diff --git a/examples/galactica/generate_galactica_1.3b.py b/examples/galactica/generate_galactica_1.3b.py
new file mode 100644
index 00000000..fe404336
--- /dev/null
+++ b/examples/galactica/generate_galactica_1.3b.py
@@ -0,0 +1,27 @@
+from flagai.model.predictor.predictor import Predictor
+from flagai.auto_model.auto_loader import AutoLoader
+import torch
+device = torch.device("cuda:3" if torch.cuda.is_available() else "cpu")
+
+loader = AutoLoader(task_name="lm",
+                    model_name="galactica-1.3b-en",
+                    model_dir="/share/projset/baaishare/baai-mrnd/xingzhaohu/")
+
+model = loader.get_model()
+model.to(device)
+model.eval()
+
+tokenizer = loader.get_tokenizer()
+
+predictor = Predictor(model, tokenizer)
+
+text = "Please write a abstract about the computer vision. \n"
+out = predictor.predict_generate_randomsample(text,
+                                              out_max_length=700,
+                                              top_k=50,
+                                              repetition_penalty=1.2,
+                                              temperature=0.7
+                                              )
+print(out)
+
+
diff --git a/flagai/auto_model/auto_loader.py b/flagai/auto_model/auto_loader.py
index a7a480f5..7fea3564 100644
--- a/flagai/auto_model/auto_loader.py
+++ b/flagai/auto_model/auto_loader.py
@@ -55,6 +55,7 @@ def __getattr__(self, name):
     "glm_title-generation": ["flagai.model.glm_model", "GLMForSeq2Seq"],
     "opt_seq2seq": ("flagai.model.opt_model", "OPTModel"),
     "opt_lm": ("flagai.model.opt_model", "OPTModel"),
+    "galactica_lm": ("flagai.model.galactica_model", "GalacticaModel"),
     "vit_classification": ("flagai.model.vision.vit", "VisionTransformer"),
     "clip_txt_img_matching": ("flagai.model.mm.clip_model", "CLIP"),
     "swinv1_classification": ("flagai.model.vision.swinv1", "SwinTransformer"),
@@ -90,6 +91,10 @@ def __getattr__(self, name):
     "glm-10b-ch": ["flagai.model.glm_model", "GLMModel", "glm", "nlp"],
     "cpm3": ["flagai.model.cpm3_model", "CPM3", "cpm3", "nlp"],
     "cpm3-train": ["flagai.model.cpm3_train_model", "CPM3", "cpm3", "nlp"],
+    "galactica-1.3b-en": ["flagai.model.galactica_model", "GalacticaModel", "galactica", "nlp", "flagai.data.tokenizer.galactica.galactica_tokenizer", "GalacticaTokenizer"],
+    "galactica-6.7b-en": ["flagai.model.galactica_model", "GalacticaModel", "galactica", "nlp", "flagai.data.tokenizer.galactica.galactica_tokenizer", "GalacticaTokenizer"],
+    "galactica-30b-en": ["flagai.model.galactica_model", "GalacticaModel", "galactica", "nlp", "flagai.data.tokenizer.galactica.galactica_tokenizer", "GalacticaTokenizer"],
+    "galactica-120b-en": ["flagai.model.galactica_model", "GalacticaModel", "galactica", "nlp", "flagai.data.tokenizer.galactica.galactica_tokenizer", "GalacticaTokenizer"],
     "vit-base-p16-224":
         ["flagai.model.vision.vit", "VisionTransformer", "vit", "vision"],
     "vit-base-p16-384":
@@ -131,6 +136,7 @@ def __getattr__(self, name):
     "altclip-bert-b": ["flagai.models.mm.AltCLIP", "AltCLIP", "altclip", "mm", "flagai.model.mm.AltCLIP",
                        "AltCLIPProcessBert"],
     "eva-clip": ["flagai.model.mm.eva_clip_model", "EVA_CLIP", "evaclip", "mm"],
+
 }
 
 
@@ -205,10 +211,14 @@ def __init__(self,
             self.model.half()
 
         if model_type == "nlp":
-            tokenizer_class = getattr(LazyImport("flagai.data.tokenizer"),
-                                      "Tokenizer")
-            self.tokenizer = tokenizer_class.from_pretrained(
-                model_name, cache_dir=download_path)
+            if brief_model_name in ["galactica", ]:
+                self.tokenizer = getattr(LazyImport(MODEL_DICT[model_name][4]),
+                                                    MODEL_DICT[model_name][5])(download_path)
+            else :
+                tokenizer_class = getattr(LazyImport("flagai.data.tokenizer"),
+                                        "Tokenizer")
