LLaMA implemented in R Tensorflow and Keras

LLaMA.R

## Setup
Sys.setenv(CUDA_VISIBLE_DEVICES='')
options(tensorflow.extract.warn_tensors_passed_asis = FALSE)

library(dplyr, warn.conflicts = FALSE)
library(purrr)
library(glue)
library(envir)

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

reticulate::use_virtualenv("./.venv", required = TRUE)

attach_eval({
  np <- reticulate::import("numpy", convert = FALSE)
  import_from(withr, with_options, local_options)
  import_from(keras$layers, Dense)
  import_from(tf$compiler$tf2xla$python$xla, dynamic_update_slice)

  nlist <- \(...) rlang::dots_list(..., .named = TRUE)
  seq_len0 <- \(x) seq.int(from = 0L, length.out = x)
})

precompute_rotarty_freqs <- function(seqlen, feature_dim, theta = 10000) {

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

  t <- tf$range(seqlen, dtype = tf$float32)
  freqs <- tf$range(start = 0, limit = 1,
                    delta = 1 / (feature_dim %/% 2),
                    dtype = tf$float32)
  tf_assert(tf$size(freqs) == feature_dim %/% 2)
  freqs <- 1 / (theta ^ freqs)

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

  # prep to recycle across head_size axis and
  # broadcast across batch_size and n_heads axes
  list(cos = tf$cos(freqs),
       sin = tf$sin(freqs)) |>
    lapply(repeat_each_twice) |>
    lapply(\(m) m[tf$newaxis, , tf$newaxis, ]) # (1, seqlen, 1, head_size)
}

apply_rotary_embedding <- function(x, freqs) {

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

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

make_mask <- function(seqlen, position_index = 0L, dtype = k_floatx()) {
  x <- tf$range(seqlen)
  i <- x[, tf$newaxis] + position_index
  j <- x[tf$newaxis, ]

  mask <- tf$where(i < j,
                   tf$constant(-Inf, dtype = dtype),
                   tf$constant(0, dtype = dtype))

  mask[tf$newaxis, tf$newaxis, , ] # (1, 1, seqlen_q, seqlen_q)
}


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

  initialize <-
    function(eps = 1e-6, ..., block_id = NULL, feeds_into = NULL) {
      super$initialize(...)
      self$eps <- eps
      self$block_id <- block_id
      self$feeds_into <- feeds_into
    }

  build <- function(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, length(input_shape))
    w_shape[length(input_shape)] <- as.integer(input_shape) |> tail(1L)

    # helper that will load
    # the pretrained-weights if we supplied `block_id` and `feeds_into`
    import_from({self}, block_id, feeds_into)
    initializer <- if (is.null(self$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 final output norm, which is not part of a TransformerBlock
      \(...) weights_path("7B/norm.weight.npy") |>
             np$load() |> np$expand_dims(0:1)

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

  rrms <- function(x) {
    # reciprocal root mean square along the last axis
    x %>%
      tf$math$square() %>%
      tf$reduce_mean(axis = -1L, keepdims = TRUE) %>%
      tf$math$add(self$eps) %>% # for numerical stability
      tf$math$rsqrt()
  }

  call <- function(x) {
    x * self$rrms(x) * self$w
  }
}

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

  initialize <- function(hidden_dim, multiple_of = 256L, ..., block_id = NULL) {
    super$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
  }

  build <- function(input_shape) {
    output_dim <- input_shape |> as.integer() |> tail(1)

    load_weight <- NULL
    if(!is.null(self$block_id))
      load_weight <- \(name) \(...) np$load(weights_path(
        "7B/layers.{self$block_id}.feed_forward.{name}.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"))

    super$build(input_shape)
  }

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

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

}


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

  initialize <- function(head_size, n_heads, ..., block_id = NULL) {
    super$initialize(...)

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

    if (is.null(block_id))
      load_weight <- function(name) NULL
    else
      load_weight <- \(name) \(...) np$load(weights_path(
        "7B/layers.{block_id}.attention.{name}.weight.npy"))$`T`

    Dense <- function(name) keras$layers$Dense(
      units = n_heads * head_size,
      use_bias = FALSE,
      kernel_initializer = load_weight(name)
    )

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

  call <- function(x, ...,
                   freqs = NULL,
                   cache = NULL,
                   cache_index = NULL,
                   mask = NULL) {

    c(batch_size, seqlen_q, n_features) %<-% tf$unstack(tf$shape(x))
    seqlen_k <- seqlen_v <- cache_index + seqlen_q

    split_heads_shape <- c(batch_size, seqlen_q, self$n_heads, self$head_size)
    q <- x |> self$wq() |> tf$reshape(split_heads_shape)
    k <- x |> self$wk() |> tf$reshape(split_heads_shape)
    v <- x |> self$wv() |> tf$reshape(split_heads_shape)

    q %<>% apply_rotary_embedding(freqs) # (bsz, seqlen_q, n_heads, head_size)
    k %<>% apply_rotary_embedding(freqs) # (bsz, seqlen_q, n_heads, head_size)

