Llama 3.1 conversion

Co-authored-by: Arthur Zucker <arthur.zucker@gmail.com>

src/transformers/modeling_rope_utils.py

@@ -129,6 +129,7 @@ def _compute_dynamic_ntk_parameters(
         Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
         post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
     """
+    # TODO (joao): use the new `original_max_position_embeddings` from rope_scaling
     if config is not None and len(rope_kwargs) > 0:
         raise ValueError(
             "Unexpected arguments: `**rope_kwargs` and `config` are mutually exclusive in "
@@ -249,6 +250,7 @@ def _compute_longrope_parameters(
         Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
         post-processing scaling factor applied to the computed cos/sin.
     """
+    # TODO (joao): use the new `original_max_position_embeddings` from rope_scaling
     # No need to keep BC with longrope, unreleased when this new pattern was created.
     if len(rope_kwargs) > 0:
         raise ValueError(
@@ -293,6 +295,50 @@ def _compute_longrope_parameters(
     return inv_freq, attention_factor
 
 
+def _compute_llama3_parameters(
+    config: PretrainedConfig, device: "torch.device", seq_len: Optional[int] = None, **rope_kwargs
+) -> Tuple["torch.Tensor", float]:
+    """
+    Computes the inverse frequencies for llama 3.1.
+
+    Args:
+        config ([`~transformers.PretrainedConfig`]):
+            The model configuration.
+        device (`torch.device`):
+            The device to use for initialization of the inverse frequencies.
+        seq_len (`int`, *optional*):
+            The current sequence length. Unused for this type of RoPE.
+        rope_kwargs (`Dict`, *optional*):
+            BC compatibility with the previous RoPE class instantiation, will be removed in v4.45.
+    Returns:
+        Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
+        post-processing scaling factor applied to the computed cos/sin.
+    """
+    # Gets the default RoPE parameters
+    inv_freq, attention_factor = _compute_default_rope_parameters(config, device, seq_len, **rope_kwargs)
+
+    factor = config.rope_scaling["factor"]  # `8` in the original implementation
+    low_freq_factor = config.rope_scaling["low_freq_factor"]  # `1` in the original implementation
+    high_freq_factor = config.rope_scaling["high_freq_factor"]  # `4` in the original implementation
+    old_context_len = config.rope_scaling["original_max_position_embeddings"]  # `8192` in the original implementation
+
+    low_freq_wavelen = old_context_len / low_freq_factor
+    high_freq_wavelen = old_context_len / high_freq_factor
+    new_freqs = []
+    for freq in inv_freq:
+        wavelen = 2 * math.pi / freq
+        if wavelen < high_freq_wavelen:
+            new_freqs.append(freq)
+        elif wavelen > low_freq_wavelen:
+            new_freqs.append(freq / factor)
+        else:
+            assert low_freq_wavelen != high_freq_wavelen
+            smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
+            new_freqs.append((1 - smooth) * freq / factor + smooth * freq)
+    inv_freq = torch.tensor(new_freqs, dtype=inv_freq.dtype, device=inv_freq.device)
+    return inv_freq, attention_factor
+
+
 # This maps the "rope_type" string field in rope config to the corresponding function to compute the RoPE parameters
 # from the model config. You can append new {'rope_type': callable} pairs to this dictionary to enable custom RoPE
 # parameterizations, as long as the callable has the same signature.
@@ -302,6 +348,7 @@ ROPE_INIT_FUNCTIONS = {
     "dynamic": _compute_dynamic_ntk_parameters,
     "yarn": _compute_yarn_parameters,
     "longrope": _compute_longrope_parameters,
+    "llama3": _compute_llama3_parameters,
 }
 
 
@@ -339,6 +386,20 @@ def _validate_linear_scaling_rope_parameters(config: PretrainedConfig):
         raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
 
