.. meta:: :description: Release notes for the Neuron Kernel Interface (NKI) component across all Neuron SDK versions :keywords: NKI, Neuron Kernel Interface, release notes, nki.language, nki.isa, kernels :date-modified: 04/09/2026 .. _nki_rn: Release Notes for Neuron Component: Neuron Kernel Interface (NKI) ================================================================== The release notes for the Neuron Kernel Interface (NKI) component. Read them for the details about the changes, improvements, and bug fixes for all release versions of the AWS Neuron SDK. .. _nki-2-29-0-rn: Neuron Kernel Interface (NKI) [0.3.0] (Neuron 2.29.0 Release) --------------------------------------------------------------------- Date of Release: 04/09/2026 AWS Neuron SDK 2.29.0 introduces NKI 0.3.0, a significant update to the Neuron Kernel Interface for General Availability. NKI 0.3.0 features NKI Standard Library (nki-stdlib), which provides developer-visible code for all NKI APIs and native language objects (e.g., NkiTensor). This release provides new exposed Trainium capabilities and features in the NKI API and re-introduces nki.language APIs. NKI 0.3.0 includes a CPU Simulator, which executes NKI kernels entirely on CPU using NumPy — enabling developers to validate kernel logic on laptops and CI environments without Trainium hardware. NKI 0.3.0 also includes the ``nki.typing`` module for declaring expected tensor shapes, a dedicated ``nki.isa.exponential`` instruction optimized for Softmax computation, matmul accumulation control, explicit memory address placement, and variable-length all-to-all collectives via ``nki.collectives.all_to_all_v``. NKI 0.3.0 includes several API breaking changes that improve correctness and consistency along with updated documentation. For the full list of changes and update examples, see the :ref:`NKI 0.3.0 Update Guide `. New Features ~~~~~~~~~~~~ * **NKI Standard Library (nki-stdlib)**: NKI 0.3.0 ships with the NKI Standard Library (nki-stdlib), which provides developer-visible code for all NKI APIs and native language objects (e.g., ``NkiTensor``). * **NKI CPU Simulator** *(Experimental)*: Executes NKI kernels entirely on CPU using NumPy, enabling local development, debugging, and functional correctness testing without Trainium hardware. Set the environment variable ``NKI_SIMULATOR=1`` to run existing kernels without code changes, or wrap the kernel call with ``nki.simulate(kernel)``. See :doc:`nki.simulate API Reference `. * **nki.language APIs** *(Experimental)*: Re-introduces ``nki.language`` APIs as convenience wrappers around ``nki.isa`` APIs, including ``nl.load``, ``nl.store``, ``nl.copy``, ``nl.matmul``, ``nl.transpose``, ``nl.softmax``, and other high-level operations. See :doc:`nki.language API Reference `. * **nki.typing module**: New module for type-annotating kernel tensor parameters. Use ``nt.tensor[shape]`` to declare expected tensor shapes. * **nki.isa.exponential**: Dedicated exponential instruction with max subtraction, faster than ``nisa.activation(op=nl.exp)`` and useful for Softmax calculation. Trn3 (NeuronCore-v4) only. See :doc:`nki.isa.exponential `. * **nki.collectives.all_to_all_v**: Variable-length all-to-all collective. Unlike ``all_to_all``, uses a metadata tensor to specify per-rank send/recv counts. See :doc:`nki.collectives API Reference `. * **Matmul accumulation**: ``nc_matmul`` and ``nc_matmul_mx`` now have an ``accumulate`` parameter that controls whether the operation overwrites or accumulates on the destination PSUM tile. The default (``accumulate=None``) auto-detects, matching Beta 2 behavior. See :doc:`nki.isa.nc_matmul `. * **Address placement**: The ``address`` parameter was added to ``nki.language.ndarray`` for explicit memory placement. See :doc:`nki.language.ndarray `. Deprecated and Removed APIs ~~~~~~~~~~~~~~~~~~~~~~~~~~~ * ``nki.isa.tensor_copy_dynamic_src`` / ``nki.isa.tensor_copy_dynamic_dst`` — Deprecated and scheduled for removal. Use ``nisa.tensor_copy()`` with ``.ap()`` and ``scalar_offset`` instead. * ``nki.jit(platform_target=...)