File size: 13,057 Bytes
0433390 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 | """
Training utilities for DFlash drafters on MLX.
Implements the training recipe from the DFlash paper:
- KV injection with target model features
- Random anchor sampling for block construction
- Sparse attention masking within blocks
- Position-dependent loss decay
"""
import math
from typing import Optional, List, Dict, Any, Tuple
import mlx.core as mx
import mlx.nn as nn
import mlx.optimizers as optim
from .model import DFlashDraftModel
class DFlashTrainer:
"""Trainer for DFlash draft models on MLX.
Trains the drafter to align block-level diffusion predictions
with a frozen autoregressive target model's outputs.
"""
def __init__(
self,
target_model,
drafter: DFlashDraftModel,
tokenizer,
max_seq_length: int = 3072,
):
self.target_model = target_model
self.drafter = drafter
self.tokenizer = tokenizer
self.max_seq_length = max_seq_length
self.mask_token_id = drafter.mask_token_id
def _prepare_training_sample(
self,
prompt: str,
response: str,
block_size: int,
) -> Dict[str, mx.array]:
"""Prepare a single training sample.
Constructs masked blocks with random anchors from target-generated
responses, matching the inference-time speculative decoding setting.
"""
# Tokenize prompt + response
prompt_ids = self.tokenizer.encode(prompt)
response_ids = self.tokenizer.encode(response)
# Truncate if too long
total_len = len(prompt_ids) + len(response_ids)
if total_len > self.max_seq_length:
response_ids = response_ids[:self.max_seq_length - len(prompt_ids)]
full_ids = prompt_ids + response_ids
full_ids_mx = mx.array(full_ids)
# Build target context features
with mx.eval_mode():
target_output = self._target_forward(full_ids_mx)
target_hidden = self.drafter.extract_context_features(
target_output["hidden_states"]
)
# Random anchor sampling for blocks
num_blocks = max(1, len(response_ids) // block_size)
block_starts = mx.random.randint(
low=len(prompt_ids),
high=len(full_ids) - block_size + 1,
shape=(num_blocks,),
)
# Create masked sequence
masked_ids = mx.array(full_ids)
labels = mx.full((len(full_ids),), -100, dtype=mx.int32) # Ignore index
for start in block_starts.tolist():
start = int(start)
end = min(start + block_size, len(full_ids))
# Anchor is first token (from target model's accepted token)
# Mask remaining positions in block
masked_ids = masked_ids.at[start + 1:end].set(self.mask_token_id)
# Labels for masked positions
labels = labels.at[start + 1:end].set(full_ids_mx[start + 1:end])
return {
"input_ids": masked_ids,
"labels": labels,
"target_hidden": target_hidden,
"prompt_length": len(prompt_ids),
}
def _target_forward(
self,
input_ids: mx.array,
) -> Dict[str, Any]:
"""Forward pass through target model to get hidden states."""
if hasattr(self.target_model, '__call__'):
result = self.target_model(input_ids)
logits = result[0] if isinstance(result, tuple) else result
else:
logits = self.target_model(input_ids)
# Extract hidden states layer by layer
hidden_states = []
hidden = input_ids
if hasattr(self.target_model, 'embed_tokens'):
hidden = self.target_model.embed_tokens(hidden)
if hasattr(self.target_model, 'layers'):
for layer in self.target_model.layers:
hidden = layer(hidden, mask=None)
hidden_states.append(hidden)
else:
hidden_states = [hidden]
return {
"logits": logits,
"hidden_states": hidden_states,
}
def _compute_loss(
self,
input_ids: mx.array,
labels: mx.array,
target_hidden: mx.array,
) -> mx.array:
"""Compute the diffusion training loss with position-dependent decay.
Implements the loss decay from the paper where tokens closer to
the anchor receive higher weights.
"""
# Embed tokens (including mask tokens)
embeddings = self.drafter.embed_tokens(input_ids)
# Build position IDs
position_ids = mx.arange(input_ids.shape[0])
# Forward through drafter
hidden_states = self.drafter(
noise_embedding=embeddings,
target_hidden=target_hidden,
position_ids=position_ids,
)
# Get logits
logits = self.drafter.get_logits(hidden_states)
# Compute cross-entropy loss for labeled positions
valid_mask = labels != -100
if not valid_mask.any():
return mx.array(0.0)
valid_logits = logits[valid_mask]
valid_labels = labels[valid_mask]
# Position-dependent weighting (exponential decay from anchor)
# Find anchor positions and compute distances
positions = mx.arange(len(labels))
# Simplified: uniform weighting for now
# Full implementation would track block boundaries
weights = mx.ones_like(valid_labels, dtype=mx.float32)
# Cross entropy
log_probs = mx.log_softmax(valid_logits, axis=-1)
nll = -log_probs[mx.arange(len(valid_labels)), valid_labels]
weighted_nll = nll * weights
return weighted_nll.mean()
def _build_batch(
self,
samples: List[Dict[str, Any]],
) -> Dict[str, mx.array]:
"""Batch multiple training samples."""
