Download to read offline
![14/29
stoi = { ch:i for i,ch in enumerate(chars) }
itos = { i:ch for i,ch in enumerate(chars) }
encode = lambda s: [stoi[c] for c in s] # encoder: take a string, output a list of integers
decode = lambda l: ''.join([itos[i] for i in l]) # decoder: take a list of integers, output a string
print(encode("hii there"))
print(decode(encode("hii there")))
This code block creates a character-to-integer mapping and its inverse (integer-to-
character mapping) for a set of characters. The stoi dictionary maps each character to a
unique integer while itos maps each integer back to its corresponding character. The
encode function takes a string as input and returns a list of integers, where each integer
corresponds to the index of the character in the chars set. The decode function takes a
list of integers and returns the original string by looking up the corresponding characters
in the itos dictionary. The code then tests the encoding and decoding functions by
encoding the string “hii there” and then decoding the resulting list of integers back into a
string.
Encoding input data
In building a GPT model, it’s important to encode the entire text dataset so that it can be
fed into the model. The following code does exactly that.
let's now encode the entire text dataset and store it into a torch.Tensor
import torch # we use PyTorch: https://pytorch.org
data = torch.tensor(encode(text), dtype=torch.long)
print(data.shape, data.dtype)
print(data[:1000]) # the 1000 characters we looked at earier will to the GPT look like this
This code imports the PyTorch library and creates a tensor called data. The tensor is filled
with the encoded text data, which is obtained by calling the encode function on the text
variable. The dtype parameter is set to torch.long to ensure that the tensor elements are
integers. The code prints the shape and data type of the data tensor. The shape attribute
tells us the size of the tensor along each dimension, while the dtype attribute tells us the
data type of the tensor elements. This information is useful for verifying that the tensor
has been created correctly and will be compatible with the GPT model. It then prints the
first 1000 elements of the data tensor, which represent the encoded text data. This is
useful for verifying that the encoding process has worked correctly and that the data has
been loaded into the tensor as expected.
Splitting up the data into train and validation sets](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-14-320.jpg)
![15/29
The following code is useful for understanding how the GPT model will process the input
data. It shows how the model will process input sequences of length block_size, and how
the input and output sequences are related to each other. This understanding can help in
designing and training the GPT model.
# Let's now split up the data into train and validation sets
n = int(0.9*len(data)) # first 90% will be train, rest val
train_data = data[:n]
val_data = data[n:]
block_size = 8
train_data[:block_size+1]
x = train_data[:block_size]
y = train_data[1:block_size+1]
for t in range(block_size):
context = x[:t+1]
target = y[t]
print(f"when input is {context} the target: {target}")
This code splits the encoded text data into training and validation sets. The first 90% of
the data is assigned to the train_data variable, while the remaining 10% is assigned to the
val_data variable. It defines the block_size variable to be 8, which determines the input
sequence size that the GPT model will process at a time. It then selects a portion of the
training data that is block_size+1 elements long and assigns it to train_data. The x
variable is assigned the first block_size elements of train_data, while the y variable is
assigned the next block_size elements of train_data, starting from the second element. In
other words, y is shifted one position relative to x. Next, the code loops over the
block_size elements of x and y, and prints out the input context and target for each
position in the input sequence. For each iteration of the loop, the context variable is set to
the first t+1 elements of x, where t ranges from 0 to block_size-1. The target variable is
set to the t-th element of y. The loop then prints out a message indicating the current input
context and target.
Generating batches of input and target data for training the GPT
torch.manual_seed(1337)
batch_size = 4 # how many independent sequences will we process in parallel?
block_size = 8 # what is the maximum context length for predictions?](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-15-320.jpg)
![16/29
def get_batch(split):
# generate a small batch of data of inputs x and targets y
data = train_data if split == 'train' else val_data
ix = torch.randint(len(data) - block_size, (batch_size,))
x = torch.stack([data[i:i+block_size] for i in ix])
y = torch.stack([data[i+1:i+block_size+1] for i in ix])
return x, y
xb, yb = get_batch('train')
print('inputs:')
print(xb.shape)
print(xb)
print('targets:')
print(yb.shape)
print(yb)
print('----')
for b in range(batch_size): # batch dimension
for t in range(block_size): # time dimension
context = xb[b, :t+1]
target = yb[b,t]
print(f"when input is {context.tolist()} the target: {target}")
This code sets the random seed of PyTorch to 1337, which ensures that the random
number generation is deterministic and reproducible. This is important for training the
GPT model and getting consistent results. It defines the batch_size and block_size
variables. batch_size specifies how many independent sequences will be processed in
parallel in each batch, while block_size specifies the maximum context length for
predictions. Then it defines a function called get_batch that generates a small batch of
data of inputs x and targets y for a given split (either train or val). The function first selects
the appropriate dataset (train_data or val_data) based on the input split. It then randomly
selects batch_size starting positions for x using torch.randint(), ensuring that each starting
position is at least block_size positions away from the end of the dataset to avoid going](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-16-320.jpg)
![17/29
out of bounds. It then constructs x and y tensors by selecting block_size elements starting
from each starting position, with y shifted one position to the right relative to x. The
function returns the x and y tensors as a tuple. It calls the get_batch() function with the
argument ‘train’ to generate a batch of training data. It then prints the shape and contents
of the x and y tensors. Finally, it loops over each element in the batch (dimension
batch_size) and each position in the input sequence (dimension block_size), and prints
out the sequence’s input context and target for each position. The context variable is set
to the first t+1 elements of xb[b,:], where t ranges from 0 to block_size-1. The target
variable is set to the t-th element of yb[b,:]. The loop then prints out a message indicating
the current input context and target.
Calculating the average loss on the training and validation datasets using a pre-
trained model
@torch.no_grad()
def estimate_loss():
out = {}
model.eval()
for split in ['train', 'val']:
losses = torch.zeros(eval_iters)
for k in range(eval_iters):
X, Y = get_batch(split)
logits, loss = model(X, Y)
losses[k] = loss.item()
out[split] = losses.mean()
model.train()
return out
This code defines a function estimate_loss() which calculates the average loss on the
training and validation datasets using a pre-trained model. It uses the @torch.no_grad()
decorator to disable gradient computation during the evaluation, and sets the model to
evaluation mode using model.eval(). Then, it iterates over the training and validation
datasets eval_iters times, computes the logits and loss for each batch using the pre-
trained model, and records the losses. Finally, it returns the average losses for the two
datasets and sets the model back to training mode using model.train(). This function is
useful for monitoring the model’s performance during training and determining when to
stop training.](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-17-320.jpg)
![18/29
Defining one head of the self-attention mechanism in a transformer model
class Head(nn.Module):
""" one head of self-attention """
def __init__(self, head_size):
super().__init__()
self.key = nn.Linear(n_embd, head_size, bias=False)
self.query = nn.Linear(n_embd, head_size, bias=False)
self.value = nn.Linear(n_embd, head_size, bias=False)
self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))
self.dropout = nn.Dropout(dropout)
def forward(self, x):
B,T,C = x.shape
k = self.key(x) # (B,T,C)
q = self.query(x) # (B,T,C)
# compute attention scores ("affinities")
wei = q @ k.transpose(-2,-1) * C**-0.5 # (B, T, C) @ (B, C, T) -> (B, T, T)
wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf')) # (B, T, T)
wei = F.softmax(wei, dim=-1) # (B, T, T)
wei = self.dropout(wei)
# perform the weighted aggregation of the values
v = self.value(x) # (B,T,C)
out = wei @ v # (B, T, T) @ (B, T, C) -> (B, T, C)
return out
This code defines a module called Head which represents one head of the self-attention
mechanism used in the GPT model. The __init__ method initializes three linear layers
(key, query, and value) that will be used to project the input tensor x into a lower-
dimensional space, which helps compute the attention scores efficiently. The forward
method takes as input a tensor x of shape (batch_size, sequence_length,](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-18-320.jpg)
![19/29
embedding_size) and computes the self-attention scores using the dot-product attention
mechanism. The attention scores are computed by taking the dot product of the query
and key projections and normalizing the result by the square root of the embedding size.
The resulting attention scores are then masked with a triangular matrix to prevent
attending to future tokens. The attention scores are then normalized with a softmax
function, multiplied by the value projection, and finally aggregated to produce the output
tensor of shape (batch_size, sequence_length, embedding_size). The dropout layer is
applied to the attention scores before the final aggregation.
Implementing the multi-head attention mechanism
class MultiHeadAttention(nn.Module):
""" multiple heads of self-attention in parallel """
def __init__(self, num_heads, head_size):
super().__init__()
self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])
self.proj = nn.Linear(n_embd, n_embd)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
out = torch.cat([h(x) for h in self.heads], dim=-1)
out = self.dropout(self.proj(out))
return out
This PyTorch module implements the multi-head attention mechanism used in building
GPT models. It contains a number of heads, each of which computes a self-attention
matrix for the input sequence. The output of each head is concatenated and projected to
the original embedding size using a linear layer and then passed through a dropout layer.
The result is a new sequence of the same length but with a larger embedding dimension
that encodes information from multiple self-attention heads. This module is used as a
building block in the GPT model.
