Skip to main content 홈 크리에이터 adu2021 skillxiv simple-gpt-normalization-strategy
simple-gpt-normalization-strategy Place RMSNorm immediately after every linear layer to stabilize activation scales at O(sqrt(d)) and reduce Hessian spectral norm. Enables 3-10x larger learning rates and faster convergence without architectural changes; improves loss by 0.08 on 7B models.
설치로 이동 Skills Marketplace 커뮤니티가 만든 AI 스킬을 발견하고 탐색하세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/ADu2021/skillXiv --skill simple-gpt-normalization-strategy명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
Zip 다운로드 다운로드 중... 이 저장소의 다른 Skills meaningful-kebab-case-name Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
action-quantization-behavior-cloning Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
adaptive-lora-personalized-ranks Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
name simple-gpt-normalization-strategy title SimpleGPT: Improving GPT via A Simple Normalization Strategy version 0.0.2 engine skillxiv-v0.0.2-claude-opus-4.6 license MIT url https://arxiv.org/abs/2602.01212 keywords ["Normalization Strategy","Training Stability","Learning Rate","Hessian Conditioning","Transformer Architecture"] description Place RMSNorm immediately after every linear layer to stabilize activation scales at O(sqrt(d)) and reduce Hessian spectral norm. Enables 3-10x larger learning rates and faster convergence without architectural changes; improves loss by 0.08 on 7B models.
SimpleGPT: Linear-Layer Normalization for Training Stability
Standard Transformer architectures place LayerNorm only before attention and MLPs, allowing activation scales to drift during training. SimpleGPT proposes a minimal change: insert RMSNorm immediately after every linear transformation (Q/K/V projections, MLPs, output projections). This simple shift dramatically improves optimization geometry, enabling much larger learning rates and faster convergence.
The key insight is that normalization right after linear projections stabilizes the activation scale that linear layers output, preventing the representation drift that limits learning rate. This is grounded in second-order optimization: the Hessian spectral norm directly limits maximum learning rate, and SimpleNorm reduces this bound significantly.
Core Concept
SimpleNorm applies a uniform pattern across the entire model:
Standard Transformer : Linear → Activation → (Attention/MLP)
SimpleGPT : Linear → RMSNorm → Activation → (Attention/MLP)
This seemingly small change has outsized impact because:
Stabilized activation scale : RMSNorm forces activations to remain at O(√d), preventing explosion or vanishingReduced Hessian curvature : Normalization smooths the loss landscape, reducing spectral norm of HessianWeight-scale invariance : Unlike unnormalized layers where curvature scales with ||W||₂², normalized layers maintain consistent curvature regardless of weight magnitude
Architecture Overview
Linear projections (Q, K, V in attention; feed-forward up/down): Apply RMSNorm immediately afterOutput projections (attention out, MLP out): Apply RMSNorm afterRMSNorm placement : Before activation functions (ReLU, GELU, etc.)Hyperparameter scaling : Increase learning rate 3-10× based on model sizetorch.compile integration : JIT compilation adds minimal overhead
Implementation
Step 1: Define RMSNorm Operation
Create an efficient RMSNorm that normalizes to unit RMS value.
class RMSNorm (nn.Module):
def __init__ (self, hidden_dim: int , eps: ):
float
1e-6
"""
Root Mean Square Layer Normalization.
Args:
hidden_dim: Dimension to normalize over
eps: Numerical stability epsilon
"""
super
self
self
def
forward
self, x: torch.Tensor
"""
Apply RMSNorm: x / sqrt(E[x^2] + eps) * weight
Args:
x: Input tensor of any shape with last dim = hidden_dim
Returns:
Normalized tensor, same shape as input
"""
2
1
True
self
return
self
Step 2: Modify Linear Layers with Post-Linear Normalization Create a composite module combining linear projection and immediate normalization.
class NormalizedLinear (nn.Module):
def __init__ (self, in_features: int , out_features: int ,
hidden_dim: Optional [int ] = None ):
"""
Linear layer with RMSNorm immediately after.
