Docs
Browse documentation

Docs / Guides

Inspect and benchmark

Read the released container directly and measure decode performance on the active machine.

Inspect a model artifact

shell
fermion inspect
fermion inspect --layer 0 --projection q
fermion inspect --sha256 --json

inspect reports geometry, parameters by tensor class, projection weight-state occupancy, byte use, per-layer occupancy spread, and KV-cache cost. --json makes the result machine-readable.

FlagEffect
--layer NRestrict inspection to one layer and skip the rest of the file.
--projection TYPESelect q, k, v, o, gate, up, or down.
--sha256Add a full-file digest.
--jsonReturn structured output.

Benchmark the current machine

shell
fermion bench
fermion bench --tokens 128 --warmup 2
fermion bench --draft auto

The benchmark prints detected hardware, model size, load time, wall throughput, and steady-state decode throughput separately. With a draft, it adds a drafted row and an output-identity check.

FlagDefaultPurpose
--tokens32Greedy generation window.
--promptfixed promptReplace the benchmark prompt.
--warmup1Warm-up generations before measurement.
--draftnoneMeasure speculative decoding too.