Next we need to add the FeedFoward module
class FeedFoward(nn.Module):
""" a simple linear layer followed by a non-linearity """
def __init__(self, n_embd):
super().__init__()](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-19-320.jpg)
![21/29
# each token directly reads off the logits for the next token from a lookup table
self.token_embedding_table = nn.Embedding(vocab_size, n_embd)
self.position_embedding_table = nn.Embedding(block_size, n_embd)
self.blocks = nn.Sequential(*[Block(n_embd, n_head=n_head) for _ in range(n_layer)])
self.ln_f = nn.LayerNorm(n_embd) # final layer norm
self.lm_head = nn.Linear(n_embd, vocab_size)
def forward(self, idx, targets=None):
B, T = idx.shape
# idx and targets are both (B,T) tensor of integers
tok_emb = self.token_embedding_table(idx) # (B,T,C)
pos_emb = self.position_embedding_table(torch.arange(T, device=device)) # (T,C)
x = tok_emb + pos_emb # (B,T,C)
x = self.blocks(x) # (B,T,C)
x = self.ln_f(x) # (B,T,C)
logits = self.lm_head(x) # (B,T,vocab_size)
if targets is None:
loss = None
else:
B, T, C = logits.shape
logits = logits.view(B*T, C)
targets = targets.view(B*T)
loss = F.cross_entropy(logits, targets)
return logits, loss
def generate(self, idx, max_new_tokens):
# idx is (B, T) array of indices in the current context
for _ in range(max_new_tokens):](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-21-320.jpg)
![22/29
# crop idx to the last block_size tokens
idx_cond = idx[:, -block_size:]
# get the predictions
logits, loss = self(idx_cond)
# focus only on the last time step
logits = logits[:, -1, :] # becomes (B, C)
# apply softmax to get probabilities
probs = F.softmax(logits, dim=-1) # (B, C)
# sample from the distribution
idx_next = torch.multinomial(probs, num_samples=1) # (B, 1)
# append sampled index to the running sequence
idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
return idx
model = BigramLanguageModel()
m = model.to(device)
# print the number of parameters in the model
print(sum(p.numel() for p in m.parameters())/1e6, 'M parameters')
# create a PyTorch optimizer
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
for iter in range(max_iters):
# every once in a while evaluate the loss on train and val sets
if iter % eval_interval == 0 or iter == max_iters - 1:
losses = estimate_loss()
print(f"step {iter}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")
# sample a batch of data
xb, yb = get_batch('train')](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-22-320.jpg)
![23/29
# evaluate the loss
logits, loss = model(xb, yb)
optimizer.zero_grad(set_to_none=True)
loss.backward()
optimizer.step()
# generate from the model
context = torch.zeros((1, 1), dtype=torch.long, device=device)
print(decode(m.generate(context, max_new_tokens=2000)[0].tolist()))
This code defines a bigram language model using PyTorch to train a GPT model.
The BigramLanguageModel class is defined as a subclass of nn.Module and contains
several layers that are used to build the model. The __init__ method initializes the model
with an embedding layer for the tokens and a separate embedding layer for the position
of the tokens. Additionally, the model has a sequence of transformer blocks, which are
defined by the Block function, and a final layer norm and linear layer to output the logits of
the next token. The forward method takes in input sequences and targets, computes the
embeddings, applies the transformer blocks, and outputs the logits of the next token
along with the loss if targets are provided.
The generate method is used to generate new sequences of text from the model. It takes
in a starting sequence and a maximum number of new tokens to generate. The method
iteratively samples the next token from the model’s predicted probability distribution and
appends it to the running sequence until the desired length is reached.
In the main part of the code, an instance of the BigramLanguageModel class is created
and moved to a specified device. The PyTorch AdamW optimizer is then created, and the
training loop begins. In each iteration, a batch of data is sampled from the training set
using the get_batch function. The model is then evaluated on this batch of data, the loss
is computed, and the gradients are backpropagated using loss.backward(). Finally, the
optimizer’s step() method is called to update the model’s parameters.
After training, the generate method is used to generate a sequence of text from the
trained model. A context tensor of zeros is created, and the generate method is called
with this context and a maximum number of new tokens to generate. The resulting
sequence of tokens is decoded using the decode function to produce a string of
generated text.
How to train an existing GPT model with your data?](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-23-320.jpg)
![25/29
If you prefer containerized execution, you can use a Dockerfile for running NeoX. To
utilize this, initially create an image named gpt-neox from the root directory of the
repository using the command
docker build -t gpt-neox -f Dockerfile ..
Additionally, you can get pre-constructed images at leogao2/gpt-neox on Docker Hub.
Following this, you can execute a container based on the created image. For example,
the command below attaches the cloned repository directory (gpt-neox) to /gpt-neox in
the container, and uses nvidia-docker to grant container access to four GPUs (numbered
0-3).
Usage
You should utilize deepy.py, a wrapper around the deepspeed launcher, to trigger all
functionalities, including inference.
There are three principal functions available to you:
1. train.py: This is for training and fine-tuning models.
2. evaluate.py: Use this to evaluate a trained model using the language model
evaluation harness.
3. generate.py: This function is for sampling text from a trained model.
You can launch these with the following command:
./deepy.py [script.py] [./path/to/config_1.yml] [./path/to/config_2.yml] ...
[./path/to/config_n.yml]
For instance, to unconditionally generate text with the GPT-NeoX-20B model, use:
./deepy.py generate.py ./configs/20B.yml
You can also optionally input a text file (e.g., prompt.txt) as the prompt. This should be a
plain .txt file with each prompt separated by newline characters. Remember to pass in the
path to an output file.
./deepy.py generate.py ./configs/20B.yml -i prompt.txt -o sample_outputs.txt
To replicate our evaluation numbers on tasks like TriviaQA and PIQA, use:
./deepy.py evaluate.py ./configs/20B.yml --eval_tasks triviaqa piqa
Configuration
GPT-NeoX operations are governed by parameters in a YAML configuration file, which is
provided to the deepy.py launcher. We have included some sample .yaml files, including
one for GPT-NeoX-20B, and example configurations for other model sizes in the configs
folder.](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-25-320.jpg)
![26/29
These files are usually all-inclusive, but not necessarily optimized. Depending on your
specific GPU setup, you might need to adjust settings such as pipe-parallel-size, model-
parallel-size for parallelism, train_micro_batch_size_per_gpu or gradient-accumulation-
steps for batch size adjustments, or the zero_optimization dict for optimizer state
parallelization.
For an in-depth guide on available features and their configuration, refer to the
configuration README. For detailed information on all possible arguments, check out
configs/neox_arguments.md.
Data preparation
Prepare your text data in the format accepted by the GPT NeoX model. This usually
involves tokenization using a tokenizer that is suitable for the GPT NeoX model.
Then, ensure you have downloaded the GPT2 tokenizer vocabulary and merge files. The
following links will lead you to them:
Training and Fine-tuning
Kickstart your training using ‘deepy.py’, which is a wrapper around DeepSpeed’s
launcher. It parallelly executes the script across multiple GPUs or nodes.
Here’s how to use it:
Execute
python ./deepy.py train.py [path/to/config1.yml] [path/to/config2.yml] ...
You can supply any number of configuration files, which will be merged when the script
runs.
Optionally, you can include a config prefix, which is a common path for all your
configuration files.
For instance. execute the following code –
python ./deepy.py train.py -d configs 125M.yml local_setup.yml
This instruction executes the ‘train.py’ script on every node of the network, with each
GPU running one instance of the script. This means every individual GPU across all
nodes will be running the ‘train.py’ script separately. The worker nodes and number of
GPUs are defined in the ‘/job/hostfile’ file (see parameter documentation), or can be
simply included as the ‘num_gpus’ argument if you’re running a single node setup.
We suggest defining the model parameters in one configuration file (like
‘configs/125M.yml’) and the data path parameters in another (like
‘configs/local_setup.yml’), for better organization, although it’s not mandatory.](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/85/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-26-320.jpg)
![14/29
stoi = { ch:i for i,ch in enumerate(chars) }
itos = { i:ch for i,ch in enumerate(chars) }
encode = lambda s: [stoi[c] for c in s] # encoder: take a string, output a list of integers
decode = lambda l: ''.join([itos[i] for i in l]) # decoder: take a list of integers, output a string
print(encode("hii there"))
print(decode(encode("hii there")))
This code block creates a character-to-integer mapping and its inverse (integer-to-
character mapping) for a set of characters. The stoi dictionary maps each character to a
unique integer while itos maps each integer back to its corresponding character. The
encode function takes a string as input and returns a list of integers, where each integer
corresponds to the index of the character in the chars set. The decode function takes a
list of integers and returns the original string by looking up the corresponding characters
in the itos dictionary. The code then tests the encoding and decoding functions by
encoding the string “hii there” and then decoding the resulting list of integers back into a
string.
Encoding input data
In building a GPT model, it’s important to encode the entire text dataset so that it can be
fed into the model. The following code does exactly that.
let's now encode the entire text dataset and store it into a torch.Tensor
import torch # we use PyTorch: https://pytorch.org
data = torch.tensor(encode(text), dtype=torch.long)
print(data.shape, data.dtype)
print(data[:1000]) # the 1000 characters we looked at earier will to the GPT look like this
This code imports the PyTorch library and creates a tensor called data. The tensor is filled
with the encoded text data, which is obtained by calling the encode function on the text
variable. The dtype parameter is set to torch.long to ensure that the tensor elements are
integers. The code prints the shape and data type of the data tensor. The shape attribute
tells us the size of the tensor along each dimension, while the dtype attribute tells us the
data type of the tensor elements. This information is useful for verifying that the tensor
has been created correctly and will be compatible with the GPT model. It then prints the
first 1000 elements of the data tensor, which represent the encoded text data. This is
useful for verifying that the encoding process has worked correctly and that the data has
been loaded into the tensor as expected.
Splitting up the data into train and validation sets](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-14-2048.jpg)
![15/29
The following code is useful for understanding how the GPT model will process the input
data. It shows how the model will process input sequences of length block_size, and how
the input and output sequences are related to each other. This understanding can help in
designing and training the GPT model.
# Let's now split up the data into train and validation sets
n = int(0.9*len(data)) # first 90% will be train, rest val
train_data = data[:n]
val_data = data[n:]
block_size = 8
train_data[:block_size+1]
x = train_data[:block_size]
y = train_data[1:block_size+1]
for t in range(block_size):
context = x[:t+1]
target = y[t]
print(f"when input is {context} the target: {target}")
This code splits the encoded text data into training and validation sets. The first 90% of
the data is assigned to the train_data variable, while the remaining 10% is assigned to the
val_data variable. It defines the block_size variable to be 8, which determines the input
sequence size that the GPT model will process at a time. It then selects a portion of the
training data that is block_size+1 elements long and assigns it to train_data. The x
variable is assigned the first block_size elements of train_data, while the y variable is
assigned the next block_size elements of train_data, starting from the second element. In
other words, y is shifted one position relative to x. Next, the code loops over the
block_size elements of x and y, and prints out the input context and target for each
position in the input sequence. For each iteration of the loop, the context variable is set to
the first t+1 elements of x, where t ranges from 0 to block_size-1. The target variable is
set to the t-th element of y. The loop then prints out a message indicating the current input
context and target.
