LLaMA in R with Keras and TensorFlow

OpenAI’s chatGPT has woke up a collective consciousness of what Massive
Language Fashions (LLMs) are able to. With that awakening comes a day by day
march of LLM information: new merchandise, new options, new fashions, new
capabilities, (and new worries). It appears we’re within the early levels of a
Cambrian explosion of LLMs and LLM powered instruments; it’s not but clear how
LLMs will affect and affect our skilled and private lives, however
it appears clear that they’ll, in a roundabout way.

Since LLMs are right here to remain, it’s worthwhile to take a while to
perceive how these fashions work from a first-principles perspective.
Beginning with the mechanics might help foster sturdy intuitions that can
inform our utilization of those fashions now and sooner or later. (Particularly if
the longer term is one the place LLMs are a staple of the info scientist’s
toolbox, as frequent as an lm() perform name).

And what higher approach is there to study than by doing. So with that
preamble, on this put up we’ll stroll via an implementation of an LLM,
LLaMA (Touvron et al. 2023)
particularly, in TensorFlow and Keras, with the purpose being to develop
understanding first, functionality second.

Why LLaMA? With the sheer quantity of LLM associated content material and information out
there, it might probably appear formidable to know the place to get began. Virtually weekly
it appears there’s a new mannequin introduced. Looking some hubs of LLM
exercise (HuggingFace,
TFHub,
reddit,
HackerNews) muddies the waters even
extra. How one can choose a particular mannequin?

Of the numerous LLM-related information gadgets up to now months, one which stands
head-and-shoulders above the group is the launch of
LLaMA,
a contemporary, foundational LLM made accessible to the general public by Meta AI in
February 2023. On frequent benchmarks, LLaMA outperforms OpenAI’s GPT-3,
whereas being considerably smaller (although nonetheless giant).

LLaMA is a good beginning place as a result of it’s a easy and trendy
structure, has wonderful efficiency on benchmarks, and is open. The
mannequin structure has had just some new concepts integrated into it since
the unique Transformer structure first described in,
“Consideration Is All You Want”
printed from Google (Vaswani et al. 2017). 4 completely different sizes of
LLaMA have been launched: 7 billion and 13 billion parameter fashions
skilled on 1 Trillion tokens, and 33 billion and 65 billion parameter
fashions skilled on 1.4 trillion tokens. This is a gigantic quantity of
coaching information these fashions have seen–the biggest 65B mannequin has been
skilled on roughly the “Chinchilla
compute-optimum” (Hoffmann et al. 2022)
variety of tokens, whereas the smaller LLaMAs are considerably
past that optimum. On this weblog put up we’ll deal with the smallest, 7B
parameter LLaMA mannequin, which you’ll be able to comfortably load domestically and run on
CPU with solely 64Gb of RAM.

Whereas not strictly needed, to comply with alongside domestically, you’ll in all probability
wish to purchase the pre-trained LLaMA weights one
approach or
one other. Notice, the
weights do include their very own license, which you’ll be able to preview
right here.

So, with out additional ado, let’s get began.

Setup

First, we’ll wish to set up the required R and Python packages, and
configure a digital atmosphere:

remotes::install_github(c("rstudio/reticulate",
                          "rstudio/tensorflow",
                          "rstudio/keras"))
# reticulate::install_python("3.10:newest")                          
reticulate::virtualenv_create("./.venv", model = "3.10:newest")
tensorflow::install_tensorflow(envname = "./.venv", model = "launch",
                               extra_packages = "tensorflow-text")

With that out of the best way, let’s load some packages and put together our R
session:

library(purrr)
library(envir)

library(tensorflow)
library(tfautograph)
library(keras)

use_virtualenv("./.venv")
choices(tensorflow.extract.warn_tensors_passed_asis = FALSE)

attach_eval({
  import_from(glue, glue)
  import_from(jsonlite, read_json)
  import_from(withr, with_dir, with_options)
  import_from(keras$layers, Dense)
  np <- reticulate::import("numpy", convert = FALSE)

  seq_len0 <- perform(x) seq.int(from = 0L, size.out = x)
})

In case you’ve acquired the pre-trained weights, it’ll be handy to
convert them from the torch checkpoint format to one thing that’s extra
framework agnostic (you solely want to do that as soon as, after all):

# reticulate::py_install("torch", pip = TRUE)
torch <- reticulate::import("torch", convert = FALSE)
with_dir("~/github/facebookresearch/llama/weights/LLaMA/7B", {
  pretrained_weights <- torch$load("consolidated.00.pth",
                                   map_location = "cpu")
  for (identify in names(pretrained_weights)) {
    filename <- sprintf("%s.npy", identify)
    array <- pretrained_weights[[name]]$numpy()
    np$save(filename, array)
    message(glue(
      "wrote: '{basename(filename)}' with form: {array$form}"))
  }
})

