Struct ServerConfigOptions 

pub struct ServerConfigOptions {

    pub context_size: Option<u64>,
    pub model_server_ctx: Option<usize>,
    pub global_default_ctx: Option<u64>,
    pub port: Option<u16>,
    pub jinja: Option<bool>,
    pub reasoning_format: Option<String>,
    pub mtp_draft_n_max: Option<u32>,
    pub mtp_draft_p_min: Option<f32>,
    pub slot_save_path: Option<PathBuf>,
    pub cache_ram_mb: Option<u64>,
    pub cache_reuse: Option<u32>,
    pub cache_type_k: Option<KvCacheType>,
    pub cache_type_v: Option<KvCacheType>,
    pub inference_params: Option<InferenceConfig>,
    pub mlock: Option<bool>,
}

Caller-supplied overrides for resolve_context_size.

All fields default to None, which means “fall through to next tier”.

Fields

context_size: Option<u64>

Override the context window size forwarded to llama-server. None lets llama-server use its built-in default.

model_server_ctx: Option<usize>

Per-model server defaults context length (from Model.server_defaults.context_length). Second tier in fallback chain.

global_default_ctx: Option<u64>

Global app setting for default context size (from Settings.default_context_size). Third tier in fallback chain.

port: Option<u16>

Bind llama-server to a specific port instead of letting the allocator choose.

jinja: Option<bool>

Override Jinja template support.

• None → auto-detect: enabled when the model has the "agent" tag.
• Some(true) → force enable regardless of tags.
• Some(false) → force disable regardless of tags.

reasoning_format: Option<String>

Override the reasoning format passed to llama-server.

• None → auto-detect from model tags (e.g. "reasoning" tag).
• Some("none") → explicitly suppress reasoning extraction even if the model has a reasoning tag.
• Some("deepseek") / Some("deepseek-legacy") → force a specific format.

mtp_draft_n_max: Option<u32>

Override the MTP draft token count.

• None → auto-detect: enabled with default n=2 when the model has the "mtp" tag.
• Some(0) → explicitly disable MTP even if the model has the "mtp" tag.
• Some(n) → enable MTP with n draft tokens.

mtp_draft_p_min: Option<f32>

Override the MTP acceptance probability threshold. Only meaningful when MTP is enabled. None uses the default (0.75).

slot_save_path: Option<PathBuf>

Directory for llama-server KV cache slot persistence (--slot-save-path).

• None — disk slot persistence disabled, no --slot-save-path flag.
• Some(dir) — enables slot save/restore. Direct pass-through, no tag-based auto-detection (unlike jinja/MTP/reasoning). Independent of cache_ram_mb/cache_reuse below.

cache_ram_mb: Option<u64>

RAM budget in MiB for llama-server’s own host-RAM prompt cache (--cache-ram). None leaves llama-server’s built-in default. Some(0) disables the cache. Direct pass-through, no tag-based auto-detection.

cache_reuse: Option<u32>

Minimum chunk size in tokens for KV-shift cache reuse past the first prefix divergence point (--cache-reuse). None leaves the feature off. Direct pass-through, no tag-based auto-detection.

cache_type_k: Option<KvCacheType>

Explicit override for the K cache element type (--cache-type-k). None resolves to the q8_0 default (see gglib_runtime::llama::args::resolve_kv_cache_types), unless GGLIB_DISABLE_KV_QUANT=1 is set.

cache_type_v: Option<KvCacheType>

Explicit override for the V cache element type (--cache-type-v). Same resolution as Self::cache_type_k. Quantizing V additionally requires Flash Attention to be active — see gglib_runtime::llama::args::kv_cache_type module docs.

inference_params: Option<InferenceConfig>

Inference parameter overrides (temperature, top-p, etc.) forwarded directly to llama-server.

mlock: Option<bool>

Whether to memory-lock the model into RAM (--mlock). None defaults to false in build_server_config().

Implementations

impl ServerConfigOptions

pub fn overlay(&self, over: &Self) -> Self

Field-wise merge: every Some in over wins, every None falls through to self.

This is the single layering primitive behind both places where two sets of options meet:

• the 3-tier cascade in UnifiedServerConfig::resolved_options, where global defaults are the base and explicit CLI/GUI overrides are over;
• per-call launch overrides layered on top of a ProcessManager’s standing template.

Note that this merges options, not resolved values — the tier chain baked into resolve_context_size (request → per-model → global → hardcoded) still runs afterwards on the merged result, so overlaying never collapses those tiers early.

over is destructured exhaustively on purpose: adding a field to this struct then fails to compile until it is given merge semantics here, rather than being silently dropped.

Trait Implementations

impl Clone for ServerConfigOptions

fn clone(&self) -> ServerConfigOptions

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ServerConfigOptions

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

Formats the value using the given formatter.

impl Default for ServerConfigOptions

fn default() -> ServerConfigOptions

Returns the “default value” for a type.