File size: 11,556 Bytes
30e9297 | 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 | """
Music Transformer Model — LLaMA-style architecture for symbolic music generation.
Key innovations combined:
- Rotary Position Embeddings (RoPE) — better long-range modeling than sinusoidal
- RMSNorm — faster than LayerNorm, used in LLaMA/Mistral
- SwiGLU activation — better than GELU/ReLU, used in LLaMA
- Grouped Query Attention (GQA) — reduces KV-cache memory by sharing KV heads
- Gradient checkpointing — cuts memory usage ~50% with ~20% speed cost
- KV-cache — O(1) per-token inference instead of O(n)
"""
import math
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
class RMSNorm(nn.Module):
"""Root Mean Square Layer Normalization (faster than LayerNorm)."""
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim))
def forward(self, x: torch.Tensor) -> torch.Tensor:
norm = x.float().pow(2).mean(-1, keepdim=True).add(self.eps).rsqrt()
return (x.float() * norm).type_as(x) * self.weight
def precompute_rope_freqs(dim: int, max_seq_len: int, theta: float = 10000.0) -> torch.Tensor:
"""Precompute RoPE frequency tensor for complex exponentials."""
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
t = torch.arange(max_seq_len, dtype=torch.float32)
freqs = torch.outer(t, freqs)
return torch.polar(torch.ones_like(freqs), freqs) # complex64
def apply_rope(xq: torch.Tensor, xk: torch.Tensor, freqs: torch.Tensor):
"""Apply rotary embeddings to query and key tensors."""
# Reshape to complex
xq_c = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
xk_c = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
# Reshape freqs for broadcasting: (seq_len,) -> (1, seq_len, 1, head_dim//2)
freqs = freqs.unsqueeze(0).unsqueeze(2)
xq_out = torch.view_as_real(xq_c * freqs).flatten(-2)
xk_out = torch.view_as_real(xk_c * freqs).flatten(-2)
return xq_out.type_as(xq), xk_out.type_as(xk)
def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
"""Repeat KV heads to match query head count for GQA."""
if n_rep == 1:
return x
bs, seq_len, n_kv_heads, head_dim = x.shape
return (
x[:, :, :, None, :]
.expand(bs, seq_len, n_kv_heads, n_rep, head_dim)
.reshape(bs, seq_len, n_kv_heads * n_rep, head_dim)
)
class GroupedQueryAttention(nn.Module):
"""
Multi-head attention with Grouped Query Attention (GQA).
Uses fewer KV heads than Q heads to reduce memory.
"""
def __init__(self, dim: int, n_heads: int, n_kv_heads: int, dropout: float = 0.1):
super().__init__()
self.n_heads = n_heads
self.n_kv_heads = n_kv_heads
self.n_rep = n_heads // n_kv_heads
self.head_dim = dim // n_heads
self.wq = nn.Linear(dim, n_heads * self.head_dim, bias=False)
self.wk = nn.Linear(dim, n_kv_heads * self.head_dim, bias=False)
self.wv = nn.Linear(dim, n_kv_heads * self.head_dim, bias=False)
self.wo = nn.Linear(n_heads * self.head_dim, dim, bias=False)
self.attn_dropout = nn.Dropout(dropout)
self.resid_dropout = nn.Dropout(dropout)
# KV-cache for inference
self.cache_k: Optional[torch.Tensor] = None
self.cache_v: Optional[torch.Tensor] = None
def forward(
self,
x: torch.Tensor,
freqs: torch.Tensor,
mask: Optional[torch.Tensor] = None,
use_cache: bool = False,
) -> torch.Tensor:
bs, seq_len, _ = x.shape
q = self.wq(x).view(bs, seq_len, self.n_heads, self.head_dim)
k = self.wk(x).view(bs, seq_len, self.n_kv_heads, self.head_dim)
v = self.wv(x).view(bs, seq_len, self.n_kv_heads, self.head_dim)
# Apply RoPE to Q and K
q_rope = q.view(bs, seq_len, self.n_heads, self.head_dim)
k_rope = k.view(bs, seq_len, self.n_kv_heads, self.head_dim)
# RoPE needs (bs, seq_len, heads, head_dim) but freqs is (seq_len, head_dim//2)
# Apply per-head
q_for_rope = q_rope.reshape(bs * self.n_heads, seq_len, self.head_dim)
k_for_rope = k_rope.reshape(bs * self.n_kv_heads, seq_len, self.head_dim)
# Simpler RoPE application
q = q.transpose(1, 2) # (bs, n_heads, seq_len, head_dim)
k = k.transpose(1, 2)
v = v.transpose(1, 2)
# Apply RoPE via cos/sin (more compatible than complex)
q, k = self._apply_rope_real(q, k, freqs)
# KV-cache for generation
if use_cache:
if self.cache_k is not None:
k = torch.cat([self.cache_k, k], dim=2)
v = torch.cat([self.cache_v, v], dim=2)
self.cache_k = k.detach()
self.cache_v = v.detach()
# GQA: repeat KV heads
k = repeat_kv(k.transpose(1, 2), self.n_rep).transpose(1, 2)
v = repeat_kv(v.transpose(1, 2), self.n_rep).transpose(1, 2)
# Scaled dot-product attention (uses Flash Attention when available)
scale = 1.0 / math.sqrt(self.head_dim)
try:
# PyTorch 2.0+ SDPA with memory-efficient backend
out = F.scaled_dot_product_attention(
q, k, v,
attn_mask=mask,
dropout_p=self.attn_dropout.p if self.training else 0.0,
is_causal=(mask is None and not use_cache),
)
except RuntimeError:
# Fallback for older PyTorch
scores = torch.matmul(q, k.transpose(-2, -1)) * scale
if mask is not None:
scores = scores + mask
elif not use_cache:
causal = torch.triu(
torch.full((seq_len, seq_len), float("-inf"), device=x.device), diagonal=1
)
scores = scores + causal
scores = F.softmax(scores, dim=-1)
scores = self.attn_dropout(scores)
out = torch.matmul(scores, v)
out = out.transpose(1, 2).contiguous().view(bs, seq_len, -1)
return self.resid_dropout(self.wo(out))
def _apply_rope_real(self, q, k, freqs):
"""Apply RoPE using real-valued sin/cos (more device-compatible)."""
