Struct ActivationHint 

pub struct ActivationHint {
    pub width: u32,
    pub height: u32,
    pub batch: u32,
    pub dtype_bytes: u32,
    pub family: ActivationFamily,
}

Per-request shape hint passed into [preflight_memory_guard] so the activation budget can scale with resolution / dtype / arch. None degrades to the previous fixed-headroom approximation (the MEMORY_BUDGET_HEADROOM baked into estimate_peak_memory’s 2 GB constant), which keeps behavior identical for callers that don’t yet have a request in scope (e.g. admin-API model loads with no resolution context).

Public because gpu_worker::ensure_model_ready_sync and gpu_worker::run_chain_blocking (both pub) take it as a parameter.

Fields

width: u32

Image-space width.

height: u32

Image-space height.

batch: u32

CFG-doubled forwards typically pass 2; non-CFG passes 1.

dtype_bytes: u32

Bytes per element (2 for bf16/fp16, 4 for f32).

family: ActivationFamily

Architecture family — drives the per-arch factor in mold_inference::device::activation_bytes.

Implementations

impl ActivationHint

pub fn from_request(req: &GenerateRequest, family_slug: &str) -> Self

Build a hint from a GenerateRequest and the manifest family slug (e.g. "flux", "sdxl"). The family slug is what activation_family_for expects — when the caller doesn’t have a strong family signal (catalog ID without an installed manifest, etc.) passing the empty string falls back to ActivationFamily::FluxDit.

pub fn budget_bytes(&self) -> u64

Compute the activation budget bytes from this hint.

Trait Implementations

impl Clone for ActivationHint

fn clone(&self) -> ActivationHint

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for ActivationHint

impl Debug for ActivationHint

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

Formats the value using the given formatter.