Migrate device headers from tt_metal/include to tt_metal/hw/inc (#36583)

Note: This PR is awaiting code review bypass. Do not review unless
you're explicitly requested via DM.

### Ticket
#31083

### Problem description
Device headers are scattered around tt_metal repo - reorganization is
needed.
We plan to put all device headers under `tt_metal/hw/inc`.

### What's changed

This PR moves device headers from `tt_metal/include` to
`tt_metal/hw/inc`.
This change was created programatically.

### Checklist

- [ ] [![All post-commit
tests](https://github.com/tenstorrent/tt-metal/actions/workflows/all-post-commit-workflows.yaml/badge.svg?branch=prybicki/migrate-tt-metal-include2)](https://github.com/tenstorrent/tt-metal/actions/workflows/all-post-commit-workflows.yaml?query=branch:prybicki/migrate-tt-metal-include2)
- [ ] [![Blackhole Post
commit](https://github.com/tenstorrent/tt-metal/actions/workflows/blackhole-post-commit.yaml/badge.svg?branch=prybicki/migrate-tt-metal-include2)](https://github.com/tenstorrent/tt-metal/actions/workflows/blackhole-post-commit.yaml?query=branch:prybicki/migrate-tt-metal-include2)
- [ ]
[![cpp-unit-tests](https://github.com/tenstorrent/tt-metal/actions/workflows/tt-metal-l2-nightly.yaml/badge.svg?branch=prybicki/migrate-tt-metal-include2)](https://github.com/tenstorrent/tt-metal/actions/workflows/tt-metal-l2-nightly.yaml?query=branch:prybicki/migrate-tt-metal-include2)
- [ ] New/Existing tests provide coverage for changes


#### 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.

- [ ] [![(Single) Choose your
pipeline](https://github.com/tenstorrent/tt-metal/actions/workflows/pipeline-select.yaml/badge.svg?branch=prybicki/migrate-tt-metal-include2)](https://github.com/tenstorrent/tt-metal/actions/workflows/pipeline-select.yaml?query=branch:prybicki/migrate-tt-metal-include2)
- [ ] `models-mandatory` preset (runs: [Device perf
regressions](https://github.com/tenstorrent/tt-metal/actions/workflows/perf-device-models.yaml)
and [Frequent model and ttnn
tests](https://github.com/tenstorrent/tt-metal/actions/workflows/fast-dispatch-full-regressions-and-models.yaml))
- [ ] `models-extended` preset (runs: the mandatory tests, plus
[Demo](https://github.com/tenstorrent/tt-metal/actions/workflows/single-card-demo-tests.yaml)
and [Model
perf](https://github.com/tenstorrent/tt-metal/actions/workflows/perf-models.yaml)
tests)
  - [ ] other selection - specify runs

- [ ] [![(T3K) Choose your
pipeline](https://github.com/tenstorrent/tt-metal/actions/workflows/pipeline-select-t3k.yaml/badge.svg?branch=prybicki/migrate-tt-metal-include2)](https://github.com/tenstorrent/tt-metal/actions/workflows/pipeline-select-t3k.yaml?query=branch:prybicki/migrate-tt-metal-include2)
- [ ] `models-mandatory` preset (runs: [Unit
tests](https://github.com/tenstorrent/tt-metal/actions/workflows/t3000-unit-tests.yaml))
- [ ] `models-extended` preset (runs: the mandatory tests, plus
[Demo](https://github.com/tenstorrent/tt-metal/actions/workflows/t3000-demo-tests.yaml)
and [Model
perf](https://github.com/tenstorrent/tt-metal/actions/workflows/t3000-model-perf-tests.yaml)
tests)
  - [ ] other selection - specify runs

