Support llama3 checkpoints with tied word embeddings

elixir-nx · Sep 20, 2024 · b01e0da · b01e0da
1 parent ed2f177
commit b01e0da
Show file tree

Hide file tree

Showing 2 changed files with 16 additions and 11 deletions.
diff --git a/lib/bumblebee/text/llama.ex b/lib/bumblebee/text/llama.ex
@@ -75,6 +75,10 @@ defmodule Bumblebee.Text.Llama do
         default: 0.02,
         doc:
           "the standard deviation of the normal initializer used for initializing kernel parameters"
+      ],
+      tie_word_embeddings: [
+        default: false,
+        doc: "whether to tie input and output embedding weights"
       ]
     ] ++
       Shared.common_options([:num_labels, :id_to_label]) ++ Shared.token_options(pad_token_id: 0)
@@ -435,6 +439,7 @@ defmodule Bumblebee.Text.Llama do
       opts =
         convert!(data,
           vocab_size: {"vocab_size", number()},
+          tie_word_embeddings: {"tie_word_embeddings", boolean()},
           max_positions: {"max_position_embeddings", number()},
           hidden_size: {"hidden_size", number()},
           num_blocks: {"num_hidden_layers", number()},
@@ -447,15 +452,16 @@ defmodule Bumblebee.Text.Llama do
           rotary_embedding_scaling_strategy:
             {"rope_scaling", optional(scaling_strategy_converter)},
           initializer_scale: {"initializer_range", number()},
-          layer_norm_epsilon: {"rms_norm_eps", number()}
+          layer_norm_epsilon: {"rms_norm_eps", number()},
+          tie_word_embeddings: {"tie_word_embeddings", boolean()}
         ) ++ Shared.common_options_from_transformers(data, spec)
 
       @for.config(spec, opts)
     end
   end
 
   defimpl Bumblebee.HuggingFace.Transformers.Model do
-    def params_mapping(_spec) do
+    def params_mapping(spec) do
       %{
         "embedder.token_embedding" => "model.embed_tokens",
         "decoder.blocks.{n}.self_attention.query" => "model.layers.{n}.self_attn.q_proj",
@@ -470,7 +476,8 @@ defmodule Bumblebee.Text.Llama do
         "decoder.blocks.{n}.ffn.output" => "model.layers.{n}.mlp.down_proj",
         "decoder.blocks.{n}.output_norm" => "model.layers.{n}.post_attention_layernorm",
         "output_norm" => "model.norm",
-        "language_modeling_head.output" => "lm_head",
+        "language_modeling_head.output" =>
+          if(spec.tie_word_embeddings, do: "model.embed_tokens", else: "lm_head"),
         "sequence_classification_head.output" => "score"
       }
     end

diff --git a/lib/bumblebee/text/t5.ex b/lib/bumblebee/text/t5.ex
@@ -10,12 +10,6 @@ defmodule Bumblebee.Text.T5 do
         tokens that can be represented in model input and output
         """
       ],
-      tie_word_embeddings: [
-        default: true,
-        doc: """
-        whether or not to tie encoder and decoder token embedding
-        """
-      ],
       hidden_size: [
         default: 512,
         doc: "the dimensionality of hidden layers"
@@ -74,6 +68,10 @@ defmodule Bumblebee.Text.T5 do
       layer_norm_epsilon: [
         default: 1.0e-6,
         doc: "the epsilon used by the layer normalization layers"
+      ],
+      tie_word_embeddings: [
+        default: true,
+        doc: "whether or not to tie encoder and decoder token embedding"
       ]
     ] ++
       Shared.common_options([:num_labels, :id_to_label]) ++
@@ -538,7 +536,6 @@ defmodule Bumblebee.Text.T5 do
       opts =
         convert!(data,
           vocab_size: {"vocab_size", number()},
-          tie_word_embeddings: {"tie_word_embeddings", boolean()},
           hidden_size: {"d_model", number()},
           attention_head_size: {"d_kv", number()},
           encoder_num_blocks: {"num_layers", number()},
@@ -551,7 +548,8 @@ defmodule Bumblebee.Text.T5 do
           activation: {"feed_forward_proj", t5_activation()},
           ffn_gated_activation: {"feed_forward_proj", ffn_gated_activation()},
           dropout_rate: {"dropout", number()},
-          initializer_scale: {"initializer_factor", number()}
+          initializer_scale: {"initializer_factor", number()},
+          tie_word_embeddings: {"tie_word_embeddings", boolean()}
         ) ++ Shared.common_options_from_transformers(data, spec)
 
       @for.config(spec, opts)