`` — Deprecated. Set the target platform via the ``NEURON_PLATFORM_TARGET_OVERRIDE`` environment variable instead. This is a breaking change. .. TODO: Create an NKI environment variables reference page and link from here. * ``nki.jit(mode=...)`` — Deprecated and ignored. The NKI Compiler now auto-detects the framework from kernel arguments. This is a breaking change. Breaking Changes ~~~~~~~~~~~~~~~~ .. note:: NKI 0.3.0 requires all NKI kernels in a model to be updated to NKI 0.3.0. Mixing NKI 0.3.0 and NKI Beta 2 kernels in the same model is not supported. For models that have not yet been updated, continue using Neuron SDK 2.28. * ``nisa.dma_copy`` — No longer supports reading directly from PSUM. Copy the PSUM tensor to SBUF first using ``nisa.tensor_copy``. * ``nisa.dma_copy`` — Enforces matching source and destination element types when using ``dge_mode=dge_mode.hwdge``. Use ``.view()`` to reinterpret types. * ``nisa.dma_copy`` — ``dst_rmw_op`` and ``unique_indices`` parameters removed. Use ``nisa.dma_compute`` instead. * ``nisa.dma_compute`` — ``scales`` and ``reduce_op`` parameters swapped positions. ``scales`` is now optional. ``unique_indices`` parameter added. Update call sites to use the new parameter order: ``nisa.dma_compute(dst, srcs, reduce_op, scales=None, unique_indices=True)``. * ``nisa.memset`` — Enforces strict type matching between ``value`` and destination dtype. x4 packed types enforce ``value=0``. Kernels that pass float values to integer-typed tensors (e.g., ``value=2.0`` instead of ``value=2``) will now raise an error at compile time. * ``nisa.sendrecv`` — ``use_gpsimd_dma`` replaced by ``dma_engine`` enum. Update existing kernels to use the new enum. * ``nisa.affine_select`` — ``offset`` moved from 3rd positional argument to keyword argument with default ``0``. * ``nisa.register_move`` — ``imm`` renamed to ``src``, now accepts ``VirtualRegister``. Update keyword argument from ``imm=`` to ``src=``. * ``nl.shared_constant`` — ``dtype`` parameter removed. Remove the ``dtype`` argument from ``nl.shared_constant`` calls; the type is now inferred from the constant. * ``nki.collectives.collective_permute_implicit_current_processing_rank_id`` — ``num_channels`` parameter removed. Remove ``num_channels`` from call sites and pass ``channel_ids`` list to ``collective_permute_implicit()`` instead. * Output tensors must use ``buffer=nl.shared_hbm``. Using ``nl.hbm`` causes compilation failures. * Raw integer enum constants no longer accepted. Use named enum members. * String buffer names no longer accepted. Use buffer objects (e.g., ``nl.sbuf``). * Keyword-only argument separator (``*``) in kernel signatures is not supported. * ``is`` / ``is not`` operators are not supported. Use ``==`` / ``!=``. * ``list`` kernel arguments are not supported. Convert to tuples. For before-and-after code examples, see the :ref:`NKI 0.3.0 Update Guide `. .. note:: The previously announced removal of the ``neuronxcc.nki.*`` namespace has been postponed to a future release. Both the ``neuronxcc.nki.*`` and ``nki.