# Find max length
max_len = max(s["input_ids"].shape[0] for s in samples)
# Pad sequences
batch_input_ids = []
batch_labels = []
batch_target_hidden = []
batch_attention_mask = []
for sample in samples:
seq_len = sample["input_ids"].shape[0]
pad_len = max_len - seq_len
# Pad input_ids with mask token
padded_ids = mx.concatenate([
sample["input_ids"],
mx.full((pad_len,), self.mask_token_id, dtype=mx.int32)
])
batch_input_ids.append(padded_ids)
# Pad labels with -100 (ignore index)
padded_labels = mx.concatenate([
sample["labels"],
mx.full((pad_len,), -100, dtype=mx.int32)
])
batch_labels.append(padded_labels)
# Attention mask (1 for real, 0 for padding)
mask = mx.concatenate([
mx.ones((seq_len,), dtype=mx.float32),
mx.zeros((pad_len,), dtype=mx.float32)
])
batch_attention_mask.append(mask)
# Target hidden (pad with zeros)
hidden = sample["target_hidden"]
if hidden.shape[1] < max_len:
pad = mx.zeros((hidden.shape[0], max_len - hidden.shape[1], hidden.shape[2]))
hidden = mx.concatenate([hidden, pad], axis=1)
batch_target_hidden.append(hidden)
return {
"input_ids": mx.stack(batch_input_ids),
"labels": mx.stack(batch_labels),
"target_hidden": mx.stack(batch_target_hidden),
"attention_mask": mx.stack(batch_attention_mask),
}
def train(
self,
dataset: str,
epochs: int = 6,
batch_size: int = 8,
lr: float = 6e-4,
warmup_ratio: float = 0.04,
grad_clip: float = 1.0,
save_every: int = 1000,
) -> DFlashDraftModel:
"""Train the DFlash drafter.
Args:
dataset: Path to dataset (JSONL with {prompt, response} pairs)
or HF dataset name with 'prompt' and 'response' columns
epochs: Number of training epochs
batch_size: Batch size
lr: Learning rate
warmup_ratio: Warmup ratio for cosine schedule
grad_clip: Gradient clipping threshold
save_every: Save checkpoint every N steps
Returns:
Trained DFlashDraftModel
"""
# Load dataset
samples = self._load_dataset(dataset)
print(f"[Trainer] Loaded {len(samples)} training samples")
# Setup optimizer
optimizer = optim.AdamW(learning_rate=lr)
# Cosine schedule with warmup
num_steps = (len(samples) // batch_size) * epochs
warmup_steps = int(num_steps * warmup_ratio)
def lr_schedule(step):
if step < warmup_steps:
return lr * (step / warmup_steps)
progress = (step - warmup_steps) / max(1, num_steps - warmup_steps)
return lr * 0.5 * (1 + math.cos(math.pi * progress))
# Training loop
step = 0
for epoch in range(epochs):
# Shuffle samples
import random
random.shuffle(samples)
epoch_losses = []
for i in range(0, len(samples), batch_size):
batch_samples = samples[i:i + batch_size]
# Prepare batch
batch = self._build_batch(batch_samples)
# Forward + backward
def loss_fn(params):
self.drafter.update(params)
loss = self._compute_loss(
batch["input_ids"],
batch["labels"],
batch["target_hidden"],
)
return loss
# Compute loss and gradients
loss, grads = mx.value_and_grad(loss_fn)(self.drafter.parameters())
# Gradient clipping
if grad_clip > 0:
grad_norm = mx.sqrt(sum(mx.sum(g * g) for g in grads.values()))
if grad_norm > grad_clip:
scale = grad_clip / grad_norm
grads = {k: v * scale for k, v in grads.items()}
# Update parameters
current_lr = lr_schedule(step)
optimizer.learning_rate = current_lr
self.drafter = optimizer.apply(grads, self.drafter)
loss_val = float(loss)
epoch_losses.append(loss_val)
if step % 10 == 0:
avg_loss = sum(epoch_losses[-10:]) / len(epoch_losses[-10:])
print(f"[Trainer] Epoch {epoch+1}/{epochs} Step {step} | "
f"Loss: {loss_val:.4f} | LR: {current_lr:.2e}")
step += 1
# Save checkpoint
if step % save_every == 0:
self._save_checkpoint(f"checkpoint_step_{step}")
avg_epoch_loss = sum(epoch_losses) / len(epoch_losses)
print(f"[Trainer] Epoch {epoch+1} complete | Avg Loss: {avg_epoch_loss:.4f}")
print("[Trainer] Training complete!")
return self.drafter
def _load_dataset(self, dataset: str) -> List[Dict[str, str]]:
"""Load dataset from path or HF Hub."""
import json
from pathlib import Path
# Try local file first
dataset_path = Path(dataset)
if dataset_path.exists():
samples = []
with open(dataset_path, "r") as f:
for line in f:
data = json.loads(line)
samples.append({
"prompt": data.get("prompt", data.get("input", "")),
"response": data.get("response", data.get("output", data.get("completion", ""))),
})
return samples
# Try Hugging Face dataset
try:
from datasets import load_dataset
ds = load_dataset(dataset, split="train")
samples = []
for item in ds:
prompt = item.get("prompt", item.get("input", item.get("question", "")))
response = item.get("response", item.get("output", item.get("answer", item.get("completion", ""))))
if prompt and response:
samples.append({"prompt": prompt, "response": response})
return samples
except Exception as e:
print(f"[Trainer] Failed to load dataset: {e}")
return []
def _save_checkpoint(self, name: str):
"""Save a training checkpoint."""
import json
from pathlib import Path
checkpoint_dir = Path("checkpoints") / name
checkpoint_dir.mkdir(parents=True, exist_ok=True)
weights = dict(self.drafter.parameters())
mx.save_safetensors(str(checkpoint_dir / "weights.safetensors"), weights)
print(f"[Trainer] Saved checkpoint to {checkpoint_dir}")
|