CompilerConfig Reference

Complete property listing for axelera.compiler.CompilerConfig. For usage examples and the compilation workflow, see Compiler Python API.

from axelera.compiler import CompilerConfig

config = CompilerConfig()
config.aipu_cores_used = 4
config.multicore_mode = "batch"

Properties

Quantization

Property	Type	Default	Description
`compiler_mode`	`CompilerMode`	`"quantize_and_lower"`	Controls what the compiler does: quantize only, lower only, or both.
`quantization_scheme`	`QuantizationScheme`	`"per_tensor_histogram"`	Quantization algorithm for activations and weights.
`quantize_dw_channel_wise`	boolean	`false`	Quantize depthwise convolutions channel-wise instead of per-tensor.
`onnx_opset_version`	integer ≥ 17	`17`	ONNX opset version used during PyTorch → ONNX conversion.
`quantization_debug`	boolean	`false`	Dump model after quantization for accuracy debugging.
`remove_io_quantization`	boolean	`false`	Remove quantize/dequantize ops from graph inputs and outputs.
`quantized_graph_export`	boolean	`true`	Export the quantized graph to a JSON file.

Graph transformations

Property	Type	Default	Description
`apply_pword_padding`	boolean	`true`	Pad input channels to a multiple of PWORD (hardware requirement).
`rewrite_concat_to_resadd`	boolean	`true`	Convert concatenation ops to pad+shift+add (no native concat on hardware).
`rewrite_dense_to_conv2d`	boolean	`true`	Canonicalise `nn.dense` to `conv2d` with 1×1 kernel.
`remove_io_padding_and_layout_transform`	boolean	`true`	Remove padding/transpose ops from the graph when the SDK pipeline handles them. Leave `true` unless running without the SDK preprocessing elements.
`apply_arithmetic_simplification`	boolean	`true`	Fast-math simplifications. May not preserve numerics exactly.
`simplify_mac_before_after_lut`	boolean	`true`	Optimize multiply-add ops around LUT operations. May not preserve numerics exactly.
`validate_operators`	boolean	`true`	Check operator compatibility with hardware before compilation.

Graph cleaner

Property	Type	Default	Description
`run_graph_cleaner`	boolean	`true`	Run the ONNX graph cleaner to remove pre/post-processing ops that belong on the host.
`graph_cleaner_split_pre_post_processing`	boolean	`true`	Split pre- and post-processing into separate passes in the cleaner.
`graph_cleaner_condition`	`GraphCleanerCondition` or null	`null`	Condition used to identify the split point.
`graph_cleaner_node`	`GraphCleanerNode` or null	`null`	Node type the condition applies to.
`graph_cleaner_threshold`	integer	`0`	Threshold value for the condition.
`graph_cleaner_dump_core_onnx`	string or null	`null`	Filename to save the core ONNX model after cleaning (for debugging).
`graph_cleaner_dump_full_opt_onnx`	string or null	`null`	Filename to save the full optimized ONNX model after cleaning.
`remove_layout_transform_from_preamble`	boolean	`false`	Remove transpose/reshape from model preamble for NHWC layouts.

Multi-core and scheduling

Property	Type	Default	Description
`aipu_cores_used`	integer 1–4	`1`	Number of AIPU cores to compile for.
`multicore_mode`	`MulticoreMode`	`"multiprocess"`	How cores share work. See Multi-Core Configuration.
`pipeline_spatial_tiles`	boolean	`true`	Software-pipeline tasks across height tiles.
`pipeline_channel_tiles`	boolean	`true`	Software-pipeline tasks across output channel tiles.
`inter_operator_async`	boolean	`true`	Asynchronous scheduling between sequential operators.
`use_list_scheduler`	boolean	`false`	Use resource-aware list scheduler for async scheduling.
`unroll_prologue_epilogue`	boolean	`false`	Unroll prologue/epilogue loops to expose more parallelism.
`use_hw_tokens`	boolean	`true`	Use hardware tokens for synchronisation.
`group_ifdw_tasks`	boolean	`false`	Hoist IFDW operations earlier for better IMC weightset utilization.
`double_buffer`	boolean	`true`	Double-buffer host↔device data transfers to hide latency.
`host_processes_used`	integer	`1`	Number of host processes for execution.

Tiling and memory planning

Property	Type	Default	Description
`max_memplan_attempts`	integer ≥ 1	`5`	Max iterations for the memory planner to find a valid configuration.
`max_tiling_attempts`	integer ≥ 1	`8`	Max attempts to tile an operator to fit in memory.
`force_h_tiling`	integer or null	`null`	Force a specific height-dimension tiling factor.
`force_oc_tiling`	integer or null	`null`	Force a specific output-channel tiling factor.
`tiling_depth`	integer ≥ 1	`1`	Max tiling depth. Set to `6` to enable depth-first scheduling; `1` disables it.
`dfs_search_constraint`	integer or null	`null`	Limit depth-first search space. Set to `1` for large networks.

