SillyTilly/Meta-Llama-3.1-405B-Instruct

1diff --git a/src/transformers/models/llama/convert_llama_weights_to_hf.py b/src/transformers/models/llama/convert_llama_weights_to_hf.py2index a0fbe4680..8b0ce2b13 1006443--- a/src/transformers/models/llama/convert_llama_weights_to_hf.py4+++ b/src/transformers/models/llama/convert_llama_weights_to_hf.py5@@ -17,10 +17,10 @@ import json6 import os7 import shutil8 import warnings9-10+from typing import List11 import torch12 13-from transformers import LlamaConfig, LlamaForCausalLM, LlamaTokenizer, PreTrainedTokenizerFast14+from transformers import LlamaConfig, LlamaForCausalLM, LlamaTokenizer, PreTrainedTokenizerFast, GenerationConfig15 from transformers.convert_slow_tokenizer import TikTokenConverter16 17 18@@ -85,8 +85,12 @@ NUM_SHARDS = {19     "65B": 8,20     "70B": 8,21     "70Bf": 8,22+    "405B": 8,23+    "405B-MP16": 16,24 }25 26+CONTEXT_LENGTH_FOR_VERSION = {"3.1": 131072, "3": 8192, "2": 4096, "1": 2048}27+28 29 def compute_intermediate_size(n, ffn_dim_multiplier=1, multiple_of=256):30     return multiple_of * ((int(ffn_dim_multiplier * int(8 * n / 3)) + multiple_of - 1) // multiple_of)31@@ -107,9 +111,10 @@ def write_model(32     input_base_path,33     model_size=None,34     safe_serialization=True,35-    llama_version=1,36+    llama_version="1",37     vocab_size=None,38     num_shards=None,39+    instruct=False,40 ):41     os.makedirs(model_path, exist_ok=True)42     tmp_model_path = os.path.join(model_path, "tmp")43@@ -125,18 +130,11 @@ def write_model(44     dims_per_head = dim // n_heads45     base = params.get("rope_theta", 10000.0)46     inv_freq = 1.0 / (base ** (torch.arange(0, dims_per_head, 2).float() / dims_per_head))47-    if base > 10000.0 and llama_version != 3:48+    if base > 10000.0 and float(llama_version) < 3:49         max_position_embeddings = 1638450     else:51-        # Depending on the Llama version, the default max_position_embeddings has different values.52-        if llama_version == 1:53-            max_position_embeddings = 204854-        elif llama_version == 2:55-            max_position_embeddings = 409656-        elif llama_version == 3:57-            max_position_embeddings = 819258-59-    vocab_size = vocab_size if vocab_size is not None else 3200060+        max_position_embeddings = CONTEXT_LENGTH_FOR_VERSION[llama_version]61+62     if params.get("n_kv_heads", None) is not None:63         num_key_value_heads = params["n_kv_heads"]  # for GQA / MQA64         num_key_value_heads_per_shard = num_key_value_heads // num_shards65@@ -144,8 +142,7 @@ def write_model(66     else:  # compatibility with other checkpoints67         num_key_value_heads = n_heads68         num_key_value_heads_per_shard = n_heads_per_shard69-        key_value_dim = dims_per_head * num_key_value_heads70-    print(num_shards, num_key_value_heads, num_key_value_heads_per_shard, key_value_dim)71+        key_value_dim = dim72 73     # permute for sliced rotary74     def permute(w, n_heads, dim1=dim, dim2=dim):75@@ -159,11 +156,9 @@ def write_model(76         loaded = torch.load(os.path.join(input_base_path, "consolidated.00.pth"), map_location="cpu")77     else:78         # Sharded79-        loaded = [80-            torch.load(os.path.join(input_base_path, file), map_location="cpu")81-            for file in os.listdir(input_base_path)82-            if file.endswith(".pth")83-        ]84+        checkpoint_list = sorted([file for file in os.listdir(input_base_path) if file.endswith(".pth")])85+        print("Loading in order:", checkpoint_list)86+        loaded = [torch.load(os.path.join(input_base_path, file), map_location="cpu") for file in checkpoint_list]87     param_count = 088     index_dict = {"weight_map": {}}89     for layer_i in range(n_layers):90@@ -263,7 +258,7 @@ def write_model(91             "lm_head.weight": loaded["output.weight"],92         }93     else:94-        concat_dim = 0 if llama_version == 3 else 195+        