Add fused minimal matmul addcmul operation

[nmauriceTT]

Ticket

#35915

Problem description

For Wan2.2, we want to fuse minimal_matmul and addcmul, like the following pattern:

spatial_ff_1BND = self.ff(spatial_normed_1BND, compute_kernel_config=self.ff_compute_kernel_config)
spatial_1BND = ttnn.addcmul(spatial_1BND, spatial_ff_1BND, c_gate_msa_1B1D)

What's changed

This adds a new dit_minimal_matmul_addcmul_fused operation that perform minimal_matmul + addcmul operation.
This operation is equivalent to:

matmul_output = minimal_matmul(tensor_input0, tensor_input1, tensor_bias)
output = addcmul(ternary_a, minimal_matmul, ternary_b, scalar) # ternary_a + scalar * minimal_matmul * ternary_b

To make minimal_matmul more extensible, I've also merged its device operation with that of minimal_matmul_split (i.e. reduce code duplication).

The kernels of dit_minimal_matmul_addcmul_fused have been implemented by modifying the minimal_matmul kernels. It is also defined by calling minimal_matmul (device operation of minimal_matmul has been updated with new parameters).

Note: Ideally, we'd like to use the addcmul_tile LLK. But it seems that unary_bcast<BroadcastType::ROW> does not work with fp32_acc_to_dst. Instead, row-broadcast of ternary_ais done throughadd_bcast_tile` (FPU). The downside is that the output should be less accurate than with addcmul.
If accuracy turns out to be a problem, then we can switch to other workaround (e.g. do broadcasting in dataflow kernels).

Performance

wan2.2_14b-720p-glx: Single GLX
(M, K, N) = (9472, 1280, 5120)

Name	Duration [ms]	Std	Gains	Speed-up
Separate	2.377 ms	0.016	N/A	N/A
Fused	2.080	0.005	-0.30	8% faster

wan2.2_14b-720p Quad GLX
(M, K, N) = (2368, 1280, 5120)

Name	Duration [ms]	Std	Gains [ms]	Speed-up
Separate	0.835	0.006	N/A	N/A
Fused	0.776	0.003	-0.059	7% faster

As a reference, here's the execution time of addcmul.

Shape	Name	Duration [ms]
Single GLX	addcmul	0.463
Quad GLX	addcmul	0.124

Fusing the operations saves us ~50-65% of the execution time addcmul.

Checklist

• 
• (failures from main)
• (unrelated errors)
• New/Existing tests provide coverage for changes
• strided_all_gather_minimal_matmul_async runs fine

Model tests

If your changes cover model-related code, you should run tests corresponding to affected models and platforms (Single card, T3K, Galaxy). "Choose your pipeline" workflows facilitate running multiple kinds of tests in a single run. Each offers models-mandatory and models-extended presets.
The former includes a minimal set of tests, to be run always. The latter extends that with additional ones - use your best judgement in deciding which is the most appropriate for your PR.

• 

  • models-mandatory preset (runs: Device perf regressions and Frequent model and ttnn tests) : Device Perf (same failures as main)
  • models-extended preset (runs: the mandatory tests, plus Demo and Model perf tests)
  • other selection - specify runs

• (failures also present on main, flux1 validated here)

  • models-mandatory preset (runs: Unit tests)
  • models-extended preset (runs: the mandatory tests, plus Demo and Model perf tests)
  • other selection - specify runs

• 

  • models-mandatory preset (runs: Quick tests)
  • models-extended preset (runs: the mandatory tests, plus Demo and Model perf tests)
  • other selection - specify runs

[Copilot]

Pull request overview

This PR introduces a fused dit_minimal_matmul_addcmul_fused operation that combines minimal_matmul and addcmul for improved performance in DiT transformer blocks (specifically targeting Wan2.2). The PR also refactors the minimal_matmul device operations by unifying the previously separate minimal_matmul and minimal_matmul_split implementations into a single device operation that supports both single and chunked outputs.