 
+def _validate_dynamic_scaling_rope_parameters(config: PretrainedConfig):
+    rope_scaling = config.rope_scaling
+    rope_type = rope_scaling["rope_type"]
+    required_keys = {"rope_type", "factor"}
+    # TODO (joao): update logic for the inclusion of `original_max_position_embeddings`
+    optional_keys = {"original_max_position_embeddings"}
+    received_keys = set(rope_scaling.keys())
+    _check_received_keys(rope_type, received_keys, required_keys, optional_keys)
+
+    factor = rope_scaling["factor"]
+    if factor is None or not isinstance(factor, float) or factor < 1.0:
+        raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
+
+
 def _validate_yarn_parameters(config: PretrainedConfig):
     rope_scaling = config.rope_scaling
     rope_type = rope_scaling["rope_type"]
@@ -374,7 +435,8 @@ def _validate_longrope_parameters(config: PretrainedConfig):
     rope_scaling = config.rope_scaling
     rope_type = rope_scaling["rope_type"]
     required_keys = {"rope_type", "short_factor", "long_factor"}
-    optional_keys = {"attention_factor", "factor"}
+    # TODO (joao): update logic for the inclusion of `original_max_position_embeddings`
+    optional_keys = {"attention_factor", "factor", "original_max_position_embeddings"}
     received_keys = set(rope_scaling.keys())
     _check_received_keys(rope_type, received_keys, required_keys, optional_keys)
 
@@ -417,13 +479,50 @@ def _validate_longrope_parameters(config: PretrainedConfig):
             )
 
 
+def _validate_llama3_parameters(config: PretrainedConfig):
+    rope_scaling = config.rope_scaling
+    rope_type = rope_scaling["rope_type"]
+    required_keys = {"rope_type", "factor", "original_max_position_embeddings", "low_freq_factor", "high_freq_factor"}
+    received_keys = set(rope_scaling.keys())
+    _check_received_keys(rope_type, received_keys, required_keys)
+
+    factor = rope_scaling["factor"]
+    if factor is None or not isinstance(factor, float) or factor < 1.0:
+        raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
+
+    low_freq_factor = rope_scaling["low_freq_factor"]
+    high_freq_factor = rope_scaling["high_freq_factor"]
+    if low_freq_factor is None or not isinstance(low_freq_factor, float):
+        raise ValueError(f"`rope_scaling`'s low_freq_factor field must be a float, got {low_freq_factor}")
+    if high_freq_factor is None or not isinstance(high_freq_factor, float):
+        raise ValueError(f"`rope_scaling`'s high_freq_factor field must be a float, got {high_freq_factor}")
+    if high_freq_factor < low_freq_factor:
+        raise ValueError(
+            "`rope_scaling`'s high_freq_factor field must be greater than low_freq_factor, got high_freq_factor="
+            f"{high_freq_factor} and low_freq_factor={low_freq_factor}"
+        )
+
+    original_max_position_embeddings = rope_scaling["original_max_position_embeddings"]
+    if original_max_position_embeddings is None or not isinstance(original_max_position_embeddings, int):
+        raise ValueError(
+            "`rope_scaling`'s original_max_position_embeddings field must be an integer, got "
+            f"{original_max_position_embeddings}"
+        )
+    if original_max_position_embeddings >= config.max_position_embeddings:
+        raise ValueError(
+            "`rope_scaling`'s original_max_position_embeddings field must be less than max_position_embeddings, got "
+            f"{original_max_position_embeddings} and max_position_embeddings={config.max_position_embeddings}"
+        )
+
+
 # Like `ROPE_INIT_FUNCTIONS`, this validation function mapping can be dynamically updated for custom RoPE types.
 ROPE_VALIDATION_FUNCTIONS = {
     "default": _validate_default_rope_parameters,
     "linear": _validate_linear_scaling_rope_parameters,
-    "dynamic": _validate_linear_scaling_rope_parameters,  # `dynamic` has the same validation pattern as `linear`
+    "dynamic": _validate_dynamic_scaling_rope_parameters,
     "yarn": _validate_yarn_parameters,
     "longrope": _validate_longrope_parameters,
+    "llama3": _validate_llama3_parameters,
 }
 