- [ ] [![(Galaxy) Choose your
pipeline](https://github.com/tenstorrent/tt-metal/actions/workflows/pipeline-select-galaxy.yaml/badge.svg?branch=prybicki/migrate-tt-metal-include2)](https://github.com/tenstorrent/tt-metal/actions/workflows/pipeline-select-galaxy.yaml?query=branch:prybicki/migrate-tt-metal-include2)
- [ ] `models-mandatory` preset (runs: [Quick
tests](https://github.com/tenstorrent/tt-metal/actions/workflows/galaxy-quick.yaml))
- [ ] `models-extended` preset (runs: the mandatory tests, plus
[Demo](https://github.com/tenstorrent/tt-metal/actions/workflows/galaxy-demo-tests.yaml)
and [Model
perf](https://github.com/tenstorrent/tt-metal/actions/workflows/galaxy-model-perf-tests.yaml)
tests)
  - [ ] other selection - specify runs

Author: prybickiTT

.github/CODEOWNERS

@@ -87,8 +87,8 @@ tt_metal/impl/program/ @abhullar-tt @jbaumanTT @tt-asaigal @tt-aho @tenstorrent/
 tt_metal/impl/sub_device/ @abhullar-tt @jbaumanTT @tt-asaigal @tt-aho @tenstorrent/codeowner-bypass
 tt_metal/impl/trace/ @abhullar-tt @jbaumanTT @tt-asaigal @tt-aho @tenstorrent/codeowner-bypass
 tt_metal/impl/tensor/ @akerteszTT @riverwuTT @tenstorrent/codeowner-bypass
-tt_metal/include/compute_kernel_api.h @davorchap @rtawfik01 @rdjogoTT @nvelickovicTT @amahmudTT @lpremovicTT @ncvetkovicTT @fvranicTT @tenstorrent/codeowner-bypass
-tt_metal/include/compute_kernel_api/ @rtawfik01 @rdjogoTT @nvelickovicTT @amahmudTT @lpremovicTT @ncvetkovicTT @fvranicTT @tenstorrent/codeowner-bypass
+tt_metal/hw/inc/api/compute/compute_kernel_api.h @davorchap @rtawfik01 @rdjogoTT @nvelickovicTT @amahmudTT @lpremovicTT @ncvetkovicTT @fvranicTT @tenstorrent/codeowner-bypass
+tt_metal/hw/inc/api/compute/ @rtawfik01 @rdjogoTT @nvelickovicTT @amahmudTT @lpremovicTT @ncvetkovicTT @fvranicTT @tenstorrent/codeowner-bypass
 tt_metal/jit_build/ @abhullar-tt @nathan-TT @jbaumanTT @kstevensTT @ruizhangTT @tenstorrent/codeowner-bypass
 tt_metal/jit_build/**/CMakeLists.txt @abhullar-tt @nathan-TT @jbaumanTT @ruizhangTT @tenstorrent/metalium-developers-infra @tenstorrent/codeowner-bypass
 tt_metal/kernels/ @abhullar-tt @tenstorrent/codeowner-bypass # these should go away or get moved to tests