+                self.tokenizer = tokenizer_class.from_pretrained(
+                    model_name, cache_dir=download_path)
 
         elif model_type == "mm":
             if model_name.startswith("altdiffusion"):
diff --git a/flagai/data/tokenizer/galactica/galactica_tokenizer.py b/flagai/data/tokenizer/galactica/galactica_tokenizer.py
new file mode 100644
index 00000000..87a28412
--- /dev/null
+++ b/flagai/data/tokenizer/galactica/galactica_tokenizer.py
@@ -0,0 +1,84 @@
+from transformers import PreTrainedTokenizerFast
+
+
+from ..tokenizer import CommandToken, Tokenizer
+
+class GalacticaTokenizer(Tokenizer):
+    def __init__(self, download_dir) -> None:
+        pass
+        self.text_tokenizer = PreTrainedTokenizerFast.from_pretrained(download_dir)
+         # parse tokens and vocabs from tokenizer
+        self._tokens = list(self.text_tokenizer.get_vocab().keys())
+        self._vocab = {k: v for k, v in self.text_tokenizer.get_vocab().items()}
+        self.num_tokens = len(self._tokens)
+
+        self._command_tokens = [
+            CommandToken('pad', '[PAD]', self.get_specialid_from_text_tokenizer('pad')),
+            CommandToken('ENC', '[CLS]', self.get_specialid_from_text_tokenizer('cls')),
+            CommandToken('MASK', '[MASK]',
+                         self.get_specialid_from_text_tokenizer('mask')),
+            CommandToken('unk', '[UNK]', self.get_specialid_from_text_tokenizer('unk')),
+            CommandToken('sep', '[SEP]', self.get_specialid_from_text_tokenizer('sep')),
+            CommandToken('eos', '[PAD]', self.get_specialid_from_text_tokenizer('pad')),
+        ]
+        
+        self.command_name_map = {tok.name: tok for tok in self._command_tokens}
+        self.command_token_map = {
+            tok.token: tok
+            for tok in self._command_tokens
+        }
+        self.command_id_map = {tok.Id: tok for tok in self._command_tokens}
+
+    def get_specialid_from_text_tokenizer(self, token):
+        if token in ["eos", "sep"]:
+            return self._vocab.get('</s>')
+        elif token == "cls":
+            return self._vocab.get('<s>')
+        elif token == "unk":
+            return self._vocab.get('<unk>')
+        elif token == "pad":
+            return self._vocab.get('<pad>')
+        elif token == "mask":
+            return self._vocab.get('<mask>')
+        else:
+            raise NameError("token not exists")
+
+    def encode_plus(self, text, max_length=512):
+        return self.text_tokenizer.encode_plus(text, truncation=True, max_length=max_length)
+
+    def decode(self, ids):
+        return self.text_tokenizer.decode(ids)
+
+    def get_vocab(self):
+        return self.text_tokenizer.get_vocab()
+    
+    def get_command_id(self, name):
+        """get command token corresponding to `name`"""
+        return self.command_name_map[name].Id
+
+    def get_command(self, name):
+        """get command token corresponding to `name`"""
+        return self.command_name_map[name]
+
+    def encode_plus(self,
+                    text,
+                    second_text=None,
+                    truncation=False,
+                    max_length=None,):
+        
+        return self.text_tokenizer.encode_plus(text, 
+                                          text_pair=second_text, 
+                                          truncation=truncation, 
+                                          max_length=max_length,
+                                          add_special_tokens=True)
+
+    def tokenize(self, **kwargs):
+        return self.text_tokenizer.tokenize(**kwargs)
+    
+    def __len__(self):
+        return len(self.text_tokenizer)
+
+if __name__ == "__main__":
+    pass 
+        
+
diff --git a/flagai/model/galactica_model.py b/flagai/model/galactica_model.py
new file mode 100644
index 00000000..98e1f9cb
--- /dev/null
+++ b/flagai/model/galactica_model.py
@@ -0,0 +1,207 @@
+# coding=utf-8
+# Copyright 2022 The Fairseq Authors and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch Galactica model."""
