Configuration & Hardware¶

Importable from syssim. These back the lower-level cost model; most users use High-level API instead.

get_hardware_info()¶

Auto-detect the current GPU and return (HardwareInfo, name). Raises RuntimeError if CUDA is unavailable or the device is unrecognized. Known devices include gh200, h100, a100, v100, a40, rtx4090, mi250, mi300, and pro6000.

hw_info, hw_name = syssim.get_hardware_info()
print(hw_name, hw_info.peak_tflops_mm)

class HardwareInfo(peak_tflops_mm, peak_tflops_math, peak_memory_bandwidth_gbps, peak_tflops_mm_conservative=None, peak_tflops_mm_fp8=None, peak_tflops_mm_fp4=None, network=None, estimator=None, sfu_peak=None, calibrated_model=None)¶

Hardware specifications for the roofline model. Peak FLOP rates are in TFLOP/s (10¹²), memory bandwidth in GB/s (10⁹). peak_tflops_mm_conservative is used for small GEMM/ATTN ops where launch overhead dominates (defaults to peak_tflops_mm).

build_estimator()¶: Return the resolved per-op estimator (cached): a TreeEstimator if calibrated_model was given, else an explicitly attached estimator, else the default RooflineEstimator.

get_peak_tflops(op_type, dtype, is_large_op=False)¶: Select the peak FLOP/s for an operator type and dtype.

get_peak_tflops_mm_for_dtype(dtype)¶: Matrix-unit peak for a dtype; FP8 dtypes use peak_tflops_mm_fp8, falling back to peak_tflops_mm when unset.

class NetworkParams¶: Network hardware parameters for the communication simulator. Single-node: nvlink_bandwidth (bytes/s, default 25e9), nvlink_count (default 12), ib_bandwidth (default 25e9). Multi-node: num_nodes (default 1), gpus_per_node (default 8). LogGP latency/overhead: loggp_nvlink_L, loggp_nvlink_o, loggp_ib_L, loggp_ib_o (seconds).

class SimulatorConfig¶: Bundle of hw_info (HardwareInfo) and cache_seq_len (int, default 0).

class ExecutionMode¶: Enum of execution modes: TRAINING (forward + backward), PREFILL (forward only, full sequence), DECODE (forward only, seq_len=1, KV-cache read).