Configuration

A scenario is one TOML tree that names every component and its parameters. Nesting mirrors ownership: a system owns its runner, model, and scheduler, or its child systems, and a runner owns its compute and network simulators.

Every polymorphic table carries a kind that selects a registered implementation. The remaining fields configure that implementation, and unknown fields are rejected. Hardware and serving parameters are required, apart from a few optional component fields such as cudagraph_capture_sizes. The only policy default is the round-robin router.

Example

The bundled single-GPU scenario (examples/scenarios/1xrtx-pro-6000--llama3_1_8b--realistic/simthesizer.toml):

[system]
kind = "single_instance"

[system.model_runner]
kind = "ring"
tp = 1
link_bw = 1.0 # Unused at TP1.
link_latency = 0.0

[system.model_runner.compute_sim]
kind = "table_lookup"
perf_dir = "../../data/rtx-pro-6000/llama3_1_8b.jsonl"

[system.model_runner.compute_sim.components]
layernorm = ["input_norm", "post_attention_norm"]
qkv_projection = ["qkv"]
o_projection = ["output_projection"]
ffn1 = ["gate_up"]
act = ["activation"]
ffn2 = ["down"]
final_layernorm = ["final_norm"]

[system.model_runner.network_sim]
kind = "flow"

[system.scheduler]
kind = "chunked_prefill"
block_size = 16
max_batch = 32
max_num_batched_tokens = 2048
device_mem = 101976440832
device_mem_utilization = 0.9
enable_prefix_caching = false

[system.model]
kind = "llama3_1"
hidden_size = 4096
intermediate_size = 14336
max_position_embeddings = 131072
num_attention_heads = 32
num_hidden_layers = 32
num_key_value_heads = 8
vocab_size = 128256

Relative paths resolve against the config file's directory.

Lifecycle

1. Deserialize builds the component tree without side effects.
2. Resolve paths rewrites relative paths once.
3. Initialize runs top-down. A system reads the runner's parallel strategy, initializes the model with it, sizes the scheduler from the model, then initializes the runner, which loads profiles and registers its topology.
4. Run seeds the workload and starts the engine.

CLI

Flag	Meaning
--config PATH	Scenario TOML
--workload PATH	Workload JSONL
--output-jsonl PATH	Result JSONL
--output-engine-trace PATH	Optional JSONL of component state over simulated time; must not already exist

Source of truth: architectures/configuration.md.