Generating batches of input and target data for training the GPT
torch.manual_seed(1337)
batch_size = 4 # how many independent sequences will we process in parallel?
block_size = 8 # what is the maximum context length for predictions?](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-15-2048.jpg)
![16/29
def get_batch(split):
# generate a small batch of data of inputs x and targets y
data = train_data if split == 'train' else val_data
ix = torch.randint(len(data) - block_size, (batch_size,))
x = torch.stack([data[i:i+block_size] for i in ix])
y = torch.stack([data[i+1:i+block_size+1] for i in ix])
return x, y
xb, yb = get_batch('train')
print('inputs:')
print(xb.shape)
print(xb)
print('targets:')
print(yb.shape)
print(yb)
print('----')
for b in range(batch_size): # batch dimension
for t in range(block_size): # time dimension
context = xb[b, :t+1]
target = yb[b,t]
print(f"when input is {context.tolist()} the target: {target}")
This code sets the random seed of PyTorch to 1337, which ensures that the random
number generation is deterministic and reproducible. This is important for training the
GPT model and getting consistent results. It defines the batch_size and block_size
variables. batch_size specifies how many independent sequences will be processed in
parallel in each batch, while block_size specifies the maximum context length for
predictions. Then it defines a function called get_batch that generates a small batch of
data of inputs x and targets y for a given split (either train or val). The function first selects
the appropriate dataset (train_data or val_data) based on the input split. It then randomly
selects batch_size starting positions for x using torch.randint(), ensuring that each starting
position is at least block_size positions away from the end of the dataset to avoid going](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-16-2048.jpg)
![17/29
out of bounds. It then constructs x and y tensors by selecting block_size elements starting
from each starting position, with y shifted one position to the right relative to x. The
function returns the x and y tensors as a tuple. It calls the get_batch() function with the
argument ‘train’ to generate a batch of training data. It then prints the shape and contents
of the x and y tensors. Finally, it loops over each element in the batch (dimension
batch_size) and each position in the input sequence (dimension block_size), and prints
out the sequence’s input context and target for each position. The context variable is set
to the first t+1 elements of xb[b,:], where t ranges from 0 to block_size-1. The target
variable is set to the t-th element of yb[b,:]. The loop then prints out a message indicating
the current input context and target.
Calculating the average loss on the training and validation datasets using a pre-
trained model
@torch.no_grad()
def estimate_loss():
out = {}
model.eval()
for split in ['train', 'val']:
losses = torch.zeros(eval_iters)
for k in range(eval_iters):
X, Y = get_batch(split)
logits, loss = model(X, Y)
losses[k] = loss.item()
out[split] = losses.mean()
model.train()
return out
This code defines a function estimate_loss() which calculates the average loss on the
training and validation datasets using a pre-trained model. It uses the @torch.no_grad()
decorator to disable gradient computation during the evaluation, and sets the model to
evaluation mode using model.eval(). Then, it iterates over the training and validation
datasets eval_iters times, computes the logits and loss for each batch using the pre-
trained model, and records the losses. Finally, it returns the average losses for the two
datasets and sets the model back to training mode using model.train(). This function is
useful for monitoring the model’s performance during training and determining when to
stop training.](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-17-2048.jpg)
![18/29
Defining one head of the self-attention mechanism in a transformer model
class Head(nn.Module):
""" one head of self-attention """
def __init__(self, head_size):
super().__init__()
self.key = nn.Linear(n_embd, head_size, bias=False)
self.query = nn.Linear(n_embd, head_size, bias=False)
self.value = nn.Linear(n_embd, head_size, bias=False)
self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))
self.dropout = nn.Dropout(dropout)
def forward(self, x):
B,T,C = x.shape
k = self.key(x) # (B,T,C)
q = self.query(x) # (B,T,C)
# compute attention scores ("affinities")
wei = q @ k.transpose(-2,-1) * C**-0.5 # (B, T, C) @ (B, C, T) -> (B, T, T)
wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf')) # (B, T, T)
wei = F.softmax(wei, dim=-1) # (B, T, T)
wei = self.dropout(wei)
# perform the weighted aggregation of the values
v = self.value(x) # (B,T,C)
out = wei @ v # (B, T, T) @ (B, T, C) -> (B, T, C)
return out
This code defines a module called Head which represents one head of the self-attention
mechanism used in the GPT model. The __init__ method initializes three linear layers
(key, query, and value) that will be used to project the input tensor x into a lower-
dimensional space, which helps compute the attention scores efficiently. The forward
method takes as input a tensor x of shape (batch_size, sequence_length,](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-18-2048.jpg)
![19/29
embedding_size) and computes the self-attention scores using the dot-product attention
mechanism. The attention scores are computed by taking the dot product of the query
and key projections and normalizing the result by the square root of the embedding size.
The resulting attention scores are then masked with a triangular matrix to prevent
attending to future tokens. The attention scores are then normalized with a softmax
function, multiplied by the value projection, and finally aggregated to produce the output
tensor of shape (batch_size, sequence_length, embedding_size). The dropout layer is
applied to the attention scores before the final aggregation.
Implementing the multi-head attention mechanism
class MultiHeadAttention(nn.Module):
""" multiple heads of self-attention in parallel """
def __init__(self, num_heads, head_size):
super().__init__()
self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])
self.proj = nn.Linear(n_embd, n_embd)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
out = torch.cat([h(x) for h in self.heads], dim=-1)
out = self.dropout(self.proj(out))
return out
This PyTorch module implements the multi-head attention mechanism used in building
GPT models. It contains a number of heads, each of which computes a self-attention
matrix for the input sequence. The output of each head is concatenated and projected to
the original embedding size using a linear layer and then passed through a dropout layer.
The result is a new sequence of the same length but with a larger embedding dimension
that encodes information from multiple self-attention heads. This module is used as a
building block in the GPT model.
Next we need to add the FeedFoward module
class FeedFoward(nn.Module):
""" a simple linear layer followed by a non-linearity """
def __init__(self, n_embd):
super().__init__()](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-19-2048.jpg)
![21/29
# each token directly reads off the logits for the next token from a lookup table
self.token_embedding_table = nn.Embedding(vocab_size, n_embd)
self.position_embedding_table = nn.Embedding(block_size, n_embd)
self.blocks = nn.Sequential(*[Block(n_embd, n_head=n_head) for _ in range(n_layer)])
self.ln_f = nn.LayerNorm(n_embd) # final layer norm
self.lm_head = nn.Linear(n_embd, vocab_size)
def forward(self, idx, targets=None):
B, T = idx.shape
# idx and targets are both (B,T) tensor of integers
tok_emb = self.token_embedding_table(idx) # (B,T,C)
pos_emb = self.position_embedding_table(torch.arange(T, device=device)) # (T,C)
x = tok_emb + pos_emb # (B,T,C)
x = self.blocks(x) # (B,T,C)
x = self.ln_f(x) # (B,T,C)
logits = self.lm_head(x) # (B,T,vocab_size)
if targets is None:
loss = None
else:
B, T, C = logits.shape
logits = logits.view(B*T, C)
targets = targets.view(B*T)
loss = F.cross_entropy(logits, targets)
return logits, loss
def generate(self, idx, max_new_tokens):
# idx is (B, T) array of indices in the current context
for _ in range(max_new_tokens):](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-21-2048.jpg)
![22/29
# crop idx to the last block_size tokens
idx_cond = idx[:, -block_size:]
# get the predictions
logits, loss = self(idx_cond)
# focus only on the last time step
logits = logits[:, -1, :] # becomes (B, C)
# apply softmax to get probabilities
probs = F.softmax(logits, dim=-1) # (B, C)
# sample from the distribution
idx_next = torch.multinomial(probs, num_samples=1) # (B, 1)
# append sampled index to the running sequence
idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
return idx
model = BigramLanguageModel()
m = model.to(device)
# print the number of parameters in the model
print(sum(p.numel() for p in m.parameters())/1e6, 'M parameters')
# create a PyTorch optimizer
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
for iter in range(max_iters):
# every once in a while evaluate the loss on train and val sets
if iter % eval_interval == 0 or iter == max_iters - 1:
losses = estimate_loss()
print(f"step {iter}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")
# sample a batch of data
xb, yb = get_batch('train')](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-22-2048.jpg)
![23/29
# evaluate the loss
logits, loss = model(xb, yb)
optimizer.zero_grad(set_to_none=True)
loss.backward()
optimizer.step()
# generate from the model
context = torch.zeros((1, 1), dtype=torch.long, device=device)
print(decode(m.generate(context, max_new_tokens=2000)[0].tolist()))
This code defines a bigram language model using PyTorch to train a GPT model.
The BigramLanguageModel class is defined as a subclass of nn.Module and contains
several layers that are used to build the model. The __init__ method initializes the model
with an embedding layer for the tokens and a separate embedding layer for the position
of the tokens. Additionally, the model has a sequence of transformer blocks, which are
defined by the Block function, and a final layer norm and linear layer to output the logits of
the next token. The forward method takes in input sequences and targets, computes the
embeddings, applies the transformer blocks, and outputs the logits of the next token
along with the loss if targets are provided.
The generate method is used to generate new sequences of text from the model. It takes
in a starting sequence and a maximum number of new tokens to generate. The method
iteratively samples the next token from the model’s predicted probability distribution and
appends it to the running sequence until the desired length is reached.
In the main part of the code, an instance of the BigramLanguageModel class is created
and moved to a specified device. The PyTorch AdamW optimizer is then created, and the
training loop begins. In each iteration, a batch of data is sampled from the training set
using the get_batch function. The model is then evaluated on this batch of data, the loss
is computed, and the gradients are backpropagated using loss.backward(). Finally, the
optimizer’s step() method is called to update the model’s parameters.
After training, the generate method is used to generate a sequence of text from the
trained model. A context tensor of zeros is created, and the generate method is called
with this context and a maximum number of new tokens to generate. The resulting
sequence of tokens is decoded using the decode function to produce a string of
generated text.
How to train an existing GPT model with your data?](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-23-2048.jpg)
![25/29
If you prefer containerized execution, you can use a Dockerfile for running NeoX. To
utilize this, initially create an image named gpt-neox from the root directory of the
repository using the command
docker build -t gpt-neox -f Dockerfile ..