docs/Doxyfile

@@ -928,83 +928,83 @@ INPUT                  = tt_metal/hw/inc/api/compile_time_args.h \
                          tt_metal/hw/inc/internal/ethernet/dataflow_api.h \
                          tt_metal/api/tt-metalium/host_api.hpp \
                          tt_metal/api/tt-metalium/experimental/profiler.hpp \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/comp.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/erf_erfc.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/erfinv.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/exp.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/gelu.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/recip.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/sqrt.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/rsqrt.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/elu.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/relu.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/i0.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/i1.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/reverseops.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/isinf_isnan.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/logical_not_noti.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/trigonometry.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/rounding.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/activations.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/clamp.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/trigonometric.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/binop_with_scalar.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/fill.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/fmod.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/rounding.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/identity.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/negative.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/dropout.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/typecast.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/prelu.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/softplus.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/softsign.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/hardtanh.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/rand.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/isinf_isnan.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/left_shift.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/right_shift.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/remainder.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/threshold.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/selu.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/where.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/addcmul.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/log1p.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/comp.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/bitwise_or.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/bitwise_and.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/bitwise_xor.h \
-                         tt_metal/include/compute_kernel_api/eltwise_unary/bitwise_not.h \
-                         tt_metal/include/compute_kernel_api/bcast.h \
-                         tt_metal/include/compute_kernel_api/cb_api.h \
-                         tt_metal/include/compute_kernel_api/compute_kernel_hw_startup.h \
-                         tt_metal/include/compute_kernel_api/eltwise_binary.h \
-                         tt_metal/include/compute_kernel_api/eltwise_binary_sfpu.h \
-                         tt_metal/include/compute_kernel_api/matmul.h \
-                         tt_metal/include/compute_kernel_api/pack.h \
-                         tt_metal/include/compute_kernel_api/pack_untilize.h \
-                         tt_metal/include/compute_kernel_api/quantization.h \
-                         tt_metal/include/compute_kernel_api/reconfig_data_format.h \
-                         tt_metal/include/compute_kernel_api/reduce.h \
-                         tt_metal/include/compute_kernel_api/reg_api.h \
-                         tt_metal/include/compute_kernel_api/tile_move_copy.h \
-                         tt_metal/include/compute_kernel_api/tilize.h \
-                         tt_metal/include/compute_kernel_api/transpose_wh.h \
-                         tt_metal/include/compute_kernel_api/untilize.h \
-                         tt_metal/include/compute_kernel_api/cumsum.h \
-                         tt_metal/include/compute_kernel_api/add_int_sfpu.h \
-                         tt_metal/include/compute_kernel_api/binop_with_scalar.h \
-                         tt_metal/include/compute_kernel_api/gcd.h \
-                         tt_metal/include/compute_kernel_api/xlogy.h \
-                         tt_metal/include/compute_kernel_api/lcm.h \
-                         tt_metal/include/compute_kernel_api/mask.h \
-                         tt_metal/include/compute_kernel_api/reshuffle.h \
-                         tt_metal/include/compute_kernel_api/binary_shift.h \
-                         tt_metal/include/compute_kernel_api/binary_max_min.h \
-                         tt_metal/include/compute_kernel_api/binary_bitwise_sfpu.h \
-                         tt_metal/include/compute_kernel_api/sub_int_sfpu.h \
-                         tt_metal/include/compute_kernel_api/mul_int_sfpu.h \
-                         tt_metal/include/compute_kernel_api.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/comp.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/erf_erfc.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/erfinv.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/exp.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/gelu.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/recip.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/sqrt.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/rsqrt.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/elu.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/relu.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/i0.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/i1.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/reverseops.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/isinf_isnan.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/logical_not_noti.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/trigonometry.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/rounding.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/activations.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/clamp.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/trigonometric.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/binop_with_scalar.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/fill.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/fmod.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/rounding.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/identity.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/negative.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/dropout.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/typecast.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/prelu.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/softplus.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/softsign.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/hardtanh.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/rand.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/isinf_isnan.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/left_shift.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/right_shift.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/remainder.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/threshold.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/selu.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/where.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/addcmul.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/log1p.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/comp.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/bitwise_or.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/bitwise_and.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/bitwise_xor.h \
+                         tt_metal/hw/inc/api/compute/eltwise_unary/bitwise_not.h \
+                         tt_metal/hw/inc/api/compute/bcast.h \
+                         tt_metal/hw/inc/api/compute/cb_api.h \
+                         tt_metal/hw/inc/api/compute/compute_kernel_hw_startup.h \
+                         tt_metal/hw/inc/api/compute/eltwise_binary.h \
+                         tt_metal/hw/inc/api/compute/eltwise_binary_sfpu.h \
+                         tt_metal/hw/inc/api/compute/matmul.h \
+                         tt_metal/hw/inc/api/compute/pack.h \
+                         tt_metal/hw/inc/api/compute/pack_untilize.h \
+                         tt_metal/hw/inc/api/compute/quantization.h \
+                         tt_metal/hw/inc/api/compute/reconfig_data_format.h \
+                         tt_metal/hw/inc/api/compute/reduce.h \
+                         tt_metal/hw/inc/api/compute/reg_api.h \
+                         tt_metal/hw/inc/api/compute/tile_move_copy.h \
+                         tt_metal/hw/inc/api/compute/tilize.h \
+                         tt_metal/hw/inc/api/compute/transpose_wh.h \
+                         tt_metal/hw/inc/api/compute/untilize.h \
+                         tt_metal/hw/inc/api/compute/cumsum.h \
+                         tt_metal/hw/inc/api/compute/add_int_sfpu.h \
+                         tt_metal/hw/inc/api/compute/binop_with_scalar.h \
+                         tt_metal/hw/inc/api/compute/gcd.h \
+                         tt_metal/hw/inc/api/compute/xlogy.h \
+                         tt_metal/hw/inc/api/compute/lcm.h \
+                         tt_metal/hw/inc/api/compute/mask.h \
+                         tt_metal/hw/inc/api/compute/reshuffle.h \
+                         tt_metal/hw/inc/api/compute/binary_shift.h \
+                         tt_metal/hw/inc/api/compute/binary_max_min.h \
+                         tt_metal/hw/inc/api/compute/binary_bitwise_sfpu.h \
+                         tt_metal/hw/inc/api/compute/sub_int_sfpu.h \
+                         tt_metal/hw/inc/api/compute/mul_int_sfpu.h \
+                         tt_metal/hw/inc/api/compute/compute_kernel_api.h \
                          tt_metal/impl/kernels/kernel_args.hpp \
                          tt_metal/include/tt_metal/metal.hpp \
                          tt_metal/api/tt-metalium/types.hpp \

