix for sharded allocater when num banks == num cores
  • PR: #5229
• MHA tutorial interactive notebook with diagrams
  • PR: #5239
• #4003: Adding a profile tutorial
  • PR: #5242
• #0: Added non-blocking read stress test
  • PR: #5243
• Revert "MHA tutorial interactive notebook with diagrams"
  • PR: #5245
• #0: Update all_gather to work for multi_link. Update falcon-40b to use 2 links for all gathers
  • PR: #5214
• #5142: Remove slow dispatch mode from workgin sweeps
  • PR: #5146
• #3003: fixed the input tensor documentation
  • PR: #5255
• #0: Temp slower resnet VM run
  • PR: #5256
• throw on fast dispatch for to_host_sharded as its not supported
  • PR: #5264
• #5253: Fix kv_past_len being passed in to rotary embedding for falcon models
  • PR: #5254
• #5233: started adding ttnn_functional_resnet
  • PR: #5240
• #3003: updated ttnn documentation to explain what features it has over tt_lib. Added standalone examples of basic usage of ttnn
  • PR: #5265
• #0: Speedup incremental builds
  • PR: #5251
• #0: Change setup.py to be git worktree friendly
  • PR: #5234
• MHA tutorial interactive notebook with diagrams
  • PR: #5277
• #3003: disable tutorial 6 from running as the unit test
  • PR: #5278
• Agrebenisan/non blocking tensor reads
  • PR: #5244
• #5275: CODEOWNERS: update to include files relevant for ttnn team
  • PR: #5276
• Fix an intermittent launch message transfer error
  • PR: #5152
• Revert "MHA tutorial interactive notebook with diagrams"
  • PR: #5282
• #0: add parens in LLK doc
  • PR: #5283
• #3003: only unit test tutorials that work on pipelines
  • PR: #5291
• #5246: Add unary math ops to ttnn
  • PR: #5259
• Vignesh/stable diffusion ttnn basic transformer block fix
  • PR: #5211
• #4854: Implement attention and rms_norm sub-module using ttnn for mis…
  • PR: #5175
• #4795: Add upblock2d to functional stable diffusion model
  • PR: #5085
• #4796: Implement Transformer2DModel using ttnn for stable_diffusion m…
  • PR: #5092
• #0: Adding llk wormhole_b0 submodule
  • PR: #5262
• #4003: Adding pyind11 to ttnn
  • PR: #5236
• #5296: Fix broken link to host_api.hpp in README.md
  • PR: #5297
• #0: Fix bug with the way we were measuring bert inference time
  • PR: #5312
• #0: Change local tt_lib._C module install from symlink to copy
  • PR: #5292
• #5233: added ability to fold batch_norm2d into conv2d
  • PR: #5317
• #5222: replace hex8_to_hex32.py with cpp to shave off some compile time -temporary fix
  • PR: #5220
• Enable tests for WHB0
  • PR: #5307
• #5137: Cleanups for newer Linux distro / toolchains
  • PR: #5162
• #5233: implemented support for converting all Resnet-18 modules using preprocess_model function
  • PR: #5325
• #3003: fix model preprocessing bug
  • PR: #5332
• #4799: Implement CrossAttnDownBlock2D sub-module using ttnn for stabl…
  • PR: #5086
• #4800: Implement UNetMidBlock2DCrossAttn using ttnn for stable_diffus…
  • PR: #5093
• #4798: Add ttnn cross attn upblock2d in functional stable diffusion m…
  • PR: #5089
• #4801: Implement Unet 2D Condition model using ttnn for stable_diffus…
  • PR: #5119
• #4965: Rename Conv2D to Conv2d and MaxPool2D to MaxPool2d to match torch
  • PR: #5219
• #0: Remove departed team member from CODEOWNERS
  • PR: #5340
• #0: add to codeowners
  • PR: #5339
• #5314: Only stall on first scheduled read after commands with side effects
  • PR: #5315
• #4965: fix bad rebase
  • PR: #5342
• #0: Add more instructions for dispatching workflow actions and a note about skipping git hooks
  • PR: #5345
• Update optimized Bert to support WH grid sizes, add sharding support for RMSNorm
  • PR: #5308
• #4642: create gtest_smoke as a sanity test suit
  • PR: #5112
• #5341: context switch if eth txq is full
  • PR: #5347
• #5323: Convolutions of small size fail during parallelization calculations
  • PR: #5324
• Npetrovic/transformer softmax
  • PR: #5298
• Fix groupnorm for narrow channels
  • PR: #5320
• #4862: added more test for ttnn bloom. Update optimized ttnn bert to match the structure of non-optimized ttnn bert
  • PR: #5336
• #0: Add an envvar parser with value detection and default value setti…
  • PR: #5367
• #4732: Clean up compute kernel apis
  • PR: #5316
• #5318: Modify Falcon7B to use attn_matmul for wormhole
  • PR: #5322
• #0: make logLocationsRecord a static function
  • PR: #5351
• #5233: run convs with auto-format
  • PR: #5364
• #5377: Avoid segfault by checking buffer !null before getting device
  • PR: #5381
• Alex/metal/pack untilize b0
  • PR: #5378
• #4487: Support block sharding in upsample
  • PR: #5361
• #5359: update python package transformers + dependencies to include Falcon
  • PR: #5360
• #3708: Add support for LN having gamma/beta in bfp8