We’ll additionally outline a helper perform so we are able to keep away from having to retype the
full path to our weights:

weights_path <- perform(filename) normalizePath(file.path(
  "~/github/facebookresearch/llama/weights/LLaMA/",
  glue(filename, .envir = dad or mum.body())), mustWork = TRUE)

And cargo the mannequin configuration parameters particular to the 7B LLaMA,
which we’ll use to construct the mannequin.

params <- read_json(weights_path("7B/params.json"))
str(params)

Listing of 6
 $ dim        : int 4096
 $ multiple_of: int 256
 $ n_heads    : int 32
 $ n_layers   : int 32
 $ norm_eps   : num 1e-06
 $ vocab_size : int -1

Tokenizer

The primary part to LLaMA is the tokenizer, which converts textual content to a
sequence of integers. The LLaMA mannequin makes use of the
SentencePiece tokenizer from
Google. SentencePiece is on the market as a TensorFlow graph operation
via
tf_text.SentencepieceTokenizer,
and likewise as a Keras layer in
keras_nlp.tokenizers.SentencepieceTokenizer.
By selection of a coin flip, we’ll use the lower-level tf_text interface.

tf_text <- reticulate::import("tensorflow_text")
tokenizer_path <- weights_path("tokenizer.mannequin")
tokenizer <- tf_text$SentencepieceTokenizer(
  tf$io$gfile$GFile(tokenizer_path, "rb")$learn(),
  add_bos = TRUE, add_eos = FALSE,
)

Let’s check it out with a immediate:

immediate <- "One of the simplest ways to draw bees"
tokenizer$tokenize(immediate)

tf.Tensor([    1   450  1900   982   304 13978   367   267], form=(8), dtype=int32)

immediate |> tokenizer$tokenize() |> tokenizer$detokenize()

tf.Tensor(b'One of the simplest ways to draw bees', form=(), dtype=string)

Let’s outline a show_tokens() helper perform and play with the
tokenizer slightly.

show_tokens <- perform(what) >
    map_chr(perform(id) >
        tokenizer$detokenize() )

  names(tokens) <- token_ids
  tokens


show_tokens(immediate)

        1       450      1900       982       304     13978       367       267
       ""     "The"    "finest"     "approach"      "to" "appeal to"      "be"      "es"

Notice that “bees” is 2 tokens. Not each token corresponds to a phrase.
For instance, one non-word token we are able to reliably count on to indicate up in a
tokenizer skilled on a corpus of English textual content is “ing.” Nonetheless, when the
“ing” token reveals up is not going to at all times comply with your intuitions, as a result of
frequent phrases get their very own token id, even when they are often decomposed into
a number of tokens.

    1  2348
   "" "ing"

        1      1985
       "" "working"

     1   8525    292
    "" "flex"  "ing"

     1   2113   9292
    ""  "gained" "king"

One other factor to notice concerning the tokenizer is that every token sequence
begins with token id 1. It is a particular beginning-of-sequence
token that we requested be added after we loaded the tokenizer with
add_bos = TRUE. There are two different such particular tokens that we are going to
encounter later: an end-of-sequence particular tokens with id 2, and an
unknown-token with id 0.

as.character(tokenizer$id_to_string(0L))

[1] "<unk>"

as.character(tokenizer$id_to_string(1L))

[1] "<s>"

as.character(tokenizer$id_to_string(2L))

[1] "</s>"

    1     0     2
   "" " ⁇ "    ""

Total, there are 32,000 tokens.

as.integer(tokenizer$vocab_size())

[1] 32000

One final commentary is that the extra often encountered tokens are
assigned decrease ids.

show_tokens(seq(50, len = 10))

 50  51  52  53  54  55  56  57  58  59
"/" "0" "1" "2" "3" "4" "5" "6" "7" "8"

show_tokens(seq(100, len = 10))

100 101 102 103 104 105 106 107 108 109
"a" "b" "c" "d" "e" "f" "g" "h" "i" "j"

show_tokens(seq(1000, len = 10))

   1000    1001    1002    1003    1004    1005    1006    1007    1008    1009
  "ied"    "ER"  "stat"   "fig"    "me"   "von" "inter"  "roid"  "ater" "their"

show_tokens(seq(10000, len = 10))

   10000    10001    10002    10003    10004    10005    10006    10007
   "ång"  "citep"    "In poor health"   "rank" "sender"   "beim"    "рак" "compat"
   10008    10009
"happens"  "diese"

show_tokens(seq(20000, len = 10))

    20000     20001     20002     20003     20004     20005     20006     20007
  "admit" "Remark"     "стя"    "Vien"      "ці"  "permut"     "cgi"    "crít"
    20008     20009
"Console"    "ctic"

show_tokens(seq(to = as.integer(tokenizer$vocab_size()) - 1, len = 10))