    if(!is.null(cache)) {
      # append k,v to respective caches; fetch full k,v from cache
      cache$k %<>% dynamic_update_slice(k, c(0L, cache_index, 0L, 0L))
      cache$v %<>% dynamic_update_slice(v, c(0L, cache_index, 0L, 0L))
      k <- cache$k[, NA:seqlen_k, , ]
      v <- cache$v[, NA:seqlen_v, , ]
    }

    v <- tf$transpose(v, c(0L, 2L, 1L, 3L)) # (bsz, n_heads, seqlen_v, head_size)
    q <- tf$transpose(q, c(0L, 2L, 1L, 3L)) # (bsz, n_heads, seqlen_q, head_size)
    k <- tf$transpose(k, c(0L, 2L, 3L, 1L)) # (bsz, n_heads, head_size, seqlen_k)

    scores <- (q %*% k) / sqrt(self$head_size) # (bsz, n_heads, seqlen_q, seqlen_k)

    # apply causal mask, so the model can't "look ahead" during training
    if (!is.null(mask))
      scores %<>% { . + mask }

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

    # adjust values tensor with attention scores
    # scores (bsz, n_heads, seqlen_q, seqlen_k)
    # v      (bsz, n_heads, seqlen_v, head_size)
    output <- scores %*% v   # (bsz, n_heads, seqlen_q, head_size)

    # combine heads back into a single features dim,
    # so Attention output_shape==input_shape
    # (needed so that you can add residuals in TransformerBlock)
    output <- output |>
      tf$transpose(c(0L, 2L, 1L, 3L)) |> # (bsz, seqlen_q, n_heads, head_size)
      tf$reshape(c(batch_size, seqlen_q,  # (bsz, seqlen_q, n_heads * head_size)
                   self$n_heads * self$head_size))

    # one more trainable linear projection for good luck
    output <- self$wo(output) # (bsz, seqlen_q, n_heads * head_size)

    if(is.null(cache))
      output
    else
      list(output, cache)
  }
}

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

  initialize <- function(attn_head_size, attn_n_heads,
                         norm_eps = k_epsilon(), ...,
                         block_id = NULL) {
    super$initialize(...)

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

  call <- function(x, ..., cache = NULL) {

    # norm and attention
    x2 <- x |>
      self$attention_norm() |>
      self$attention(..., cache = cache)

    # maybe unpack cache returned by Attention
    if(!is.null(cache))
      c(x2, cache) %<-% x2

    x <- x + x2 # add residual

    # norm and swiglu projection
    x2 <- x %>%
      self$feed_forward_norm() %>%
      self$feed_forward()

    x <- x + x2 # residual again

    if(is.null(cache)) x else list(x, cache)
  }
}


TransformerDecoder(keras$Model) %py_class% {

  initialize <- function(vocab_size, n_blocks, n_heads, head_size, norm_eps) {
    super$initialize()
    self$head_size <- head_size
    self$n_heads <- n_heads

    self$tok_embeddings <- keras$layers$Embedding(
      input_dim = vocab_size,
      output_dim = n_heads*head_size,
      embeddings_initializer =
        \(...) np$load(weights_path("7B/tok_embeddings.weight.npy")))

    self$blocks <- lapply(seq_len0(n_blocks), function(block_id) {
      TransformerBlock(attn_head_size = head_size,
                       attn_n_heads = n_heads,
                       norm_eps = norm_eps,
                       block_id = block_id)
    })
    self$norm <- RMSNorm(block_id = -1, eps = norm_eps)
    self$output_proj <- Dense(
      vocab_size, use_bias = FALSE,
      kernel_initializer = \(...)
          np$load(weights_path("7B/output.weight.npy"))$`T`)

    self$freqs <- precompute_rotarty_freqs(feature_dim = head_size,
                                           seqlen = 2048L)
  }

  call <- function(tokens) {
    c(bsz, seqlen) %<-% tf$unstack(tf$shape(tokens))

    freqs <- self$freqs |> lapply(\(f) f[, NA:seqlen, , ])
    mask <- make_mask(seqlen)

    x <- tokens |>
      self$tok_embeddings()

    for (block in self$blocks)
      x <- block(x, freqs = freqs, mask = mask)

    local_options(c(tensorflow.extract.warn_negatives_pythonic = FALSE))

    x |>
      self$norm() |>
      _[, -1, ] |>
      self$output_proj()

  }

  call_with_cache <- function(tokens, cache, position) {

    c(batch_size, seqlen) %<-% tf$unstack(tf$shape(tokens))

