import dataclasses
import os
import datasets
import tqdm
import tokenizers
import torch
import torch.nn as nn
import torch.nn.useful as F
import torch.optim.lr_scheduler as lr_scheduler
from torch import Tensor
# Load the tokenizer
tokenizer = tokenizers.Tokenizer.from_file(“bpe_50K.json”)
# Load the dataset
dataset = datasets.load_dataset(“HuggingFaceFW/fineweb”, “sample-10BT”, cut up=“practice”)
# Construct the mannequin
@dataclasses.dataclass
class LlamaConfig:
“”“Outline Llama mannequin hyperparameters.”“”
vocab_size: int = 50000 # Measurement of the tokenizer vocabulary
max_position_embeddings: int = 2048 # Most sequence size
hidden_size: int = 768 # Dimension of hidden layers
intermediate_size: int = 4*768 # Dimension of MLP’s hidden layer
num_hidden_layers: int = 12 # Variety of transformer layers
num_attention_heads: int = 12 # Variety of consideration heads
num_key_value_heads: int = 3 # Variety of key-value heads for GQA
def rotate_half(x: Tensor) -> Tensor:
“”“Rotates half the hidden dims of the enter.
It is a helper perform for rotary place embeddings (RoPE).
For a tensor of form (…, d), it returns a tensor the place the final
d/2 dimensions are rotated by swapping and negating.
Args:
x: Enter tensor of form (…, d)
Returns:
Tensor of identical form with rotated final dimension
““”
x1, x2 = x.chunk(2, dim=–1)
return torch.cat((–x2, x1), dim=–1) # Concatenate with rotation
class RotaryPositionEncoding(nn.Module):
“”“Rotary place encoding.”“”
def __init__(self, dim: int, max_position_embeddings: int) -> None:
“”“Initialize the RotaryPositionEncoding module
Args:
dim: The hidden dimension of the enter tensor to which RoPE is utilized
max_position_embeddings: The utmost sequence size of the enter tensor
““”
tremendous().__init__()
self.dim = dim
self.max_position_embeddings = max_position_embeddings
# compute a matrix of ntheta_i
N = 10_000.0
inv_freq = 1.0 / (N ** (torch.arange(0, dim, 2) / dim))
inv_freq = torch.cat((inv_freq, inv_freq), dim=–1)
place = torch.arange(max_position_embeddings)
sinusoid_inp = torch.outer(place, inv_freq)
# save cosine and sine matrices as buffers, not parameters
self.register_buffer(“cos”, sinusoid_inp.cos())
self.register_buffer(“sin”, sinusoid_inp.sin())
def ahead(self, x: Tensor) -> Tensor:
“”“Apply RoPE to tensor x
Args:
x: Enter tensor of form (batch_size, seq_length, num_heads, head_dim)
Returns:
Output tensor of form (batch_size, seq_length, num_heads, head_dim)
““”
batch_size, seq_len, num_heads, head_dim = x.form
dtype = x.dtype
# remodel the cosine and sine matrices to 4D tensor and the identical dtype as x
cos = self.cos.to(dtype)[:seq_len].view(1, seq_len, 1, –1)
sin = self.sin.to(dtype)[:seq_len].view(1, seq_len, 1, –1)
# apply RoPE to x
output = (x * cos) + (rotate_half(x) * sin)
return output
class LlamaAttention(nn.Module):
“”“Grouped-query consideration with rotary embeddings.”“”
def __init__(self, config: LlamaConfig) -> None:
tremendous().__init__()
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.hidden_size // self.num_heads
self.num_kv_heads = config.num_key_value_heads # GQA: H_kv < H_q
# hidden_size should be divisible by num_heads
assert (self.head_dim * self.num_heads) == self.hidden_measurement
# Linear layers for Q, Okay, V projections
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
self.k_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
self.v_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
def ahead(self, hidden_states: Tensor, rope: RotaryPositionEncoding, attn_mask: Tensor) -> Tensor:
bs, seq_len, dim = hidden_states.measurement()
# Undertaking inputs to Q, Okay, V
query_states = self.q_proj(hidden_states).view(bs, seq_len, self.num_heads, self.head_dim)
key_states = self.k_proj(hidden_states).view(bs, seq_len, self.num_kv_heads, self.head_dim)
value_states = self.v_proj(hidden_states).view(bs, seq_len, self.num_kv_heads, self.head_dim)
# Apply rotary place embeddings
query_states = rope(query_states)
key_states = rope(key_states)
# Transpose tensors from BSHD to BHSD dimension for scaled_dot_product_attention
query_states = query_states.transpose(1, 2)
key_states = key_states.transpose(1, 2)
value_states = value_states.transpose(1, 2)
# Use PyTorch’s optimized consideration implementation
# setting is_causal=True is incompatible with setting specific consideration masks
attn_output = F.scaled_dot_product_attention(
query_states,
key_states,
value_states,
attn_mask=attn_mask,
dropout_p=0.0,
enable_gqa=True,