31990 31991 31992 31993 31994 31995 31996 31997 31998 31999
  "ὀ"  "げ"  "べ"  "边"  "还"  "黃"  "왕"  "收"  "弘"  "给"

Transferring on, the following step after tokenization is embedding. An embedding
layer is successfully a dictionary lookup that converts an integer (token
id) to a 1-d float array. For this we are able to use the usual keras
Embedding layer.

tok_embeddings <- keras$layers$Embedding(
  input_dim = tokenizer$vocab_size(),
  output_dim = params$dim,
  embeddings_initializer =
    (...) np$load(weights_path("7B/tok_embeddings.weight.npy"))
)

tok_embeddings(3L) |> str()

<tf.Tensor: form=(4096), dtype=float32, numpy=…>

immediate |> # "One of the simplest ways to draw bees"
  tokenizer$tokenize() |>
  tok_embeddings() |>
  str()

<tf.Tensor: form=(8, 4096), dtype=float32, numpy=…>

TransformerBlock

As soon as it’s tokenized and embedded, the enter then passes via the majority
of the mannequin, a sequence of repeating TransformerBlock layers. The 7B
mannequin has 32 of those TransformerBlock layers, whereas the 65B mannequin has
80 of them.

weights_path("7B/params.json")  |> read_json() |> _$n_layers

[1] 32

weights_path("65B/params.json") |> read_json() |> _$n_layers

[1] 80

Here’s what the transformer block appears to be like like:

TransformerBlock(keras$layers$Layer) %py_class% {
  initialize <- perform(attn_head_size, attn_n_heads,
                         norm_eps = k_epsilon(), ...,
                         block_id = NULL) {
    tremendous$initialize(...)

    self$consideration <- Consideration(attn_head_size, attn_n_heads,
                                block_id = block_id)

    self$feed_forward <- FeedForward(
      hidden_dim = 4 * attn_head_size * attn_n_heads,
      block_id = block_id)

    self$attention_norm <- RMSNorm(eps = norm_eps,
                                   block_id = block_id,
                                   feeds_into = "consideration")
    self$feed_forward_norm <- RMSNorm(eps = norm_eps,
                                      block_id = block_id,
                                      feeds_into = "ffn")
  }

  name <- perform(x) >
      self$attention_norm() 
}

Whereas there may be not loads of code, there are loads of concepts packed in
there. This block varieties the principle trunk of the mannequin, so it’s value
taking the time to undergo it slowly.

We implement the TransformerBlock as a subclassed
keras.layers.Layer. That is provides us some niceties like the flexibility to
compose with different Keras layers, however these are largely irrelevant to the
function of this weblog put up; we might simply as simply implement this as,
for instance, a vanilla R6 class. Our TransformerBlock class has two
strategies: initialize, known as after we first create the block, and
name, known as after we run the ahead cross of the block.

In initialize, we create 4 layers: an Consideration layer, a
FeedForward layer, and a pair of RMSNorm layers. We’ll take a detailed take a look at
every of those quickly, however even earlier than we achieve this, we are able to see how they match
collectively by trying on the TransformerBlock$name() technique.

The name technique has just a few easy concepts. In no explicit order, the
first one to look at is the composition sample of including residuals.

x2 <- x |> ...
x <- x + x2 # add residual x to x2

It is a frequent sample that helps with mannequin coaching, and particularly
to assist with the vanishing gradient
drawback. It’s
a skip-connection within the other-wise linear sequence of matrix
transformations. It reinjects info (throughout the ahead cross), and
gradients (throughout again propagation), again into the trunk. You possibly can assume
of those residual connections as liberating the learnable layers in-between
(the ... within the pseudo code) from the burden of getting to
“pass-through” or “protect” info in x, permitting the weights to
as a substitute deal with studying transformations which are, (in corporatese
vernacular), value-adding.

The following composition sample to notice is the repeating utilization of a
normalization layer:

x2 <- x |> norm() |> ...
x <- x + x2

There are lots of sorts of normalization layers, however to barely
over-generalize, they will all be regarded as a stabilizer that helps
with coaching. Like their deep-learning cousins the regularizers, their
essential perform is to maintain values passing via in a smart vary–in
the ball park of (-1, 1), sometimes. We’ll take a more in-depth take a look at
RMSNorm quickly.

Stripped of two methods which are largely there to assist the mannequin practice,
residuals and normalization, the core of the TransformerBlock is simply
this:

x |> consideration() |> feed_forward()

In a second we’ll see that that feed_foward is a barely fancier
variation of a traditional sequence of Dense layer. Earlier than we get
there we are able to we safely skip forward to distill the next instinct: a
TransformerBlock is principally an Consideration layer adopted by just a few
(fancy) dense layers, with some easy composition patterns (methods)
that assist with coaching. Consideration is the guts of the mannequin: it’s the
most fascinating, and likewise essentially the most concerned.

