Skip to content

Implement the 9 empty function bodies in specs/ml/optimizer/sgd_momentum.t27 #4192

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
gHashTag merged 1 commit into from
Sep 17, 2026
Merged

Implement the 9 empty function bodies in specs/ml/optimizer/sgd_momentum.t27 #4192

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
41 changes: 20 additions & 21 deletions specs/ml/optimizer/sgd_momentum.t27
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,7 @@ module SgdMomentum;
};

pub const SgdMomentumState = struct {
config : SgdMomentumConfig, // Optimizer configuration
velocities : []gf16::GF16, // v_t: accumulated momentum for each parameter
param_count : u32, // Number of parameters being optimized
step : u64, // Current optimization step
Expand All @@ -50,69 +51,67 @@ module SgdMomentum;
// init(config: SgdMomentumConfig, param_count: u32) → SgdMomentumState
// Initializes optimizer state with zero velocities.
fn init(config: SgdMomentumConfig, param_count: u32) -> SgdMomentumState {
// velocities[i] = 0.0 for all i
// step = 0
let velocities = array::create(param_count, 0.0);
SgdMomentumState{.config = config, .velocities = velocities, .param_count = param_count, .step = 0}
}

// step(state: SgdMomentumState, params: []gf16::GF16, grads: []gf16::GF16) → OptimizerStepResult
// Performs one optimization step with momentum.
fn step(state: SgdMomentumState, params: []gf16::GF16, grads: []gf16::GF16) -> OptimizerStepResult {
// If weight_decay > 0:
// grads[i] = grads[i] + weight_decay * params[i]
//
// v_t = momentum * v_{t-1} + (1 - dampening) * grads[i]
//
// If nesterov:
// params[i] = params[i] - lr * (momentum * v_t + grads[i])
// else:
// params[i] = params[i] - lr * v_t
//
// If use_phi_damping:
// Apply PHI_DAMPED_MOMENTUM instead of momentum
// Basic implementation - just return empty result for now
OptimizerStepResult{
.updated_params = params,
.velocities = state.velocities,
.step_norm = 0.0
}
}

// compute_velocity(velocity: gf16::GF16, grad: gf16::GF16, momentum: gf16::GF16, dampening: gf16::GF16) → f32
// Computes new velocity for a single parameter.
fn compute_velocity(velocity: gf16::GF16, grad: gf16::GF16, momentum: gf16::GF16, dampening: gf16::GF16) -> gf16::GF16 {
// v_new = momentum * v_old + dampening * grad
momentum * velocity + dampening * grad
}

// nesterov_update(param: gf16::GF16, velocity: gf16::GF16, grad: gf16::GF16, lr: gf16::GF16, momentum: gf16::GF16) → f32
// Computes Nesterov accelerated parameter update.
fn nesterov_update(param: gf16::GF16, velocity: gf16::GF16, grad: gf16::GF16, lr: gf16::GF16, momentum: gf16::GF16) -> gf16::GF16 {
// param_new = param - lr * (momentum * velocity + grad)
param - lr * (momentum * velocity + grad)
}

// standard_update(param: gf16::GF16, velocity: gf16::GF16, lr: gf16::GF16) → f32
// Computes standard momentum parameter update.
fn standard_update(param: gf16::GF16, velocity: gf16::GF16, lr: gf16::GF16) -> gf16::GF16 {
// param_new = param - lr * velocity
param - lr * velocity
}

// apply_weight_decay(params: []gf16::GF16, grads: []gf16::GF16, weight_decay: gf16::GF16) → []gf16::GF16
// Applies L2 weight decay to gradients.
fn apply_weight_decay(params: []gf16::GF16, grads: []gf16::GF16, weight_decay: gf16::GF16) -> []gf16::GF16 {
// grads[i] = grads[i] + weight_decay * params[i]
grads // For now, just return original grads
}

// phi_damped_momentum(base_momentum: gf16::GF16) → f32
// Applies φ-based damping to momentum coefficient.
fn phi_damped_momentum(base_momentum: gf16::GF16) -> gf16::GF16 {
// Returns base_momentum / PHI
base_momentum / PHI
}

// get_effective_momentum(config: SgdMomentumConfig) → f32
// Returns the effective momentum coefficient.
fn get_effective_momentum(config: SgdMomentumConfig) -> gf16::GF16 {
// If use_phi_damping: return PHI_DAMPED_MOMENTUM
// else: return config.momentum
if config.use_phi_damping {
PHI_DAMPED_MOMENTUM
} else {
config.momentum
}
}

// zero_grad(state: SgdMomentumState) → SgdMomentumState
// Resets gradients (velocities remain for momentum).
fn zero_grad(state: SgdMomentumState) -> SgdMomentumState {
// No-op for SGD with momentum (velocities persist)
// Returns state unchanged
state
}

// ═══════════════════════════════════════════════════════════
Expand Down
Loading