Additionally, you can get pre-constructed images at leogao2/gpt-neox on Docker Hub.
Following this, you can execute a container based on the created image. For example,
the command below attaches the cloned repository directory (gpt-neox) to /gpt-neox in
the container, and uses nvidia-docker to grant container access to four GPUs (numbered
0-3).
Usage
You should utilize deepy.py, a wrapper around the deepspeed launcher, to trigger all
functionalities, including inference.
There are three principal functions available to you:
1. train.py: This is for training and fine-tuning models.
2. evaluate.py: Use this to evaluate a trained model using the language model
evaluation harness.
3. generate.py: This function is for sampling text from a trained model.
You can launch these with the following command:
./deepy.py [script.py] [./path/to/config_1.yml] [./path/to/config_2.yml] ...
[./path/to/config_n.yml]
For instance, to unconditionally generate text with the GPT-NeoX-20B model, use:
./deepy.py generate.py ./configs/20B.yml
You can also optionally input a text file (e.g., prompt.txt) as the prompt. This should be a
plain .txt file with each prompt separated by newline characters. Remember to pass in the
path to an output file.
./deepy.py generate.py ./configs/20B.yml -i prompt.txt -o sample_outputs.txt
To replicate our evaluation numbers on tasks like TriviaQA and PIQA, use:
./deepy.py evaluate.py ./configs/20B.yml --eval_tasks triviaqa piqa
Configuration
GPT-NeoX operations are governed by parameters in a YAML configuration file, which is
provided to the deepy.py launcher. We have included some sample .yaml files, including
one for GPT-NeoX-20B, and example configurations for other model sizes in the configs
folder.](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-25-2048.jpg)
![26/29
These files are usually all-inclusive, but not necessarily optimized. Depending on your
specific GPU setup, you might need to adjust settings such as pipe-parallel-size, model-
parallel-size for parallelism, train_micro_batch_size_per_gpu or gradient-accumulation-
steps for batch size adjustments, or the zero_optimization dict for optimizer state
parallelization.
For an in-depth guide on available features and their configuration, refer to the
configuration README. For detailed information on all possible arguments, check out
configs/neox_arguments.md.
Data preparation
Prepare your text data in the format accepted by the GPT NeoX model. This usually
involves tokenization using a tokenizer that is suitable for the GPT NeoX model.
Then, ensure you have downloaded the GPT2 tokenizer vocabulary and merge files. The
following links will lead you to them:
Training and Fine-tuning
Kickstart your training using ‘deepy.py’, which is a wrapper around DeepSpeed’s
launcher. It parallelly executes the script across multiple GPUs or nodes.
Here’s how to use it:
Execute
python ./deepy.py train.py [path/to/config1.yml] [path/to/config2.yml] ...
You can supply any number of configuration files, which will be merged when the script
runs.
Optionally, you can include a config prefix, which is a common path for all your
configuration files.
For instance. execute the following code –
python ./deepy.py train.py -d configs 125M.yml local_setup.yml
This instruction executes the ‘train.py’ script on every node of the network, with each
GPU running one instance of the script. This means every individual GPU across all
nodes will be running the ‘train.py’ script separately. The worker nodes and number of
GPUs are defined in the ‘/job/hostfile’ file (see parameter documentation), or can be
simply included as the ‘num_gpus’ argument if you’re running a single node setup.
We suggest defining the model parameters in one configuration file (like
‘configs/125M.yml’) and the data path parameters in another (like
‘configs/local_setup.yml’), for better organization, although it’s not mandatory.](https://image.slidesharecdn.com/leewayhertz-240702111344-dec7a74e/75/leewayhertz-com-How-to-build-a-GPT-model-1-pdf-26-2048.jpg)
More Related Content
Similar to leewayhertz.com-How to build a GPT model (1).pdf
![UiPath Automation Suite Installation (Hands-On) [2/3]](https://cdn.slidesharecdn.com/ss_thumbnails/automationsuitecommunitysession2-251015095633-a6d862f1-thumbnail.jpg?width=600ounds&width=560&fit=bounds)
leewayhertz.com-How to build a GPT model (1).pdf
- 1. 1/29 How to builda GPT model? leewayhertz.com/build-a-gpt-model Introduced by OpenAI, powerful Generative Pre-trained Transformer (GPT) language models have opened up new frontiers in Natural Language Processing (NLP). The integration of GPT models into virtual assistants and chatbots boosts their capabilities, which has resulted in a surge in demand for GPT models. According to a report published by Allied Market Research, titled “Global NLP Market,” the global NLP market size was valued at $11.1 billion in 2020 and is estimated to reach $341.5 billion by 2030, growing at a CAGR of 40.9% from 2021 to 2030. Interestingly, the demand for GPT models are a major contributor to this growth. GPT models are a collection of deep learning-based language models created by the OpenAI team. Without supervision, these models can perform various NLP tasks like question-answering, textual entailment, text summarization, etc. These language models require very few or no examples to understand tasks. They perform equivalent to or even better than state-of-the-art models trained in a supervised fashion. The GPT series from OpenAI has radically transformed the landscape of artificial intelligence. The latest addition to the series, GPT-4, has further expanded the horizons for AI applications. This article will take you on a journey through the innovative realm of GPT-4. We’ll delve into its notable advancements of GPT models while exploring how this state-of-the-art model is reshaping our interactions with AI across diverse sectors. This article deeply delves into all aspects of GPT models and discusses the steps required to build a GPT model from scratch. What is a GPT model?
- 2. 2/29 Overview of GPTmodels Use cases of GPT models Working mechanism of GPT models How to choose the right GPT model for your needs? Prerequisites to build a GPT model How to create a GPT model? – Steps for building a GPT model How to train an existing GPT model with your data? Leverage LeewayHertz’s AI development services to build a GPT model Things to consider while building a GPT model What is a GPT model? GPT stands for Generative Pre-trained Transformer, the first generalized language model in NLP. Previously, language models were only designed for single tasks like text generation, summarization or classification. GPT is the first generalized language model ever created in the history of natural language processing that can be used for various NLP tasks. Now let us explore the three components of GPT, namely Generative, Pre- Trained, and Transformer and understand what they mean. Generative: Generative models are statistical models used to generate new data. These models can learn the relationships between variables in a data set to generate new data points similar to those in the original data set. Pre-trained: These models have been pre-trained using a large data set which can be used when it is difficult to train a new model. Although a pre-trained model might not be perfect, it can save time and improve performance. Transformer: The transformer model, an artificial neural network created in 2017, is the most well-known deep learning model capable of handling sequential data such as text. Many tasks like machine translation and text classification are performed using transformer models. GPT can perform various NLP tasks with high accuracy depending on the large datasets it was trained on and its architecture of billion parameters, allowing it to understand the logical connections within the data. GPT models, like the latest version GPT-3, have been pre-trained using text from five large datasets, including Common Crawl and WebText2. The corpus contains nearly a trillion words, allowing GPT-3 to perform NLP tasks quickly and without any examples of data. Overview of GPT models GPT models, short for Generative Pretrained Transformers, are advanced deep learning models designed for generating human-like text. These models, developed by OpenAI, have seen several iterations: GPT-1, GPT-2, GPT-3, and most recently, GPT-4.
