Struct ModelLaunchSpec 

pub struct ModelLaunchSpec {
    pub id: u32,
    pub name: String,
    pub file_path: PathBuf,
    pub tags: Vec<String>,
    pub architecture: Option<String>,
    pub quantization: Option<String>,
    pub context_length: Option<u64>,
    pub server_defaults: Option<ServerConfig>,
    pub file_size_bytes: u64,
    pub kv_elems_per_token: Option<KvElemsPerToken>,
    pub kv_memory_is_partial: bool,
    pub model_sampling: ModelSamplingDefaults,
}

Launch specification for running a model.

Contains all information needed to actually launch a model, including the file path. Separate from ModelSummary to avoid leaking filesystem details in catalog operations.

Fields

id: u32

Database ID of the model.

name: String

Model name.

file_path: PathBuf

Absolute path to the GGUF file.

tags: Vec<String>

Tags/labels associated with the model.

architecture: Option<String>

Model architecture (for runtime configuration).

quantization: Option<String>

Quantization label (Q4_K_M), when the catalog recorded one.

Carried purely so the launch can name what it loaded — see crate::domain::LaunchNarration. None for models whose GGUF metadata did not identify a quantization.

context_length: Option<u64>

Maximum context length the model supports.

server_defaults: Option<ServerConfig>

Per-model server defaults (e.g., context_length for launch).

file_size_bytes: u64

Total on-disk size of the model weights in bytes, summed across all shards for multi-part GGUFs.

Used to budget host memory at launch (see crate::server_config::compute_auto_cache_ram_mb). 0 when the size could not be determined — callers must treat that as “unknown” rather than “free”.

kv_elems_per_token: Option<KvElemsPerToken>

Estimated K/V element counts consumed per token of context, derived from the model’s GGUF metadata (see crate::domain::estimate_kv_elems_per_token). Type-agnostic — callers convert to bytes via crate::domain::kv_bytes_per_token once the launch’s resolved K/V cache types are known.

None when the metadata doesn’t carry the layer/head counts needed; callers substitute a conservative allowance.

kv_memory_is_partial: bool

True when the model’s KV memory retains only part of the token history — sliding-window, hybrid, or recurrent attention (see crate::domain::kv_memory_is_partial).

Such models cannot be resumed from llama-server’s disk slot files: the save/restore path does not carry the context checkpoints they need, so a “successful” restore still forces a full prompt re-prefill. Callers disable the disk slot layer for these models and rely on the in-RAM prompt cache, which does preserve checkpoints.

model_sampling: ModelSamplingDefaults

What this model’s own GGUF declares about sampler defaults (see crate::domain::ModelSamplingDefaults).

llama.cpp applies these over its own build defaults for every field no CLI flag sets, and reports the result as /props’s default_generation_settings. Carried to the running target so the proxy’s baseline check can tell a model’s own recommendation from a pin bump, rather than reporting the first as the second.

Derived from the metadata already on the catalog row, the same way kv_elems_per_token and kv_memory_is_partial are.

Trait Implementations

impl Clone for ModelLaunchSpec

fn clone(&self) -> ModelLaunchSpec

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ModelLaunchSpec

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.