File size: 11,098 Bytes
1b08b01 50706f8 1b08b01 ea344ec 13f6128 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 |
---
title: SmolLM2-135M Text Generator
emoji: 🐨
colorFrom: yellow
colorTo: blue
sdk: gradio
sdk_version: 6.0.1
app_file: app.py
pinned: false
short_description: A Llama based SmolLM2-135M Transformer (Decoder only)
---
HuggingFace space for inference demo: https://huggingface.co/spaces/Sualeh77/smollm2-135m-trained-on-tinyShakespear-forfun
# SmolLM2-135M Implementation
A from-scratch PyTorch implementation of the SmolLM2-135M language model, following the LLaMA architecture with modern optimizations.
## Overview
This repository contains a complete implementation of SmolLM2-135M, a 135 million parameter decoder-only transformer model. The implementation includes:
- **Model Architecture** (`model.py`): Complete model definition with KV cache support
- **Training Script** (`train.py`): PyTorch Lightning training with WSD scheduler
- **Gradio App** (`app.py`): Interactive web interface for text generation
## Model Architecture (`model.py`)
### Architecture Components
The model follows the LLaMA-style decoder-only transformer architecture with the following key components:
#### 1. **SmolConfig** (Configuration Class)
A dataclass that stores all model hyperparameters:
```python
@dataclass
class SmolConfig:
vocab_size: int = 49152 # Vocabulary size
hidden_size: int = 576 # Hidden dimension
intermediate_size: int = 1536 # MLP intermediate dimension
num_hidden_layers: int = 30 # Number of transformer layers
num_attention_heads: int = 9 # Number of query heads
num_key_value_heads: int = 3 # Number of key/value heads (GQA)
max_position_embeddings: int = 8192 # Maximum sequence length
rope_theta: float = 100000.0 # RoPE base frequency
rms_norm_eps: float = 1e-5 # RMSNorm epsilon
attention_bias: bool = False # Whether to use bias in attention
mlp_bias: bool = False # Whether to use bias in MLP
dtype: torch.dtype = torch.bfloat16
```
**Key Features:**
- `head_dim` property: Automatically computes head dimension (hidden_size // num_attention_heads = 64)
- `from_hf()` class method: Loads configuration from HuggingFace model config
#### 2. **RMSNorm** (Root Mean Square Normalization)
Replaces LayerNorm with a more efficient normalization:
```python
class RMSNorm(nn.Module):
def forward(self, x):
norm = x.pow(2).mean(dim=-1, keepdim=True)
x = x * torch.rsqrt(norm + self.eps)
return self.weight * x
```
**Benefits:**
- More efficient than LayerNorm (no mean subtraction)
- Used throughout the model for pre-norm architecture
#### 3. **RoPE** (Rotary Positional Embeddings)
Rotary Position Embeddings applied to query and key tensors:
```python
def build_rope_cache(seq_len, head_dim, base, device, dtype):
# Computes cosine and sine caches for RoPE
inv_freq = 1.0 / (base ** (freq_seq / half_dim))
freqs = torch.outer(t, inv_freq)
cos = freqs.cos()[None, None, :, :]
sin = freqs.sin()[None, None, :, :]
return cos, sin
def apply_rope(x, cos, sin):
# Applies rotary transformation to input tensor
x1, x2 = x[..., :half], x[..., half:]
x1_rot = x1 * cos - x2 * sin
x2_rot = x1 * sin + x2 * cos
return torch.cat([x1_rot, x2_rot], dim=-1)
```
**Key Features:**
- Relative positional encoding (no absolute position embeddings)
- Applied only to Q and K (not V)
- Supports efficient caching for inference
#### 4. **MultiHeadSelfAttention** (Grouped Query Attention)
Implements GQA (Grouped Query Attention) where:
- **Query heads**: 9 (full attention)
- **Key/Value heads**: 3 (shared across query heads)
```python
class MultiHeadSelfAttention(nn.Module):
def forward(self, x, cos, sin, past_key_value=None, use_cache=False):
# 1. Project to Q, K, V
q = self.q_proj(x) # (B, T, n_heads * head_dim)
k = self.k_proj(x) # (B, T, n_kv_heads * head_dim)
v = self.v_proj(x) # (B, T, n_kv_heads * head_dim)
# 2. Apply RoPE to Q and K
q = apply_rope(q, cos, sin)
k = apply_rope(k, cos, sin)
# 3. KV Cache support (for inference)
if past_key_value:
k = torch.cat([past_k, k], dim=2)
v = torch.cat([past_v, v], dim=2)
# 4. GQA: Expand K/V if needed
if n_kv_heads < n_heads:
k = k.repeat_interleave(repeat_factor, dim=1)
v = v.repeat_interleave(repeat_factor, dim=1)
# 5. Compute attention scores
scores = (q @ k.transpose(-2, -1)) / sqrt(head_dim)
scores = scores + causal_mask # Causal masking
# 6. Softmax and weighted sum
probs = F.softmax(scores, dim=-1)
out = probs @ v
return out, present_key_value
```
**Key Features:**
- **KV Cache**: Efficient inference by caching past key-value pairs
- **GQA**: Reduces memory by sharing K/V heads (3:1 ratio)
- **Causal Masking**: Prevents attending to future tokens
- **RoPE Integration**: Positional encoding via rotary embeddings
#### 5. **SmolMLP** (SwiGLU Activation)
Implements the SwiGLU (Swish-Gated Linear Unit) MLP:
```python
class SmolMLP(nn.Module):
def forward(self, x):
# fc1 outputs 2 * intermediate_size
x = self.fc1(x) # (B, T, 2 * 1536) = (B, T, 3072)
x1, x2 = x.chunk(2, dim=-1) # Split into two parts
# SwiGLU: SiLU(x1) * x2
return self.fc2(F.silu(x1) * x2)
```