- 3. 3/29 Introduced in 2018,GPT-1 was the first in this series, using a unique Transformer architecture to vastly improve language generation capabilities. It was built with 117 million parameters and trained on a mix of datasets from Common Crawl and BookCorpus. GPT-1 could generate fluent and coherent language given some context. However, it had limitations, including the tendency to repeat text and difficulties with complex dialogue and long-term dependencies. OpenAI then released GPT-2 in 2019. This model was much larger, with 1.5 billion parameters, and was trained on an even larger and diverse dataset. Its main strength was the ability to generate realistic text sequences and human-like responses. However, GPT-2 struggled with maintaining context and coherence over longer passages. The introduction of GPT-3 in 2020 marked a huge leap forward. With a staggering 175 billion parameters, GPT-3 was trained on vast datasets and could generate nuanced responses across various tasks. It could generate text, write code, create art, and more, making it a valuable tool for many applications like chatbots and language translation. However, GPT-3 wasn’t perfect and had its share of biases and inaccuracies. Following GPT-3, OpenAI introduced an upgraded version, GPT-3.5, and eventually released GPT-4 in March 2023. GPT-4 is the latest and most advanced of OpenAI’s language models which is multi modal. It can generate more accurate statements and handle images as inputs, allowing for captions, classifications, and analyses. GPT-4 also showcases creative capabilities like composing songs or writing screenplays. It comes in two variants, differing in their context window size: gpt-4-8K and gpt-4-32K. GPT-4’s ability to understand complex prompts and demonstrate human-like performance on various tasks is a significant leap forward. Yet, as with all powerful tools, there are valid concerns about potential misuse and ethical implications. It’s crucial to keep these factors in mind when exploring the capabilities and applications of GPT models. Discover GPT Model Expertise Dive into GPT model building. Ready to level up your AI? Let’s collaborate. Use cases of GPT models GPT models are known for their versatile applications, providing immense value in various sectors. Here, we will discuss three key use cases: Understanding Human Language, Content Generation for UI Design, and Applications in Natural Language Processing. Understanding human language using NLP
- 4. 4/29 GPT models areinstrumental in enhancing the computer’s ability to understand and process human language. This encompasses two main areas: Human Language Understanding (HLU): HLU refers to the machine’s ability to comprehend the meaning of sentences and phrases, effectively translating human knowledge into machine-readable format. This is achieved using deep neural networks or feed-forward neural networks and involves a complex mix of statistical, probabilistic, decision tree, fuzzy set, and reinforcement learning techniques. Developing models in this area is challenging and requires substantial expertise, time, and resources. Natural Language Processing (NLP): NLP focuses on interpreting and analyzing written or spoken human language. It involves training computers to understand language, rather than programming them with pre-set rules or instructions. Key applications of NLP include information retrieval, classification, summarization, sentiment analysis, document generation, and question answering. It also plays a pivotal role in data mining, sentiment analysis, and computational tasks. Generating content for user interface design GPT models can be employed to generate content for user interface design. For example, they can assist in creating web pages where users can upload various forms of content with just a few clicks. This ranges from adding basic elements like captions, titles, descriptions, and alt tags, to incorporating interactive components like buttons, quizzes, and cards. This automation reduces the need for additional development resources and investment. Applications in computer vision systems for image recognition GPT models are not only limited to processing text. When combined with computer vision systems, they can perform tasks such as image recognition. These systems can identify and remember specific elements within an image, like faces, colors, and landmarks. GPT- 3, with its transformer architecture, can handle such tasks effectively. Enhancing customer support with AI-powered chatbots GPT models are revolutionizing customer support by powering AI chatbots. These chatbots, armed with GPT-4, can understand and respond to customer queries with increased precision. They can simulate human-like conversations, providing detailed responses, and instant support around the clock. This significantly enhances customer service by providing quick, accurate responses, leading to improved customer satisfaction and loyalty. Bridging language barriers with accurate translation Language translation is another area where GPT-4 excels. Its advanced language understanding capabilities enable it to translate text between various languages accurately. GPT-4 can grasp the nuances of different languages and provide translations
- 5. 5/29 that retain theoriginal meaning and context. This feature can be incredibly useful in facilitating cross-cultural communication and making information accessible to a global audience. Streamlining code generation GPT-4’s ability to understand and generate programming language code has made it a valuable tool for developers. It can produce code snippets based on a developer’s input, significantly speeding up the coding process and reducing the chance of errors. By understanding the context and nuances of different programming languages, GPT-4 can assist in more complex coding tasks, thus contributing to more efficient and streamlined software development. Transforming education with personalized tutoring The education sector can greatly benefit from the implementation of GPT-4. It can generate educational content tailored to a learner’s needs, providing personalized tutoring and learning assistance. From explaining complex concepts in a simple manner to providing support with homework, GPT-4 can make learning more engaging and accessible. Its ability to adapt to different learning styles and pace can contribute to a more personalized and effective learning experience. Assisting in creative writing In the realm of creative writing, GPT-4 can be an invaluable assistant. It can provide writers with creative suggestions, help overcome writer’s block, and even generate entire stories or poems. By understanding the context and maintaining the flow of the narrative, GPT-4 can produce creative pieces that are coherent and engaging. This can be a valuable tool for writers, stimulating creativity, and enhancing productivity. Working mechanism of GPT models GPT is an AI language model based on transformer architecture that is pre-trained, generative, unsupervised, and capable of performing well in zero/one/few-shot multitask settings. It predicts the next token (an instance of a sequence of characters) from a sequence of tokens for NLP tasks, it has not been trained on. After seeing only a few examples, it can achieve the desired outcomes in certain benchmarks, including machine translation, Q&A and cloze tasks. GPT models calculate the likelihood of a word appearing in a text given that it appears in another text primarily based on conditional probability. For example, in the sentence, “Margaret is organizing a garage sale… Perhaps we could purchase that old…” the word chair is more likely appropriate than the word ‘elephant’. Also, transformer models use multiple units called attention blocks that learn which parts of a text sequence to be focused on. One transformer might have multiple attention blocks, each learning different aspects of a language.
- 6. 6/29 LeewayHertz Output Probabilities Feed Forward Multi- Head Attention Multi- Head Attention Nx Nx Positional Encoding Positional Encoding Outputs (shiftedright) Inputs Add & Norm Add & Norm Add & Norm Add & Norm Feed Forward Add & Norm Linear Softmax Masked Multi-Head Attention Input Embedding Output Embedding Transformer architecture A transformer architecture has two main segments: an encoder that primarily operates on the input sequence and a decoder that operates on the target sequence during training and predicts the next item. For example, a transformer might take a sequence of English words and predict the French word in the correct translation until it is complete. The encoder determines which parts of the input should be emphasized. For example, the encoder can read a sentence like “The quick brown fox jumped.” It then calculates the embedding matrix (embedding in NLP allows words with similar meanings to have a similar representation) and converts it into a series of attention vectors. Now, what is an attention vector? You can view an attention vector in a transformer model as a special calculator, which helps the model understand which parts of any given information are most important in making a decision. Suppose you have been asked multiple questions in an exam that you must answer using different information pieces. The attention vector helps you to pick the most important information to answer each question. It works in the same way in the case of a transformer model.
- 7. 7/29 The multi-head attentionblock initially produces these attention vectors. They are then normalized and passed into a fully connected layer. Normalization is again done before being passed to the decoder. During training, the encoder works directly on the target output sequence. Let us say that the target output is the French translation of the English sentence “The quick brown fox jumped.” The decoder computes separate embedding vectors for each French word of the sentence. Additionally, the positional encoder is applied in the form of sine and cosine functions. Also, masked attention is used, which means that the first word of the French sentence is used, whereas all other words are masked. This allows the transformer to learn to predict the next French words. These outputs are then added and normalized before being passed on to another attention block which also receives the attention vectors generated by the encoder. Alongside, GPT models employ some data compression while consuming millions upon millions of sample texts to convert words into vectors which are nothing but numerical representations. The language model then unpacks the compressed text into human- friendly sentences. The model’s accuracy is improved by compressing and decompressing text. This also allows it to calculate the conditional probability of each word. GPT models can perform well in “few shots” settings and respond to text samples that have been seen before. They only require a few examples to produce pertinent responses because they have been trained on many text samples. Besides, GPT models have many capabilities, such as generating unprecedented-quality synthetic text samples. If you prime the model with an input, it will generate a long continuation. GPT models outperform other language models trained on domains such as Wikipedia, news, and books without using domain-specific training data. GPT learns language tasks such as reading comprehension, summarization and question answering from the text alone, without task-specific training data. These tasks’ scores (“score” refers to a numerical value the model assigns to represent the likelihood or probability of a given output or result) are not the best, but they suggest unsupervised techniques with sufficient data and computation that could benefit the tasks. Here is a comprehensive comparison of GPT models with other language models. Feature GPT BERT (Bidirectional Encoder Representations from Transformers) ELMo (Embeddings from Language Models) Pretraining approach Unidirectional language modeling Bidirectional language modeling (masked language modeling and next sentence prediction) Unidirectional language modeling Pretraining data Large amounts of text from the internet Large amounts of text from the internet A combination of internal and external corpus
- 8. 8/29 Feature GPT BERT (Bidirectional EncoderRepresentations from Transformers) ELMo (Embeddings from Language Models) Architecture Transformer network Transformer network Deep bi-directional LSTM network Outputs Context-aware token-level embeddings Context-aware token-level and sentence-level embeddings Context-aware word- level embeddings Fine-tuning approach Multi-task fine- tuning (e.g., text classification, sequence labeling) Multi-task fine-tuning (e.g., text classification, question answering) Fine-tuning on individual tasks Advantages Can generate text, high flexibility in fine-tuning, large model size Strong performance on a variety of NLP tasks, considering the context in both directions Generates task- specific features, considers context from the entire input sequence Limitations Can generate biased or inaccurate text, requires large amounts of data Limited to fine-tuning and requires task-specific architecture modifications; requires large amounts of data Limited context and task-specific; requires task-specific architecture modifications How to choose the right GPT model for your needs? Choosing the right GPT model for your project depends on several factors, including the complexity of the tasks you want the model to handle, the type of language you want to generate, and the size of your available dataset. If you need a model that can generate simple text responses, such as replying to customer inquiries, GPT-1 could be a sufficient choice. It’s capable of accomplishing straightforward tasks without requiring extensive data or computational resources. However, if your project involves more complex language generation like conducting deep analyses of vast amounts of web content, recommending reading material, or generating stories, then GPT-3 would be a more suitable option. GPT-3 has the capacity to process and learn from billions of web pages, providing more nuanced and sophisticated outputs. In terms of data requirements, the size of your available dataset should be a key consideration. GPT-3, with its larger capacity for learning, tends to work best with big datasets. If you don’t have large amounts of data available for training, GPT-3 might not be the most efficient choice.
- 9. 9/29 In contrast, GPT-1and GPT-2 are more manageable models that can be trained effectively with smaller datasets. These versions could be more fitting for projects with limited data resources or for small-scale tasks. Looking ahead, there’s GPT-4. While details about its specific capabilities and requirements aren’t yet widely available, it’s likely that this newer iteration will offer enhanced performance and may require even larger datasets and more computational resources. Always consider the complexity of your task, your resource availability, and the specific benefits each GPT model offers when choosing the right one for your project. Prerequisites to build a GPT model To build a GPT (Generative Pretrained Transformer) model, the following tools and resources are required: A deep learning framework, such as TensorFlow or PyTorch, to implement the model and train it on large amounts of data. A large amount of training data, such as text from books, articles, or websites to train the model on language patterns and structure. A high-performance computing environment, such as GPUs or TPUs, for accelerating the training process. Knowledge of deep learning concepts, such as neural networks and natural language processing (NLP), to design and implement the model. Tools for data pre-processing and cleaning, such as Numpy, Pandas, or NLTK, to prepare the training data for input into the model. Tools for evaluating the model, such as perplexity or BLEU scores, to measure its performance and make improvements. An NLP library, such as spaCy or NLTK, for tokenizing, stemming and performing other NLP tasks on the input data. Besides, you need to understand the following deep learning concepts to build a GPT model: Neural networks: As GPT models implement neural networks, you must thoroughly understand how they work and their implementation techniques in a deep learning framework. Natural language Processing (NLP): For GPT modeling processes, tokenization, stemming, and text generation, NLP techniques are widely used. So, it is necessary to have a fundamental understanding of NLP techniques and their applications. Transformers: GPT models work based on transformer architecture, so understanding it and its role in language processing and generation is important. Attention mechanisms: Knowledge of how attention mechanisms work is essential to enhance the performance of the GPT model. Pretraining: It is essential to apply the concept of pretraining to the GPT model to improve its performance on NLP tasks.