+import random
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from flagai.model.layers.activations import ACT2FN
+from flagai.model.gpt2_model import GPT2Model, GPT2Stack, GPT2Config
+
+
+class OPTLearnedPositionalEmbedding(nn.Embedding):
+    """
+    This module learns positional embeddings up to a fixed maximum size.
+    """
+
+    def __init__(self, num_embeddings: int, embedding_dim: int):
+        # OPT is set up so that if padding_idx is specified then offset the embedding ids by 2
+        # and adjust num_embeddings appropriately. Other models don't have this hack
+        self.offset = 2
+        super().__init__(num_embeddings + self.offset, embedding_dim)
+
+    def forward(self, attention_mask: torch.LongTensor, past_key_values_length: int = 0):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        attention_mask = attention_mask.long()
+
+        # create positions depending on attention_mask
+        positions = (torch.cumsum(attention_mask, dim=1).type_as(attention_mask) * attention_mask).long() - 1
+        # cut positions if `past_key_values_length` is > 0
+        positions = positions[:, past_key_values_length:]
+
+        return super().forward(positions + self.offset)
+
+class OPTStack(GPT2Stack):
+    def __init__(self, config: GPT2Config):
+        super(OPTStack, self).__init__(config)
+        self.wpe = OPTLearnedPositionalEmbedding(config.n_positions, config.hidden_size)
+        self.ln_f = None
+        if config.do_layer_norm_before:
+            self.ln_f = nn.LayerNorm(config.hidden_size)
+
+        if config.n_embd != config.hidden_size:
+            self.project_out = nn.Linear(config.hidden_size, config.n_embd, bias=False)
+        else:
+            self.project_out = None
+
+        if config.n_embd != config.hidden_size:
+            self.project_in = nn.Linear(config.n_embd, config.hidden_size, bias=False)
+        else:
+            self.project_in = None
+
+    def get_position_embeddings(self, **kwargs):
+        pass
+        padding_mask = kwargs["padding_mask"]
+        past_length = kwargs["past_length"]
+        position_embeds = self.wpe(padding_mask, past_length)
+        return position_embeds
+
+def trans_opt_to_gpt_config(opt_config_json):
+    trans_config_json = {}
+    trans_key = {
+        "ffn_dim": "n_inner",
+        "hidden_size": "hidden_size",
+        "max_position_embeddings": "n_positions",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+        "vocab_size": "vocab_size",
+        "activation_function": "activation_function",
+        "checkpoint_activations": "checkpoint_activations",
+        "word_embed_proj_dim": "n_embd",
+        "do_layer_norm_before": "do_layer_norm_before",
+    }
+    for k, v in opt_config_json.items():
+        if k in trans_key:
+            trans_config_json[trans_key[k]] = v
+
+    return trans_config_json
+
+class GalacticaModel(GPT2Model):
+
+    def __init__(self, config, **kwargs):
+        config = trans_opt_to_gpt_config(config)
+        super(GalacticaModel, self).__init__(config, **kwargs)
+        self.transformer = OPTStack(self.config_gpt)
+
+    def load_weights(self, checkpoint_path):
+        checkpoint = torch.load(checkpoint_path,
+                                map_location=torch.device("cpu"))
+        if "module" in checkpoint:
+            # ddp
+            checkpoint = checkpoint["module"]
+
+        checkpoint_ = {}
+        for k, v in checkpoint.items():
+            if k[:6] == "model.":
+                checkpoint_[k[6:]] = v
+            else :
+                checkpoint_[k] = v
+
+        checkpoint = self.transpose_weight(checkpoint_)
+        self.load_state_dict(checkpoint, strict=False)
+        self.lm_head.weight.data = nn.Parameter(self.transformer.wte.weight.data)
+
+        return checkpoint
+
+    def transpose_weight(self, checkpoints):
+
+        checkponts_ = {
+            "transformer.wte.weight": checkpoints["decoder.embed_tokens.weight"],
+            "transformer.wpe.weight": checkpoints["decoder.embed_positions.weight"],
