API reference
AxoSim CLI
Complete options and defaults for axosim.
Command contract
The tables include parser defaults, choices, required arguments, flag actions, and compatibility options hidden from ordinary help. Every command and subcommand accepts -h or --help. Flag defaults refer to their destination value: store_true sets it true, while store_false sets it false. For example, --blocking-transfer changes non_blocking from its default true to false. Named presets may override parser defaults; explicitly supplied options take precedence. Use the installed command’s --help when working with another revision.
axosim make-demo-data
write deterministic demo NPZ shards
| Argument | Type/action | Default | Required | Choices |
|---|---|---|---|---|
--output |
str / store |
None |
yes | — |
--samples |
int / store |
8 |
no | — |
--shards |
int / store |
1 |
no | — |
--time-steps |
int / store |
32 |
no | — |
--input-dim |
int / store |
64 |
no | — |
--seed |
int / store |
0 |
no | — |
axosim convert-neuronio
convert raw NeuronIO pickle files to deterministic shards
| Argument | Type/action | Default | Required | Choices |
|---|---|---|---|---|
--input |
str / store |
None |
yes | — |
--output |
str / store |
None |
yes | — |
--shard-size |
int / store |
128 |
no | — |
--window-size |
int / store |
None |
no | — |
--window-stride |
int / store |
None |
no | — |
--ignore-start |
int / store |
0 |
no | — |
axosim repack-shards
repack shards as sliceable .npy arrays
| Argument | Type/action | Default | Required | Choices |
|---|---|---|---|---|
--input |
str / store |
None |
yes | — |
--output |
str / store |
None |
yes | — |
--input-dtype |
str / store |
'int8' |
no | ['int8', 'float32'] |
axosim list-presets
list named train/evaluate presets
No additional arguments are required; this command prints the named recipes as JSON.
axosim evaluate
evaluate a model with the corrected full-trace metric path
| Argument | Type/action | Default | Required | Choices |
|---|---|---|---|---|
--preset |
str / store |
None |
no | ['fulltrace-paper-metric', 'legacy-baseline-smoke'] |
--data |
str / store |
None |
yes | — |
--output |
str / store |
None |
yes | — |
--input-dim |
int / store |
1278 |
no | — |
--memory-units |
int / store |
30 |
no | — |
--branches |
int / store |
32 |
no | — |
--model-kind |
str / store |
'baseline' |
no | ['baseline', 'official', 'axomamba', 'mamba-official', 'mamba-pytorch', 'branch-mamba-pytorch', 'branch-trace-rnn'] |
--model-config |
str / store |
'configs/branch_elm_30_official.json' |
no | — |
--batch-size |
int / store |
8 |
no | — |
--seed |
int / store |
0 |
no | — |
--device |
str / store |
'cpu' |
no | — |
--checkpoint |
str / store |
None |
no | — |
--window-size |
int / store |
None |
no | — |
--window-stride |
int / store |
None |
no | — |
--ignore-start |
int / store |
0 |
no | — |
--cache-shards |
int / store |
1 |
no | — |
--soma-units |
str / store |
'millivolts' |
no | ['millivolts', 'normalized'] |
--metric-ignore-start |
int / store |
0 |
no | — |
--metric-mask-mode |
str / store |
'ignore-start' |
no | ['ignore-start', 'official-overlap'] |
--metric-stitch-burn-in |
int / store |
150 |
no | — |
--soma-affine-calibration |
flag / store_true |
False |
no | — |
--pin-memory |
flag / store_true |
False |
no | — |
--blocking-transfer |
flag / store_false |
True |
no | — |
--registry |
str / store |
None |
no | — |
--no-registry |
flag / store_const |
None |
no | — |
--model-kind: use ‘official’ for the paper-style model, ‘axomamba’ for the promoted Mamba surrogate, ‘mamba-official’ for the generic official mamba-ssm backend, or ‘branch-trace-rnn’ for the RNN comparison path; ‘baseline’ is legacy/debug
--metric-mask-mode: Compatibility option hidden from default help.
--metric-stitch-burn-in: Compatibility option hidden from default help.