Args:
in_features: Input dimension
out_features: Output dimension
hidden_dim: Dimension for RMSNorm (default: out_features)
"""
super ().__init__()
self .linear = nn.Linear(in_features, out_features, bias=True )
self .norm = RMSNorm(hidden_dim or out_features)
def forward (self, x: torch.Tensor ) -> torch.Tensor:
"""Apply linear projection then normalize."""
x = self .linear(x)
x = self .norm(x)
return x
class NormalizedAttention (nn.Module):
def __init__ (self, hidden_dim: int , num_heads: int ):
"""Attention with normalized Q/K/V projections."""
super ().__init__()
self .hidden_dim = hidden_dim
self .num_heads = num_heads
head_dim = hidden_dim // num_heads
self .q_proj = NormalizedLinear(hidden_dim, hidden_dim, hidden_dim)
self .k_proj = NormalizedLinear(hidden_dim, hidden_dim, hidden_dim)
self .v_proj = NormalizedLinear(hidden_dim, hidden_dim, hidden_dim)
self .out_proj = NormalizedLinear(hidden_dim, hidden_dim, hidden_dim)
def forward (self, hidden_states: torch.Tensor ) -> torch.Tensor:
"""Standard attention with normalized projections."""
batch_size, seq_len, _ = hidden_states.shape
queries = self .q_proj(hidden_states).view(
batch_size, seq_len, self .num_heads, -1
).transpose(1 , 2 )
keys = self .k_proj(hidden_states).view(
batch_size, seq_len, self .num_heads, -1
).transpose(1 , 2 )
values = self .v_proj(hidden_states).view(
batch_size, seq_len, self .num_heads, -1
).transpose(1 , 2 )
attn_scores = torch.matmul(queries, keys.transpose(-2 , -1 ))
attn_scores = attn_scores / math.sqrt(self .hidden_dim // self .num_heads)
attn_weights = torch.softmax(attn_scores, dim=-1 )
attn_output = torch.matmul(attn_weights, values)
attn_output = attn_output.transpose(1 , 2 ).contiguous()
attn_output = attn_output.view(batch_size, seq_len, -1 )
output = self .out_proj(attn_output)
return output
Step 3: Implement MLP with Normalized Projections Create feed-forward network with normalized intermediate layers.
class NormalizedMLP (nn.Module):
def __init__ (self, hidden_dim: int , intermediate_dim: int ,
activation: str = "gelu" ):
"""
Feed-forward network with normalized projections.
Args:
hidden_dim: Input/output dimension
intermediate_dim: Hidden layer dimension (typically 4x)
activation: Activation function name
"""
super ().__init__()
self .up_proj = NormalizedLinear(hidden_dim, intermediate_dim, intermediate_dim)
self .down_proj = NormalizedLinear(intermediate_dim, hidden_dim, hidden_dim)
if activation == "gelu" :
self .activation = torch.nn.GELU()
elif activation == "relu" :
self .activation = torch.nn.ReLU()
else :
self .activation = torch.nn.SiLU()
def forward (self, x: torch.Tensor ) -> torch.Tensor:
"""Up-project, activate, down-project with normalization."""
x = self .up_proj(x)
x = self .activation(x)
x = self .down_proj(x)
return x
Step 4: Assemble into Complete Transformer Layer Integrate normalized components into standard transformer architecture.
class SimpleNormTransformerLayer (nn.Module):
def __init__ (self, hidden_dim: int , num_heads: int , intermediate_dim: int ):
"""Transformer layer using SimpleNorm pattern."""
super ().__init__()
self .self_attn = NormalizedAttention(hidden_dim, num_heads)
self .mlp = NormalizedMLP(hidden_dim, intermediate_dim)
self .norm1 = RMSNorm(hidden_dim)
self .norm2 = RMSNorm(hidden_dim)
def forward (self, x: torch.Tensor ) -> torch.Tensor:
"""Transformer layer: pre-norm attention + pre-norm MLP + residuals."""
