Whatever your take on Large Language Models (LLMs) – are they beneficial? dangerous? a short-lived fashion, like crypto? – they are here, now. And that means, it is a good thing to know (at a level one needs to decide for oneself) how they work. On this same day, I am publishing What are Large Language Models? What are they not?, intended for a more general audience. In this post, I’d like to address deep learning practitioners, walking through a torch implementation of GPT-2 (Radford et al. 2019), the second in OpenAI’s succession of ever-larger models trained on ever-more-vast text corpora. You’ll see that a complete model implementation fits in fewer than 250 lines of R code.
Sources, resources
The code I’m going to present is found in the minhub repository. This repository deserves a mention of its own. As emphasized in the README,
minhub is a collection of minimal implementations of deep learning models, inspired by minGPT. All models are designed to be self-contained, single-file, and devoid of external dependencies, making them easy to copy and integrate into your own projects.
Evidently, this makes them excellent learning material; but that is not all. Models also come with the option to load pre-trained weights from Hugging Face’s model hub. And if that weren’t enormously convenient already, you don’t have to worry about how to get tokenization right: Just download the matching tokenizer from Hugging Face, as well. I’ll show how this works in the final section of this post. As noted in the minhub README, these facilities are provided by packages hfhub and tok.
As realized in minhub, gpt2.R is, mostly, a port of Karpathy’s MinGPT. Hugging Face’s (more sophisticated) implementation has also been consulted. For a Python code walk-through, see https://amaarora.github.io/posts/2020-02-18-annotatedGPT2.html. This text also consolidates links to blog posts and learning materials on language modeling with deep learning that have become “classics” in the short time since they were written.
A minimal GPT-2
Overall architecture
The original Transformer (Vaswani et al. 2017) was built up of both an encoder and a decoder stack, a prototypical use case being machine translation. Subsequent developments, dependent on envisaged primary usage, tended to forego one of the stacks. The first GPT, which differs from GPT-2 only in relative subtleties, kept only the decoder stack. With “self-attention” wired into every decoder block, as well as an initial embedding step, this is not a problem – external input is not technically different from successive internal representations.
Here is a screenshot from the initial GPT paper (Radford and Narasimhan 2018), visualizing the overall architecture. It is still valid for GPT-2. Token as well as position embedding are followed by a twelve-fold repetition of (identical in structure, though not sharing weights) transformer blocks, with a task-dependent linear layer constituting model output.
In gpt2.R, this global structure and what it does is defined in nn_gpt2_model(). (The code is more modularized – so don’t be confused if code and screenshot don’t perfectly match.)
First, in initialize(), we have the definition of modules:
self$transformer <- nn_module_dict(list(
wte = nn_embedding(vocab_size, n_embd),
wpe = nn_embedding(max_pos, n_embd),
drop = nn_dropout(pdrop),
h = nn_sequential(!!!map(
1:n_layer,
\(x) nn_gpt2_transformer_block(n_embd, n_head, n_layer, max_pos, pdrop)
)),
ln_f = nn_layer_norm(n_embd, eps = 1e-5)
))
self$lm_head <- nn_linear(n_embd, vocab_size, bias = FALSE)The two top-level components in this model are the transformer and lm_head, the output layer. This code-level distinction has an important semantic dimension, with two aspects standing out. First, and quite directly, transformer’s definition communicates, in a succinct way, what it is that constitutes a Transformer. What comes thereafter – lm_head, in our case – may vary. Second, and importantly, the distinction reflects the essential underlying idea, or essential operationalization, of natural language processing in deep learning. Learning consists of two steps, the first – and indispensable one – being to learn about language (this is what LLMs do), and the second, much less resource-consuming, one consisting of adaptation to a concrete task (such as question answering, or text summarization).
To see in what order (and how often) things happen, we look inside forward():
tok_emb <- self$transformer$wte(x)
pos <- torch_arange(1, x$size(2))$to(dtype = "long")$unsqueeze(1)
pos_emb <- self$transformer$wpe(pos)
x <- self$transformer$drop(tok_emb + pos_emb)
x <- self$transformer$h(x)
x <- self$transformer$ln_f(x)
x <- self$lm_head(x)
xAll modules in transformer are called, and thus executed, once; this includes h – but h itself is a sequential module made up of transformer blocks.
Since these blocks are the core of the model, we’ll look at them next.