 

src/transformers/models/llama/configuration_llama.py

@@ -73,25 +73,28 @@ class LlamaConfig(PretrainedConfig):
             End of stream token id.
         pretraining_tp (`int`, *optional*, defaults to 1):
             Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
-            document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to understand more about it. This value is
-            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
-            issue](https://github.com/pytorch/pytorch/issues/76232).
+            document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to
+            understand more about it. This value is necessary to ensure exact reproducibility of the pretraining
+            results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232).
         tie_word_embeddings (`bool`, *optional*, defaults to `False`):
             Whether to tie weight embeddings
         rope_theta (`float`, *optional*, defaults to 10000.0):
             The base period of the RoPE embeddings.
         rope_scaling (`Dict`, *optional*):
-            Dictionary containing the scaling configuration for the RoPE embeddings. IMPORTANT: RoPE scaling expects
-            `max_position_embeddings` to remain unchanged -- some methods, like 'longrope', require the original value
-            to determine which scaling to apply.
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
             Expected contents:
                 `rope_type` (`str`):
-                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope'],
-                    with 'default' being the original RoPE implementation.
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
                 `factor` (`float`, *optional*):
                     Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
                     most scaling types, a `factor` of x will enable the model to handle sequences of length x *
-                    `max_position_embeddings`.
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
                 `attention_factor` (`float`, *optional*):
                     Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
                     computation. If unspecified, it defaults to value recommended by the implementation, using the
@@ -104,12 +107,16 @@ class LlamaConfig(PretrainedConfig):
                     ramp function. If unspecified, it defaults to 1.
                 `short_factor` (`List[float]`, *optional*):
                     Only used with 'longrope'. The scaling factor to be applied to short contexts (<
-                    `max_position_embeddings` * `factor`). Must be a list of numbers with the same length as the hidden
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                     size divided by the number of attention heads divided by 2
                 `long_factor` (`List[float]`, *optional*):
-                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
-                    `max_position_embeddings` * `factor`). Must be a list of numbers with the same length as the hidden
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
                     size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
         attention_bias (`bool`, *optional*, defaults to `False`):
             Whether to use a bias in the query, key, value and output projection layers during self-attention.
         attention_dropout (`float`, *optional*, defaults to 0.0):

src/transformers/models/llama/convert_llama_weights_to_hf.py

@@ -17,10 +17,11 @@ import json
 import os
 import shutil
 import warnings
+from typing import List
 
 import torch
 
-from transformers import LlamaConfig, LlamaForCausalLM, LlamaTokenizer, PreTrainedTokenizerFast
+from transformers import GenerationConfig, LlamaConfig, LlamaForCausalLM, LlamaTokenizer, PreTrainedTokenizerFast
 from transformers.convert_slow_tokenizer import TikTokenConverter
 
 
@@ -85,8 +86,12 @@ NUM_SHARDS = {
     "65B": 8,
     "70B": 8,
     "70Bf": 8,
+    "405B": 8,
+    "405B-MP16": 16,
 }
 
+CONTEXT_LENGTH_FOR_VERSION = {"3.1": 131072, "3": 8192, "2": 4096, "1": 2048}
+
 
 def compute_intermediate_size(n, ffn_dim_multiplier=1, multiple_of=256):
     return multiple_of * ((int(ffn_dim_multiplier * int(8 * n / 3)) + multiple_of - 1) // multiple_of)
@@ -107,9 +112,10 @@ def write_model(
     input_base_path,
     model_size=None,
     safe_serialization=True,
-    llama_version=1,
+    llama_version="1",
     vocab_size=None,
     num_shards=None,
+    instruct=False,
 ):
     os.makedirs(model_path, exist_ok=True)
     tmp_model_path = os.path.join(model_path, "tmp")
@@ -125,18 +131,11 @@ def write_model(
     dims_per_head = dim // n_heads
     base = params.get("rope_theta", 10000.0)
     inv_freq = 1.0 / (base ** (torch.arange(0, dims_per_head, 2).float() / dims_per_head))
-    if base > 10000.0 and llama_version != 3:
+    if base > 10000.0 and float(llama_version) < 3:
         max_position_embeddings = 16384
     else:
-        # Depending on the Llama version, the default max_position_embeddings has different values.
-        if llama_version == 1:
-            max_position_embeddings = 2048
-        elif llama_version == 2:
-            max_position_embeddings = 4096
-        elif llama_version == 3:
-            max_position_embeddings = 8192
-
-    vocab_size = vocab_size if vocab_size is not None else 32000
+        max_position_embeddings = CONTEXT_LENGTH_FOR_VERSION[llama_version]
+
     if params.get("n_kv_heads", None) is not None:
         num_key_value_heads = params["n_kv_heads"]  # for GQA / MQA
         num_key_value_heads_per_shard = num_key_value_heads // num_shards
@@ -144,8 +143,7 @@ def write_model(
     else:  # compatibility with other checkpoints
         num_key_value_heads = n_heads
         num_key_value_heads_per_shard = n_heads_per_shard
-        key_value_dim = dims_per_head * num_key_value_heads
-    print(num_shards, num_key_value_heads, num_key_value_heads_per_shard, key_value_dim)
+        key_value_dim = dim
 
