gladius-training / kernel /warm_memory.py
amuzetnoM's picture
GLADIUS training package: kernel + omega + synthase + checkpoint (step 529)
63e99b4 verified
"""
GLADIUS v2.0 β€” Warm Memory: Share + EBLoRA + Locas Synthesis
The dragon. Three papers forged into one mechanism:
- Locas (2602.05085): GLU-FFN structure, principled initialization, merge capability
- Share (2602.06043): Evolving shared subspace, incremental integration
- EBLoRA (2602.00722): Spectral balancing, Stiefel manifold constraint
This replaces the stub WarmMemory in memory.py.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
class LocasAdapter(nn.Module):
"""
Locas-style GLU-FFN adapter.
Same structure as base model's SwiGLU layers, but low-rank.
Can be merged INTO base model weights (permanentize).
Initialized from base model parameters for fast convergence.
"""
def __init__(self, hidden_dim: int, rank: int):
super().__init__()
self.hidden_dim = hidden_dim
self.rank = rank
# Low-rank GLU-FFN: hidden_dim β†’ rank β†’ hidden_dim
self.gate_proj = nn.Linear(hidden_dim, rank, bias=False)
self.up_proj = nn.Linear(hidden_dim, rank, bias=False)
self.down_proj = nn.Linear(rank, hidden_dim, bias=False)
# Scale (starts small, grows as adapter learns)
self.scale = nn.Parameter(torch.tensor(0.01))
self._init_weights()
def _init_weights(self):
# Small random init (Locas principled init happens externally
# when base model layers are available)
nn.init.normal_(self.gate_proj.weight, std=0.01)
nn.init.normal_(self.up_proj.weight, std=0.01)
nn.init.zeros_(self.down_proj.weight) # Zero init β†’ starts as identity
def init_from_base(self, base_gate: nn.Linear, base_up: nn.Linear):
"""
Locas principled initialization: extract top-k singular vectors
from base model's FFN projections.
"""
with torch.no_grad():
# Gate: top-rank directions of base gate projection
try:
U, S, V = torch.linalg.svd(base_gate.weight, full_matrices=False)
self.gate_proj.weight.data = V[:self.rank, :]
except torch._C._LinAlgError:
pass # Keep random init
# Up: top-rank directions of base up projection
try:
U, S, V = torch.linalg.svd(base_up.weight, full_matrices=False)
self.up_proj.weight.data = V[:self.rank, :]
except torch._C._LinAlgError:
pass # Keep random init
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""GLU-FFN forward. Returns residual to add to hidden state."""
gate = F.silu(self.gate_proj(x))
up = self.up_proj(x)
return self.down_proj(gate * up) * self.scale
def get_weight_matrix(self) -> torch.Tensor:
"""
Reconstruct the effective weight matrix for spectral analysis.
Returns the linearized adapter: W_eff β‰ˆ down @ (gate βŠ™ up) [simplified]
For spectral analysis, we use the dominant path: down @ up
"""
return self.down_proj.weight @ self.up_proj.weight # (hidden, hidden)
class SpectralBalancer:
"""
EBLoRA-inspired spectral balancing.
Monitors the condition number of the adapter and rebalances
when singular values become too skewed (which causes forgetting).
"""
def __init__(self, condition_threshold: float = 10.0):
self.condition_threshold = condition_threshold
self.history = []
def condition_number(self, adapter: LocasAdapter) -> float:
"""Compute Οƒ_max / Οƒ_min of the effective weight matrix."""
with torch.no_grad():
W = adapter.get_weight_matrix()
try:
S = torch.linalg.svdvals(W)
except torch._C._LinAlgError:
return 1.0 # Can't compute β€” assume healthy
S_nonzero = S[S > 1e-8]
if len(S_nonzero) < 2:
return 1.0
return (S_nonzero[0] / S_nonzero[-1]).item()
def needs_rebalance(self, adapter: LocasAdapter) -> bool:
cn = self.condition_number(adapter)
self.history.append(cn)
return cn > self.condition_threshold
def rebalance(self, adapter: LocasAdapter):
"""
Force spectral balance by normalizing singular values.