- 10. 10/29 Generative models: Understandingthe basic concepts and methods of generative models is essential to understand how they can be applied to build your own GPT model. Language modeling: GPT models work based on large amounts of text data. So, a clear understanding of language modeling is required to apply it for GPT model training. Optimization: An understanding of optimization algorithms, such as stochastic gradient descent, is required to optimize the GPT model during training. Alongside this, you need proficiency in any of the following programming languages with a solid understanding of programming concepts, such as object-oriented programming, data structures, and algorithms, to build a GPT model. Python: The most commonly used programming language in deep learning and AI. It has several libraries, such as TensorFlow, PyTorch, and Numpy, used for building and training GPT models. R: A popular programming language for data analysis and statistical modeling, with several packages for deep learning and AI. Julia: A high-level, high-performance programming language well-suited for numerical and scientific computing, including deep learning. How to create a GPT model? A step-by-step guide In this section, with code snippets, we will show steps to build a GPT (Generative Pre- trained Transformer) model from scratch using the PyTorch library and transformer architecture. The code is organized into several sections performing the following tasks sequentially: Data preprocessing: The first section of the code preprocesses the input text data by tokenizing it into a list of words, encoding each word into a unique integer, and generating sequences of fixed length using a sliding window approach. Model configuration: This section of the code defines the configuration parameters for the GPT model, including the number of transformer layers, the number of attention heads, the size of the hidden layers, and the size of the vocabulary. Model architecture: This section of the code defines the architecture of the GPT model using PyTorch modules. The model consists of an embedding layer, followed by a stack of transformer layers, and a linear layer that outputs the probability distribution over the vocabulary for the next word in the sequence. Training loop: This section of the code defines the training loop for the GPT model. It uses the Adam optimizer to minimize the cross-entropy loss between the sequence’s predicted and actual next words. The model is trained on batches of data generated from the preprocessed text data.
- 11. 11/29 Text generation: Thefinal section of the code demonstrates how to use the trained GPT model to generate new text. It initializes the context with a given seed sequence and iteratively generates new words by sampling from the probability distribution output by the model for the next word in the sequence. The generated text is decoded back into words and printed to the console. We will use this dataset – https://raw.githubusercontent.com/karpathy/char- rnn/master/data/tinyshakespeare/input.txt to train a model based on the transformer architecture. The full code can be downloaded from here. Building a GPT model involves the following steps: Importing libraries The first step is to import the necessary libraries for building a neural network using PyTorch, which includes importing the necessary modules and functions. import torch import torch.nn as nn from torch.nn import functional as F In this code snippet, the developer is importing the PyTorch library, which is a popular deep learning framework used for building neural networks. The developer then imports the nn module from the torch library which contains classes and functions for defining and training neural networks. Defining hyperparameters The next step is to define various hyperparameters for building a GPT model. These hyperparameters are essential for training and fine-tuning the GPT model.These hyperparameters will determine the model’s performance, speed, and capacity, and the developer can experiment with different values to optimize the model’s behavior. # hyperparameters batch_size = 16 # how many independent sequences will we process in parallel? block_size = 32 # what is the maximum context length for predictions? max_iters = 5000 eval_interval = 100 learning_rate = 1e-3 device = 'cuda' if torch.cuda.is_available() else 'cpu' eval_iters = 200
- 12. 12/29 n_embd = 64 n_head= 4 n_layer = 4 dropout = 0.0 The hyperparameters defined in this code snippet are: batch_size: This parameter determines the number of independent sequences that will be processed in parallel during training. A larger batch size can speed up training, but it requires more memory. block_size: This parameter sets the maximum context length for predictions. The GPT model generates predictions based on the context it receives as input, and this parameter sets the maximum length of that context. max_iters: This parameter sets the maximum number of training iterations for the GPT model. eval_interval: This parameter sets the number of training iterations, after which the model’s performance will be evaluated. learning_rate: This parameter determines the learning rate for the optimizer during training. device: This parameter sets the device (CPU or GPU) on which the GPT model will be trained. eval_iters: This parameter sets the number of training iterations, after which the model’s performance will be evaluated and saved. n_embd: This parameter sets the number of embedding dimensions for the GPT model. The embedding layer maps the input sequence into a high-dimensional space, and this parameter determines the size of that space. n_head: This parameter sets the number of attention heads in the multi-head attention layer of the GPT model. The attention mechanism allows the model to focus on specific parts of the input sequence. n_layer: This parameter sets the number of layers in the GPT model. dropout: This parameter sets the dropout probability for the GPT model. Dropout is a regularization technique that randomly drops out some of the neural network’s nodes during training to prevent overfitting. Reading input file torch.manual_seed(1337) # wget https://raw.githubusercontent.com/karpathy/char- rnn/master/data/tinyshakespeare/input.txt with open('input.txt', 'r', encoding='utf-8') as f: text = f.read()
- 13. 13/29 In this codesnippet, the developer is setting a manual seed for PyTorch’s random number generator using torch.manual_seed(). This is done to ensure that the results of the GPT model are reproducible. The argument passed to torch.manual_seed() is an arbitrary number (1337 in this case) that serves as the seed for the random number generator. By setting a fixed seed, the developer can ensure that the same sequence of random numbers is generated every time the code is run, which in turn ensures that the GPT model is trained and tested on the same data. Next, the developer is reading in a text file using Python’s built-in open() function and reading its contents using the read() method. The text file contains the input text that will be used to train the GPT model. The text data can be preprocessed further, for instance, by cleaning the text, tokenizing it, and creating a vocabulary, depending on the requirements of the GPT model. Once the text data is preprocessed, it can be passed through the GPT model to generate predictions. Identifying unique characters that occur in a text chars = sorted(list(set(text))) vocab_size = len(chars) In this code snippet, we are creating a vocabulary for the GPT model. First, we create a sorted list of unique characters present in the text data using the set() function and list() constructor. The set() function returns a collection of unique elements from the text, and the list() constructor converts that set into a list. The sorted() function sorts the list alphabetically, creating a sorted list of unique characters present in the text. Next, we are getting the length of the chars list using the len() function. This gives the number of unique characters in the text and serves as the vocabulary size for the GPT model. The vocabulary size is an important hyperparameter that determines the capacity of the GPT model. The larger the vocabulary size, the more expressive the model can be, but it also increases the model’s complexity and training time. The vocabulary size is typically chosen based on the size of the input text and the nature of the problem being solved. Once the vocabulary is created, the characters in the text data can be mapped to integer values and passed through the GPT model to generate predictions. Creating mapping The first step is to create a mapping between characters and integers, which is necessary for building a language model such as GPT. For the model to work with text data, it needs to be able to represent each character as a numerical value, which is what the following code accomplishes. create a mapping from characters to integers
- 14. 14/29 stoi = {ch:i for i,ch in enumerate(chars) } itos = { i:ch for i,ch in enumerate(chars) } encode = lambda s: [stoi[c] for c in s] # encoder: take a string, output a list of integers decode = lambda l: ''.join([itos[i] for i in l]) # decoder: take a list of integers, output a string print(encode("hii there")) print(decode(encode("hii there"))) This code block creates a character-to-integer mapping and its inverse (integer-to- character mapping) for a set of characters. The stoi dictionary maps each character to a unique integer while itos maps each integer back to its corresponding character. The encode function takes a string as input and returns a list of integers, where each integer corresponds to the index of the character in the chars set. The decode function takes a list of integers and returns the original string by looking up the corresponding characters in the itos dictionary. The code then tests the encoding and decoding functions by encoding the string “hii there” and then decoding the resulting list of integers back into a string. Encoding input data In building a GPT model, it’s important to encode the entire text dataset so that it can be fed into the model. The following code does exactly that. let's now encode the entire text dataset and store it into a torch.Tensor import torch # we use PyTorch: https://pytorch.org data = torch.tensor(encode(text), dtype=torch.long) print(data.shape, data.dtype) print(data[:1000]) # the 1000 characters we looked at earier will to the GPT look like this This code imports the PyTorch library and creates a tensor called data. The tensor is filled with the encoded text data, which is obtained by calling the encode function on the text variable. The dtype parameter is set to torch.long to ensure that the tensor elements are integers. The code prints the shape and data type of the data tensor. The shape attribute tells us the size of the tensor along each dimension, while the dtype attribute tells us the data type of the tensor elements. This information is useful for verifying that the tensor has been created correctly and will be compatible with the GPT model. It then prints the first 1000 elements of the data tensor, which represent the encoded text data. This is useful for verifying that the encoding process has worked correctly and that the data has been loaded into the tensor as expected. Splitting up the data into train and validation sets
- 15. 15/29 The following codeis useful for understanding how the GPT model will process the input data. It shows how the model will process input sequences of length block_size, and how the input and output sequences are related to each other. This understanding can help in designing and training the GPT model. # Let's now split up the data into train and validation sets n = int(0.9*len(data)) # first 90% will be train, rest val train_data = data[:n] val_data = data[n:] block_size = 8 train_data[:block_size+1] x = train_data[:block_size] y = train_data[1:block_size+1] for t in range(block_size): context = x[:t+1] target = y[t] print(f"when input is {context} the target: {target}") This code splits the encoded text data into training and validation sets. The first 90% of the data is assigned to the train_data variable, while the remaining 10% is assigned to the val_data variable. It defines the block_size variable to be 8, which determines the input sequence size that the GPT model will process at a time. It then selects a portion of the training data that is block_size+1 elements long and assigns it to train_data. The x variable is assigned the first block_size elements of train_data, while the y variable is assigned the next block_size elements of train_data, starting from the second element. In other words, y is shifted one position relative to x. Next, the code loops over the block_size elements of x and y, and prints out the input context and target for each position in the input sequence. For each iteration of the loop, the context variable is set to the first t+1 elements of x, where t ranges from 0 to block_size-1. The target variable is set to the t-th element of y. The loop then prints out a message indicating the current input context and target. Generating batches of input and target data for training the GPT torch.manual_seed(1337) batch_size = 4 # how many independent sequences will we process in parallel? block_size = 8 # what is the maximum context length for predictions?