docs/source/tt-metalium/tt_metal/labs/matmul/lab1/lab1.rst

@@ -1163,10 +1163,10 @@ Then, adjust the code to perform matrix multiplication, by making the following
    so your code only needs to generate indices in the right order.
 
 #. Update the compute kernel to perform matrix multiplication rather than elementwise addition.
-   To initialize the Tensix Engine for matrix multiplication, you will need to use the ``mm_init`` function provided in ``tt_metal/include/compute_kernel_api/matmul.h``.
+   To initialize the Tensix Engine for matrix multiplication, you will need to use the ``mm_init`` function provided in ``tt_metal/hw/inc/api/compute/matmul.h``.
    Do not use any other initialization functions for matrix multiplication (specifically do **not** use ``binary_op_init_common``, because that function is only
    applicable to elementwise operations, not to matrix multiplication).
-   To multiply two tiles, you will need to use the ``matmul_tiles`` function provided in ``tt_metal/include/compute_kernel_api/matmul.h``.
+   To multiply two tiles, you will need to use the ``matmul_tiles`` function provided in ``tt_metal/hw/inc/api/compute/matmul.h``.
    This function accumulates the result into the destination register; i.e. it adds to the existing values in the register rather than overwriting existing content.
    By judiciously choosing when to call ``tile_regs_acquire``, which initializes all tiles in the destination register array to zero, and when to call
    ``tile_regs_commit``, which signals that the compute core is done writing to the destination register,

docs/source/tt-metalium/tt_metal/labs/matmul/lab2/lab2.rst

@@ -752,7 +752,7 @@ Follow these steps to complete the exercise:
      into the destination register requires a new operation:
 
      - ``copy_tile(in_cb_id, in_tile_index, dst_tile_index)`` defined in
-       ``tt_metal/include/compute_kernel_api/tile_move_copy.h`` copies a tile from the intermediate
+       ``tt_metal/hw/inc/api/compute/tile_move_copy.h`` copies a tile from the intermediate
        CB to the destination register array at the specified index.
      - Before calling this function, you need to call ``copy_tile_to_dst_init_short(in_cb_id)``
        to set up the Tensix Engine for the copy operation.

models/demos/deepseek_v3_b1/kernel_includes/tt_metal/include/compute_kernel_api/add_rsqrt.h