Changes:

• Added new fused operation dit_minimal_matmul_addcmul_fused that computes output = residual + scalar * matmul(input, weight) * gate
• Unified minimal_matmul and minimal_matmul_split device operations, changing return types from Tensor to std::vector<Tensor>
• Extended kernels to support fused ternary (addcmul) operations with new circular buffers and runtime parameters

Reviewed changes

Copilot reviewed 28 out of 28 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
dit_minimal_matmul_addcmul_fused/*	New operation implementation with nanobind bindings and comprehensive documentation
minimal_matmul_device_operation.*	Unified device operation supporting both single and split outputs, added ternary fusion parameters
minimal_matmul_program_factory.*	Added circular buffers for ternary inputs, extended runtime argument handling
minimal_matmul.cpp, minimal_matmul_split.cpp	Updated to use unified device operation returning vector of tensors
kernels/compute.cpp	Added add_bias_and_addcmul_block function implementing fused bias and addcmul logic
kernels/dm_in*.cpp	Extended dataflow kernels to read and process ternary input tensors
kernels/matmul_dataflow_common.hpp	Added read_ternary_blocks_sync helper for reading ternary tensors
minimal_matmul_split_* (deleted)	Removed duplicate device operation files now unified with minimal_matmul
CMakeLists.txt	Updated build configuration to remove split-specific files and add new fused operation
test_dit_minimal_matmul_addcmul_fused.py	Comprehensive tests covering basic functionality, Wan2.2 shapes, and different scalar values

[Copilot]

Pull request overview

Copilot reviewed 27 out of 27 changed files in this pull request and generated 5 comments.

[nmauriceTT]

/codeowners ping

[tenstorrent-github-bot]

CodeOwners Group Analysis

This PR requires approval from one member of each of the following groups:

Summary: 1 pending groups, 7 approved groups

Group Information:

• ⏳ tenstorrent/metalium-developers-ops-data-movement (Team) - Members: Naif Tarafdar, Saad Jameel, Juan Camilo Vega, Ligang Long, Nour Ardo, Adrian Morrison, Ilia Taraban, Sheran Cardoza | Pending approval

  📁 Files owned by this team (2 files)

  • ttnn/cpp/ttnn/operations/experimental/ccl/strided_all_gather_minimal_matmul_async/device/strided_all_gather_minimal_matmul_async_op.cpp
  • ttnn/cpp/ttnn/operations/experimental/ccl/strided_all_gather_minimal_matmul_async/device/strided_all_gather_minimal_matmul_async_program.cpp

---

• ✅ tenstorrent/metalium-developers-infra (Team) - Members: Raymond Kim, Michael Chiou, Bryan Keith, Bryan Wilder Field Lozano, Andrew Fuller, William Ly, Kannika Kabilar, Anthony Kirby, Rose Li, Subin Lee, Evan Banerjee, NSexton, David Popov, Aditi Rajesh Shah, Jacek Jakub Lakis, Iris Wang, jessica yuan, Hasan Baig | Approved by: Evan Banerjee, NSexton, Colman Glagovich (shared)

  📁 Files owned by this team (3 files)

  • ttnn/CMakeLists.txt
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/CMakeLists.txt
  • ttnn/cpp/ttnn/operations/experimental/transformer/CMakeLists.txt

---

• ✅ tenstorrent/metalium-developers-mmfusedreduce (Team) - Members: Borys Bradel, Vasisht Suresh, Edwin Lee, Nikhil Soraba, Ryan Miller, Izajasz Wrosz | Approved by: Borys Bradel

  📁 Files owned by this team (1 files)

  • tests/ttnn/docs_examples/test_matrix_multiplication_examples.py

---

• ✅ tenstorrent/metalium-developers-ttnn-core (Team) - Members: Pavlo Hilei, Brian Liu, Joseph Chu, Artem Yerofieiev, Diego Gomez | Approved by: Evan Banerjee, NSexton, Borys Bradel, NSexton (shared)