Transformer block
Here’s how, in nn_gpt2_transformer_block(), each of the twelve blocks is defined.
self$ln_1 <- nn_layer_norm(n_embd, eps = 1e-5)
self$attn <- nn_gpt2_attention(n_embd, n_head, n_layer, max_pos, pdrop)
self$ln_2 <- nn_layer_norm(n_embd, eps = 1e-5)
self$mlp <- nn_gpt2_mlp(n_embd, pdrop)On this level of resolution, we see that self-attention is computed afresh at every stage, and that the other constitutive ingredient is a feed-forward neural network. In addition, there are two modules computing layer normalization, the type of normalization employed in transformer blocks. Different normalization algorithms tend to distinguish themselves from one another in what they average over; layer normalization (Ba, Kiros, and Hinton 2016) – surprisingly, maybe, to some readers – does so per batch item. That is, there is one mean, and one standard deviation, for each unit in a module. All other dimensions (in an image, that would be spatial dimensions as well as channels) constitute the input to that item-wise statistics computation.
Continuing to zoom in, we will look at both the attention- and the feed-forward network shortly. Before, though, we need to see how these layers are called. Here is all that happens in forward():
x <- x + self$attn(self$ln_1(x))
x + self$mlp(self$ln_2(x))These two lines deserve to be read attentively. As opposed to just calling each consecutive layer on the previous one’s output, this inserts skip (also termed residual) connections that, each, circumvent one of the parent module’s principal stages. The effect is that each sub-module does not replace, but just update what is passed in with its own view on things.
Transformer block up close: Self-attention
Of all modules in GPT-2, this is by far the most intimidating-looking. But the basic algorithm employed here is the same as what the classic “dot product attention paper” (Bahdanau, Cho, and Bengio 2014) proposed in 2014: Attention is conceptualized as similarity, and similarity is measured via the dot product. One thing that can be confusing is the “self” in self-attention. This term first appeared in the Transformer paper (Vaswani et al. 2017), which had an encoder as well as a decoder stack. There, “attention” referred to how the decoder blocks decided where to focus in the message received from the encoding stage, while “self-attention” was the term coined for this technique being applied inside the stacks themselves (i.e., between a stack’s internal blocks). With GPT-2, only the (now redundantly-named) self-attention remains.
Resuming from the above, there are two reasons why this might look complicated. For one, the “triplication” of tokens introduced, in Transformer, through the “query - key - value” frame1. And secondly, the additional batching introduced by having not just one, but several, parallel, independent attention-calculating processes per layer (“multi-head attention”). Walking through the code, I’ll point to both as they make their appearance.
We again start with module initialization. This is how nn_gpt2_attention() lists its components:
# key, query, value projections for all heads, but in a batch
self$c_attn <- nn_linear(n_embd, 3 * n_embd)
# output projection
self$c_proj <- nn_linear(n_embd, n_embd)
# regularization
self$attn_dropout <- nn_dropout(pdrop)
self$resid_dropout <- nn_dropout(pdrop)
# causal mask to ensure that attention is only applied to the left in the input sequence
self$bias <- torch_ones(max_pos, max_pos)$
bool()$
tril()$
view(c(1, 1, max_pos, max_pos)) |>
nn_buffer()Besides two dropout layers, we see:
- A linear module that effectuates the above-mentioned triplication. Note how this is different from just having three identical versions of a token: Assuming all representations were initially mostly equivalent (through random initialization, for example), they will not remain so once we’ve begun to train the model.
- A module, called
c_proj, that applies a final affine transformation. We will need to look at usage to see what this module is for. - A buffer – a tensor that is part of a module’s state, but exempt from training – that makes sure that attention is not applied to previous-block output that “lies in the future.” Basically, this is achieved by masking out future tokens, making use of a lower-triangular matrix.
As to forward(), I am splitting it up into easy-to-digest pieces.
As we enter the method, the argument, x, is shaped just as expected, for a language model: batch dimension times sequence length times embedding dimension.
x$shape
[1] 1 24 768Next, two batching operations happen: (1) triplication into queries, keys, and values; and (2) making space such that attention can be computed for the desired number of attention heads all at once. I’ll explain how after listing the complete piece.