attn_input = self .norm1(x)
attn_output = self .self_attn(attn_input)
x = x + attn_output
mlp_input = self .norm2(x)
mlp_output = self .mlp(mlp_input)
x = x + mlp_output
return x
Step 5: Training Configuration with Increased Learning Rates Set up training with larger learning rates enabled by SimpleNorm.
def train_with_simple_norm (
model: nn.Module,
train_dataset,
base_learning_rate: float = 1e-3 ,
scale_factor: float = 5.0 ,
num_epochs: int = 10
):
"""
Train model with SimpleNorm using increased learning rates.
Args:
model: Model using NormalizedLinear layers
train_dataset: Training data
base_learning_rate: Starting learning rate
scale_factor: Multiplier for LR (typically 3-10x)
num_epochs: Training epochs
"""
learning_rate = base_learning_rate * scale_factor
optimizer = torch.optim.AdamW(
model.parameters(),
lr=learning_rate,
betas=(0.9 , 0.95 ),
weight_decay=0.01
)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer,
T_max=num_epochs * len (train_dataset),
eta_min=learning_rate * 0.1
)
model = torch.compile (model)
for epoch in range (num_epochs):
total_loss = 0
for batch_idx, (input_ids, labels) in enumerate (train_dataset):
logits = model(input_ids)
loss = torch.nn.functional.cross_entropy(
logits.view(-1 , logits.shape[-1 ]),
labels.view(-1 )
)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0 )
optimizer.step()
scheduler.step()
total_loss += loss.item()
avg_loss = total_loss / len (train_dataset)
print (f"Epoch {epoch} : loss={avg_loss:.4 f} , lr={scheduler.get_last_lr()[0 ]:.2 e} " )
return model
def validate_convergence_speedup (model_simple: nn.Module,
model_standard: nn.Module,
validation_data ):
"""
Compare convergence speed: SimpleNorm vs standard.
Returns:
Speedup ratio (iterations to convergence)
"""
def measure_convergence (model, target_loss=0.1 ):
steps = 0
loss = float ('inf' )
while loss > target_loss:
steps += 1
if steps > 10000 :
break
return steps
steps_simple = measure_convergence(model_simple)
steps_standard = measure_convergence(model_standard)
speedup = steps_standard / steps_simple
print (f"SimpleNorm converges {speedup:.2 f} x faster" )
return speedup
Practical Guidance When to use SimpleGPT normalization:
Model training where convergence speed is critical
Any Transformer-based architecture (language, vision, multimodal)
Fine-tuning scenarios where larger learning rates help
Systems where 3-8% training time reduction is valuable
Already-converged models (no re-training needed)
Systems with highly optimized training pipelines where changes are risky
Inference-only scenarios (SimpleNorm has negligible inference cost)
Learning rate too large: Even with SimpleNorm, overly aggressive LR causes divergence; start conservative and increase
Skipping layer normalization: Must place RMSNorm after every linear layer for stability benefits
Not adjusting warmup: Warmup schedule may need adjustment for larger learning rates
torch.compile incompatibility: Ensure all operations are supported by compiler
Hyperparameter Guidelines:
Parameter Range Tuning Learning rate scale 3x-10x Start at 3x; increase if stable; 10x for aggressive schedules Weight decay 0.01-0.05 Standard value; increase if regularization needed Gradient clipping 1.0 Keep standard clipping; SimpleNorm doesn't eliminate need Warmup epochs 0.5-2% May increase slightly; SimpleNorm more stable earlier
Reference Key results: 0.08 point loss improvement on 7B Llama3 models; 3-10× larger learning rates enabled; validated across nanoGPT (120M), Llama2 (7B), and Llama3 (8B). Minimal training overhead with torch.compile. No architectural changes required.