*`` namespaces continue to be supported in this release. Other Changes ~~~~~~~~~~~~~ * ``nki.isa.dma_engine`` alias repurposed as the ``dma_engine`` enum for DMA transfer engine selection. * ``nki.isa.iota`` — ``offset`` now optional with default ``0``. * ``nki.isa.core_barrier`` — ``engine`` default changed from ``unknown`` to ``gpsimd`` (no behavioral change). * ``nki.language.num_programs`` — ``axes`` default changed from ``None`` to ``0``. * ``nki.language.program_id`` — ``axis`` now defaults to ``0``. * ``nki.language.ndarray`` — ``buffer`` default changed from ``None`` to ``nl.sbuf``. * ``nki.language.zeros`` — ``buffer`` default changed from ``None`` to ``nl.sbuf``. * ``nki.language.sequential_range`` — ``stop`` and ``step`` now have default values (``None`` and ``1``). Bug Fixes ~~~~~~~~~ * Fixed incorrect axis handling in ``nisa.tensor_reduce``. Beta 2 incorrectly allowed ``axis=1`` to refer to the last free dimension even for 3D/4D tensors. NKI 0.3.0 corrects this so that axis values correspond to the actual tensor dimensions. * Fixed ``nisa.range_select`` silently overriding user-specified parameters. The ``on_false_value`` and ``reduce_cmd`` parameters were incorrectly ignored by the compiler — ``on_false_value`` was always set to ``-3.4028235e+38`` and ``reduce_cmd`` was always set to ``reset_reduce``, regardless of the values passed in. NKI 0.3.0 honors the ``reduce_cmd`` parameter and documents the ``FP32_MIN`` hardware constraint for ``on_false_value``. Known Issues ~~~~~~~~~~~~ **Math Operations** * ``nki.language.divide`` is not supported — Division is not available as a hardware instruction. As a workaround, multiply by the reciprocal: ``nl.multiply(x, nl.reciprocal(y))``. * ``nki.language.fmod`` and ``nki.language.mod`` are not supported — Modulo operations are not available as hardware instructions. These APIs work in simulation but will fail when compiled for Trainium hardware. * ``nki.language.power`` does not support scalar exponents — ``nl.power(tile, scalar)`` is not supported. Use ``nl.power(tile, tile)`` instead, where both operands are tiles. **Broadcasting** * Binary operations do not support broadcasting — Operations like ``nl.add(a, b)`` require both operands to have the same shape. Broadcasting (e.g., adding a ``(128, 1)`` tile to a ``(128, 512)`` tile) is not yet supported. * ``nki.language.softmax`` and ``nki.language.rms_norm`` fail on hardware — These functions rely on internal broadcasting between full-size and reduced-size tiles, which is not supported on hardware. They work correctly in simulation. **Random Number Generation** * ``nki.language.random_seed`` requires a tensor, not a scalar — Pass a ``[1, 1]`` tensor on SBUF instead of a Python integer. For example: ``nl.random_seed(nl.full((1, 1), 42, dtype=nl.int32, buffer=nl.sbuf))``. * ``nki.language.rand`` and ``nki.language.random_seed`` engine behavior — On NeuronCore-v4+ (Trn3+), ``rand`` uses ``nisa.rand2`` on the Vector Engine. On earlier NeuronCores, ``rand`` uses ``nisa.rng`` which may run on a different engine than ``random_seed``, potentially causing ``random_seed`` to have no effect on ``rand`` output. **Matrix Operations** * ``nki.language.matmul`` without ``transpose_x=True`` is not supported — Calling ``nl.matmul(x, y)`` without setting ``transpose_x=True`` will fail. As a workaround, always use ``nl.matmul(x, y, transpose_x=True)`` and pre-arrange data accordingly. **Data Movement** * ``nki.language.store`` does not support PSUM tiles directly — Storing a tile that resides in PSUM requires manually copying it to SBUF first using ``nisa.tensor_copy``. A future release will handle this automatically. * ``nki.language.copy`` uses lossy FP32 casting — ``nl.copy`` uses the Scalar Engine which internally casts through FP32, which is lossy for integer types with values exceeding FP32 precision (e.g., int32 values > 2^23). Additionally, cross-buffer copies (e.g., PSUM to SBUF) are not supported. **Control Flow** * ``nki.language.dynamic_range`` loop variable cannot be used in index arithmetic — The induction variable of a ``dynamic_range`` loop is a scalar, not a register. It cannot be used as a ``scalar_offset`` in access patterns or in arithmetic expressions for computing tile offsets. Use ``nl.affine_range`` or ``nl.static_range`` if you need to compute offsets from the loop