File size: 6,035 Bytes
187a491 | 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 | """PyTorch Lightning module for training the BoccaccioAI model."""
from __future__ import annotations
import logging
from typing import Any
import torch
import lightning as L
from src.model.config import BoccaccioConfig
from src.model.transformer import BoccaccioForCausalLM
from src.training.scheduler import get_cosine_schedule_with_warmup
log = logging.getLogger(__name__)
class BoccaccioLightningModule(L.LightningModule):
"""Lightning wrapper around :class:`BoccaccioForCausalLM`.
Parameters
----------
config:
Model architecture configuration.
training_config:
Dictionary with training hyper-parameters. Expected keys include
``learning_rate``, ``weight_decay``, ``beta1``, ``beta2``,
``warmup_steps``, ``num_tokens``, ``micro_batch_size``,
``gradient_accumulation_steps``, ``sequence_length``, and
optionally ``compile_model``.
"""
def __init__(self, config: BoccaccioConfig, training_config: dict[str, Any]) -> None:
super().__init__()
self.config = config
self.training_config = training_config
self.model = BoccaccioForCausalLM(config)
if training_config.get("compile_model", False):
self.model = torch.compile(self.model)
num_params = sum(p.numel() for p in self.model.parameters() if p.requires_grad)
log.info("Trainable parameters: %s (%.2fM)", num_params, num_params / 1e6)
# ------------------------------------------------------------------
# Training / validation steps
# ------------------------------------------------------------------
def training_step(self, batch: dict[str, torch.Tensor], batch_idx: int) -> torch.Tensor:
input_ids = batch["input_ids"]
labels = batch["labels"]
attention_mask = batch.get("attention_mask")
outputs = self.model(input_ids=input_ids, labels=labels, attention_mask=attention_mask)
loss = outputs["loss"]
self.log("train/loss", loss, prog_bar=True, on_step=True, on_epoch=False)
self.log("train/ppl", torch.exp(loss), prog_bar=True, on_step=True, on_epoch=False)
self.log(
"train/lr",
self.optimizers().param_groups[0]["lr"],
on_step=True,
on_epoch=False,
)
return loss
def validation_step(self, batch: dict[str, torch.Tensor], batch_idx: int) -> torch.Tensor:
input_ids = batch["input_ids"]
labels = batch["labels"]
attention_mask = batch.get("attention_mask")
outputs = self.model(input_ids=input_ids, labels=labels, attention_mask=attention_mask)
loss = outputs["loss"]
self.log("val/loss", loss, prog_bar=True, on_step=True, on_epoch=False)
self.log("val/ppl", torch.exp(loss), prog_bar=True, on_step=True, on_epoch=False)
return loss
# ------------------------------------------------------------------
# Optimizer and scheduler
# ------------------------------------------------------------------
def configure_optimizers(self) -> dict[str, Any]:
decay_params: list[torch.nn.Parameter] = []
no_decay_params: list[torch.nn.Parameter] = []
for name, param in self.model.named_parameters():
if not param.requires_grad:
continue
# Biases, embedding weights, and RMSNorm weights do not get
# weight-decay; only Linear *weight* matrices are decayed.
if name.endswith(".bias"):
no_decay_params.append(param)
elif "embed_tokens" in name:
no_decay_params.append(param)
elif "norm" in name:
no_decay_params.append(param)
else:
decay_params.append(param)
optimizer = torch.optim.AdamW(
[
{"params": decay_params, "weight_decay": self.training_config["weight_decay"]},
{"params": no_decay_params, "weight_decay": 0.0},
],
lr=self.training_config["learning_rate"],
betas=(
self.training_config["beta1"],
self.training_config["beta2"],
),
)
# Calculate total steps for cosine schedule.
# Pre-training: based on num_tokens (binary data, no padding).
# Fine-tuning: based on num_examples * num_epochs / batch_size
# (each example is one batch item, padded to sequence_length).
batch_size = self.training_config["micro_batch_size"]
num_epochs = self.training_config.get("num_epochs", 1)
if self.training_config.get("num_examples", 0) > 0:
# Fine-tuning: optimizer steps = examples * epochs / (batch_size * grad_accum)
# The scheduler with interval="step" is called per optimizer step,
# not per forward pass, so we must divide by gradient_accumulation_steps.
num_examples = self.training_config["num_examples"]
grad_accum = self.training_config.get("gradient_accumulation_steps", 1)
total_steps = (num_examples * num_epochs) // (batch_size * grad_accum)
else:
# Pre-training: steps = tokens * epochs / batch_tokens
batch_tokens = (
batch_size
* self.training_config["gradient_accumulation_steps"]
* self.training_config["sequence_length"]
)
num_tokens = self.training_config.get("num_tokens", 0)
if num_tokens > 0:
total_steps = (num_tokens * num_epochs) // batch_tokens
else:
total_steps = 5000 # fallback
scheduler = get_cosine_schedule_with_warmup(
optimizer,
warmup_steps=self.training_config["warmup_steps"],
total_steps=total_steps,
)
return {
"optimizer": optimizer,
"lr_scheduler": {
"scheduler": scheduler,
"interval": "step",
},
}
|