- 16. 16/29 def get_batch(split): # generatea small batch of data of inputs x and targets y data = train_data if split == 'train' else val_data ix = torch.randint(len(data) - block_size, (batch_size,)) x = torch.stack([data[i:i+block_size] for i in ix]) y = torch.stack([data[i+1:i+block_size+1] for i in ix]) return x, y xb, yb = get_batch('train') print('inputs:') print(xb.shape) print(xb) print('targets:') print(yb.shape) print(yb) print('----') for b in range(batch_size): # batch dimension for t in range(block_size): # time dimension context = xb[b, :t+1] target = yb[b,t] print(f"when input is {context.tolist()} the target: {target}") This code sets the random seed of PyTorch to 1337, which ensures that the random number generation is deterministic and reproducible. This is important for training the GPT model and getting consistent results. It defines the batch_size and block_size variables. batch_size specifies how many independent sequences will be processed in parallel in each batch, while block_size specifies the maximum context length for predictions. Then it defines a function called get_batch that generates a small batch of data of inputs x and targets y for a given split (either train or val). The function first selects the appropriate dataset (train_data or val_data) based on the input split. It then randomly selects batch_size starting positions for x using torch.randint(), ensuring that each starting position is at least block_size positions away from the end of the dataset to avoid going
- 17. 17/29 out of bounds.It then constructs x and y tensors by selecting block_size elements starting from each starting position, with y shifted one position to the right relative to x. The function returns the x and y tensors as a tuple. It calls the get_batch() function with the argument ‘train’ to generate a batch of training data. It then prints the shape and contents of the x and y tensors. Finally, it loops over each element in the batch (dimension batch_size) and each position in the input sequence (dimension block_size), and prints out the sequence’s input context and target for each position. The context variable is set to the first t+1 elements of xb[b,:], where t ranges from 0 to block_size-1. The target variable is set to the t-th element of yb[b,:]. The loop then prints out a message indicating the current input context and target. Calculating the average loss on the training and validation datasets using a pre- trained model @torch.no_grad() def estimate_loss(): out = {} model.eval() for split in ['train', 'val']: losses = torch.zeros(eval_iters) for k in range(eval_iters): X, Y = get_batch(split) logits, loss = model(X, Y) losses[k] = loss.item() out[split] = losses.mean() model.train() return out This code defines a function estimate_loss() which calculates the average loss on the training and validation datasets using a pre-trained model. It uses the @torch.no_grad() decorator to disable gradient computation during the evaluation, and sets the model to evaluation mode using model.eval(). Then, it iterates over the training and validation datasets eval_iters times, computes the logits and loss for each batch using the pre- trained model, and records the losses. Finally, it returns the average losses for the two datasets and sets the model back to training mode using model.train(). This function is useful for monitoring the model’s performance during training and determining when to stop training.
- 18. 18/29 Defining one headof the self-attention mechanism in a transformer model class Head(nn.Module): """ one head of self-attention """ def __init__(self, head_size): super().__init__() self.key = nn.Linear(n_embd, head_size, bias=False) self.query = nn.Linear(n_embd, head_size, bias=False) self.value = nn.Linear(n_embd, head_size, bias=False) self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size))) self.dropout = nn.Dropout(dropout) def forward(self, x): B,T,C = x.shape k = self.key(x) # (B,T,C) q = self.query(x) # (B,T,C) # compute attention scores ("affinities") wei = q @ k.transpose(-2,-1) * C**-0.5 # (B, T, C) @ (B, C, T) -> (B, T, T) wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf')) # (B, T, T) wei = F.softmax(wei, dim=-1) # (B, T, T) wei = self.dropout(wei) # perform the weighted aggregation of the values v = self.value(x) # (B,T,C) out = wei @ v # (B, T, T) @ (B, T, C) -> (B, T, C) return out This code defines a module called Head which represents one head of the self-attention mechanism used in the GPT model. The __init__ method initializes three linear layers (key, query, and value) that will be used to project the input tensor x into a lower- dimensional space, which helps compute the attention scores efficiently. The forward method takes as input a tensor x of shape (batch_size, sequence_length,
- 19. 19/29 embedding_size) and computesthe self-attention scores using the dot-product attention mechanism. The attention scores are computed by taking the dot product of the query and key projections and normalizing the result by the square root of the embedding size. The resulting attention scores are then masked with a triangular matrix to prevent attending to future tokens. The attention scores are then normalized with a softmax function, multiplied by the value projection, and finally aggregated to produce the output tensor of shape (batch_size, sequence_length, embedding_size). The dropout layer is applied to the attention scores before the final aggregation. Implementing the multi-head attention mechanism class MultiHeadAttention(nn.Module): """ multiple heads of self-attention in parallel """ def __init__(self, num_heads, head_size): super().__init__() self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)]) self.proj = nn.Linear(n_embd, n_embd) self.dropout = nn.Dropout(dropout) def forward(self, x): out = torch.cat([h(x) for h in self.heads], dim=-1) out = self.dropout(self.proj(out)) return out This PyTorch module implements the multi-head attention mechanism used in building GPT models. It contains a number of heads, each of which computes a self-attention matrix for the input sequence. The output of each head is concatenated and projected to the original embedding size using a linear layer and then passed through a dropout layer. The result is a new sequence of the same length but with a larger embedding dimension that encodes information from multiple self-attention heads. This module is used as a building block in the GPT model. Next we need to add the FeedFoward module class FeedFoward(nn.Module): """ a simple linear layer followed by a non-linearity """ def __init__(self, n_embd): super().__init__()
- 20. 20/29 self.net = nn.Sequential( nn.Linear(n_embd,4 * n_embd), nn.ReLU(), nn.Linear(4 * n_embd, n_embd), nn.Dropout(dropout), ) def forward(self, x): return self.net(x) class Block(nn.Module): """ Transformer block: communication followed by computation """ def __init__(self, n_embd, n_head): # n_embd: embedding dimension, n_head: the number of heads we'd like super().__init__() head_size = n_embd // n_head self.sa = MultiHeadAttention(n_head, head_size) self.ffwd = FeedFoward(n_embd) self.ln1 = nn.LayerNorm(n_embd) self.ln2 = nn.LayerNorm(n_embd) def forward(self, x): x = x + self.sa(self.ln1(x)) x = x + self.ffwd(self.ln2(x)) return x Model training and text generation class BigramLanguageModel(nn.Module): def __init__(self): super().__init__()
- 21. 21/29 # each tokendirectly reads off the logits for the next token from a lookup table self.token_embedding_table = nn.Embedding(vocab_size, n_embd) self.position_embedding_table = nn.Embedding(block_size, n_embd) self.blocks = nn.Sequential(*[Block(n_embd, n_head=n_head) for _ in range(n_layer)]) self.ln_f = nn.LayerNorm(n_embd) # final layer norm self.lm_head = nn.Linear(n_embd, vocab_size) def forward(self, idx, targets=None): B, T = idx.shape # idx and targets are both (B,T) tensor of integers tok_emb = self.token_embedding_table(idx) # (B,T,C) pos_emb = self.position_embedding_table(torch.arange(T, device=device)) # (T,C) x = tok_emb + pos_emb # (B,T,C) x = self.blocks(x) # (B,T,C) x = self.ln_f(x) # (B,T,C) logits = self.lm_head(x) # (B,T,vocab_size) if targets is None: loss = None else: B, T, C = logits.shape logits = logits.view(B*T, C) targets = targets.view(B*T) loss = F.cross_entropy(logits, targets) return logits, loss def generate(self, idx, max_new_tokens): # idx is (B, T) array of indices in the current context for _ in range(max_new_tokens):
- 22. 22/29 # crop idxto the last block_size tokens idx_cond = idx[:, -block_size:] # get the predictions logits, loss = self(idx_cond) # focus only on the last time step logits = logits[:, -1, :] # becomes (B, C) # apply softmax to get probabilities probs = F.softmax(logits, dim=-1) # (B, C) # sample from the distribution idx_next = torch.multinomial(probs, num_samples=1) # (B, 1) # append sampled index to the running sequence idx = torch.cat((idx, idx_next), dim=1) # (B, T+1) return idx model = BigramLanguageModel() m = model.to(device) # print the number of parameters in the model print(sum(p.numel() for p in m.parameters())/1e6, 'M parameters') # create a PyTorch optimizer optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate) for iter in range(max_iters): # every once in a while evaluate the loss on train and val sets if iter % eval_interval == 0 or iter == max_iters - 1: losses = estimate_loss() print(f"step {iter}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}") # sample a batch of data xb, yb = get_batch('train')
- 23. 23/29 # evaluate theloss logits, loss = model(xb, yb) optimizer.zero_grad(set_to_none=True) loss.backward() optimizer.step() # generate from the model context = torch.zeros((1, 1), dtype=torch.long, device=device) print(decode(m.generate(context, max_new_tokens=2000)[0].tolist())) This code defines a bigram language model using PyTorch to train a GPT model. The BigramLanguageModel class is defined as a subclass of nn.Module and contains several layers that are used to build the model. The __init__ method initializes the model with an embedding layer for the tokens and a separate embedding layer for the position of the tokens. Additionally, the model has a sequence of transformer blocks, which are defined by the Block function, and a final layer norm and linear layer to output the logits of the next token. The forward method takes in input sequences and targets, computes the embeddings, applies the transformer blocks, and outputs the logits of the next token along with the loss if targets are provided. The generate method is used to generate new sequences of text from the model. It takes in a starting sequence and a maximum number of new tokens to generate. The method iteratively samples the next token from the model’s predicted probability distribution and appends it to the running sequence until the desired length is reached. In the main part of the code, an instance of the BigramLanguageModel class is created and moved to a specified device. The PyTorch AdamW optimizer is then created, and the training loop begins. In each iteration, a batch of data is sampled from the training set using the get_batch function. The model is then evaluated on this batch of data, the loss is computed, and the gradients are backpropagated using loss.backward(). Finally, the optimizer’s step() method is called to update the model’s parameters. After training, the generate method is used to generate a sequence of text from the trained model. A context tensor of zeros is created, and the generate method is called with this context and a maximum number of new tokens to generate. The resulting sequence of tokens is decoded using the decode function to produce a string of generated text. How to train an existing GPT model with your data?