# freqs shape: (seq_len, head_dim//2)
seq_len = q.shape[2]
freqs = freqs[:seq_len]
cos_f = freqs.cos().unsqueeze(0).unsqueeze(0) # (1, 1, seq_len, head_dim//2)
sin_f = freqs.sin().unsqueeze(0).unsqueeze(0)
def rotate_half(x):
x1, x2 = x.chunk(2, dim=-1)
return torch.cat((-x2, x1), dim=-1)
q = q * cos_f.repeat(1, 1, 1, 2) + rotate_half(q) * sin_f.repeat(1, 1, 1, 2)
k = k * cos_f.repeat(1, 1, 1, 2) + rotate_half(k) * sin_f.repeat(1, 1, 1, 2)
return q, k
def reset_cache(self):
self.cache_k = None
self.cache_v = None
class SwiGLU(nn.Module):
"""SwiGLU activation — superior to GELU/ReLU, used in LLaMA."""
def __init__(self, dim: int, hidden_dim: int, dropout: float = 0.1):
super().__init__()
self.w1 = nn.Linear(dim, hidden_dim, bias=False)
self.w2 = nn.Linear(hidden_dim, dim, bias=False)
self.w3 = nn.Linear(dim, hidden_dim, bias=False)
self.dropout = nn.Dropout(dropout)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.dropout(self.w2(F.silu(self.w1(x)) * self.w3(x)))
class TransformerBlock(nn.Module):
"""Single transformer block with pre-norm architecture."""
def __init__(self, dim: int, n_heads: int, n_kv_heads: int, hidden_dim: int, dropout: float):
super().__init__()
self.attention = GroupedQueryAttention(dim, n_heads, n_kv_heads, dropout)
self.feed_forward = SwiGLU(dim, hidden_dim, dropout)
self.norm1 = RMSNorm(dim)
self.norm2 = RMSNorm(dim)
def forward(
self,
x: torch.Tensor,
freqs: torch.Tensor,
mask: Optional[torch.Tensor] = None,
use_cache: bool = False,
) -> torch.Tensor:
# Pre-norm residual connections
x = x + self.attention(self.norm1(x), freqs, mask, use_cache)
x = x + self.feed_forward(self.norm2(x))
return x
class MusicTransformer(nn.Module):
"""
LLaMA-style Transformer for music generation.
Combines: RoPE + GQA + SwiGLU + RMSNorm + gradient checkpointing.
~5M parameters with default config — suitable for training on consumer GPUs.
"""
def __init__(self, config):
super().__init__()
self.config = config
self.token_emb = nn.Embedding(config.vocab_size, config.dim)
self.dropout = nn.Dropout(config.dropout)
self.layers = nn.ModuleList([
TransformerBlock(
config.dim, config.n_heads, config.n_kv_heads,
config.hidden_dim, config.dropout,
)
for _ in range(config.n_layers)
])
self.norm = RMSNorm(config.dim)
self.output = nn.Linear(config.dim, config.vocab_size, bias=False)
# Weight tying — reduces params and improves generalization
self.token_emb.weight = self.output.weight
# Precompute RoPE frequencies
head_dim = config.dim // config.n_heads
freqs = self._precompute_freqs(head_dim, config.max_seq_len, config.rope_theta)
self.register_buffer("freqs", freqs, persistent=False)
self.grad_checkpoint = False
self._init_weights()
def _precompute_freqs(self, dim: int, max_seq_len: int, theta: float) -> torch.Tensor:
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
t = torch.arange(max_seq_len, dtype=torch.float32)
return torch.outer(t, freqs)
def _init_weights(self):
"""Xavier-style initialization for stable training."""
for module in self.modules():
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(
self,
input_ids: torch.Tensor,
targets: Optional[torch.Tensor] = None,
use_cache: bool = False,
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
bs, seq_len = input_ids.shape
h = self.dropout(self.token_emb(input_ids))
freqs = self.freqs[:seq_len].to(h.device)
for layer in self.layers:
if self.grad_checkpoint and self.training:
h = torch.utils.checkpoint.checkpoint(
layer, h, freqs, None, use_cache, use_reentrant=False
)
else:
h = layer(h, freqs, use_cache=use_cache)
h = self.norm(h)
logits = self.output(h)
loss = None
if targets is not None:
loss = F.cross_entropy(
logits.view(-1, logits.size(-1)),
targets.view(-1),
ignore_index=0, # Ignore padding
)
return logits, loss
def reset_caches(self):
for layer in self.layers:
layer.attention.reset_cache()
def count_parameters(self) -> int:
return sum(p.numel() for p in self.parameters() if p.requires_grad)
@classmethod
def from_config(cls, model_config) -> "MusicTransformer":
return cls(model_config)
|