Decouples magnitude from direction (EBLoRA core principle):
1. SVD the effective weight
2. Soft-clamp singular values toward the mean
3. Reconstruct
"""
with torch.no_grad():
# Work on down @ up (the dominant linear path)
W = adapter.down_proj.weight @ adapter.up_proj.weight # (H, H)
# Diagonal epsilon β€” regularize ill-conditioned matrices
# Prevents SVD convergence failure from degenerate/repeated singular values
eps = 1e-6 * torch.eye(W.shape[0], W.shape[1], device=W.device, dtype=W.dtype)
W = W + eps
try:
U, S, Vh = torch.linalg.svd(W, full_matrices=False)
except torch._C._LinAlgError:
# Matrix still too sick β€” skip this rebalance cycle
return
# Filter near-zero singular values
mask = S > 1e-6
if mask.sum() < 2:
return # Nothing meaningful to balance
S_active = S[mask]
# Soft-clamp: move singular values toward geometric mean
log_S = torch.log(S_active)
log_mean = log_S.mean()
# Shrink toward mean by 50%
balanced_log_S = log_mean + 0.5 * (log_S - log_mean)
S_balanced = S.clone()
S_balanced[mask] = torch.exp(balanced_log_S)
# Reconstruct balanced weight
rank = adapter.rank
# Safe reconstruction: use min of available dims
k = min(rank, S_balanced.shape[0], U.shape[1], Vh.shape[0])
sqrt_S = torch.sqrt(S_balanced[:k].clamp(min=1e-8))
new_down_full = U[:, :k] @ torch.diag(sqrt_S) # (H, k)
new_up_full = torch.diag(sqrt_S) @ Vh[:k, :] # (k, H)
# Write back (truncate to adapter rank)
r = min(rank, k)
adapter.down_proj.weight.data[:, :r] = new_down_full[:, :r]
adapter.up_proj.weight.data[:r, :] = new_up_full[:r, :]
class SubspaceTracker:
"""
Share-inspired evolving subspace tracker.
Maintains a compact representation of what the warm memory "knows."
New knowledge is checked against this subspace:
- If it projects well β†’ already known β†’ small update
- If large residual β†’ novel knowledge β†’ evolve subspace
"""
def __init__(self, hidden_dim: int, rank: int, novelty_threshold: float = 0.1):
self.hidden_dim = hidden_dim
self.rank = rank
self.novelty_threshold = novelty_threshold
# The subspace basis (orthonormal)
self.basis = torch.zeros(rank, hidden_dim)
# Importance of each basis direction
self.importance = torch.zeros(rank)
self.initialized = False
def initialize_from_adapter(self, adapter: LocasAdapter):
"""Extract subspace from current adapter state."""
with torch.no_grad():
W = adapter.get_weight_matrix()
U, S, Vh = torch.linalg.svd(W, full_matrices=False)
k = min(self.rank, len(S))
self.basis[:k] = Vh[:k]
self.importance[:k] = S[:k]
self.initialized = True
def compute_novelty(self, gradient: torch.Tensor) -> tuple[float, torch.Tensor]:
"""
Measure how much of the gradient is NOT captured by current subspace.
"""
if not self.initialized:
return float('inf'), gradient
# Flatten gradient and basis to same dimensionality
g_flat = gradient.flatten()
basis_flat = self.basis.reshape(self.rank, -1).to(g_flat.device)
# Ensure compatible dimensions: truncate/pad basis if needed
g_dim = g_flat.shape[0]
b_dim = basis_flat.shape[1]
if g_dim != b_dim:
# Project gradient down to basis dimensionality
g_proj = g_flat[:b_dim] if g_dim > b_dim else F.pad(g_flat, (0, b_dim - g_dim))
else:
g_proj = g_flat
projection = basis_flat @ g_proj # (rank,)
reconstructed = projection @ basis_flat # (b_dim,)
residual = g_proj - reconstructed
novelty = residual.norm().item()
return novelty, residual.reshape(self.hidden_dim)
def evolve(self, new_direction: torch.Tensor, importance: float):
"""
Integrate a new direction into the subspace.
Replaces the least important existing direction.
"""
# Normalize
new_dir_flat = new_direction.flatten()
new_dir_flat = new_dir_flat / (new_dir_flat.norm() + 1e-8)
# Replace least important direction
least_idx = self.importance.argmin().item()
self.basis[least_idx] = new_dir_flat[:self.hidden_dim] # Truncate if needed
self.importance[least_idx] = importance
# Decay all importance values slightly (recency bias)
self.importance *= 0.99
class RealWarmMemory(nn.Module):
"""
Full warm memory implementation: Locas + Share + EBLoRA.
Architecture: Locas GLU-FFN adapter (per transformer layer)
Evolution: Share subspace tracking for novelty detection
Stability: EBLoRA spectral balancing
This is the dragon, tamed.