)
# Transpose output tensor from BHSD to BSHD dimension, reshape to 3D, after which challenge output
attn_output = attn_output.transpose(1, 2).reshape(bs, seq_len, self.hidden_size)
attn_output = self.o_proj(attn_output)
return attn_output
class LlamaMLP(nn.Module):
“”“Feed-forward community with SwiGLU activation.”“”
def __init__(self, config: LlamaConfig) -> None:
tremendous().__init__()
# Two parallel projections for SwiGLU
self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
self.act_fn = F.silu # SwiGLU activation perform
# Undertaking again to hidden measurement
self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
def ahead(self, x: Tensor) -> Tensor:
# SwiGLU activation: multiply gate and up-projected inputs
gate = self.act_fn(self.gate_proj(x))
up = self.up_proj(x)
return self.down_proj(gate * up)
class LlamaDecoderLayer(nn.Module):
“”“Single transformer layer for a Llama mannequin.”“”
def __init__(self, config: LlamaConfig) -> None:
tremendous().__init__()
self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=1e–5)
self.self_attn = LlamaAttention(config)
self.post_attention_layernorm = nn.RMSNorm(config.hidden_size, eps=1e–5)
self.mlp = LlamaMLP(config)
def ahead(self, hidden_states: Tensor, rope: RotaryPositionEncoding, attn_mask: Tensor) -> Tensor:
# First residual block: Self-attention
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
attn_outputs = self.self_attn(hidden_states, rope=rope, attn_mask=attn_mask)
hidden_states = attn_outputs + residual
# Second residual block: MLP
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.mlp(hidden_states) + residual
return hidden_states
class LlamaModel(nn.Module):
“”“The total Llama mannequin with none pretraining heads.”“”
def __init__(self, config: LlamaConfig) -> None:
tremendous().__init__()
self.rotary_emb = RotaryPositionEncoding(
config.hidden_size // config.num_attention_heads,
config.max_position_embeddings,
)
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
self.norm = nn.RMSNorm(config.hidden_size, eps=1e–5)
def ahead(self, input_ids: Tensor, attn_mask: Tensor) -> Tensor:
# Convert enter token IDs to embeddings
hidden_states = self.embed_tokens(input_ids)
# Course of by way of all transformer layers, then the ultimate norm layer
for layer in self.layers:
hidden_states = layer(hidden_states, rope=self.rotary_emb, attn_mask=attn_mask)
hidden_states = self.norm(hidden_states)
# Return the ultimate hidden states
return hidden_states
class LlamaForPretraining(nn.Module):
def __init__(self, config: LlamaConfig) -> None:
tremendous().__init__()
self.base_model = LlamaModel(config)
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
def ahead(self, input_ids: Tensor, attn_mask: Tensor) -> Tensor:
hidden_states = self.base_model(input_ids, attn_mask)
return self.lm_head(hidden_states)
def create_causal_mask(seq_len: int, machine: torch.machine, dtype: torch.dtype = torch.float32) -> Tensor:
“”“Create a causal masks for self-attention.
Args:
seq_len: Size of the sequence
machine: System to create the masks on
dtype: Information kind of the masks
Returns:
Causal masks of form (seq_len, seq_len)
““”
masks = torch.full((seq_len, seq_len), float(‘-inf’), machine=machine, dtype=dtype)
.triu(diagonal=1)
return masks
def create_padding_mask(batch, padding_token_id, machine: torch.machine, dtype: torch.dtype = torch.float32) -> Tensor:
“”“Create a padding masks for a batch of sequences for self-attention.
Args:
batch: Batch of sequences, form (batch_size, seq_len)
padding_token_id: ID of the padding token
Returns:
Padding masks of form (batch_size, 1, seq_len, seq_len)
““”
padded = torch.zeros_like(batch, machine=machine, dtype=dtype)
.masked_fill(batch == padding_token_id, float(‘-inf’))
masks = padded[:,:,None] + padded[:,None,:]
return masks[:, None, :, :]
# Generator perform to create padded sequences of fastened size
class PretrainingDataset(torch.utils.information.Dataset):
def __init__(self, dataset: datasets.Dataset, tokenizer: tokenizers.Tokenizer,
seq_length: int, machine: torch.machine = None):
self.dataset = dataset
self.tokenizer = tokenizer
self.machine = machine
self.seq_length = seq_length
self.bot = tokenizer.token_to_id(“[BOT]”)
self.eot = tokenizer.token_to_id(“[EOT]”)
self.pad = tokenizer.token_to_id(“[PAD]”)
def __len__(self):
return len(self.dataset)
def __getitem__(self, index):
“”“Get a sequence of token ids from the dataset. [BOT] and [EOT] tokens
are added. Clipped and padded to the sequence size.