- 24. 24/29 The previous segmentprovided an introduction on how to construct a GPT model from the ground up. Now, let’s delve into the process of enhancing a pre-existing model using your unique data. This is known as ‘fine-tuning’, a process that refines a base or ‘foundation’ model for specific tasks or datasets. OpenAI offers a range of foundation models that one can leverage, with GPT-NeoX being a notable example. If you are interested in fine-tuning GPT-NeoX with your data, the following steps will guide you through the process. The complete code for the GPT-NeoX can be downloaded from here – https://github.com/EleutherAI/gpt-neox Pre-requisites There are some environmental setup required for GPT-NeoX as well dependencies to be set prior to using the model. Here are the details – Setting up your host To begin, ensure your environment is equipped with Python 3.8 and a suitable version of PyTorch 1.8 or higher. Please be aware that GPT-NeoX relies on certain libraries that may not be compatible with Python 3.10 and above. Python 3.9 seems to function, but our codebase is primarily designed and tested with Python 3.8. To set up the additional required dependencies, execute the following from the repository root: pip install -r requirements/requirements.txt python ./megatron/fused_kernels/setup.py install # optional if not using fused kernels The codebase used here is based on DeeperSpeed, which is a custom version of the DeepSpeed library. DeeperSpeed is a specialized fork of Microsoft’s DeepSpeed library that’s customized to the needs of the GPT-NeoX project. It comes with additional changes tailored specifically for GPT-NeoX by EleutherAI. We highly recommend using an environment isolation tool like Anaconda or a virtual machine prior to proceeding. This is crucial because not doing so could potentially disrupt other repositories that are dependent on DeepSpeed. Flash Attention For utilizing Flash-Attention, begin by installing the extra dependencies specified in ./requirements/requirements-flashattention.txt. Then, adjust the attention type in your configuration as needed (refer to configs). This modification can enhance performance considerably over standard attention, especially on certain GPU architectures like Ampere GPUs (like A100s). Please refer to the repository for further information. Containerized setup
- 25. 25/29 If you prefercontainerized execution, you can use a Dockerfile for running NeoX. To utilize this, initially create an image named gpt-neox from the root directory of the repository using the command docker build -t gpt-neox -f Dockerfile .. Additionally, you can get pre-constructed images at leogao2/gpt-neox on Docker Hub. Following this, you can execute a container based on the created image. For example, the command below attaches the cloned repository directory (gpt-neox) to /gpt-neox in the container, and uses nvidia-docker to grant container access to four GPUs (numbered 0-3). Usage You should utilize deepy.py, a wrapper around the deepspeed launcher, to trigger all functionalities, including inference. There are three principal functions available to you: 1. train.py: This is for training and fine-tuning models. 2. evaluate.py: Use this to evaluate a trained model using the language model evaluation harness. 3. generate.py: This function is for sampling text from a trained model. You can launch these with the following command: ./deepy.py [script.py] [./path/to/config_1.yml] [./path/to/config_2.yml] ... [./path/to/config_n.yml] For instance, to unconditionally generate text with the GPT-NeoX-20B model, use: ./deepy.py generate.py ./configs/20B.yml You can also optionally input a text file (e.g., prompt.txt) as the prompt. This should be a plain .txt file with each prompt separated by newline characters. Remember to pass in the path to an output file. ./deepy.py generate.py ./configs/20B.yml -i prompt.txt -o sample_outputs.txt To replicate our evaluation numbers on tasks like TriviaQA and PIQA, use: ./deepy.py evaluate.py ./configs/20B.yml --eval_tasks triviaqa piqa Configuration GPT-NeoX operations are governed by parameters in a YAML configuration file, which is provided to the deepy.py launcher. We have included some sample .yaml files, including one for GPT-NeoX-20B, and example configurations for other model sizes in the configs folder.
- 26. 26/29 These files areusually all-inclusive, but not necessarily optimized. Depending on your specific GPU setup, you might need to adjust settings such as pipe-parallel-size, model- parallel-size for parallelism, train_micro_batch_size_per_gpu or gradient-accumulation- steps for batch size adjustments, or the zero_optimization dict for optimizer state parallelization. For an in-depth guide on available features and their configuration, refer to the configuration README. For detailed information on all possible arguments, check out configs/neox_arguments.md. Data preparation Prepare your text data in the format accepted by the GPT NeoX model. This usually involves tokenization using a tokenizer that is suitable for the GPT NeoX model. Then, ensure you have downloaded the GPT2 tokenizer vocabulary and merge files. The following links will lead you to them: Training and Fine-tuning Kickstart your training using ‘deepy.py’, which is a wrapper around DeepSpeed’s launcher. It parallelly executes the script across multiple GPUs or nodes. Here’s how to use it: Execute python ./deepy.py train.py [path/to/config1.yml] [path/to/config2.yml] ... You can supply any number of configuration files, which will be merged when the script runs. Optionally, you can include a config prefix, which is a common path for all your configuration files. For instance. execute the following code – python ./deepy.py train.py -d configs 125M.yml local_setup.yml This instruction executes the ‘train.py’ script on every node of the network, with each GPU running one instance of the script. This means every individual GPU across all nodes will be running the ‘train.py’ script separately. The worker nodes and number of GPUs are defined in the ‘/job/hostfile’ file (see parameter documentation), or can be simply included as the ‘num_gpus’ argument if you’re running a single node setup. We suggest defining the model parameters in one configuration file (like ‘configs/125M.yml’) and the data path parameters in another (like ‘configs/local_setup.yml’), for better organization, although it’s not mandatory.
- 27. 27/29 Leverage LeewayHertz’s AIdevelopment services to build a GPT model LeewayHertz offers specialized GPT model development services, catering to the unique needs of businesses. LeewayHertz’s approach is multi-faceted and tailored to ensure businesses fully leverage the potential of AI. Here are some services that LeewayHertz offers to businesses willing to leverage GPT models: Generative AI consulting LeewayHertz provides expert consulting services to help businesses strategize the adoption of GPT models in line with their goals. Their profound technical expertise extends to foundational models and the broader spectrum of generative AI, enabling them to meticulously craft solutions that precisely meet clients’ requirements in accordance with their unique use cases. Data analysis for GPT models LeewayHertz excels in data analysis, a critical step in GPT model development. Whether dealing with structured datasets or unstructured text, our analysts are adept at extracting and processing data to uncover insights. This process is vital for training and refining GPT models to ensure they deliver accurate and relevant results. Custom GPT model development Recognizing the diverse needs of different industries, LeewayHertz specializes in creating custom, domain-specific GPT models using clients’ proprietary data. This process involves assessing the client’s industry and objectives, selecting an appropriate foundational model, and fine-tuning it with proprietary data. This ensures the model is not only powerful but also directly aligned with the client’s business needs. Development of GPT-based solutions LeewayHertz uses foundational models like GPT-4, and GPT 3.5 Turbo to build innovative solutions such as chatbots, recommendation systems, and predictive tools. These solutions are intelligent, creative, and adaptable, designed to tackle complex challenges in various business contexts. Integration into workflows An essential part of our service is the seamless integration of GPT-based solutions into clients’ existing tech infrastructures. This ensures minimal disruption to ongoing operations, allowing businesses to benefit from AI advancements without hindering their current processes. Ongoing upgrade and maintenance
- 28. 28/29 Understanding the dynamicnature of technology, LeewayHertz offers continuous maintenance and upgrade services. This ensures that the custom solution remains cutting-edge, providing ongoing value and innovation to keep businesses competitive. LeewayHertz’s comprehensive approach in building GPT models involves in-depth consultation, specialized data analysis, custom model development, innovative solution creation, seamless integration, and ongoing support. This holistic approach ensures that businesses can effectively harness the power of generative AI to meet their specific objectives and challenges. Things to consider while building a GPT model Removing bias and toxicity As we strive to build powerful generative AI models, we must be aware of the tremendous responsibility that comes with it. It is crucial to acknowledge that models such as GPT are trained on vast and unpredictable data from the internet, which can lead to biases and toxic language in the final product. As AI technology evolves, responsible practices become increasingly important. We must ensure that our AI models are developed and deployed ethically and with social responsibility in mind. Prioritizing responsible AI practices is vital in reducing the risks of biased and toxic content while fully unlocking the potential of generative AI to create a better world. It is necessary to take a proactive approach to ensure that the output generated by AI models is free from bias and toxicity. This includes filtering training datasets to eliminate potentially harmful content and implementing watchdog models to monitor output in real- time. Furthermore, leveraging first-party data to train and fine-tune AI models can significantly enhance their quality. This allows customization to meet specific use cases, improving overall performance. Improving hallucination It is essential to acknowledge that while GPT models can generate convincing arguments, they may not always be based on factual accuracy. Within the developer community, this issue is known as “hallucination,” which can reduce the reliability of the output produced by these AI models. To overcome this challenge, you need to consider the measures as taken by OpenAI and other vendors, including data augmentation, adversarial training, improved model architectures, and human evaluation to enhance the accuracy of the output and decrease the risk of hallucination and ensure output generated by the model is as precise and dependable as possible. Preventing data leakage Establishing transparent policies is crucial to prevent developers from passing sensitive information into GPT models, which could be incorporated into the model and resurfaced in a public context. By implementing such policies, we can prevent the unintentional disclosure of sensitive information, safeguard the privacy and security of individuals and
- 29. 29/29 organizations, and avoidany negative consequences. This is essential to remain vigilant in safeguarding against potential risks associated with the use of GPT models and take proactive measures to mitigate them. Incorporating queries and actions Current generative models can provide answers based on their initial large training data set or smaller “fine-tuning” data sets, which are not real-time and historical. However, the next generation of models will take a significant leap forward. These models will possess the capability to identify when to seek information from external sources such as a database or Google or trigger actions in external systems, transforming generative models from isolated oracles to fully connected conversational interfaces with the world. By enabling this new level of connectivity, we can unlock a new set of use cases and possibilities for these models, creating a more dynamic and seamless user experience that provides real-time, relevant information and insights. Endnote GPT models are a significant milestone in the history of AI development, which is a part of a larger LLM trend that will grow in the future. Furthermore, OpenAI’s groundbreaking move to provide API access is part of its model-as-a-service business scheme. Additionally, GPT’s language-based capabilities allow for creating innovative products as it excels at tasks such as text summarization, classification, and interaction. GPT models are expected to shape the future internet and how we use technology and software. Building a GPT model may be challenging, but with the right approach and tools, it becomes a rewarding experience that opens up new opportunities for NLP applications.