    # Sanity check: after the initial seeding of cache with the prompt, we
    # should only be running inference on one token at a time.
    tf_assert(position == 0 | seqlen == 1)

    if(is.numeric(position) && position == 0L) {
      # initial cache seeding
      mask <- make_mask(seqlen)
      freqs <- self$freqs |> lapply(\(f) f[, NA:seqlen, , ])
    } else {
      # inference with one token
      position %<>% as_tensor(dtype = "int32")
      freqs <- self$freqs |> lapply(\(f) f[, position, , ])
      mask <- NULL
    }

    blocks <- self$blocks
    stopifnot(is.list(cache), length(cache) == length(blocks))

    x <- tokens |>
      self$tok_embeddings()

    for (i in seq_along(blocks)) {
      c(x, cache[[i]]) %<-% blocks[[i]](x, cache = cache[[i]],
                                        cache_index = position,
                                        freqs = freqs,
                                        mask = mask)
    }

    local_options(c(tensorflow.extract.warn_negatives_pythonic = FALSE))

    output <- x |>
      self$norm() |>
      _[,-1,] |>
      self$output_proj()

    list(output, cache)
  }

  .make_cache <- function(prompt_tokens, max_seqlen = 2048L) {
    c(batch_size, seqlen) %<-% tf$unstack(tf$shape(prompt_tokens))
    import_from({self}, head_size, n_heads)

    max_seqlen <- min(max_seqlen + seqlen, 2048L)
    cache_shape <- c(batch_size, max_seqlen, n_heads, head_size)
    cache <- lapply(seq_along(self$blocks), \(.) {
      list(k = tf$zeros(cache_shape), v = tf$zeros(cache_shape))
    })

    tokens_with_preallocated_space <-
      tf$zeros(c(batch_size, max_seqlen), dtype = "int32") |>
      dynamic_update_slice(update = prompt_tokens, indices = c(0L, 0L))

    # run first forward pass to seed cache with initial prompt
    # return (propmt_tokens, next_token_probs, cache)
    c(tokens_with_preallocated_space,
      self$call_with_cache(prompt_tokens, cache = cache, position = 0L))
  }

  private$sampler_fn <- \(logits) logits |>
      tf$argmax(axis = -1L, output_type = "int32") |>
      tf$expand_dims(-1L)

  sampler %<-active% function(fn) {
    if(missing(fn))
      private$sampler_fn
    else
      private$sampler_fn <- fn
  }

  generate <- function(prompt, max_len = 20L) {
    max_len %<>% as_tensor("int32")
    prompt %<>% as_tensor()

    # accept either tokens or a string
    if (prompt$dtype$name == "string") {

      if(length(dim(prompt)) == 0) # ensure a batch dim
        prompt %<>% .[tf$newaxis]
      tokens <- tokenizer$tokenize(prompt)$to_tensor()

    } else {

      tokens <- prompt
      if(length(dim(prompt)) == 1) # ensure a batch dim
        tokens %<>% .[tf$newaxis, ]

    }

    c(batch_size, initial_prompt_len) %<-% tf$unstack(tf$shape(tokens))
    max_seqlen <- min(max_len + initial_prompt_len, 2048L)

    c(tokens, next_token_probs, cache) %<-% self$.make_cache(tokens, max_len)
    i <- initial_prompt_len
    autograph({
      # enable `if` and `for` to accept tensors
      for (i in tf$range(initial_prompt_len, max_seqlen, dtype = "int32")) {
        next_token <- self$sampler(next_token_probs)
        tokens %<>% dynamic_update_slice(next_token, c(0L, i))
        if (any(next_token == 2L))
          break # end-of-sequence token
        c(next_token_probs, cache) %<-%
          self$call_with_cache(next_token, cache, i)
      }
    })

    tokens %<>% .[, NA:(i+1)] # drop unused preallocated space

    if(prompt$dtype$name == "string")
      # return string if supplied a string
      tokenizer$detokenize(tokens)
    else
      tokens
  }
}



# ---- load

weights_path <- function(rel_path) {
  normalizePath(
    file.path(
      "~/github/facebookresearch/llama/weights/LLaMA/",
      glue::glue(rel_path, .envir = parent.frame())
    ),
    mustWork = TRUE
  )
}

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

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


llama <- TransformerDecoder(vocab_size = tokenizer$vocab_size(),
                            n_blocks = params$n_layers,
                            n_heads = params$n_heads,
                            head_size = params$dim %/% params$n_heads,
                            norm_eps = params$norm_eps)

prompt <- "The best way to attract bees"

test_generate <- function() {
  prompt |>
    tokenizer$tokenize() |>
    llama$generate(as_tensor(17L)) |>
    tokenizer$detokenize() |>
    as.character() |>
    strwrap(60) |> writeLines()
}


test_generate()
## expected output with the argmax() sampler:
#    The best way to attract bees to your garden is to plant a
#    variety of flowers that bloom at different times.

# Timings on CPU:
print(system.time(test_generate()))
#   user  system elapsed
# 99.562   0.149  89.057

# Compile to XLA
llama$generate %<>% tf_function(jit_compile = TRUE)

# First call includes tracing time
print(system.time(test_generate()))
#   user  system elapsed
# 64.944   0.809  55.314

# Second call is pure graph mode
print(system.time(generate()))
#    user  system elapsed
# 28.754   0.120  18.453

For details and explanations, see the associated blog post: https://blogs.rstudio.com/ai/posts/2023-05-25-llama-tensorflow-keras/