File size: 7,436 Bytes
3d1cafb | 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 | """
Julian Model Layers.
Core building blocks: RMSNorm, RoPE, Attention, FFN.
"""
import math
from typing import Optional, Tuple
import jax
import jax.numpy as jnp
import flax.linen as nn
from flax.linen import initializers
from .config import JulianConfig
class RMSNorm(nn.Module):
"""Root Mean Square Layer Normalization."""
dim: int
eps: float = 1e-6
@nn.compact
def __call__(self, x: jnp.ndarray) -> jnp.ndarray:
weight = self.param(
"weight",
initializers.ones,
(self.dim,)
)
# RMS norm
variance = jnp.mean(x ** 2, axis=-1, keepdims=True)
x = x * jax.lax.rsqrt(variance + self.eps)
return x * weight
def precompute_rope_frequencies(
dim: int,
max_seq_len: int,
theta: float = 10000.0,
) -> Tuple[jnp.ndarray, jnp.ndarray]:
"""Precompute RoPE sin/cos frequencies."""
# Frequency for each dimension pair
freqs = 1.0 / (theta ** (jnp.arange(0, dim, 2).astype(jnp.float32) / dim))
# Position indices
positions = jnp.arange(max_seq_len)
# Outer product: [seq_len, dim/2]
angles = jnp.outer(positions, freqs)
# Sin and cos
sin = jnp.sin(angles)
cos = jnp.cos(angles)
return sin, cos
def apply_rope(
x: jnp.ndarray,
sin: jnp.ndarray,
cos: jnp.ndarray,
) -> jnp.ndarray:
"""Apply rotary position embeddings."""
# x shape: [batch, seq_len, n_heads, head_dim]
# sin/cos shape: [seq_len, head_dim/2]
seq_len = x.shape[1]
sin = sin[:seq_len]
cos = cos[:seq_len]
# Split x into pairs
x1 = x[..., ::2] # Even indices
x2 = x[..., 1::2] # Odd indices
# Rotate
# [batch, seq, heads, dim/2]
sin = sin[None, :, None, :] # Add batch and head dims
cos = cos[None, :, None, :]
rotated_x1 = x1 * cos - x2 * sin
rotated_x2 = x1 * sin + x2 * cos
# Interleave back
rotated = jnp.stack([rotated_x1, rotated_x2], axis=-1)
rotated = rotated.reshape(x.shape)
return rotated
class Attention(nn.Module):
"""Multi-head self-attention with RoPE."""
config: JulianConfig
@nn.compact
def __call__(
self,
x: jnp.ndarray,
sin: jnp.ndarray,
cos: jnp.ndarray,
mask: Optional[jnp.ndarray] = None,
deterministic: bool = True,
) -> jnp.ndarray:
batch_size, seq_len, _ = x.shape
config = self.config
# QKV projections
q = nn.Dense(
config.d_model,
use_bias=config.use_bias,
kernel_init=initializers.normal(config.initializer_range),
name="q_proj",
)(x)
k = nn.Dense(
config.d_model,
use_bias=config.use_bias,
kernel_init=initializers.normal(config.initializer_range),
name="k_proj",
)(x)
v = nn.Dense(
config.d_model,
use_bias=config.use_bias,
kernel_init=initializers.normal(config.initializer_range),
name="v_proj",
)(x)
# Reshape for multi-head attention
# [batch, seq, d_model] -> [batch, seq, n_heads, head_dim]
q = q.reshape(batch_size, seq_len, config.n_heads, config.head_dim)
k = k.reshape(batch_size, seq_len, config.n_heads, config.head_dim)
v = v.reshape(batch_size, seq_len, config.n_heads, config.head_dim)
# Apply RoPE to Q and K
q = apply_rope(q, sin, cos)
k = apply_rope(k, sin, cos)
# Transpose for attention: [batch, n_heads, seq, head_dim]
q = jnp.transpose(q, (0, 2, 1, 3))
k = jnp.transpose(k, (0, 2, 1, 3))
v = jnp.transpose(v, (0, 2, 1, 3))
# Scaled dot-product attention in bfloat16 for memory efficiency
scale = 1.0 / math.sqrt(config.head_dim)
# Force bfloat16 for attention computation (major memory savings)
q = q.astype(jnp.bfloat16)
k = k.astype(jnp.bfloat16)
v = v.astype(jnp.bfloat16)
attn_weights = jnp.einsum("bhqd,bhkd->bhqk", q, k) * scale
# Apply causal mask
if mask is not None:
attn_weights = jnp.where(mask, attn_weights, jnp.finfo(jnp.bfloat16).min)
attn_weights = jax.nn.softmax(attn_weights.astype(jnp.float32), axis=-1)
attn_weights = attn_weights.astype(jnp.bfloat16)
# Dropout
if not deterministic:
attn_weights = nn.Dropout(
rate=config.attention_dropout,
deterministic=deterministic,
)(attn_weights)
# Apply attention to values
attn_output = jnp.einsum("bhqk,bhkd->bhqd", attn_weights, v)
# Reshape back: [batch, n_heads, seq, head_dim] -> [batch, seq, d_model]
attn_output = jnp.transpose(attn_output, (0, 2, 1, 3))
attn_output = attn_output.reshape(batch_size, seq_len, config.d_model)
# Output projection
output = nn.Dense(
config.d_model,
use_bias=config.use_bias,
kernel_init=initializers.normal(config.initializer_range),
name="o_proj",
)(attn_output)
return output
class FeedForward(nn.Module):
"""SwiGLU Feed-Forward Network."""
config: JulianConfig
@nn.compact
def __call__(
self,
x: jnp.ndarray,
deterministic: bool = True,
) -> jnp.ndarray:
config = self.config
# Gate and up projections
gate = nn.Dense(
config.d_ff,
use_bias=config.use_bias,
kernel_init=initializers.normal(config.initializer_range),
name="gate_proj",
)(x)
up = nn.Dense(
config.d_ff,
use_bias=config.use_bias,
kernel_init=initializers.normal(config.initializer_range),
name="up_proj",
)(x)
# SwiGLU activation
hidden = jax.nn.silu(gate) * up
# Dropout
if not deterministic:
hidden = nn.Dropout(
rate=config.dropout,
deterministic=deterministic,
)(hidden)
# Down projection
output = nn.Dense(
config.d_model,
use_bias=config.use_bias,
kernel_init=initializers.normal(config.initializer_range),
name="down_proj",
)(hidden)
return output
class TransformerBlock(nn.Module):
"""Single transformer decoder block."""
config: JulianConfig
@nn.compact
def __call__(
self,
x: jnp.ndarray,
sin: jnp.ndarray,
cos: jnp.ndarray,
mask: Optional[jnp.ndarray] = None,
deterministic: bool = True,
) -> jnp.ndarray:
config = self.config
# Pre-norm attention
residual = x
x = RMSNorm(config.d_model, config.rms_norm_eps, name="input_layernorm")(x)
x = Attention(config, name="self_attn")(x, sin, cos, mask, deterministic)
if not deterministic:
x = nn.Dropout(rate=config.dropout, deterministic=deterministic)(x)
x = residual + x
# Pre-norm FFN
residual = x
x = RMSNorm(config.d_model, config.rms_norm_eps, name="post_attention_layernorm")(x)
x = FeedForward(config, name="mlp")(x, deterministic)
if not deterministic:
x = nn.Dropout(rate=config.dropout, deterministic=deterministic)(x)
x = residual + x
return x
|