variable. **Multi-Core (LNC2)** * LNC2 requires identical control flow across cores — When running with Logical NeuronCore 2 (LNC2), the NKI compiler expects each physical NeuronCore to execute identical control flow. Programs with dynamic control flow that differs across cores may deadlock or produce incorrect results. This constraint is not enforced at compile time. .. _nki-2-28-0-rn: Neuron Kernel Interface (NKI) (Beta 2 - 0.2.0) [2.28] (Neuron 2.28.0 Release) ----------------------------------------------------------------------------- Date of Release: 02/26/2026 New Features ~~~~~~~~~~~~ * LNC (Large Neuron Core) multi-core support: * **Shared buffers and canonical outputs**: The compiler now tracks :doc:`shared_hbm ` tensors declared in kernels and canonicalizes LNC kernel outputs into a consistent form. This is foundational infrastructure for multi-core kernel compilation. See :doc:`LNC Overview `. * **Private HBM tensors**: Users can declare tensors private to a single NeuronCore using the :doc:`private_hbm ` memory type, distinct from regular and shared HBM. * **Intra-LNC collectives**: New ISA instruction types for multi-core collective operations such as cross-core reductions and broadcasts. See full API listing under :doc:`nki.collectives ` below. * New ``nki.isa`` APIs: * :doc:`nki.isa.nonzero_with_count ` — returns nonzero element indices and their count, useful for sparse computation and dynamic masking * ``nki.isa.exponential`` — computes element-wise exponential on tensors. See :doc:`nki.isa.activation `. * New :doc:`nki.collectives ` module, enabling collective communication across multiple NeuronCores directly from NKI kernels: * :doc:`nki.collectives.all_reduce ` * :doc:`nki.collectives.all_gather ` * :doc:`nki.collectives.reduce_scatter ` * :doc:`nki.collectives.all_to_all ` * :doc:`nki.collectives.collective_permute ` * :doc:`nki.collectives.collective_permute_implicit ` * :doc:`nki.collectives.collective_permute_implicit_reduce ` * :doc:`nki.collectives.rank_id ` * New ``dtypes``: * :doc:`nki.language.float8_e4m3fn ` — for FP8 inference and training workloads * New NKI language features: * ``no_reorder`` blocks — use ``with no_reorder(): ...`` to prevent the compiler from reordering instructions within a block, for kernels where instruction ordering affects correctness * ``__call__`` special method support — callable objects (classes with ``__call__``) can now be used as functions within NKI kernels * ``tensor.view`` method — tensors now support ``.view()`` for reshaping * :doc:`Shared constants ` can now be passed to kernels as string arguments, not just tensor objects Improvements ~~~~~~~~~~~~ * Updated ``nki.isa`` APIs: * :doc:`nki.isa.dma_transpose ` now supports indirect addressing * :doc:`nki.isa.dma_copy ` now supports ``unique_indices`` parameter * :doc:`nki.isa.register_alloc ` now accepts an optional tensor argument to pre-fill the allocated register with initial values * Compiler output improvements: * The compiler no longer truncates diagnostic output; users now receive the full set of warnings and errors Breaking Changes ~~~~~~~~~~~~~~~~ * :doc:`nki.isa.nc_matmul ` parameter ``psumAccumulateFlag`` has been removed. This parameter had no effect on compilation or execution. Simply remove it from your kernel code. * :doc:`nki.isa.nc_matmul ` parameter ``is_moving_zero`` has been renamed to ``is_moving_onezero`` to match hardware semantics, consistent with the companion ``is_stationary_onezero`` parameter. Kernels that passed ``is_moving_zero`` by name should update to ``is_moving_onezero``. * ``nki.tensor`` has moved to ``nki.meta.tensor``. Users should update their imports accordingly. .. note:: The previously announced removal of the ``neuronxcc.nki.*`` namespace has been postponed from Neuron 2.28 to Neuron 2.29. Both the ``neuronxcc.nki.