  • PR: #5376
• #4003: Skip sweep tests if not available
  • PR: #5392
• #4003: use faster TMs in optimized ttnn whisper
  • PR: #5384
• #4732: Clean up compute_kernel_api
  • PR: #5375
• More optimizations for group_attn_matmul
  • PR: #5385
• #5233: updated resnet18 to run residual connections
  • PR: #5390
• #3003: added more meaningful errors to ttnn. Updated getitem to run on device in the cases when it can
  • PR: #5403
• #5233: simplified the logic in tracer
  • PR: #5370
• #3003: include ttl operations and necessary types under ttnn.ttl
  • PR: #5405
• #0: Add note about no merge commits in main
  • PR: #5349
• #0: Add timeout in profiler regression workflow
  • PR: #5355
• codeowners update
  • PR: #5407
• #5365: Add device argument to determine grid size based on target
  • PR: #5366
• disable whisper until further investigation, see issue #5430
  • PR: #5431
• #3003: fixed ttnn convs
  • PR: #5432
• #3886: Fix build error for C++ tests in debug mode
  • PR: #5434
• #4954: Support depth 32 in maxpool writer
  • PR: #4956
• #0: Pass output cb to pack init functions
  • PR: #5418
• #0: skipping DeviceLoadBlankKernels on remote devices
  • PR: #5437
• #5359: transformers: update version and relax pcc asserts
  • PR: #5421
• #3003: guidelines for adding new op
  • PR: #5440
• Don't assume user has one entry in their $PYTHONPATH
  • PR: #5250
• FP32 tensor support for matmul
  • PR: #5414
• #3003: updated tutorial 001 to describe the tensor more comprehensively before showing the add
  • PR: #5441
• Onboard additional metal code owners
  • PR: #5445
• #5402: Add redesigned host-side sw command queue, it can be configured i…
  • PR: #5382
• #3003: fixed docs
  • PR: #5455
• Alex/metal/enable conv tests on b0
  • PR: #5425
• #5356: git bisect script to find broken commits
  • PR: #5348
• #0: Update data_format.cpp file
  • PR: #5399
• Add skip to full grid matmul whb0
  • PR: #5461
• #3003: simplified the logic in ttnn/operations/matmul.py. Added dataclasses instead of tuples for CoreGrid and ShardShape
  • PR: #5450
• #5204: adding moreh's test suit. removing an absolute assertion.
  • PR: #5373
• Npetrovic/lt gt ne fix
  • PR: #5304
• #0: Move device id attribute from tensor to DeviceStorage
  • PR: #5467
• #3003: fixed scheduled pipeline
  • PR: #5466
• Npetrovic/transformer concat sweeps ttnn
  • PR: #5208
• #3003: added support for running ttnn.matmul using 1D_systolic_array. Also, added support for passsing in the program config directly
  • PR: #5468
• #5247: Add unary ops to ttnn
  • PR: #5279
• #4854: Implement feed-forward sub-module using ttnn for mistral model
  • PR: #5176
• #5326: Add tensor manip ops to ttnn
  • PR: #5410
• #4775: Add lamb optimizer sweep
  • PR: #5039
• disable model perf measurements on VMs
  • PR: #5464
• #0: Add more explicit checks for Q and KV heads sharding for group_attn_matmul
  • PR: #5481
• #4003: added manage_device decorator. Renamed ttnn.open to ttnn.open_device and ttnn.close to ttnn.close_device
  • PR: #5473
• #5442: Add math ops to ttnn
  • PR: #5465
• #5343: Add binary ops to ttnn
  • PR: #5443
• #5474: Add backward op support for expm1 and exp2
  • PR: #5476
• #3003: renamed module.(hpp/cpp) to init.(hpp/cpp). Added stubs for matmul
  • PR: #5475
• #5216: Fix broken link for developer's docs in CONTRIBUTING.md
  • PR: #5488
• #5334: Add unary math ops to ttnn
  • PR: #5404
• #5492: Assert before exiting completion queue thread to ensure users …
  • PR: #5497
• #4003: updated create_sharded_memory_config
  • PR: #5503
• #5099: Add dprint/watcher enabled to kernel hash
  • PR: #5502
• A collection of small DPrint features/fixes
  • PR: #5463
• #0: Fix rebase issue from previous commit
  • PR: #5505
• #5056: Impl moreh_groupnorm
  • PR: #5490
• #4003: fix docs
  • PR: #5507
• #5113: Added generation function which excludes ranges
  • PR: #5371
• Barsic/bw sweeps
  • PR: #5141
• #4416: Add post-commit workflow to check build for all available CONFIGs
  • PR: #5495
• #0: Update ATOL to 0.0081 for roberta
  • PR: #5493
• #3420: remove S/R EthernetConfig, move out of experimental ns
  • PR: #5470
• #4420: eager EnqueueHostToDeviceTransfer impl
  • PR: #5293
• #5446: Remove restart command and associated functionality
  • PR: #5501
• #4982: Upgrade to checkout v4 so we can use new node 20 and get rid of warnings
  • PR: #5524
• #4003: removed pad_to_tile and unpad_to_tile
  • PR: #5528
• #0: Add fail-fast: false for build-and-upload
  • PR: #5527
• #4003: changed ttnn.ttl to ttnn.experimental
  • PR: #5533
• #5531: Remove ppa registration for Git
  • PR: #5532
• #4620: Add host<->L1 bw/latency tests
  • PR: #5537
• #4443: distribute data transfer between brisc and ncrisc
  • PR: #5534