     # permute for sliced rotary
     def permute(w, n_heads, dim1=dim, dim2=dim):
@@ -159,11 +157,9 @@ def write_model(
         loaded = torch.load(os.path.join(input_base_path, "consolidated.00.pth"), map_location="cpu")
     else:
         # Sharded
-        loaded = [
-            torch.load(os.path.join(input_base_path, file), map_location="cpu")
-            for file in sorted(os.listdir(input_base_path))
-            if file.endswith(".pth")
-        ]
+        checkpoint_list = sorted([file for file in os.listdir(input_base_path) if file.endswith(".pth")])
+        print("Loading in order:", checkpoint_list)
+        loaded = [torch.load(os.path.join(input_base_path, file), map_location="cpu") for file in checkpoint_list]
     param_count = 0
     index_dict = {"weight_map": {}}
     for layer_i in range(n_layers):
@@ -263,7 +259,7 @@ def write_model(
             "lm_head.weight": loaded["output.weight"],
         }
     else:
-        concat_dim = 0 if llama_version == 3 else 1
+        concat_dim = 0 if llama_version in ["3", "3.1"] else 1
         state_dict = {
             "model.norm.weight": loaded[0]["norm.weight"],
             "model.embed_tokens.weight": torch.cat(
@@ -282,6 +278,18 @@ def write_model(
     write_json(index_dict, os.path.join(tmp_model_path, "pytorch_model.bin.index.json"))
     ffn_dim_multiplier = params["ffn_dim_multiplier"] if "ffn_dim_multiplier" in params else 1
     multiple_of = params["multiple_of"] if "multiple_of" in params else 256
+
+    if llama_version in ["3", "3.1"]:
+        bos_token_id = 128000
+
+        if instruct:
+            eos_token_id = [128001, 128008, 128009]
+        else:
+            eos_token_id = 128001
+    else:
+        bos_token_id = 1
+        eos_token_id = 2
+
     config = LlamaConfig(
         hidden_size=dim,
         intermediate_size=compute_intermediate_size(dim, ffn_dim_multiplier, multiple_of),
@@ -292,11 +300,21 @@ def write_model(
         vocab_size=vocab_size,
         rope_theta=base,
         max_position_embeddings=max_position_embeddings,
-        bos_token_id=128000 if llama_version == 3 else 1,
-        eos_token_id=128001 if llama_version == 3 else 2,
+        bos_token_id=bos_token_id,
+        eos_token_id=eos_token_id,
     )
     config.save_pretrained(tmp_model_path)
 
+    if instruct:
+        generation_config = GenerationConfig(
+            do_sample=True,
+            temperature=0.6,
+            top_p=0.9,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+        )
+        generation_config.save_pretrained(tmp_model_path)
+
     # Make space so we can load the model properly now.
     del state_dict
     del loaded
@@ -313,7 +331,7 @@ def write_model(
 