Memory hierarchy

Property	Type	Default	Description
`enable_buffer_promotion`	boolean	`true`	Allow buffers to be promoted from DDR → L2 → L1 for faster access.
`split_buffer_promotion`	boolean	`false`	Split promotion into two passes (L1 and L2) with DFS scheduling in between.
`io_memory_pool`	string	`"global.ddr"`	Initial home for I/O buffers. One of: `global.ddr`, `global.l2`, `global.l1`.
`constant_memory_pool`	string	`"global.ddr"`	Initial home for constant buffers.
`workspace_memory_pool`	string	`"global.ddr"`	Initial home for workspace buffers.
`l1_constraint`	integer or null	`null`	Max L1 memory the compiler may use (bytes).
`l2_constraint`	integer or null	`null`	Max L2 memory the compiler may use (bytes).
`ddr_constraint`	integer or null	`null`	Max DDR memory the compiler may use (bytes).
`l1_size_used`	integer	`4194304`	L1 memory available to the compiler (bytes). Default: 4 MB.
`l2_size_used`	integer	`33554432`	L2 memory available to the compiler (bytes). Default: 32 MB.
`ddr_size_used`	integer	`1073741824`	DDR memory available to the compiler (bytes). Default: 1 GB.
`use_sysdma`	boolean	`false`	Use System DMA for DDR→L2 weight transfers (Core DMA handles L2→L1).
`dma_dual_channel`	boolean	`true`	Enable dual-channel DMA optimization.
`page_memory`	boolean	`true`	Apply paging to reduce L2/DDR fragmentation.
`elf_in_ddr`	boolean	`true`	Store ELF file in DDR rather than L2.
`stream_tasklist`	boolean	`true`	Stream tasklist chunks into AIPU memory on-the-fly.
`dpu_constants_home`	string	`"global.l2"`	Where DPU constants reside. One of: `global.ddr`, `global.l2`.

IMC and MVM

Property	Type	Default	Description
`mvm_utilization_limit`	float 0.125–1.0	`1.0`	Fraction of MVM array active MACs. Reduce to lower power consumption.
`enable_icr`	boolean	`true`	In-Core Replication for layers with small output-channel counts.
`icrx_force`	integer	`-1`	Force a specific ICR factor. `-1` = automatic.
`icrx_parallel_block_threshold`	integer 1–4	`4`	Maximum number of parallel IMC blocks for which ICR is still applied.
`icrx_max_factor`	integer 2–8	`8`	Maximum ICR factor along the X image direction.
`enable_swicr`	boolean	`true`	Subword ICR for first layers with small input-channel counts.
`imc_double_buffer_pipeline`	boolean	`false`	Double-buffer weight loading in software-pipelined sections.
`imc_double_buffer_sequential`	boolean	`false`	Double-buffer weight loading in sequential IR sections.
`dpu_allocation_algorithm`	`DPUAllocationAlgorithm`	`"try_all"`	Register-allocation algorithm for the DPU vector unit.
`softmax_neutral_value`	float	`-100000.0`	Padding value for softmax inputs. Chosen so the softmax LUT outputs zero for padding elements without affecting non-padding numerics.

Output and paths

Property	Type	Default	Description
`output_dir`	path string	—	Directory for compiler outputs and deployment artifacts.
`remove_output_dir`	boolean	`false`	Clean the output directory before compilation.
`model_name`	string	`""`	Model name used in logging and to select model-specific optimization presets.
`save_error_artifact`	boolean	`false`	Save a ZIP archive with the lowered model and error messages on failure.
`randomize_onnx_model`	boolean	`true`	Randomise cached ONNX model weights before saving the error artifact.

Hardware and runtime

Property	Type	Default	Description
`frequency`	integer 20 MHz–800 MHz	`800000000`	Device clock frequency in Hz.
`resources_used`	float (0, 1]	`1.0`	Fraction of memory resources the compiled model may use.
`input_dmabuf`	boolean	`false`	Use DMA for input data transfer.
`output_dmabuf`	boolean	`false`	Use DMA for output data transfer.

Profiling and debugging

Property	Type	Default	Description
`profiling_levels`	list of `ProfilingLevel`	`[]`	Profiling trace levels to enable. Enabling multiple simultaneously can reduce accuracy.
`profiling_drop_percentile`	float	`0.25`	Drop top and bottom N% of profiling samples to remove outliers.
`trace_tvm_passes`	boolean	`false`	Trace TVM pass execution (start/end times, parent-child relationships). Output: `pass_dependency_graph.json`.
`propagate_span_information`	boolean	`true`	Propagate source-span information through the compiler.
`model_debug_save_dir`	path string	—	Directory to save the quantized/optimized model for debugging.
`quantization_debug`	boolean	`false`	Dump the quantized model for accuracy measurement and debugging.