With the framing in place, let’s undergo and take a more in-depth take a look at
RMSNorm, FeedForward, after which with the muse in place, we’ll
flip our consideration to Consideration.

RMSNorm

RMSNorm(keras$layers$Layer) %py_class% {
  initialize <-
    perform(eps = 1e-6, ..., block_id = NULL, feeds_into = NULL) {
      tremendous$initialize(...)
      self$eps <- eps
      self$block_id <- block_id
      self$feeds_into <- feeds_into
    }

  construct <- perform(input_shape) {
    # input_shape == (batch_size, seqlen, params$dim)
    # self$w will broadcast over batch_size and seqlen dims.
    # w_shape == (1, 1, params$dim)
    w_shape <- rep(1L, size(input_shape))
    w_shape[length(input_shape)] <- as.integer(input_shape) |> tail(1L)

    # outline a neighborhood perform that can load
    # the pretrained-weights if we provided `block_id` and `feeds_into`
    import_from({self}, block_id, feeds_into)
    initializer <-if (is.null(block_id))
      "ones"
      else if (block_id >=0) {
        (...) weights_path("7B/layers.{block_id}.{feeds_into}_norm.weight.npy") |>
               np$load() |> np$expand_dims(0:1)
      } else if(block_id == -1)
        # load weights for the ultimate output normalization layer, which isn't
        # a part of a TransformerBlock
        (...) weights_path("7B/norm.weight.npy") |>
               np$load() |> np$expand_dims(0:1)

    self$w <- self$add_weight(form = w_shape,
                              initializer = initializer,
                              trainable = TRUE)
  }

  rrms <- perform(x) {
    # reciprocal root imply sq. alongside the final axis
    x %>% # (batch_size, seqlen, n_features)
      tf$math$sq.() %>%
      tf$reduce_mean(axis = -1L, keepdims = TRUE) %>% # (batch_size, seqlen, 1)
      tf$math$add(self$eps) %>% # for numerical stability
      tf$math$rsqrt()
  }

  name <- perform(x) {
    x * self$rrms(x) * self$w
  }
}

RMSnorm() has a single trainable tensor w. Within the ahead cross, every
worth within the enter is multiplied by the reciprocal-root-mean-square of
all of the values within the characteristic axis and by w. Definitely a mouthful, however
only a easy sequence of arithmetic transformations ultimately,
designed for the categorical function of adjusting the vary of values
passing via.

Let’s kick the tires on it:

norm <- RMSNorm()
m <- matrix(c(0, 1,
              2, 3), nrow = 2)
norm(m)

tf.Tensor(
[[0.         1.4142132 ]
 [0.44721353 1.3416406 ]], form=(2, 2), dtype=float32)

tf.Tensor(
[[0.         1.4142137 ]
 [0.44721362 1.3416408 ]], form=(2, 2), dtype=float32)

tf.Tensor(
[[0.        1.4142137]
 [0.4472136 1.3416408]], form=(2, 2), dtype=float32)

FeedForward

Subsequent up is FeedForward()

FeedForward(keras$layers$Layer) %py_class% {

  initialize <- perform(hidden_dim, multiple_of = 256L,
                         ..., block_id = NULL) {
    tremendous$initialize()

    if(!is.null(multiple_of)) {
      hidden_dim <- hidden_dim %>%
        { as.integer( . * (2/3)) } %>%
        { (. + multiple_of - 1) %/% multiple_of } %>%
        { . * multiple_of }
    }

    self$hidden_dim <- hidden_dim
    self$block_id <- block_id
  }

  construct <- perform(input_shape) {
    output_dim <- input_shape |> as.integer() |> tail(1)

    if(is.null(self$block_id))
      load_weight <- (...) NULL
    else
      load_weight <- (identify) (...) np$load(weights_path(
        "7B/layers.{self$block_id}.feed_forward.{identify}.weight.npy"))$`T`

    self$w1 <- Dense(self$hidden_dim, use_bias = FALSE,
                     kernel_initializer = load_weight("w1"))
    self$w2 <- Dense(output_dim, use_bias = FALSE,
                     kernel_initializer = load_weight("w2"))
    self$w3 <- Dense(self$hidden_dim, use_bias = FALSE,
                     kernel_initializer = load_weight("w3"))

    tremendous$construct(input_shape)
  }

  name <- perform(x) {
    import_from({self}, w1, w2, w3)
    import_from(tf$nn, silu)

    x %>%
      { silu(w1(.)) * w3(.) } %>% # SwiGLU
      w2()
  }

}

FeedForward consists of three Dense layers. initialize does some
easy arithmetic, munging on the enter worth hidden_dim to make sure the
dimension is a performant a number of of 256, and construct is generally boiler plate
for creating the layers and loading the weights.