--registry: append a one-line JSONL record; defaults to
--no-registry: When selected, sets registry to 'off'.
axosim diagnose
write biology-oriented surrogate fidelity diagnostics
| Argument | Type/action | Default | Required | Choices |
|---|---|---|---|---|
--preset |
str / store |
None |
no | ['fulltrace-paper-metric', 'legacy-baseline-smoke'] |
--data |
str / store |
None |
yes | — |
--output-dir |
str / store |
None |
yes | — |
--input-dim |
int / store |
1278 |
no | — |
--memory-units |
int / store |
30 |
no | — |
--branches |
int / store |
32 |
no | — |
--model-kind |
str / store |
'baseline' |
no | ['baseline', 'official', 'axomamba', 'mamba-official', 'mamba-pytorch', 'branch-mamba-pytorch', 'branch-trace-rnn'] |
--model-config |
str / store |
'configs/branch_elm_30_official.json' |
no | — |
--batch-size |
int / store |
8 |
no | — |
--seed |
int / store |
0 |
no | — |
--device |
str / store |
'cpu' |
no | — |
--checkpoint |
str / store |
None |
no | — |
--window-size |
int / store |
None |
no | — |
--window-stride |
int / store |
None |
no | — |
--ignore-start |
int / store |
0 |
no | — |
--cache-shards |
int / store |
1 |
no | — |
--metric-ignore-start |
int / store |
0 |
no | — |
--metric-mask-mode |
str / store |
'ignore-start' |
no | ['ignore-start', 'official-overlap'] |
--metric-stitch-burn-in |
int / store |
150 |
no | — |
--soma-affine-calibration |
flag / store_true |
False |
no | — |
--max-samples |
int / store |
None |
no | — |
--pin-memory |
flag / store_true |
False |
no | — |
--blocking-transfer |
flag / store_false |
True |
no | — |
--metric-mask-mode: Compatibility option hidden from default help.
--metric-stitch-burn-in: Compatibility option hidden from default help.
axosim train
train a model with the current experiment workflow
| Argument | Type/action | Default | Required | Choices |
|---|---|---|---|---|
--preset |
str / store |
None |
no | ['axomamba-probe', 'legacy-baseline-smoke', 'mamba-official-probe', 'mamba2-official-probe', 'official-fulltrace-reference', 'paper-random-fulltrace'] |
--data |
str / store |
None |
yes | — |
--checkpoint |
str / store |
None |
yes | — |
--metrics |
str / store |
None |
yes | — |
--input-dim |
int / store |
1278 |
no | — |
--memory-units |
int / store |
30 |
no | — |
--branches |
int / store |
32 |
no | — |
--model-kind |
str / store |
'baseline' |
no | ['baseline', 'official', 'axomamba', 'mamba-official', 'mamba-pytorch', 'branch-mamba-pytorch', 'branch-trace-rnn'] |
--model-config |
str / store |
'configs/branch_elm_30_official.json' |
no | — |
--epochs |
int / store |
1 |
no | — |
--batch-size |
int / store |
8 |
no | — |
--learning-rate |
float / store |
0.0005 |
no | — |
--burn-in |
int / store |
0 |
no | — |
--seed |
int / store |
0 |
no | — |
--device |
str / store |
'cpu' |
no | — |
--window-size |
int / store |
None |
no | — |
--window-stride |
int / store |
None |
no | — |
--ignore-start |
int / store |
0 |
no | — |
--window-sampling |
str / store |
'random-full-trace' |
no | ['deterministic', 'random-full-trace', 'official-full-trace'] |
--epoch-samples |
int / store |
None |
no | — |
--shard-reuse-batches |
int / store |
1 |
no | — |
--file-load-fraction |
float / store |
0.3 |
no | — |
--full-trace-length |
int / store |
None |
no | — |
--sample-window-reads |
flag / store_true |
False |
no | — |
--shuffle-mode |
str / store |
'shard' |
no | ['sample', 'shard', 'none'] |
--cache-shards |
int / store |
1 |
no | — |
--val-data |
str / store |
None |
no | — |
--val-batch-size |
int / store |
None |
no | — |
--val-cache-shards |
int / store |
1 |
no | — |
--val-window-size |
int / store |
None |
no | — |
--val-window-stride |
int / store |
None |
no | — |
--val-ignore-start |
int / store |
0 |
no | — |
--val-soma-units |
str / store |
'millivolts' |
no | ['millivolts', 'normalized'] |
--val-metric-ignore-start |
int / store |
500 |
no | — |
--val-metric-mask-mode |
str / store |
'ignore-start' |
no | ['ignore-start', 'official-overlap'] |
--val-metric-stitch-burn-in |
int / store |
150 |
no | — |
--val-soma-affine-calibration |
flag / store_true |
True |
no | — |
--no-val-soma-affine-calibration |
flag / store_false |
