Crate ferrum_testkit 

Testing utilities for the ferrum inference engine.

• Mocks: MockModelExecutor, MockSampler, MockKvCacheManager, MockTokenizer
• Configurable: ConfigurableModelExecutor (specific token sequences, EOS)
• Bench: BenchmarkResult, percentile calculation, JSON output
• Paged: PagedAttentionExecutor with real paged KV cache

All components are hardware-independent (CPU-only, no GPU required).

Re-exports

pub use paged_executor::PagedAttentionExecutor;

pub use paged_executor::PagedExecutorConfig;

Modules

bench

Benchmark result types and utilities.

paged_executor

Model executor that uses PagedAttention KV cache.

Structs

ConfigurableModelExecutor

Model executor that produces a configurable sequence of tokens.

MockKvCacheHandle

Mock KV cache handle — tracks block metadata without allocating real memory.

MockKvCacheManager

Mock KV cache manager — tracks allocations in memory, simulates block limits.

MockModelExecutor

Mock model executor that simulates prefill/decode with configurable latency. No model weights, no GPU — pure async simulation.

MockSampler

Greedy sampler — always picks the token with highest logit. Deterministic, no temperature or top-k.

MockTensor

A mock tensor that stores shape and optional f32 data. No GPU, no Candle — pure Rust.

MockTensorFactory

Mock tensor factory implementing TensorFactory without any ML backend.

MockTokenizer

Mock tokenizer: splits on whitespace, assigns sequential token IDs. EOS token is vocab_size - 1.