The novelty of FeedForward() is within the name() technique, the place relatively
than composing the Dense layers in a traditional sequential mannequin
with, say, ReLU activations in between and possibly some dropout, the
layers are composed to type a “SwiGLU” unit. The publication by Shazeer (2020)
of SwiGLU and different variations on GLU is an exemplar of the categories
of explorations and enhancements across the Transformer structure
since its preliminary publication in
2017; a gradual accretion of
enhancements that has introduced us to at the moment. The Feedforward$name() is
only a single SwiGLU adopted by a linear projection. In its essence,
it’s a intelligent composition of three (realized) linear projections, an
element-wise multiplication, and a silu()
activation
perform.

Maybe essentially the most shocking commentary to make right here is the relative
dearth of activation features, and even non-linearities, not simply in
FeedForward, however total. The silu() on this feedforward, the
reciprocal-root-mean-square in RMSnorm(), and a softmax() in
Consideration() are the one non-linear transformations in the entire
sequence of TransformerBlocks. Every part else is a linear
transformation!

Consideration

Lastly, let’s flip our consideration to Consideration().

Consideration(keras$layers$Layer) %py_class% {
  initialize <- perform(head_size, n_heads,
                         ..., block_id = NULL) {
    tremendous$initialize(...)

    self$head_size <- head_size
    self$n_heads <- n_heads

    if (is.null(block_id))
      load_weight <- perform(identify) NULL
    else
      load_weight <- (identify) (...) np$load(weights_path(
        "7B/layers.{block_id}.consideration.{identify}.weight.npy"))$`T`

    Dense <- perform(identify) keras$layers$Dense(
      models = n_heads * head_size,
      use_bias = FALSE,
      kernel_initializer = load_weight(identify)
    )

    self$wq <- Dense("wq")
    self$wk <- Dense("wk")
    self$wv <- Dense("wv")
    self$wo <- Dense("wo")
  }

  name <- perform(x) {
    c(batch_size, seqlen, n_features) %<-% tf$unstack(tf$form(x))

    # 1. undertaking (linear remodel) x into
    #    question, key, and worth tensors
    # 2. reshape q okay v, splitting out the final dim (n_features)
    #    into n_heads impartial subspaces,
    #    every with dimension head_size.
    #    (n_features == head_size * n_heads)
    split_heads_shape <- c(batch_size, seqlen,
                           self$n_heads, self$head_size)
    q <- x |> self$wq() |> tf$reshape(split_heads_shape)
    okay <- x |> self$wk() |> tf$reshape(split_heads_shape)
    v <- x |> self$wv() |> tf$reshape(split_heads_shape)

    # embed positional info in question and key
    # (bsz, seqlen, n_heads, head_size)
    q %<>% apply_rotary_embedding()
    okay %<>% apply_rotary_embedding()

    # reshape:
    #   transfer heads out of the final 2 axes,
    #   so later matmuls are carried out throughout the subspaces (heads)
    #   between (seqlen, head_size) axes
    v <- tf$transpose(v, c(0L, 2L, 1L, 3L)) # (bsz, n_heads, seqlen, head_size)
    q <- tf$transpose(q, c(0L, 2L, 1L, 3L)) # (bsz, n_heads, seqlen, head_size)
    okay <- tf$transpose(okay, c(0L, 2L, 3L, 1L)) # (bsz, n_heads, head_size, seqlen)

    # calculate and normalize consideration scores
    scores <- q %*% okay                       # (bsz, n_heads, seqlen, seqlen)
    scores <- scores / sqrt(self$head_size) # scale

    # apply causal masks, so the mannequin cannot "look forward" throughout coaching
    masks <- make_mask(seqlen, dtype = scores$dtype)
    scores %<>% { . + masks }

    scores <- tf$nn$softmax(scores, axis = -1L)

    # alter values tensor with consideration scores
                      # scores (bsz, n_heads, seqlen, seqlen)
                      # v      (bsz, n_heads, seqlen, head_size)
    output <- scores %*% v   # (bsz, n_heads, seqlen, head_size)

    # mix heads again right into a single options dim,
    # so Consideration output_shape==input_shape
    output <- output |>
      tf$transpose(c(0L, 2L, 1L, 3L)) |> # (bsz, seqlen, n_heads, head_size)
      tf$reshape(tf$form(x))            # (bsz, seqlen, n_heads * head_size)

    # yet another trainable linear projection for good luck
    output <- self$wo(output) # (bsz, seqlen, n_heads * head_size)

    output
  }
}

Consideration in LLaMA is comparable however not an identical to the Consideration
described within the authentic Transformers
paper (and accessible as a keras
builtin below keras$layers$MultiHeadAttention()). The core novelty is
the addition of the apply_rotary_embedding() perform, which we’ll
describe shortly. The extra novelty is balanced by the simplicity
from the truth that the layer is performing self-attention—we don’t want
to cross in several question, key, and worth tensors (or cause about what
meaning), for the reason that similar enter serves all three roles. Notice that the
standard MultiHeadAttention() layer is roofed fairly completely in
the 2nd Version of Deep Studying with R,
together with a full implementation of consideration in base R.