# batch size, sequence length, embedding dimensionality (n_embd)
c(b, t, c) %<-% x$shape
# calculate query, key, values for all heads in batch and move head forward to be the batch dim
c(q, k, v) %<-% ((self$c_attn(x)$
split(self$n_embd, dim = -1)) |>
map(\(x) x$view(c(b, t, self$n_head, c / self$n_head))) |>
map(\(x) x$transpose(2, 3)))First, the call to self$c_attn() yields query, key, and value vectors for each embedded input token. split() separates the resulting matrix into a list. Then map() takes care of the second batching operation. All of the three matrices are re-shaped, adding a fourth dimension. This fourth dimension takes care of the attention heads. Note how, as opposed to the multiplying process that triplicated the embeddings, this divides up what we have among the heads, leaving each of them to work with a subset inversely proportional to the number of heads used. Finally, map(\(x) x$transpose(2, 3) mutually exchanges head and sequence-position dimensions.
Next comes the computation of attention itself.
# causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
att <- q$matmul(k$transpose(-2, -1)) * (1 / sqrt(k$size(-1)))
att <- att$masked_fill(self$bias[, , 1:t, 1:t] == 0, -Inf)
att <- att$softmax(dim = -1)
att <- self$attn_dropout(att)First, similarity between queries and keys is computed, matrix multiplication effectively being a batched dot product. (If you’re wondering about the final division term in line one, this scaling operation is one of the few aspects where GPT-2 differs from its predecessor. Check out the paper if you’re interested in the related considerations.) Next, the aforementioned mask is applied, resultant scores are normalized, and dropout regularization is used to encourage sparsity.
Finally, the computed attention2 needs to be passed on to the ensuing layer. This is where the value vectors come in – those members of this trinity that we haven’t yet seen in action.
y <- att$matmul(v) # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
y <- y$transpose(2, 3)$contiguous()$view(c(b, t, c)) # re-assemble all head outputs side by side
# output projection
y <- self$resid_dropout(self$c_proj(y))
yConcretely, what the matrix multiplication does here is weight the value vectors by the attention, and add them up. This happens for all attention heads at the same time, and really represents the outcome of the algorithm as a whole.
Remaining steps then restore the original input size. This involves aligning the results for all heads one after the other, and then, applying the linear layer c_proj to make sure these results are not treated equally and/or independently, but combined in a useful way. Thus, the projection operation hinted at here really is a made up of a mechanical step (view()) and an “intelligent” one (transformation by c_proj()).
Transformer block up close: Feed-forward network (MLP)
Compared to the first, the attention module, there really is not much to say about the second core component of the transformer block (nn_gpt2_mlp()). It really is “just” an MLP – no “tricks” involved. Two things deserve pointing out, though.
First, you may have heard about the MLP in a transformer block working “position-wise,” and wondered what is meant by this. Consider what happens in such a block:
x <- x + self$attn(self$ln_1(x))
x + self$mlp(self$ln_2(x))The MLP receives its input (almost) directly from the attention module. But that, as we saw, was returning tensors of size [batch size, sequence length, embedding dimension]. Inside the MLP – cf. its forward() – the number of dimensions never changes:
x |>
self$c_fc() |> # nn_linear(n_embd, 4 * n_embd)
self$act() |> # nn_gelu(approximate = "tanh")
self$c_proj() |> # nn_linear(4 * n_embd, n_embd)
self$dropout() # nn_dropout(pdrop)Thus, these transformations are applied to all elements in the sequence, independently.
Second, since this is the only place where it appears, a note on the activation function employed. GeLU stands for “Gaussian Error Linear Units,” proposed in (Hendrycks and Gimpel 2020). The idea here is to combine ReLU-like activation effects with regularization/stochasticity. In theory, each intermediate computation would be weighted by its position in the (Gaussian) cumulative distribution function – effectively, by how much bigger (smaller) it is than the others. In practice, as you see from the module’s instantiation, an approximation is used.
And that’s it for GPT-2’s main actor, the repeated transformer block. Remain two things: what happens before, and what happens thereafter.