*`` and ``nki.*`` namespaces continue to be supported in this release. We encourage customers to migrate to the ``nki.*`` namespace using the :doc:`NKI Beta 2 Migration Guide `. Bug Fixes ~~~~~~~~~ * Fixed incorrect default value for ``on_false_value`` in ``nki.isa.range_select``. The default was ``0.0`` instead of negative infinity (``-inf``). This caused ``range_select`` to write zeros for out-of-range elements instead of the expected negative-infinity sentinel, which could produce incorrect results in downstream reductions (e.g., max-pooling or top-k). See :doc:`nki.isa.range_select `. * Fixed default value parsing for keyword-only arguments in NKI kernels. When a Python function used keyword-only arguments with default values (arguments after ``*`` in the signature), the NKI compiler did not associate the defaults with their corresponding parameter names. This caused keyword-only arguments to appear as required even when they had defaults, leading to "missing argument" errors during kernel compilation. * Fixed wrong default for ``reduce_cmd`` in :doc:`nki.isa.activation `. The default was incorrectly set to ``ZeroAccumulate`` instead of ``Idle``, causing the accumulator to be zeroed before every activation call even when no reduction was requested. * Fixed missing ALU operators (``rsqrt``, ``abs``, ``power``) in :doc:`nki.isa.tensor_scalar ` and :doc:`nki.isa.tensor_tensor `. Passing these operators previously raised an "unsupported operator" error. See :doc:`NKI Language Guide `. * Fixed ``float8_e4m3fn`` to ``float8_e4m3`` conversion for kernel inputs and outputs. When a tensor with dtype ``float8_e4m3fn`` was passed to the compiler, the automatic conversion to ``float8_e4m3`` could fail with a size-check error. The conversion now validates sizes correctly before casting. See :doc:`nki.language.float8_e4m3 `. * Fixed dynamic for loop incorrectly incrementing the loop induction variable. In loops with a runtime-determined trip count (``sequential_range`` with non-constant bounds), the compiler generated incorrect increment code, causing the loop counter to never advance and the loop to run indefinitely or produce incorrect iteration values. See :doc:`nki.language.sequential_range `. * Fixed reshape of ``shared_hbm`` and ``private_hbm`` tensors failing partition size check. Reshape only recognized plain ``hbm`` memory as exempt from partition-dimension size validation. Tensors allocated in ``shared_hbm`` or ``private_hbm`` (used for cross-kernel and kernel-private storage) incorrectly triggered a "partition size mismatch" error when reshaped. See :doc:`nki.language.shared_hbm ` and :doc:`nki.language.private_hbm `. * Fixed bias shape checking in :doc:`nki.isa.activation `. The ``bias`` parameter was not validated for shape correctness. A bias tensor with a free dimension other than 1 (e.g., shape ``(128, 64)`` instead of ``(128, 1)``) was accepted without validation, which could produce incorrect results. The compiler now raises an error if the bias free dimension is not 1. * Fixed incorrect line numbers in stack traces and error reporting. An off-by-one error in the line offset calculation caused all reported line numbers to be shifted by one. Additionally, error location was sometimes lost when errors propagated across file boundaries. * Fixed invalid keyword arguments being silently ignored instead of raising an error. When calling an NKI API with a misspelled or unsupported keyword argument, the argument was ignored without warning. The compiler now validates all keyword argument names against the function signature and raises an ``unexpected keyword argument`` error for unrecognized names. * Fixed ``nki.jit`` in auto-detection mode returning an uncalled kernel object instead of executing the kernel. When ``nki.jit`` was used without specifying a framework