The Guild command line

Run guild --help from your built binary for its authoritative command surface. These commands are listed in the pinned README:

Command	Purpose
`guild pull <model>`	Download a registry model
`guild import <path>`	Import an existing local model
`guild list`	List installed models
`guild show <model>`	Inspect a model in the store
`guild run <model>`	Run interactively
`guild serve <model>`	Start the native OpenAI-compatible server
`guild bench <model>`	Benchmark a model
`guild ps`	Show running model processes

Containers and optimization

The GEXEC contract describes the machine-specific optimizer:

guild optimize model.gpack --mode auto -o model.gexec
guild inspect model.gexec --plan
guild run model.gpack --gexec model.gexec

model.gpack is your actual local portable model container. GEXEC is an execution cache, not a replacement for its source weights. Missing or incompatible GEXEC data falls back to portable execution with a logged reason.

Decode strategies

guild run <model> --decode-strategy adaptive --spec-max-draft 4

The scheduler can choose normal, mtp, lookup, and retrieval when eligible. MTP requires the appropriate draft weights. Lookup proposes continuations from token history; retrieval uses a bounded session-local store. Every proposal is verified by the target model.

Legacy --spec and strategy-specific options are documented in the pinned adaptive decoding report. Sampling parity across Qwen verification-window sizes in that report uses GUILD_IQ_MT_MIN=1 to align CPU reduction behavior.