To develop an understanding of the mechanics in a layer like this, it’s
useful to quickly unsee a few of the minutia that may act as a fog
obscuring the essence of the operation. On this occasion, if we
quickly strip out the transpose()s and reshape()s (as intelligent and
very important as they’re), that is what’s left:

name <- perform(x) > self$wv()

  # rotate q,okay to inject place info.
  # cross q,okay to calculate an consideration rating for every token pair.
  scores <- rotate(q) %*% rotate(okay)   

Returning to the transpose()s and reshapes(), you may observe that
their function is to make it in order that the eye calculations are
carried out throughout n_heads impartial subspaces, relatively than in a
single bigger area. The identical reasoning drives this resolution as that
driving utilization of depthwise-separable convolutions in picture fashions.
Empirically, for the fastened compute funds, factoring options into
impartial subspaces performs higher than doing the identical core
operations in single bigger characteristic area. As with all issues, there may be
a steadiness to strike between n_heads (the variety of subspaces) and
head_dim (the scale of every subspace). The LLaMA authors have struck
the steadiness like this on the numerous mannequin sizes:

lapply(c("7B", "13B", "30B", "65B"), (dimension) {
  p <- read_json(weights_path("{dimension}/params.json"))
  with(p, listing(llama_size = dimension,
               n_heads = n_heads,
               head_dim = dim %/% n_heads))
}) |> dplyr::bind_rows()

# A tibble: 4 × 3
  llama_size n_heads head_dim
  <chr>        <int>    <int>
1 7B              32      128
2 13B             40      128
3 30B             52      128
4 65B             64      128

Subsequent lets flip our consideration to the causal consideration masks.

make_mask <- perform(seqlen, dtype = k_floatx()) {
  x <- tf$vary(seqlen)
  masks <- tf$the place(x[, tf$newaxis] < x[tf$newaxis, ],
                   tf$fixed(-Inf, dtype = dtype),
                   tf$fixed(0, dtype = dtype))

  # broadcast over batch and heads dim
  masks[tf$newaxis, tf$newaxis, , ] # (1, 1, seqlen, seqlen)
}

The masks is a strictly higher triangular matrix crammed with -Inf
values. Including the masks to the eye scores prevents the mannequin from
with the ability to “look forward” and see the eye rating for a token
pairing it hasn’t seen but at a selected place within the sequence.
This want for a masks is finest regarded as a vestige from coaching,
an equipment that the mannequin wanted to study with and now it might probably’t perform with out.
Throughout coaching, gradients are calculated for predictions from all
token positions in a sequence, together with predictions tokens the place the right
reply is proper there, because the very subsequent token in similar sequence. The masks
prevents the mannequin from with the ability to cheat and look forward into the longer term,
one thing it gained’t have the ability to do as soon as it’s we’re working it for inference.

tf.Tensor(
[[[[  0. -inf -inf -inf -inf]
   [  0.   0. -inf -inf -inf]
   [  0.   0.   0. -inf -inf]
   [  0.   0.   0.   0. -inf]
   [  0.   0.   0.   0.   0.]]]], form=(1, 1, 5, 5), dtype=float32)

Rotary Place Embedding

Subsequent lets flip our consideration to apply_rotary_embedding(). This core
innovation was printed by Su et al. (2022) within the paper titled
“RoFormer: Enhanced Transformer with Rotary Place Embedding”.

Some context:

• The naked Consideration() mechanism doesn’t go away any chance for a
  token’s place in a sequence to have an effect on the eye scores, since
  solely token-pairs are scored. Consideration treats its enter like a
  bag-of-tokens.

• The place of a token in a sequence is clearly essential, and the
  consideration layer ought to have entry to that info.

• Absolutely the place of a token in a sequence is much less essential
  than the relative place between tokens. (Particularly so for lengthy
  sequences).

Which leads us into the advanced airplane. If we think about the options as
advanced numbers, we are able to rotate them, and we are able to calculate angles between
them. From the Roformers paper:

Particularly, incorporating the relative place embedding is
easy: merely rotate the affine-transformed phrase embedding
vector by quantity of angle multiples of its place index and thus
interprets the instinct behind Rotary Place Embedding

Increasing barely: the rotation matrix is designed in order that
subsequently, after rotating our q and okay token sequence embedding
the identical approach, the angle between token options is a perform of the
relative distance between these tokens within the token sequence. The
relative angle between two tokens is invariant to absolutely the
place of these tokens within the full sequence.

In brief, the rotation injects positional info. The that means or
interpretability of that positional info, or how it’s meant to
be used, and even extracted from the results of q %*% okay, is left to the
mannequin to study.