"""
def __init__(self, config, num_layers: int | None = None):
super().__init__()
from .config import KernelConfig
self.config = config
self.hidden_dim = config.hidden_dim
self.rank = config.warm_rank
num_layers = num_layers or config.num_layers
# Per-layer Locas adapters
self.adapters = nn.ModuleList([
LocasAdapter(config.hidden_dim, config.warm_rank)
for _ in range(num_layers)
])
# Shared spectral balancer
self.balancer = SpectralBalancer(config.warm_condition_threshold)
# Per-layer subspace trackers (not nn.Module β€” just tracking)
self.trackers = [
SubspaceTracker(config.hidden_dim, config.warm_rank, config.warm_novelty_threshold)
for _ in range(num_layers)
]
# Update counter
self.register_buffer('update_count', torch.tensor(0, dtype=torch.long))
def forward(self, x: torch.Tensor, layer_idx: int = 0) -> torch.Tensor:
"""Apply warm memory adapter for a specific layer."""
if layer_idx < len(self.adapters):
return x + self.adapters[layer_idx](x)
return x
def forward_all(self, x: torch.Tensor) -> torch.Tensor:
"""Apply all adapters sequentially (for simple use)."""
for adapter in self.adapters:
x = x + adapter(x)
return x
@torch.no_grad()
def consolidate(self, hot_keys: torch.Tensor, hot_values: torch.Tensor,
importance_scores: torch.Tensor):
"""
Real consolidation: hot memory β†’ warm adapters.
1. Filter by importance
2. Check novelty against subspace
3. Update adapters with spectral balancing
"""
self.update_count += 1
# Average hot memory signal
signal = hot_values.mean(dim=0)
if signal.norm() < 1e-6:
return # Nothing meaningful in hot memory
for i, (adapter, tracker) in enumerate(zip(self.adapters, self.trackers)):
# Initialize tracker if needed
if not tracker.initialized:
tracker.initialize_from_adapter(adapter)
# Compute pseudo-gradient from hot memory content
W_grad = torch.outer(signal, signal) # Rank-1 outer product
# Check novelty
novelty, residual = tracker.compute_novelty(W_grad)
if novelty > tracker.novelty_threshold:
# Novel knowledge β†’ evolve subspace
tracker.evolve(residual, novelty)
# Small update to adapter (learning rate controlled)
lr = 0.001
r = min(adapter.rank, residual.shape[0])
h = min(self.hidden_dim, adapter.up_proj.weight.shape[1])
adapter.up_proj.weight.data[:r, :h] += lr * residual[:r].unsqueeze(1).expand(r, h) * 0.01
# UPDATE DOWN_PROJ β€” the missing link!
# Without this, warm memory tracks knowledge but can't express it.
# down_proj maps rank-space β†’ hidden-space (the output pathway)
# CRITICAL: once down_proj is non-zero, backprop gradients flow through
# the adapter during training, enabling the optimizer to refine warm memory.
signal_norm = signal / (signal.norm() + 1e-8)
r_down = min(adapter.rank, adapter.down_proj.weight.shape[1])
h_down = min(self.hidden_dim, adapter.down_proj.weight.shape[0])
down_update = signal_norm[:h_down].unsqueeze(1) * residual[:r_down].unsqueeze(0)
# Use 10x the up_proj rate to bootstrap the output pathway
adapter.down_proj.weight.data[:h_down, :r_down] += lr * down_update * 0.1
# Ramp scale toward 0.3 (from init 0.01) so adapter output is meaningful
target_scale = 0.3
adapter.scale.data += lr * 0.1 * (target_scale - adapter.scale.data)
# Periodic spectral rebalancing (ALWAYS check, not just on consolidation)
if self.update_count % max(1, self.config.warm_balance_frequency // 10) == 0:
cn = self.balancer.condition_number(adapter)
if cn > self.config.warm_condition_threshold:
self.balancer.rebalance(adapter)
def condition_number(self) -> float:
"""Average condition number across all adapter layers."""
cns = [self.balancer.condition_number(a) for a in self.adapters]
return sum(cns) / len(cns)
def checkpoint(self, path: str):
"""Save warm memory state."""
state = {
'adapters': self.state_dict(),
'update_count': self.update_count.item(),
'subspace_states': [
{'basis': t.basis.clone(), 'importance': t.importance.clone()}
for t in self.trackers
],
'spectral_history': self.balancer.history[-100:], # Last 100
}
torch.save(state, path)
def restore(self, path: str):
"""Load warm memory state."""
state = torch.load(path, weights_only=False)
self.load_state_dict(state['adapters'], strict=False)
self.update_count.fill_(state['update_count'])
for tracker, ss in zip(self.trackers, state.get('subspace_states', [])):
tracker.basis = ss['basis']
tracker.importance = ss['importance']
tracker.initialized = True
self.balancer.history = state.get('spectral_history', [])