@@ -4,8 +4,8 @@
 
 #pragma once
 
-#include "compute_kernel_api.h"
-#include "compute_kernel_api/common.h"
+#include "api/compute/compute_kernel_api.h"
+#include "api/compute/common.h"
 #ifdef TRISC_MATH
 #include "../../hw/ckernels/blackhole/metal/llk_api/llk_sfpu/llk_math_eltwise_unary_sfpu_add_rsqrt.h"
 #endif

models/demos/deepseek_v3_b1/kernel_includes/tt_metal/include/compute_kernel_api/custom_mm.h

@@ -4,7 +4,7 @@
 
 #pragma once
 
-#include "compute_kernel_api/common.h"
+#include "api/compute/common.h"
 #ifdef TRISC_MATH
 #include "../../hw/ckernels/blackhole/metal/llk_api/llk_math_custom_mm_api.h"
 #endif

models/demos/deepseek_v3_b1/kernel_includes/tt_metal/include/compute_kernel_api/deepseek_moe_gate.h

@@ -4,9 +4,9 @@
 
 #pragma once
 
-#include "compute_kernel_api.h"
-#include "compute_kernel_api/common.h"
-#include "compute_kernel_api/transpose_wh.h"
+#include "api/compute/compute_kernel_api.h"
+#include "api/compute/common.h"
+#include "api/compute/transpose_wh.h"
 #ifdef TRISC_MATH
 #ifdef ARCH_BLACKHOLE
 #include "../../hw/ckernels/blackhole/metal/llk_api/llk_sfpu/llk_math_deepseek_moe_gate_topk_single_face.h"

models/demos/deepseek_v3_b1/kernel_includes/tt_metal/include/compute_kernel_api/rmsnorm.h

@@ -4,8 +4,8 @@
 
 #pragma once
 
-#include "compute_kernel_api.h"
-#include "compute_kernel_api/common.h"
+#include "api/compute/compute_kernel_api.h"
+#include "api/compute/common.h"
 #ifdef TRISC_MATH
 #include "../../hw/ckernels/blackhole/metal/llk_api/llk_math_rmsnorm_bcast_scalar_dest_reuse_api.h"
 #endif

models/demos/deepseek_v3_b1/micro_ops/deepseek_moe_gate/kernels/deepseek_moe_gate_compute.cpp

@@ -7,12 +7,12 @@
 #define REDUCE_OP PoolType::SUM
 #define REDUCE_DIM ReduceDim::REDUCE_ROW
 
-#include "compute_kernel_api.h"
-#include "compute_kernel_api/common.h"
-#include "compute_kernel_api/tile_move_copy.h"
-#include "compute_kernel_api/pack.h"
-#include "compute_kernel_api/reconfig_data_format.h"
-#include "compute_kernel_api/eltwise_binary.h"
+#include "api/compute/compute_kernel_api.h"
+#include "api/compute/common.h"
+#include "api/compute/tile_move_copy.h"
+#include "api/compute/pack.h"
+#include "api/compute/reconfig_data_format.h"
+#include "api/compute/eltwise_binary.h"
 #include "../../../kernel_includes/tt_metal/include/compute_kernel_api/deepseek_moe_gate.h"
 
 void kernel_main() {

models/demos/deepseek_v3_b1/micro_ops/dram_streaming_matmul/kernels/compute.cpp

@@ -4,7 +4,7 @@
 
 #include <cstdint>
 
-#include "compute_kernel_api/tile_move_copy.h"
+#include "api/compute/tile_move_copy.h"
 #include "../../../kernel_includes/tt_metal/include/compute_kernel_api/custom_mm.h"
 
 // Fused SiLU activation support (only when FUSE_SILU is defined)
@@ -13,7 +13,7 @@
 // - ITERATIONS: minimum 2 required for SFPU, then scales (m<=4->2, m=8->4, m>=16->8)
 // Total: significant speedup vs default silu_tile()
 #ifdef FUSE_SILU
-#include "compute_kernel_api.h"  // for silu_tile_init() and llk_math_eltwise_unary_sfpu_silu
+#include "api/compute/compute_kernel_api.h"  // for silu_tile_init() and llk_math_eltwise_unary_sfpu_silu
 #endif
 
 /**