Command-Line Interface¶
Installing SysSim provides the syssim command. The compute subcommands
(run, memory, summary, sweep) share these flags:
--hardware PATH (required), --tp, --dp, --cp, --sp, --micro-batch,
--global-batch, --dtype {fp16,bf16,fp8}, --recompute {selective,full},
--format {table,json,yaml}. Pipeline parallelism (pp) is Python-API only.
run¶
Full simulation report (step time, MFU, memory, bottlenecks).
syssim run MODEL --hardware HW \
--tp 2 --dp 2 --micro-batch 1 --global-batch 8
memory¶
Peak per-GPU memory only (skips step-time estimation).
syssim memory MODEL --hardware HW --micro-batch 1 --global-batch 8
summary¶
Print the resolved model / hardware / parallelism / training configuration.
syssim summary MODEL --hardware HW --micro-batch 1 --global-batch 8
sweep¶
Sweep one or more config axes and print the best row by a metric.
syssim sweep MODEL --hardware HW --micro-batch 1 --global-batch 8 \
--over parallelism.tp=1,2,4 --metric mfu
--over path=v1,v2,... may be repeated; --metric defaults to mfu.
profile¶
Build the inputs for a calibrated estimator (see Calibrated Estimator). Two modes:
# Layer profiling (needs GPUs): writes <out>/profile.parquet
syssim profile --out data/gh200 --num-workers 4
syssim profile --dry-run # preview the job list, no GPU
# Network profiling: measures NCCL collectives, derives topology
syssim profile --out data/gh200 --network --gpus-per-node 4
calibrate¶
Fit per-family residual trees from profile.parquet (CPU-only).
syssim calibrate --data data/gh200 --hardware HW
Writes gemm_model.lgb, elementwise_model.lgb, reduction_model.lgb, and
manifest.json into the data directory, ready to pass as calibrated_model= in
HardwareConfig.