concat_dim = 0 if llama_version in ['3', '3.1'] else 196         state_dict = {97             "model.norm.weight": loaded[0]["norm.weight"],98             "model.embed_tokens.weight": torch.cat(99@@ -282,6 +277,18 @@ def write_model(100     write_json(index_dict, os.path.join(tmp_model_path, "pytorch_model.bin.index.json"))101     ffn_dim_multiplier = params["ffn_dim_multiplier"] if "ffn_dim_multiplier" in params else 1102     multiple_of = params["multiple_of"] if "multiple_of" in params else 256103+104+    if llama_version in ['3', '3.1']:105+        bos_token_id = 128000106+107+        if instruct:108+            eos_token_id = [128001, 128008, 128009]109+        else:110+            eos_token_id = 128001111+    else:112+        bos_token_id = 1113+        eos_token_id = 2114+115     config = LlamaConfig(116         hidden_size=dim,117         intermediate_size=compute_intermediate_size(dim, ffn_dim_multiplier, multiple_of),118@@ -292,11 +299,21 @@ def write_model(119         vocab_size=vocab_size,120         rope_theta=base,121         max_position_embeddings=max_position_embeddings,122-        bos_token_id=128000 if llama_version == 3 else 1,123-        eos_token_id=128001 if llama_version == 3 else 2,124+        bos_token_id=bos_token_id,125+        eos_token_id=eos_token_id,126     )127     config.save_pretrained(tmp_model_path)128 129+    if instruct:130+        generation_config = GenerationConfig(131+            do_sample=True,132+            temperature=0.6,133+            top_p=0.9,134+            bos_token_id=bos_token_id,135+            eos_token_id=eos_token_id,136+        )137+        generation_config.save_pretrained(tmp_model_path)138+139     # Make space so we can load the model properly now.140     del state_dict141     del loaded142@@ -313,7 +330,7 @@ def write_model(143 144 145 class Llama3Converter(TikTokenConverter):146-    def __init__(self, vocab_file, num_reserved_special_tokens=256, **kwargs):147+    def __init__(self, vocab_file, special_tokens=None, instruct=False, model_max_length=None, **kwargs):148         super().__init__(vocab_file, **kwargs)149         tokenizer = self.converted()150         chat_template = (151@@ -327,34 +344,27 @@ class Llama3Converter(TikTokenConverter):152             "{% endfor %}"153             "{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}"154         )155-        num_reserved_special_tokens = 256156-        special_tokens = [157-            "<|begin_of_text|>",158-            "<|end_of_text|>",159-            "<|reserved_special_token_0|>",160-            "<|reserved_special_token_1|>",161-            "<|reserved_special_token_2|>",162-            "<|reserved_special_token_3|>",163-            "<|start_header_id|>",164-            "<|end_header_id|>",165-            "<|reserved_special_token_4|>",166-            "<|eot_id|>",  # end of turn167-        ] + [f"<|reserved_special_token_{i}|>" for i in range(5, num_reserved_special_tokens - 5)]168         tokenizer.add_special_tokens(special_tokens)169 170         self.tokenizer = PreTrainedTokenizerFast(171             tokenizer_object=tokenizer,172             bos_token="<|begin_of_text|>",173-            eos_token="<|end_of_text|>",174-            chat_template=chat_template,175+            eos_token="<|end_of_text|>" if not instruct else "<|eot_id|>",176+            chat_template=chat_template if instruct else None,177             model_input_names=["input_ids", "attention_mask"],178+            model_max_length=model_max_length,179         )180 181 182-def write_tokenizer(tokenizer_path, input_tokenizer_path, llama_version=2):183+def write_tokenizer(tokenizer_path, input_tokenizer_path, llama_version="2", special_tokens=None, instruct=False):184     tokenizer_class = LlamaTokenizer if LlamaTokenizerFast is None else LlamaTokenizerFast185-    if llama_version == 3:186-        tokenizer = Llama3Converter(input_tokenizer_path).tokenizer187+    if llama_version in ["3", "3.1"]:188+        tokenizer = Llama3Converter(189+            input_tokenizer_path,190+            special_tokens,191+            instruct,192+            model_max_length=CONTEXT_LENGTH_FOR_VERSION[llama_version]193+        ).tokenizer194     else:195         tokenizer = tokenizer_class(input_tokenizer_path)196     print(f"Saving a {tokenizer_class.