**Key Features:**
- **SwiGLU**: `SiLU(x1) * x2` activation (better than ReLU/GELU)
- **No bias**: Following LLaMA architecture
- **Efficient**: Single matrix multiplication with split
#### 6. **SmolBlock** (Transformer Block)
Combines attention and MLP with pre-norm and residual connections:
```python
class SmolBlock(nn.Module):
def forward(self, x, cos, sin, past_key_value=None, use_cache=False):
# Pre-norm attention with residual
attn_out, present_kv = self.attn(
self.attn_norm(x), cos, sin,
past_key_value=past_key_value, use_cache=use_cache
)
x = x + attn_out
# Pre-norm MLP with residual
x = x + self.mlp(self.mlp_norm(x))
return x, present_kv
```
**Architecture:**
- **Pre-norm**: Normalization before attention/MLP (not after)
- **Residual connections**: Skip connections for gradient flow
- **KV Cache passthrough**: Supports efficient inference
#### 7. **SmolLM2** (Main Model)
Top-level model that combines all components:
```python
class SmolLM2(nn.Module):
def __init__(self, config):
self.embed_tokens = nn.Embedding(vocab_size, hidden_size)
self.layers = nn.ModuleList([SmolBlock(config) for _ in range(30)])
self.norm = RMSNorm(hidden_size)
self.lm_head = nn.Linear(hidden_size, vocab_size, bias=False)
# Weight tying: share embeddings and output weights
self.lm_head.weight = self.embed_tokens.weight
def forward(self, input_ids, past_key_values=None, use_cache=False):
# 1. Token embeddings
x = self.embed_tokens(input_ids)
# 2. Build RoPE cache
cos, sin = build_rope_cache(...)
# 3. Pass through transformer layers
present_key_values = []
for layer in self.layers:
x, present_kv = layer(x, cos, sin, past_key_value, use_cache)
if use_cache:
present_key_values.append(present_kv)
# 4. Final norm and language modeling head
x = self.norm(x)
logits = self.lm_head(x)
return logits, present_key_values
```
**Key Features:**
- **Weight Tying**: Embeddings and output weights are shared (reduces parameters)
- **KV Cache Support**: Full support for efficient autoregressive generation
- **30 Layers**: Deep transformer stack for capacity
#### 8. **Generate Method** (Text Generation)
Autoregressive text generation with KV cache:
```python
@torch.no_grad()
def generate(self, input_ids, max_new_tokens=100, temperature=1.0,
top_k=None, top_p=None, eos_token_id=None):
generated = input_ids
past_key_values = None
for _ in range(max_new_tokens):
# Forward pass with KV cache
logits, past_key_values = self.forward(
generated[:, -1:] if past_key_values else generated,
past_key_values=past_key_values,
use_cache=True
)
# Sample next token with temperature, top-k, top-p
next_token_logits = logits[:, -1, :] / temperature
# Apply top-k and top-p filtering
probs = F.softmax(next_token_logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
generated = torch.cat([generated, next_token], dim=1)
if eos_token_id and (next_token == eos_token_id).all():
break
return generated
```
**Key Features:**
- **KV Cache**: Only processes new tokens (not entire sequence)
- **Sampling**: Supports temperature, top-k, and top-p (nucleus) sampling
- **Efficient**: O(1) per token after initial forward pass
### Model Specifications
| Parameter | Value |
|-----------|-------|
| **Total Parameters** | ~135M |
| **Hidden Size** | 576 |
| **Layers** | 30 |
| **Attention Heads** | 9 (Q), 3 (K/V) |
| **Head Dimension** | 64 |
| **Intermediate Size** | 1536 |
| **Vocabulary Size** | 49,152 |
| **Max Sequence Length** | 8,192 |
| **RoPE Theta** | 100,000 |
| **Activation** | SwiGLU (SiLU-gated) |
| **Normalization** | RMSNorm |
| **Weight Tying** | Yes (embeddings = output) |
### Key Design Choices
1. **GQA (Grouped Query Attention)**: 3:1 ratio reduces memory by 66% for K/V cache
2. **Pre-norm Architecture**: More stable training than post-norm
3. **RMSNorm**: Faster and simpler than LayerNorm
4. **RoPE**: Relative positional encoding, no learned embeddings
5. **SwiGLU**: Better activation than ReLU/GELU
6. **Weight Tying**: Reduces parameters and improves generalization
7. **No Biases**: Following LLaMA, reduces parameters slightly
### Usage Example
```python
from model import SmolConfig, SmolLM2
from transformers import AutoConfig
# Load config from HuggingFace
hf_config = AutoConfig.from_pretrained("HuggingFaceTB/SmolLM2-135M")
config = SmolConfig.from_hf(hf_config)
# Create model
model = SmolLM2(config)
# Forward pass (training)
input_ids = torch.randint(0, config.vocab_size, (2, 512))
logits, _ = model(input_ids, use_cache=False)
# Text generation (inference with KV cache)
prompt_ids = tokenizer.encode("Hello, how are you?")
generated = model.generate(
prompt_ids,
max_new_tokens=100,
temperature=0.8,
top_k=50
)
```
## Training
See `README_TRAINING.md` for detailed training instructions.
## Inference
See `app.py` for the Gradio web interface or use the `generate()` method directly.
## References
- [SmolLM2 Paper](https://arxiv.org/abs/2406.02528)
- [LLaMA Architecture](https://arxiv.org/abs/2302.13971)
- [RoPE: Rotary Position Embedding](https://arxiv.org/abs/2104.09864)
- [SwiGLU Activation](https://arxiv.org/abs/2002.05202)
|