From words to codes: Token and position embeddings
Admittedly, if you tokenize the input dataset as required (using the matching tokenizer from Hugging Face – see below), you do not really end up with words. But still, the well-established fact holds: Some change of representation has to happen if the model is to successfully extract linguistic knowledge. Like many Transformer-based models, the GPT family encodes tokens in two ways. For one, as word embeddings. Looking back to nn_gpt2_model(), the top-level module we started this walk-through with, we see:
wte = nn_embedding(vocab_size, n_embd)This is useful already, but the representation space that results does not include information about semantic relations that may vary with position in the sequence – syntactic rules, for example, or phrase pragmatics. The second type of encoding remedies this. Referred to as “position embedding,” it appears in nn_gpt2_model() like so:
wpe = nn_embedding(max_pos, n_embd)Another embedding layer? Yes, though this one embeds not tokens, but a pre-specified number of valid positions (ranging from 1 to 1024, in GPT’s case). In other words, the network is supposed to learn what position in a sequence entails. This is an area where different models may vary vastly. The original Transformer employed a form of sinusoidal encoding; a more recent refinement is found in, e.g., GPT-NeoX (Su et al. 2021).
Once both encodings are available, they are straightforwardly added (see nn_gpt2_model()$forward()):
tok_emb <- self$transformer$wte(x)
pos <- torch_arange(1, x$size(2))$to(dtype = "long")$unsqueeze(1)
pos_emb <- self$transformer$wpe(pos)
x <- self$transformer$drop(tok_emb + pos_emb)The resultant tensor is then passed to the chain of transformer blocks.
Output
Once the transformer blocks have been applied, the last mapping is taken care of by lm_head:
x <- self$lm_head(x) # nn_linear(n_embd, vocab_size, bias = FALSE)This is a linear transformation that maps internal representations back to discrete vocabulary indices, assigning a score to every index. That being the model’s final action, it is left to the sample generation process is to decide what to make of these scores. Or, put differently, that process is free to choose among different established techniques. We’ll see one – pretty standard – way in the next section.
This concludes model walk-through. I have left out a few details (such as weight initialization); consult gpt.R if you’re interested.
End-to-end-usage, using pre-trained weights
It’s unlikely that many users will want to train GPT-2 from scratch. Let’s see, thus, how we can quickly set this up for sample generation.
Create model, load weights, get tokenizer
The Hugging Face model hub lets you access (and download) all required files (weights and tokenizer) directly from the GPT-2 page. All files are versioned; we use the most recent version.
identifier <- "gpt2"
revision <- "e7da7f2"
# instantiate model and load Hugging Face weights
model <- gpt2_from_pretrained(identifier, revision)
# load matching tokenizer
tok <- tok::tokenizer$from_pretrained(identifier)
model$eval()tokenize
Decoder-only transformer-type models don’t need a prompt. But usually, applications will want to pass input to the generation process. Thanks to tok, tokenizing that input couldn’t be more convenient:
torch_tensor
Columns 1 to 11 2949 7077 318 10893 319 262 5527 11 2489 286 262
Columns 12 to 22 3595 318 257 20596 9546 2644 31779 2786 3929 287 10804
Columns 23 to 24 13 31428
[ CPULongType{1,24} ]Generate samples
Sample generation is an iterative process, the model’s last prediction getting appended to the – growing – prompt.
prompt_length <- idx$size(-1)
for (i in 1:30) { # decide on maximal length of output sequence
# obtain next prediction (raw score)
with_no_grad({
logits <- model(idx + 1L)
})
last_logits <- logits[, -1, ]
# pick highest scores (how many is up to you)
c(prob, ind) %<-% last_logits$topk(50)
last_logits <- torch_full_like(last_logits, -Inf)$scatter_(-1, ind, prob)
# convert to probabilities
probs <- nnf_softmax(last_logits, dim = -1)
# probabilistic sampling
id_next <- torch_multinomial(probs, num_samples = 1) - 1L
# stop if end of sequence predicted
if (id_next$item() == 0) {
break
}
# append prediction to prompt
idx <- torch_cat(list(idx, id_next), dim = 2)
}To see the output, just use tok$decode():
tok$decode(as.integer(idx))[1] "No duty is imposed on the rich, rights of the poor is a hollow phrase...
Enough languishing in custody. Equality is over"To experiment with text generation, just copy the self-contained file, and try different sampling-related parameters. (And prompts, of course!)
As always, thanks for reading!
Photo by Marjan Blan on Unsplash
If this terminology is unfamiliar, you’ll find a nice (and very popular) introduction here.↩︎
I am italicizing the word so as to hint at a special way of using the term. While the expression in itself does sound rather strange, attention is often employed to signify the state reached after normalizing the – usually seen as “raw” – scores.↩︎