Struct BenchServeCommand 

pub struct BenchServeCommand {

    pub base_url: String,
    pub model: String,
    pub tokenizer: PathBuf,
    pub target_backend: Option<BenchTargetBackend>,
    pub http_connection_mode: BenchHttpConnectionMode,
    pub concurrency: u32,
    pub concurrency_sweep: Vec<u32>,
    pub request_rate: Option<f64>,
    pub dataset: String,
    pub random_input_len: usize,
    pub random_output_len: usize,
    pub ignore_eos: bool,
    pub enable_thinking: Option<bool>,
    pub sharegpt_path: Option<PathBuf>,
    pub shared_prefix_len: usize,
    pub shared_suffix_len: usize,
    pub num_prompts: u32,
    pub warmup_requests: u32,
    pub n_repeats: u32,
    pub goodput: Option<Slo>,
    pub timeout: f64,
    pub fail_on_error: bool,
    pub max_error_rate: Option<f64>,
    pub require_ci: bool,
    pub seed: Option<u64>,
    pub output: String,
    pub out: Option<PathBuf>,
    pub hw_id: Option<String>,
    pub commit_sha: Option<String>,
    pub tag: Option<String>,
}

Fields

base_url: String

Base URL of the ferrum (or other OpenAI-compatible) server.

model: String

Model identifier sent in request body’s model field. Use the local path so vllm bench serve numbers compare 1:1.

tokenizer: PathBuf

Path to the model directory containing tokenizer.json. Used to generate exact-length random token sequences.

target_backend: Option<BenchTargetBackend>

Backend of the server under test. This is independent of the HTTP client’s own compile-time accelerator features.

http_connection_mode: BenchHttpConnectionMode

HTTP connection lifecycle used by the benchmark client.

concurrency: u32

Closed-loop concurrency (single cell). Default when no other mode is given. Alias: --max-concurrency (legacy vLLM naming).

concurrency_sweep: Vec<u32>

Closed-loop concurrency sweep. Overrides --concurrency. E.g. --concurrency-sweep 1,4,16,32 runs four closed-loop cells.

request_rate: Option<f64>

Open-loop arrival rate (req/s, Poisson). When set, overrides the closed-loop modes — this is the goodput-relevant scenario.

dataset: String

Dataset: random (tokenizer-aware), sharegpt (load from JSONL), shared-prefix (1024-tok shared prefix + unique suffix). PLAYBOOK § 2 Scenario A (sharegpt) / Scenario C (shared-prefix).

random_input_len: usize

Number of tokens per random prompt (--dataset random only).

random_output_len: usize

Max output tokens per request.

ignore_eos: bool

Send vLLM-compatible ignore_eos=true so fixed-output benchmark requests run until max_tokens instead of stopping on model EOS.

enable_thinking: Option<bool>

Typed chat-template thinking control. Omitted from payloads unless set.

sharegpt_path: Option<PathBuf>

Path to a ShareGPT-format JSONL file (--dataset sharegpt). Each line should be a {"conversations": [{"from": "...", "value": "..."}, ...]} object (HF anon8231489123/ShareGPT_Vicuna format).

shared_prefix_len: usize

Shared prefix length in tokens (--dataset shared-prefix only).

shared_suffix_len: usize

Per-request unique suffix length in tokens (--dataset shared-prefix).

num_prompts: u32

Total prompts sent per run (warmup is counted separately).

warmup_requests: u32

Warmup requests sent before measurement begins each run. Discarded from the metrics; PLAYBOOK § 0.3 mandates ≥ 10 for committed reports.

n_repeats: u32

Independent repeats per cell. PLAYBOOK § 0.4: ≥ 3 unlocks stddev + CI95; n < 3 emits mean only.

goodput: Option<Slo>

SLO triple for goodput. Format: ttft:500 tpot:50 e2el:30000 (or comma-separated). Goodput is reported only when all three are set.

timeout: f64

Per-request HTTP timeout in seconds.

fail_on_error: bool

Exit non-zero when any measured request errors. Warmup failures always fail. Release gates must set this.

max_error_rate: Option<f64>

Maximum measured request error rate allowed when error enforcement is active.

require_ci: bool

Require n_repeats >= 3 so reports include CI/stddev evidence.

seed: Option<u64>

Deterministic prompt-generation seed. Repeat i uses a stable derivation.

output: String

Output format: json (BenchReport), jsonl (append one BenchReport per line — used by scripts/compare-commits.sh), md (human-readable markdown).

out: Option<PathBuf>

Output file path. For jsonl, the file is opened append-mode so repeated invocations against the same DB accumulate rows. Alias: --result-file (legacy).

hw_id: Option<String>

Override env.hw_id (defaults to auto-detected).

commit_sha: Option<String>

Override env.commit_sha (defaults to git rev-parse --short HEAD).

tag: Option<String>

Tag string written into the report’s model field’s suffix — useful for the bench_vs_vllm.sh script which tags ferrum/vllm.

Trait Implementations

impl Args for BenchServeCommand

fn group_id() -> Option<Id>

Report the ArgGroup::id for this set of arguments

fn augment_args<'b>(__clap_app: Command) -> Command

Append to Command so it can instantiate Self via FromArgMatches::from_arg_matches_mut

fn augment_args_for_update<'b>(__clap_app: Command) -> Command

Append to Command so it can instantiate self via FromArgMatches::update_from_arg_matches_mut

impl Clone for BenchServeCommand

fn clone(&self) -> BenchServeCommand

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl FromArgMatches for BenchServeCommand

fn from_arg_matches(__clap_arg_matches: &ArgMatches) -> Result<Self, Error>

Instantiate Self from ArgMatches, parsing the arguments as needed.

fn from_arg_matches_mut( __clap_arg_matches: &mut ArgMatches, ) -> Result<Self, Error>

Instantiate Self from ArgMatches, parsing the arguments as needed.

fn update_from_arg_matches( &mut self, __clap_arg_matches: &ArgMatches, ) -> Result<(), Error>

Assign values from ArgMatches to self.

fn update_from_arg_matches_mut( &mut self, __clap_arg_matches: &mut ArgMatches, ) -> Result<(), Error>

Assign values from ArgMatches to self.