Struct VramGuard 

pub struct VramGuard { /* private fields */ }

VRAM reservation guard.

Acquires a VRAM reservation on creation and releases it on drop. Enforces C-VRAM-001: no allocation beyond budget.

Implementations

impl VramGuard

pub fn acquire(budget_mb: usize, task: &str) -> Result<Self, GpuError>

Acquire a VRAM reservation.

Checks the ledger and reserves budget_mb of VRAM. Returns GpuError::InsufficientMemory if not enough VRAM available.

pub fn acquire_wait( budget_mb: usize, task: &str, timeout_secs: u64, ) -> Result<Self, GpuError>

Acquire with waiting: poll until VRAM is available or timeout.

pub fn update_actual(&mut self, actual_mb: usize) -> Result<(), GpuError>

Update the actual measured VRAM after GPU initialization.

Call this after CudaTrainer::new() + weight upload to record the real VRAM usage (may be less than budgeted).

pub fn budget_mb(&self) -> usize

Budget that was reserved (MB).

pub fn gpu_uuid(&self) -> &str

GPU UUID this guard is for.

pub fn status(&self) -> Result<String, GpuError>

Read the current GPU status display.

pub fn check_overshoot(&self) -> Option<(usize, usize)>

Check if actual VRAM usage overshoots the budget.

Returns Some((actual, budget)) if overshoot detected, None if within budget.

Contract: vram-guard-v1 / budget_overshoot

pub fn auto_estimate_budget(param_count: usize) -> usize

Auto-estimate the VRAM budget for a model based on parameter count.

Uses rule of thumb: ~2 bytes per parameter for f16 weights + 20% overhead.

Contract: vram-guard-v1 / auto_budget_estimate