Skip to content

flide verify

flide verify checks Flide's guarantees on your model, your requests and your GPU. It reads a file of requests, one per line with state and questions, as flide run takes them, and runs three checks.

Check What it compares
fork every question of every request against the same question asked alone, the state and that question only
load every request alone against the same requests shuffled into rounds of 2, 8 and 32, as a loaded server batches them
session with --session-cache-gb, each state grown in four steps, every step answered with the session cache on and off

Each check reports the largest gap between the probabilities, the top answers kept and the answers that keep their bits. A top answer that moves counts against a check only where the reference's top two options are at least 0.05 apart. With --batch-invariant, any bit that moves under load fails it too, since that mode promises the same bits at any load. The command exits with status 1 when a check fails.

flide verify --model Qwen/Qwen3.5-0.8B --requests requests.jsonl --session-cache-gb 4

On 32 requests from the Decision Index, RACE, MultiRC and QuALITY with Qwen3.5-0.8B on an H200 it prints

fork     PASS  105 questions against each asked alone: largest probability gap 3.07e-02, 105/105 top answers kept, 27/105 the same bits
load     PASS  315 answers in rounds of 2, 8, 32 against each request alone: largest probability gap 3.69e-02, 315/315 top answers kept, 215/315 the same bits
session  PASS  384 answers over 4 growing steps, 29952 state tokens resumed: largest probability gap 3.18e-04 against the cache off, 384/384 top answers kept

and with --batch-invariant every gap is 0 and every answer keeps its bits.

Option Default What it does
--model M required a Hub repository id or a local directory
--requests FILE required JSON lines, one request each
--limit N 32 requests to read
--format F from the checkpoint the model family
--format-arg KEY=VALUE an option of the family
--batch-invariant off checks the batch-invariant mode, bit for bit
--session-cache-gb G 0 also checks the session cache, this large
--device D cuda:0 the GPU to run on