Advanced — internal paths

These are resolved automatically by the SDK. Override only if your installation is non-standard or you are running the compiler outside the SDK environment.

Property	Type	Default	Description
`compiler_dir`	path string	—	Root directory of the compiler package, used to locate internal resources and dependencies.
`runtime_dir`	path string	—	Directory for the Axelera runtime.
`device_dir`	path string	—	Directory for device resources.

Advanced — memory layout constants

Hardware memory map values. Do not change unless directed by Axelera support — incorrect values will prevent models from running.

Property	Type	Default	Description
`l1_size_reserved`	integer	`524288`	L1 memory reserved for the system (bytes). Default: 512 KB.
`l2_size_reserved`	integer	`1245184`	L2 memory reserved for the system (bytes).
`l2_size_reserved_tasklist`	integer	`1048576`	L2 memory reserved per core for the tasklist (bytes). Default: 1 MB.
`ddr_size_max`	integer	`1073741824`	Total DDR memory size (bytes). Default: 1 GB.
`ddr_size_reserved`	integer	`33554432`	DDR memory reserved for the system (bytes). Default: 32 MB.
`l1_virtual_address`	integer	`206175207424`	Virtual address of L1 memory. Copied from `mmap_config.h`.
`l1_core0_physical_address`	integer	`402653184`	Physical address of L1 memory in core 0. Copied from `memorymap.h`. Required because the AIPU simulator does not support virtual memory.
`dpu_instructions_home`	string	`"default"`	Where DPU instructions are placed. One of: `"default"`, `"l2"`.
`dwpu_instructions_home`	string	`"default"`	Where DWPU instructions are placed. One of: `"default"`, `"l2"`.
`ignore_weight_buffers`	boolean	`true`	Exclude weight buffers when determining tiling factors during memory scheduling.

Enum definitions

`CompilerMode`

Value	Description
`"quantize_and_lower"`	Quantize the model then compile to hardware binary (default).
`"quantize_only"`	Quantize only — produces a model that runs on CPU for accuracy validation.
`"lower_only"`	Compile a pre-quantized model to hardware binary without re-quantizing.

`MulticoreMode`

Value	Description
`"multiprocess"`	Each core runs a separate OS process (default).
`"multithread"`	Cores share a process with multiple threads.
`"batch"`	Different items in a batch run on different cores simultaneously.
`"cooperative"`	Cores cooperate on a single inference.
`"pipeline"`	Model is split across cores as a pipeline stage.

See Multi-Core Configuration for when to use each mode.

`QuantizationScheme`

Value	Description
`"per_tensor_histogram"`	Activations quantized per-tensor with histogram observer; weights per-channel with min-max (default).
`"per_tensor_min_max"`	Activations quantized per-tensor with min-max observer; weights per-channel with min-max.
`"hybrid_per_tensor_per_channel"`	Activations per-tensor (histogram), except depth-wise convolution inputs (per-channel min-max); weights per-channel min-max.

`DPUAllocationAlgorithm`

Value	Description
`"try_all"`	Try all algorithms in sequence until one succeeds (default).
`"graph"`	Graph-coloring register allocator.
`"lazy"`	Lazy (greedy) allocator.
`"backjump_recursive"`	Backtracking recursive allocator.

`GraphCleanerCondition`

Value	Description
`"maximum_weight_tensor_size"`	Split on the node with the largest weight tensor.
`"maximum_weight_tensor_first_dimension_size"`	Split on the node with the largest first weight dimension.

`GraphCleanerNode`

Value	Description
`"MatMul"`	Apply the condition to MatMul nodes.
`"Gemm"`	Apply the condition to Gemm nodes.
`"Clip"`	Apply the condition to Clip nodes.

`ProfilingLevel`

Trace-line type identifiers used in the profiling output file.

[?] · [B] · [PB] · [PE] · [K] · [M] · [T]

Properties​

Quantization​

Graph transformations​

Graph cleaner​

Multi-core and scheduling​

Tiling and memory planning​

Memory hierarchy​

IMC and MVM​

Output and paths​

Hardware and runtime​

Profiling and debugging​

Advanced — internal paths​

Advanced — memory layout constants​

Enum definitions​

CompilerMode​

MulticoreMode​

QuantizationScheme​

DPUAllocationAlgorithm​

GraphCleanerCondition​

GraphCleanerNode​

ProfilingLevel​