True |
no | — |
--best-checkpoint |
str / store |
None |
no | — |
--max-train-batches |
int / store |
None |
no | — |
--lr-schedule |
str / store |
'constant' |
no | ['constant', 'cosine'] |
--lr-schedule-steps |
int / store |
None |
no | — |
--optimizer |
str / store |
'adam' |
no | ['adam', 'adamw'] |
--weight-decay |
float / store |
0.0 |
no | — |
--l1-lambda |
float / store |
0.0 |
no | — |
--spike-loss-weight |
float / store |
0.5 |
no | — |
--soma-loss-weight |
float / store |
0.5 |
no | — |
--sparse-soma-loss-weight |
float / store |
0.0 |
no | — |
--sparse-soma-high-voltage-quantile |
float / store |
0.9 |
no | — |
--sparse-soma-high-dvdt-quantile |
float / store |
0.9 |
no | — |
--sparse-soma-input-event-quantile |
float / store |
0.95 |
no | — |
--sparse-soma-spike-window |
int / store |
5 |
no | — |
--sparse-soma-post-event-window |
int / store |
5 |
no | — |
--sera-soma-loss-weight |
float / store |
0.0 |
no | — |
--sera-soma-min-weight |
float / store |
0.05 |
no | — |
--sera-soma-relevance-power |
float / store |
1.0 |
no | — |
--soma-slope-loss-weight |
float / store |
0.0 |
no | — |
--grad-clip-norm |
float / store |
0.0 |
no | — |
--train-update-log-interval |
int / store |
0 |
no | — |
--train-update-log |
str / store |
None |
no | — |
--prefetch-batches |
int / store |
0 |
no | — |
--pin-memory |
flag / store_true |
False |
no | — |
--blocking-transfer |
flag / store_false |
True |
no | — |
--registry |
str / store |
None |
no | — |
--no-registry |
flag / store_const |
None |
no | — |
--model-kind: use ‘official’ for the paper-style model, ‘axomamba’ for the promoted Mamba surrogate, ‘mamba-official’ for the generic official mamba-ssm backend, or ‘branch-trace-rnn’ for the RNN comparison path; ‘baseline’ is legacy/debug
--sample-window-reads: read random/official windows by sample instead of caching full compressed shards
--val-metric-mask-mode: Compatibility option hidden from default help.
--val-metric-stitch-burn-in: Compatibility option hidden from default help.
--registry: append a one-line JSONL record; defaults to
--no-registry: When selected, sets registry to 'off'.
axosim benchmark
time model forward passes on random input
| Argument | Type/action | Default | Required | Choices |
|---|---|---|---|---|
--input-dim |
int / store |
1278 |
no | — |
--memory-units |
int / store |
30 |
no | — |
--branches |
int / store |
32 |
no | — |
--model-kind |
str / store |
'baseline' |
no | ['baseline', 'official', 'axomamba', 'mamba-official', 'mamba-pytorch', 'branch-mamba-pytorch', 'branch-trace-rnn'] |
--model-config |
str / store |
'configs/branch_elm_30_official.json' |
no | — |
--batch-size |
int / store |
1 |
no | — |
--time-steps |
int / store |
500 |
no | — |
--runs |
int / store |
10 |
no | — |
--device |
str / store |
'cpu' |
no | — |
--compile |
flag / store_true |
False |
no | — |
--output |
str / store |
None |
no | — |
--model-kind: use ‘official’ for the paper-style model, ‘mamba-official’ for the official mamba-ssm backend, or ‘branch-trace-rnn’ for the RNN comparison path; ‘baseline’ is legacy/debug
axosim inference-benchmark
benchmark deployment-oriented inference throughput across batch and sequence scales
| Argument | Type/action | Default | Required | Choices |
|---|---|---|---|---|
--input-dim |
int / store |
1278 |
no | — |
--memory-units |
int / store |
30 |
no | — |
--branches |
int / store |
32 |
no | — |
--model-kind |
str / store |
'baseline' |
no | ['baseline', 'official', 'axomamba', 'mamba-official', 'mamba-pytorch', 'branch-mamba-pytorch', 'branch-trace-rnn'] |
--model-config |
str / store |
'configs/branch_elm_30_official.json' |
no | — |
--checkpoint |
str / store |
None |
no | — |
--batch-sizes |
str / store |
'1,8,32,128' |
no | — |
--time-steps |
str / store |
'500,1000,6000' |
no | — |
--runs |
int / store |
20 |
no | — |
--warmup-runs |
int / store |
5 |
no | — |
--precision |
str / store |
'float32' |
no | ['float32', 'float16', 'bfloat16'] |
--device |
str / store |
'cpu' |
no | — |
--compile |
flag / store_true |
False |
no | — |
--accuracy-metrics |
str / store |
None |
no | — |
--output |
str / store |
None |
yes | — |