+                       }
+
+        if "decoder.project_in.weight" in checkpoints:
+            checkponts_["transformer.project_in.weight"] = checkpoints["decoder.project_in.weight"]
+            checkponts_["transformer.project_out.weight"] = checkpoints["decoder.project_out.weight"]
+
+        if "decoder.final_layer_norm.weight" in checkpoints:
+            checkponts_["transformer.ln_f.weight"] = checkpoints["decoder.final_layer_norm.weight"]
+            checkponts_["transformer.ln_f.bias"] = checkpoints["decoder.final_layer_norm.bias"]
+
+        q_weight = None
+        k_weight = None
+        v_weight = None
+        q_bias = None
+        k_bias = None
+        v_bias = None
+        for k, v in checkpoints.items():
+            # first ln
+            if "decoder.layers" in k and "self_attn_layer_norm" in k:
+                layer_id = k.split(".")[2]
+                weight_or_bias = k.split(".")[-1]
+                checkponts_[f"transformer.h.{layer_id}.ln_1.{weight_or_bias}"] = v
+                continue
+
+            # qkv
+            if "self_attn.k_proj.weight" in k:
+                k_weight = v
+                continue
+            if "self_attn.k_proj.bias" in k:
+                k_bias = v
+                continue
+
+            if "self_attn.v_proj.weight" in k:
+                v_weight = v
+                continue
+            if "self_attn.v_proj.bias" in k:
+                v_bias = v
+                continue
+
+            if "self_attn.q_proj.weight" in k:
+                q_weight = v
+                qkv_weight = torch.cat([q_weight, k_weight, v_weight], dim=0)
+                layer_id = k.split(".")[2]
+                checkponts_[f"transformer.h.{layer_id}.attn.c_attn.weight"] = qkv_weight
+                continue
+
+            if "self_attn.q_proj.bias" in k:
+                q_bias = v
+                qkv_bias = torch.cat([q_bias, k_bias, v_bias], dim=0)
+                layer_id = k.split(".")[2]
+                checkponts_[f"transformer.h.{layer_id}.attn.c_attn.bias"] = qkv_bias
+                continue
+
+            # att out
+            if "decoder.layers" in k and "self_attn.out_proj" in k:
+                layer_id = k.split(".")[2]
+                weight_or_bias = k.split(".")[-1]
+                checkponts_[f"transformer.h.{layer_id}.attn.c_proj.{weight_or_bias}"] = v
+                continue
+
+            # fc1
+            if "decoder.layers" in k and "fc1" in k:
+                layer_id = k.split(".")[2]
+                weight_or_bias = k.split(".")[-1]
+                checkponts_[f"transformer.h.{layer_id}.mlp.c_fc.{weight_or_bias}"] = v
+                continue
+
+            # fc2
+            if "decoder.layers" in k and "fc2" in k:
+                layer_id = k.split(".")[2]
+                weight_or_bias = k.split(".")[-1]
+                checkponts_[f"transformer.h.{layer_id}.mlp.c_proj.{weight_or_bias}"] = v
+                continue
+
+            # second ln
+            if "decoder.layers" in k and "final_layer_norm" in k:
+                layer_id = k.split(".")[2]
+                weight_or_bias = k.split(".")[-1]
+                checkponts_[f"transformer.h.{layer_id}.ln_2.{weight_or_bias}"] = v
+                continue
+
+        return checkponts_
\ No newline at end of file
diff --git a/flagai/model/predictor/predictor.py b/flagai/model/predictor/predictor.py
index 937e3bc5..03cf7ff4 100644
--- a/flagai/model/predictor/predictor.py
+++ b/flagai/model/predictor/predictor.py
@@ -322,7 +322,8 @@ def predict_generate_randomsample(self,
                                     device)
 
         elif "gpt" in self.class_name.lower(
-        ) or "opt" in self.class_name.lower():
+        ) or "opt" in self.class_name.lower() \
+            or "galactica" in self.class_name.lower():
             return gpt_random_sample_use_cache(self.model, self.tokenizer,
                                                text, input_max_length,
                                                out_max_length, top_k, top_p,