__name__} to {tokenizer_path}.")197@@ -362,6 +372,37 @@ def write_tokenizer(tokenizer_path, input_tokenizer_path, llama_version=2):198     return tokenizer199 200 201+DEFAULT_LLAMA_SPECIAL_TOKENS = {202+    "3": [203+        "<|begin_of_text|>",204+        "<|end_of_text|>",205+        "<|reserved_special_token_0|>",206+        "<|reserved_special_token_1|>",207+        "<|reserved_special_token_2|>",208+        "<|reserved_special_token_3|>",209+        "<|start_header_id|>",210+        "<|end_header_id|>",211+        "<|reserved_special_token_4|>",212+        "<|eot_id|>",  # end of turn213+    ]214+    + [f"<|reserved_special_token_{i}|>" for i in range(5, 256 - 5)],215+    "3.1": [216+        "<|begin_of_text|>",217+        "<|end_of_text|>",218+        "<|reserved_special_token_0|>",219+        "<|reserved_special_token_1|>",220+        "<|finetune_right_pad_id|>",221+        "<|reserved_special_token_2|>",222+        "<|start_header_id|>",223+        "<|end_header_id|>",224+        "<|eom_id|>",  # end of message225+        "<|eot_id|>",  # end of turn226+        "<|python_tag|>",227+    ]228+    + [f"<|reserved_special_token_{i}|>" for i in range(3, 256 - 8)],229+}230+231+232 def main():233     parser = argparse.ArgumentParser()234     parser.add_argument(235@@ -383,9 +424,9 @@ def main():236     # Different Llama versions used different default values for max_position_embeddings, hence the need to be able to specify which version is being used.237     parser.add_argument(238         "--llama_version",239-        choices=[1, 2, 3],240-        default=1,241-        type=int,242+        choices=["1", "2", "3", "3.1"],243+        default="1",244+        type=str,245         help="Version of the Llama model to convert. Currently supports Llama1 and Llama2. Controls the context size",246     )247     parser.add_argument(248@@ -394,11 +435,34 @@ def main():249         type=int,250         help="The number of individual shards used for the model. Does not have to be the same as the number of consolidated_xx.pth",251     )252+    parser.add_argument(253+        "--special_tokens",254+        default=None,255+        type=List[str],256+        help="The list of special tokens that should be added to the model.",257+    )258+    parser.add_argument(259+        "--instruct",260+        default=False,261+        type=bool,262+        help="Whether the model is an instruct model or not. Will affect special tokens for llama 3.1.",263+    )264     args = parser.parse_args()265     if args.model_size is None and args.num_shards is None:266         raise ValueError("You have to set at least `num_shards` if you are not giving the `model_size`")267+    if args.special_tokens is None:268+        args.special_tokens = DEFAULT_LLAMA_SPECIAL_TOKENS[str(args.llama_version)]269+270     spm_path = os.path.join(args.input_dir, "tokenizer.model")271-    vocab_size = len(write_tokenizer(args.output_dir, spm_path, llama_version=args.llama_version))272+    vocab_size = len(273+        write_tokenizer(274+            args.output_dir,275+            spm_path,276+            llama_version=args.llama_version,277+            special_tokens=args.special_tokens,278+            instruct=args.instruct279+        )280+    )281     if args.model_size != "tokenizer_only":282         write_model(283             model_path=args.output_dir,284@@ -408,6 +472,7 @@ def main():285             llama_version=args.llama_version,286             vocab_size=vocab_size,287             num_shards=args.num_shards,288+            instruct=args.instruct289         )290 291 292