Repeat a deployment without changing the model
Fix the model download to a commit, preserve the environment, and distinguish configuration differences from measured improvements.
Reviewed 2026-09-08 · Advanced
A model name is not an immutable version
A repository default branch and a release tag can change. A saved observation of its commit is useful evidence, but does not force a later command to download that commit. Use the full commit identifier in the download step and keep the local destination separate for each repository and revision.
In the deployment workbench, selecting a published GGUF file or complete shard group fills the model commit. The generated Hugging Face CLI command uses --revision and --local-dir; the following engine command reads that local path rather than fetching the default branch. Run the download first and stop if it fails. An unavailable commit is not permission to substitute a newer one.
Model pinning is only one part of reproducibility
The model commit does not fix the runtime, Python packages, driver, operating system, device configuration, or sampling behavior. Record their actual versions and the exact artifact. This site does not install or freeze your environment, inspect local modifications, or guarantee bit-identical generated text.
For a split model, every required shard must be accounted for. The workbench accepts complete, ordered manifests of up to 64 files and includes every filename in the fixed-revision download command. Missing, duplicated, or mixed groups are rejected. All downloads must succeed before opening the first shard. MLX and vLLM repository downloads include all files at the commit, which can require more disk space than one weight variant. Review access terms and installation instructions before running any command.
Keep configuration changes separate from performance claims
Save a deployment record before changing settings, then another after the test. Select both records to see exactly which configuration, artifact facts, or user reports differ. A changed success status is not evidence of faster generation or lower peak memory.
Use the measurement notebook for actual observations. Keep the model revision, runtime version, hardware, precision, context, real prompt token count, concurrency, load state and workload policy aligned before combining samples. Unknown values stay unknown. If you time decoding from the first token to the last, count only the tokens produced within that interval; do not divide the entire output length by a different timing interval.