Right here is the code:

apply_rotary_embedding <- perform(x) {
  c(., seqlen, ., head_size) %<-%
    tf$unstack(tf$form(x))

  rotation_matrix <- compute_rotation_matrix(seqlen, head_size)

  x %>%
    view_as_complex() %>%
    { . * rotation_matrix } %>%
    view_as_real()

}

compute_rotation_matrix <-
  perform(seqlen, feature_dim, theta = 10000) {
    # `feature_dim` right here goes to be consideration$head_size
    # `seqlen` goes to match the token sequence size.

    t <- tf$vary(seqlen, dtype = tf$float32)
    freqs <- tf$vary(begin = 0, restrict = 1, delta = 1 / (feature_dim %/% 2),
                      dtype = tf$float32)
    tf_assert(tf$dimension(freqs) == feature_dim %/% 2)
    freqs <- 1.0 / (theta ^ freqs)

    # outer product; (seqlen, head_size/2)
    freqs <- tf$einsum('a,b->ab', t, freqs)

    rot_mat <- tf$advanced(tf$cos(freqs), tf$sin(freqs))

    # the positional embedding shall be broadcast throughout batch and heads dim
    rot_mat[tf$newaxis, , tf$newaxis, ] #(1, seqlen, 1, headdim/2)
  }

view_as_complex <- perform(x) {
  tf$advanced(x[all_dims(), `::2`],
             x[all_dims(), `2::2`])
}

view_as_real <- perform(x) {
  # xs = (..., f);  xs2 = (..., f*2)
  xs <- tf$form(x)
  xs2 <- tf$concat(listing(xs[1:(length(xs)-1)],
                        xs[length(xs), drop = FALSE] * 2L),
                   axis = 0L)

  x2 <- tf$stack(listing(Re(x), Im(x)), axis = -1L)

  # (..., f, 2) -> (..., f*2)
  tf$reshape(x2, xs2)
}

As you may see, to think about the embedding options as present within the
advanced airplane, we merely deal with adjoining pairs of floats within the
underlying array as the true and imaginary a part of a posh quantity. We
rotate the embeddings within the advanced airplane, then return to imagining
the options as present in the true airplane. Once more, the job of
decoding the that means of the options after rotation is left to the
mannequin to study.

We will rapidly affirm that the rotary embeddings solely rotate options
and don’t scale them:

close to <- perform (x, y, tol = 1e-6) abs(x - y) < tol
all(close to(1, Mod(compute_rotation_matrix(2048L, 128L))))

tf.Tensor(True, form=(), dtype=bool)

There’s yet another trick to look at earlier than shifting on: due to a few of
the mathematical properties of the rotation matrix, it’s potential to
keep away from doing a full advanced multiply operation and nonetheless arrive on the
similar outcome. Additionally, for the reason that rotation matrix by no means modifications, it makes
sense to solely compute it as soon as and cache it, like so:

precomputed_rotation_matrix <- compute_rotation_matrix(
  seqlen = 2048L, # LLaMA max seqlen
  feature_dim = with(params, dim %/% n_heads)  # head_size
)

apply_rotary_embedding_faster <- perform(x) {

  rotate_every_two <- perform(x) {
    x1 <- x[all_dims(), `::2`]
    x2 <- x[all_dims(), `2::2`]
    x_ <- tf$stack(listing(-x2, x1), axis = -1L)
    tf$reshape(x_, tf$form(x))
  }

  repeat_each_twice <- perform(x) {
    tf$`repeat`(x, 2L, axis = -1L)
  }

  seqlen <- tf$form(x)[2]
  rot <- precomputed_rotation_matrix[, NA:seqlen, , ]

  cos <- Re(rot) |> repeat_each_twice()
  sin <- Im(rot) |> repeat_each_twice()

  (x * cos) + (rotate_every_two(x) * sin)
}

rand <- tf$random$uniform(form(3, 8, params$n_heads, 128))
all(apply_rotary_embedding(rand) ==
    apply_rotary_embedding_faster(rand))

tf.Tensor(True, form=(), dtype=bool)

apply_rotary_embedding <- apply_rotary_embedding_faster

Lastly, observe that the rotary positional embeddings are utilized inside
every Consideration layer. That is completely different from the unique Transformer
implementation, the place a positional embedding was solely added as soon as on the
head of the mannequin. Just like residual connections, you may consider the
presence of those repeated injections of positional info as
relieving the remaining trainable layers from the burden of allocating
a few of their weights to the duty of “passing via” or “preserving”
the positional info for later layers.

