AIConfigurator Rust Core

Phase 1 Rust core for estimating forward-pass latency from AIC model metadata and perf files.

This crate intentionally does not change the existing Python SDK. It gives Rust callers, especially Dynamo Mocker, a reusable estimator that loads metadata once and then serves per-iteration estimates without Python/GIL overhead.

The v1 input is a per-attention-DP-rank list of AIC-owned Rust ForwardPassMetrics, aligned with Dynamo FPM v1. That keeps this first iteration close to existing Dynamo telemetry while avoiding a direct AIC dependency on Dynamo crates.

use aiconfigurator_core::{
    create_engine_step_estimator, BackendKind, DataType, EngineConfig, ForwardPassMetrics,
    ScheduledRequestMetrics, ENGINE_CONFIG_SCHEMA_VERSION,
};

let estimator = create_engine_step_estimator(EngineConfig {
    schema_version: ENGINE_CONFIG_SCHEMA_VERSION,
    model_name: "Qwen/Qwen3-8B".to_string(),
    model_arch: None,
    system_name: "h100_sxm".to_string(),
    backend: BackendKind::Vllm,
    backend_version: Some("0.19.0".to_string()),
    tp_size: 1,
    pp_size: 1,
    moe_tp_size: None,
    moe_ep_size: None,
    attention_dp_size: None,
    weight_dtype: Some(DataType::Bfloat16),
    moe_dtype: None,
    activation_dtype: Some(DataType::Bfloat16),
    kv_cache_dtype: Some(DataType::Bfloat16),
    kv_block_size: None,
    extra: Default::default(),
})?;

let latency = estimator.forward_pass_time(&[ForwardPassMetrics {
    scheduled_requests: ScheduledRequestMetrics {
        num_decode_requests: 1,
        sum_decode_kv_tokens: 2048,
        ..Default::default()
    },
    ..Default::default()
}])?;

Current scope:

• Prefill, decode, and mixed prefill plus decode steps for all model families represented in AIC's checked-in model configs.
• FPM v1 aggregate scheduled request fields per attention-DP rank as the estimator input.
• AIC-style Hugging Face model config JSON files.
• AIC gemm_perf.txt, context_attention_perf.txt, generation_attention_perf.txt, moe_perf.txt, context_mla_perf.txt, and generation_mla_perf.txt CSV files.
• Explicit Git LFS pointer detection for perf data that has not been pulled locally.
• A minimal C ABI used by Python tests to opt into the Rust estimator without requiring PyO3 or changing the default Python SDK path.

Known Phase 1 limits:

• Request-level FPM v2 fields, KV block modeling, exact Python parity for every specialized op, P2P pipeline composition, WideEP accuracy work, and production Python bindings are intentionally left for later PRs.
• Tests use fixture perf files so they can run without the large AIC perf databases.