  📁 Files owned by this team (8 files)

  • tests/ttnn/nightly/unit_tests/operations/experimental/test_dit_minimal_matmul_addcmul_fused.py
  • ttnn/CMakeLists.txt
  • ttnn/cpp/ttnn/operations/experimental/experimental_nanobind.cpp
  • ttnn/cpp/ttnn/operations/experimental/transformer/CMakeLists.txt
  • ttnn/cpp/ttnn/operations/experimental/transformer/dit_minimal_matmul_addcmul_fused/dit_minimal_matmul_addcmul_fused.cpp
  • ttnn/cpp/ttnn/operations/experimental/transformer/dit_minimal_matmul_addcmul_fused/dit_minimal_matmul_addcmul_fused.hpp
  • ttnn/cpp/ttnn/operations/experimental/transformer/dit_minimal_matmul_addcmul_fused/dit_minimal_matmul_addcmul_fused_nanobind.cpp
  • ttnn/cpp/ttnn/operations/experimental/transformer/dit_minimal_matmul_addcmul_fused/dit_minimal_matmul_addcmul_fused_nanobind.hpp

---

• ✅ tests/ttnn/ (Group) - Members: Borys Bradel, Dongjin Na, Jaehoon Jung | Approved by: Borys Bradel

  📁 Files owned by this group (1 files)

  • tests/ttnn/nightly/unit_tests/operations/experimental/test_dit_minimal_matmul_addcmul_fused.py

---

• ✅ ttnn//nanobind** (Group) - Members: NSexton | Approved by: NSexton

  📁 Files owned by this group (3 files)

  • ttnn/cpp/ttnn/operations/experimental/experimental_nanobind.cpp
  • ttnn/cpp/ttnn/operations/experimental/transformer/dit_minimal_matmul_addcmul_fused/dit_minimal_matmul_addcmul_fused_nanobind.cpp
  • ttnn/cpp/ttnn/operations/experimental/transformer/dit_minimal_matmul_addcmul_fused/dit_minimal_matmul_addcmul_fused_nanobind.hpp

---

• ✅ ttnn/cpp/ttnn/operations/experimental/minimal_matmul//CMakeLists.txt** (Group) - Members: Colman Glagovich, Jonathan Su | Approved by: Colman Glagovich, Evan Banerjee, NSexton (shared)

  📁 Files owned by this group (1 files)

  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/CMakeLists.txt

---

• ✅ ttnn/cpp/ttnn/operations/experimental/minimal_matmul/ (Group) - Members: Colman Glagovich, Jonathan Su | Approved by: Colman Glagovich

  📁 Files owned by this group (16 files)

  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/kernels/compute.cpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/kernels/dm_in0_sender.cpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/kernels/dm_in1_sender_out.cpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/kernels/matmul_dataflow_common.hpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/minimal_matmul_device_operation.cpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/minimal_matmul_device_operation.hpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/minimal_matmul_device_operation_types.hpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/minimal_matmul_program_factory.cpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/minimal_matmul_program_factory.hpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/minimal_matmul_split_device_operation.cpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/minimal_matmul_split_device_operation.hpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/minimal_matmul_split_program_factory.cpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/device/minimal_matmul_split_program_factory.hpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/minimal_matmul.cpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/minimal_matmul_split.cpp
  • ttnn/cpp/ttnn/operations/experimental/minimal_matmul/minimal_matmul_split.hpp

Note: At least one approval from each group is sufficient.

[tenstorrent-github-bot]

Hi Borys Bradel (@bbradelTT), Colman Glagovich (@cglagovichTT), Edwin Lee (@edwinleeTT), Izajasz Wrosz (@iwroszTT), Jonathan Su (@jonathansuTT), NSexton (@nsextonTT), this PR Add fused minimal matmul addcmul operation by Nathan Maurice (@nmauriceTT) needs your approval/review to merge this.

[github-actions]

⚠️ Clang-Tidy found issue(s) with the introduced code (1/1)