““”
seq = self.dataset[index][“text”]
tokens: listing[int] = [self.bot] + self.tokenizer.encode(seq).ids + [self.eot]
# pad to focus on sequence size
toklen = len(tokens)
if toklen < self.seq_length+1:
pad_length = self.seq_length+1 – toklen
tokens += [self.pad] * pad_size
# return the sequence
x = torch.tensor(tokens[:self.seq_length], dtype=torch.int64, machine=self.machine)
y = torch.tensor(tokens[1:self.seq_length+1], dtype=torch.int64, machine=self.machine)
return x, y
# Create pretraining mannequin with default config
model_config = LlamaConfig()
machine = torch.machine(“cuda”) if torch.cuda.is_available() else torch.machine(“cpu”)
mannequin = LlamaForPretraining(model_config).to(machine)
# print the mannequin measurement
print(f“Mannequin parameters measurement: {sum(p.numel() for p in mannequin.parameters()) / 1024**2:.2f} M”)
print(f“Mannequin buffers measurement: {sum(p.numel() for p in mannequin.buffers()) / 1024**2:.2f} M”)
# Coaching parameters
epochs = 3
learning_rate = 1e–3
batch_size = 8
seq_length = 512
num_warmup_steps = 1000
PAD_TOKEN_ID = tokenizer.token_to_id(“[PAD]”)
# DataLoader, optimizer, scheduler, and loss perform
mannequin.practice()
dataloader = torch.utils.information.DataLoader(
PretrainingDataset(dataset, tokenizer, seq_length, machine),
batch_size=batch_measurement
)
optimizer = torch.optim.AdamW(
mannequin.parameters(), lr=learning_rate, betas=(0.9, 0.95), eps=1e–8, weight_decay=0.01
)
num_training_steps = len(dataloader) * epochs
warmup_scheduler = lr_scheduler.LinearLR(
optimizer,
start_factor=0.1, end_factor=1.0, total_iters=num_warmup_steps
)
cosine_scheduler = lr_scheduler.CosineAnnealingLR(
optimizer,
T_max=num_training_steps – num_warmup_steps,
eta_min=0
)
scheduler = lr_scheduler.SequentialLR(
optimizer,
schedulers=[warmup_scheduler, cosine_scheduler],
milestones=[num_warmup_steps]
)
loss_fn = nn.CrossEntropyLoss(ignore_index=PAD_TOKEN_ID)
# search for final checkpoint
if os.path.exists(“llama_pretraining_checkpoint.pth”):
checkpoint = torch.load(“llama_pretraining_checkpoint.pth”)
begin_epoch = checkpoint[“epoch”]
begin_batch = checkpoint[“batch”]
mannequin.load_state_dict(checkpoint[“model”])
optimizer.load_state_dict(checkpoint[“optimizer”])
scheduler.load_state_dict(checkpoint[“scheduler”])
del checkpoint
print(f“Resuming coaching from epoch {begin_epoch} and batch {begin_batch}”)
else:
begin_epoch = 0
begin_batch = 0
# begin coaching
for epoch in vary(begin_epoch, epochs):
dataloader = torch.utils.information.DataLoader(
PretrainingDataset(
dataset.skip(begin_batch * batch_size),
tokenizer,
seq_length,
machine,
),
batch_size=batch_measurement
)
pbar = tqdm.tqdm(dataloader, desc=f“Epoch {epoch+1}/{epochs}”)
for batch_id, batch in enumerate(pbar):
if (begin_batch + batch_id) % 1000 == 0:
# checkpoint the mannequin and optimizer state
torch.save({
“mannequin”: mannequin.state_dict(),
“optimizer”: optimizer.state_dict(),
“scheduler”: scheduler.state_dict(),
“epoch”: epoch,
“batch”: batch_id + begin_batch,
}, f“llama_pretraining_checkpoint.pth”)
# get batched information
input_ids, target_ids = batch
# create consideration masks: causal masks + padding masks
attn_mask = create_causal_mask(input_ids.form[1], machine) +
create_padding_mask(input_ids, PAD_TOKEN_ID, machine)
# extract output from mannequin
logits = mannequin(input_ids, attn_mask)
# compute loss: cross-entropy between logits and goal, ignoring padding tokens
loss = loss_fn(logits.view(–1, logits.measurement(–1)), target_ids.view(–1))
# backward with loss and apply gradient clipping
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(mannequin.parameters(), 1.0)
optimizer.step()
scheduler.step()
pbar.set_postfix(loss=loss.merchandise())
pbar.replace(1)
begin_batch = 0 # reset for subsequent epoch
pbar.shut()
# Save the mannequin
torch.save(mannequin.state_dict(), “llama_pretraining_model.pth”)
torch.save(mannequin.base_model.state_dict(), “llama_model.pth”)