mode (e.g., ``@nki.jit`` with no ``mode`` argument), the auto-detection path constructed the appropriate framework-specific kernel object but returned it without calling it. The user received a kernel object instead of the computed result, requiring an extra manual invocation. See :doc:`nki.jit `. * Fixed stale kernel object state between trace invocations. When tracing the same kernel multiple times (e.g., with different input shapes), compiler state was not fully reset between invocations, causing name collisions and incorrect results. The trace state is now fully reset before each invocation. * Improved 'removed during code migration' error messages with clear descriptions of unimplemented features. APIs not available in this release (``nki.baremetal``, ``nki.benchmark``, ``nki.profile``, ``nki.simulate_kernel``) previously raised a generic ``NotImplementedError("removed during code migration")`` message. Each now raises a specific message naming the unsupported API. Additionally, calling an ``nki.jit`` kernel with no arguments now raises a clear error instead. See :doc:`NKI Beta 2 Migration Guide `. * Fixed nested ``nki_jit`` decorators not being allowed. The NKI compiler only recognized ``@nki.jit``-decorated functions when they were plain function objects. Nested decorators (e.g., ``@my_wrapper @nki.jit``) wrapped the function in a non-function object, causing the compiler to skip it. The compiler now correctly unwraps decorator chains to find the underlying kernel function. See :doc:`nki.jit `. Known Issues ~~~~~~~~~~~~ * ``nki.isa.range_select``: The ``on_false_value`` and ``reduce_cmd`` parameters are incorrectly ignored by the NKI compiler. The ``on_false_value`` is always set to ``(-3.4028235e+38)`` and ``reduce_cmd`` is always set to ``reduce_cmd.reset_reduce``, regardless of the values passed in. .. _nki-2-27-0-rn: Neuron Kernel Interface (NKI) (Beta 2 - 0.1.0) [2.27] (Neuron 2.27.0 Release) ----------------------------------------------------------------------------- Date: 12/25/2025 Improvements ~~~~~~~~~~~~~~~ * new ``nki.language`` APIs: * ``nki.language.device_print`` * new ``nki.isa`` APIs: * ``nki.isa.dma_compute`` * ``nki.isa.nki.isa.quantize_mx`` * ``nki.isa.nki.isa.nc_matmul`` * ``nki.isa.nki.isa.nc_n_gather`` [used to be ``nl.gather_flattened`` with free partition limited to 512] * ``nki.isa.rand2`` * ``nki.isa.rand_set_state`` * ``nki.isa.rand_get_state`` * ``nki.isa.set_rng_seed`` * ``nki.isa.rng`` * new ``dtypes``: * ``nki.language.float8_e5m2_x4`` * ``nki.language.float4_e2m1fn_x4`` * ``nki.language.float8_e4m3fn_x4`` * changes to existing APIs: * several ``nki.language`` APIs have been removed in NKI Beta 2 * all nki.isa APIs have ``dst`` as an input param * all nki.isa APIs removed ``dtype`` and ``mask`` support * ``nki.isa.memset`` — removed ``shape`` positional arg , since we have ``dst`` * ``nki.isa.affine_select`` — instead of ``pred``, we now take ``pattern`` and ``cmp_op`` params * ``nki.isa.iota`` — ``expr`` replaced with ``pattern`` and ``offset`` * ``nki.isa.nc_stream_shuffle`` - ``src`` and ``dst`` order changed * docs improvements: * restructured NKI Documentation to align with workflows * added :doc:`Trainium3 Architecture Guide for NKI ` * added :doc:`About Neuron Kernel Interface (NKI) ` * added :doc:`NKI Environment Setup Guide ` * added :doc:`Get Started with NKI ` * added :doc:`NKI Language Guide ` * added :doc:`About the NKI Compiler ` * added :doc:`About NKI Beta 2 Migration ` * added :doc:`MXFP Matrix Multiplication with NKI ` * updated :doc:`Matrix Multiplication Tutorial ` * updated :doc:`Profile a NKI Kernel ` * updated :doc:`NKI APIs ` * updated :doc:`NKI Library docs ` * removed NKI Error Guide Known Issues ~~~~~~~~~~~~ * ``nki.isa.nki.isa.nc_matmul`` - ``is_moving_onezero`` was incorrectly named ``is_moving_zero`` in this release * NKI ISA semantic checks are not available with Beta 2, workaround is to reference the API docs * NKI Collectives are not available with Beta 2 * ``nki.benchmark`` and ``nki.profile`` are not available with Beta 2 ---- .. _nki-2-26-0-rn: Neuron Kernel Interface (NKI) (Beta) [2.26] (Neuron 2.26.0 Release) -------------------------------------------------------------------- Date: 09/18/2025 Improvements ~~~~~~~~~~~~~~~ * new ``nki.language`` APIs: * ``nki.language.gelu_apprx_sigmoid`` - Gaussian Error Linear Unit activation function with sigmoid approximation. * ``nki.language.tile_size.total_available_sbuf_size`` to get total available SBUF size * new ``nki.isa`` APIs: * ``nki.isa.select_reduce`` - selectively copy elements with max reduction * ``nki.isa.sequence_bounds`` - compute sequence bounds of segment IDs * ``nki.isa.dma_transpose`` * ``axes`` param to define 4D transpose for some supported cases * ``dge_mode`` to specify Descriptor Generation Engine (DGE). * ``nl.gelu_apprx_sigmoid`` op support on ``nki.isa.activation`` * fixes / improvements: * ``nki.language.store`` supports PSUM buffer with extra additional copy inserted. * docs/tutorial improvements: * ``nki.isa.dma_transpose`` API doc and example * ``nki.simulate_kernel`` example improvement * use ``nl.fp32.min`` in tutorial code instead of a magic number * better error reporting: * indirect indexing on transpose * mask expressions ---- .. _nki-2-24-0-rn: Neuron Kernel Interface (NKI) (Beta) [2.24] (Neuron 2.24.0 Release) -------------------------------------------------------------------- Date: 06/24/2025 Improvements ~~~~~~~~~~~~~~~ * ``sqrt`` valid data range extended for accuracy improvement with wider numerical values support. * ``nki.language.gather_flattened`` new API * ``nki.isa.nc_match_replace8`` additional param ``dst_idx`` * improved docs/examples on ``nki.isa.nc_match_replace8``, ``nki.isa.nc_stream_shuffle`` * improved error messages ---- .. _nki-2-23-0-rn: Neuron Kernel Interface (NKI) (Beta) [2.23] (Neuron 2.23.0 Release) -------------------------------------------------------------------- Date: 05/20/2025 Improvements ~~~~~~~~~~~~~~~ * ``nki.isa.range_select`` (for trn2) new instruction * ``abs``, ``power`` ops supported on to nki.isa tensor instruction * ``abs`` op supported on ``nki.isa.activation`` instruction * GpSIMD engine support added to ``add``, ``multiply`` in 32bit integer to nki.isa tensor operations * ``nki.isa.tensor_copy_predicated`` support for reversing predicate. * ``nki.isa.tensor_copy_dynamic_src``, ``tensor_copy_dynamic_dst`` engine selection. * ``nki.isa.dma_copy`` additional support with ``dge_mode``, ``oob_mode``, and in-place add ``rmw_op``. * ``+=, -=, /=, *=`` operators now work consistently across loop types, PSUM, and SBUF, * fixed simulation for instructions: ``nki.language.rand``, ``random_seed``, ``nki.isa.dropout`` * fixed simulation masking behavior * Added warning when the block dimension is used for SBUF and PSUM tensors, see: :ref:`NKI Block Dimension Migration Guide ` ---- .. _nki-2-22-0-rn: Neuron Kernel Interface (NKI) (Beta) [2.22] (Neuron 2.22.0 Release) -------------------------------------------------------------------- Date: 04/03/2025 Improvements ~~~~~~~~~~~~~~~ * New modules and APIs: * ``nki.profile`` * ``nki.isa`` new APIs: * ``tensor_copy_dynamic_dst`` * ``tensor_copy_predicated`` * ``max8``, ``nc_find_index8``, ``nc_match_replace8`` * ``nc_stream_shuffle`` * ``nki.language`` new APIs: ``mod``, ``fmod``, ``reciprocal``, ``broadcast_to``, ``empty_like`` * Improvements: * ``nki.isa.nc_matmul`` now supports PE tiling feature * ``nki.isa.activation`` updated to support reduce operation and ``reduce`` commands * ``nki.isa.engine`` enum * ``engine`` parameter added to more ``nki.isa`` APIs that support engine selection (ie, ``tensor_scalar``, ``tensor_tensor``, ``memset``) * Documentation for ``nki.kernels`` have been moved to the GitHub: https://aws-neuron.github.io/nki-samples. The source code can be viewed at https://github.com/aws-neuron/nki-samples. * These kernels are still shipped as part of Neuron package in ``neuronxcc.nki.kernels`` module * Documentation updates: * Kernels public repository https://aws-neuron.github.io/nki-samples * Updated :doc:`profiling guide ` to use ``nki.profile`` instead of ``nki.benchmark`` * NKI ISA Activation functions table now have :ref:`valid input data ranges` listed * NKI ISA Supported Math operators now have :ref:`supported engine` listed * Clarify ``+=`` syntax support/limitation ---- .. _nki-2-21-0-rn: Neuron Kernel Interface (NKI) (Beta) [2.21] (Neuron 2.21.0 Release) -------------------------------------------------------------------- Date: 12/16/2024 Improvements ~~~~~~~~~~~~~~~ * New modules and APIs: * ``nki.compiler`` module with Allocation Control and Kernel decorators, see guide for more info. * ``nki.isa``: new APIs (``activation_reduce``, ``tensor_partition_reduce``, ``scalar_tensor_tensor``, ``tensor_scalar_reduce``, ``tensor_copy``, ``tensor_copy_dynamic_src``, ``dma_copy``), new activation functions(``identity``, ``silu``, ``silu_dx``), and target query APIs (``nc_version``, ``get_nc_version``). * ``nki.language``: new APIs (``shared_identity_matrix``, ``tan``, ``silu``, ``silu_dx``, ``left_shift``, ``right_shift``, ``ds``, ``spmd_dim``, ``nc``). * New ``datatype ``: ``float8_e5m2`` * New ``kernels`` (``allocated_fused_self_attn_for_SD_small_head_size``, ``allocated_fused_rms_norm_qkv``) added, kernels moved to public repository. * Improvements: * Semantic analysis checks for nki.isa APIs to validate supported ops, dtypes, and tile shapes. * Standardized naming conventions with keyword arguments for common optional parameters. * Transition from function calls to kernel decorators (``jit``, ``benchmark``, ``baremetal``, ``simulate_kernel``). * Documentation updates: * Tutorial for :doc:`SPMD usage with multiple Neuron Cores on Trn2 ` ---- .. _nki-2-20-1-rn: Neuron Kernel Interface (NKI) (Beta) (Neuron 2.20.1 Release) ------------------------------------------------------------- Date: 12/03/2024 Improvements ~~~~~~~~~~~~~~~ * NKI support for Trainium2, including full integration with Neuron Compiler. Users can directly shard NKI kernels across multiple Neuron Cores from an SPMD launch grid. See :doc:`tutorial ` for more info. See :doc:`Trainium2 Architecture Guide ` for an initial version of the architecture specification (more details to come in future releases). * New calling convention in NKI kernels, where kernel output tensors are explicitly returned from the kernel instead of pass-by-reference. See any :doc:`NKI tutorial ` for code examples. ---- .. _nki-2-20-0-rn: Neuron Kernel Interface (NKI) (Beta) [2.20] (Neuron 2.20.0 Release) -------------------------------------------------------------------- Date: 09/16/2024 Improvements ~~~~~~~~~~~~~~~ * This release includes the beta launch of the Neuron Kernel Interface (NKI) (Beta). NKI is a programming interface enabling developers to build optimized compute kernels on top of Trainium and Inferentia. NKI empowers developers to enhance deep learning models with new capabilities, performance optimizations, and scientific innovation. It natively integrates with PyTorch and JAX, providing a Python-based programming environment with Triton-like syntax and tile-level semantics offering a familiar programming experience for developers. Additionally, to enable bare-metal access precisely programming the instructions used by the chip, this release includes a set of NKI APIs (``nki.isa``) that directly emit Neuron Instruction Set Architecture (ISA) instructions in NKI kernels.