Positional embeddings are a wealthy topic that additionally comes up in different
deep studying architectures, like denoising diffusion (Falbel and Keydana 2023),
so time spent understanding them higher is time effectively
spent. For the needs of this weblog put up we’ve lined the factors
wanted and we’ll transfer on to tying all items collectively. To go deeper and
develop a extra mathematically knowledgeable perceive of RoPE, two wonderful
beginning factors are:

1. The unique paper by Su et al. (2022)

2. This weblog put up by
   Biderman et al. (2021)

Tying all of it collectively

With Tokenizer, Embedding, TransformerBlock (RMSNorm,
Consideration FeedForward and apply_rotary_embedding) all lined,
it’s time to tie all of the items collectively right into a Transformer mannequin. We
might do that utilizing %py_class% like with the opposite layers above, however
it’s simply as simple to maneuver over to utilizing the Keras practical API at this
level.

layer_transformer_block <- create_layer_wrapper(TransformerBlock)
layer_rms_norm <- create_layer_wrapper(RMSNorm)

# enter to the mannequin shall be output from the tokenizer
enter <- layer_input(form(NA)) #, dtype = "int32")

x <- enter |>
  tok_embeddings()  # instantiated earlier within the blog-post

for(block_id in seq_len0(params$n_layers)) >
    layer_transformer_block(attn_head_size = params$dim %/% params$n_heads,
                            attn_n_heads = params$n_heads,
                            norm_eps = params$norm_eps,
                            block_id = block_id)


# closing output projection into logits of output tokens
x <- x |>
  layer_rms_norm(block_id = -1, eps = params$norm_eps) |>
  layer_dense(
    tokenizer$vocab_size(), use_bias = FALSE,
    kernel_initializer = (...) np$load(weights_path("7B/output.weight.npy"))$`T`
  )

# slice out the logits for the final token
with_options(c(tensorflow.extract.warn_negatives_pythonic = FALSE), {
  output <- x[, -1, ]
})

llama <- keras_model(enter, output) %>%
  compile(jit_compile = TRUE)

The enter to the mannequin is tokenized textual content and the output is the
(unnormalized) possibilities for every token in tokenizer$vocab_size()
being the following token within the sequence.

next_token_probs <- immediate %>%
  tokenizer$tokenize() %>%
  llama()

next_token_probs

tf.Tensor(
[[-2.4503722e+00 -3.4463339e+00  1.3200411e+01 ...  4.8804146e-01
  -1.3277926e+00  9.9985600e-03]], form=(1, 32000), dtype=float32)

Sampling methods for choosing a token from the token logits is a
wealthy matter, (additionally lined completely within the Deep Studying with
R ebook), however this weblog put up is lengthy sufficient
already. So for now, let’s simply take the argmax().

sampler <- (logits) tf$argmax(logits, axis = -1L, output_type = "int32")

(next_token <- sampler(next_token_probs))

tf.Tensor([304], form=(1), dtype=int32)

tokenizer$detokenize(next_token) |> as.character()

[1] "to"

Let’s run it for just a few tokens and let LLaMa end the sentence:

prompt_tokens <- tokenizer$tokenize("One of the simplest ways to draw bees")

for (i in 1:20) {

  next_token_probs <- prompt_tokens |> llama()
  next_token <- sampler(next_token_probs)

  prompt_tokens %<>% { tf$concat(c(., next_token), axis = -1L) }

  # finish of sentence
  if (as.logical(next_token == tokenizer$string_to_id(".")))
    break
}

prompt_tokens |>
  tokenizer$detokenize() |>
  as.character() |>
  strwrap(60) |> writeLines()

One of the simplest ways to draw bees to your backyard is to plant a
number of flowers that bloom at completely different occasions.

Wrapping up

On this weblog put up we’ve walked via the LLaMA structure
applied in R TensorFlow, together with learn how to load pretrained weights,
after which run the mannequin to generate a sentence. Notice, a lot of the code in
this weblog put up is tailor-made for didactic functions. Whereas the
implementation of the LLaMA structure lined on this weblog put up is
applicable for coaching, there are just a few modifications you’ll wish to
make earlier than doing loads of textual content era. These embrace issues like:

• Within the Consideration layer, caching the okay and v tensors. Then,
  after the primary ahead cross with the preliminary immediate, solely feeding
  the mannequin the one new token from the sampler(), relatively than
  feeding the mannequin all of the tokens of the complete immediate on every ahead
  cross.

• Solely producing the causal masks make_mask() and rotary_matrix
  slices as soon as per ahead cross, as a substitute of inside every Consideration
  name.

• Updating the TransformerBlock to be cache-aware and to cross
  via the suitable arguments to Consideration()

• Wrapping all the extra book-keeping logic in a customized
  TransformerDecoder() class.

The modifications required to implement these optimizations for inference
balloon the code dimension and are largely about book-keeping, so we gained’t go
via them on this weblog put up. Nonetheless, yow will discover a fuller
implementation of LLaMA in R Tensorflow, together with a cache-aware
generate() technique that solely feeds the mannequin one token at a time throughout
the principle inference loop, (and compiles to XLA!),
right here.

That’s all for now. Thanks for studying and glad travels to all
exploring this thrilling LLM terrain!

Picture by Sébastien Goldberg on Unsplash