 
 class Llama3Converter(TikTokenConverter):
-    def __init__(self, vocab_file, num_reserved_special_tokens=256, **kwargs):
+    def __init__(self, vocab_file, special_tokens=None, instruct=False, model_max_length=None, **kwargs):
         super().__init__(vocab_file, **kwargs)
         tokenizer = self.converted()
         chat_template = (
@@ -327,34 +345,24 @@ class Llama3Converter(TikTokenConverter):
             "{% endfor %}"
             "{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}"
         )
-        num_reserved_special_tokens = 256
-        special_tokens = [
-            "<|begin_of_text|>",
-            "<|end_of_text|>",
-            "<|reserved_special_token_0|>",
-            "<|reserved_special_token_1|>",
-            "<|reserved_special_token_2|>",
-            "<|reserved_special_token_3|>",
-            "<|start_header_id|>",
-            "<|end_header_id|>",
-            "<|reserved_special_token_4|>",
-            "<|eot_id|>",  # end of turn
-        ] + [f"<|reserved_special_token_{i}|>" for i in range(5, num_reserved_special_tokens - 5)]
         tokenizer.add_special_tokens(special_tokens)
 
         self.tokenizer = PreTrainedTokenizerFast(
             tokenizer_object=tokenizer,
             bos_token="<|begin_of_text|>",
-            eos_token="<|end_of_text|>",
-            chat_template=chat_template,
+            eos_token="<|end_of_text|>" if not instruct else "<|eot_id|>",
+            chat_template=chat_template if instruct else None,
             model_input_names=["input_ids", "attention_mask"],
+            model_max_length=model_max_length,
         )
 
 
-def write_tokenizer(tokenizer_path, input_tokenizer_path, llama_version=2):
+def write_tokenizer(tokenizer_path, input_tokenizer_path, llama_version="2", special_tokens=None, instruct=False):
     tokenizer_class = LlamaTokenizer if LlamaTokenizerFast is None else LlamaTokenizerFast
-    if llama_version == 3:
-        tokenizer = Llama3Converter(input_tokenizer_path).tokenizer
+    if llama_version in ["3", "3.1"]:
+        tokenizer = Llama3Converter(
+            input_tokenizer_path, special_tokens, instruct, model_max_length=CONTEXT_LENGTH_FOR_VERSION[llama_version]
+        ).tokenizer
     else:
         tokenizer = tokenizer_class(input_tokenizer_path)
     print(f"Saving a {tokenizer_class.__name__} to {tokenizer_path}.")
@@ -362,6 +370,37 @@ def write_tokenizer(tokenizer_path, input_tokenizer_path, llama_version=2):
     return tokenizer
 
 
+DEFAULT_LLAMA_SPECIAL_TOKENS = {
+    "3": [
+        "<|begin_of_text|>",
+        "<|end_of_text|>",
+        "<|reserved_special_token_0|>",
+        "<|reserved_special_token_1|>",
+        "<|reserved_special_token_2|>",
+        "<|reserved_special_token_3|>",
+        "<|start_header_id|>",
+        "<|end_header_id|>",
+        "<|reserved_special_token_4|>",
+        "<|eot_id|>",  # end of turn
+    ]
+    + [f"<|reserved_special_token_{i}|>" for i in range(5, 256 - 5)],
+    "3.1": [
+        "<|begin_of_text|>",
+        "<|end_of_text|>",
+        "<|reserved_special_token_0|>",
+        "<|reserved_special_token_1|>",
+        "<|finetune_right_pad_id|>",
+        "<|reserved_special_token_2|>",
+        "<|start_header_id|>",
+        "<|end_header_id|>",
+        "<|eom_id|>",  # end of message
+        "<|eot_id|>",  # end of turn
+        "<|python_tag|>",
+    ]
+    + [f"<|reserved_special_token_{i}|>" for i in range(3, 256 - 8)],
+}
+
+
 def main():
     parser = argparse.ArgumentParser()
     parser.add_argument(
@@ -383,9 +422,9 @@ def main():
     # Different Llama versions used different default values for max_position_embeddings, hence the need to be able to specify which version is being used.
     parser.add_argument(
         "--llama_version",
-        choices=[1, 2, 3],
-        default=1,
-        type=int,
+        choices=["1", "2", "3", "3.1"],
+        default="1",
+        type=str,
         help="Version of the Llama model to convert. Currently supports Llama1 and Llama2. Controls the context size",
     )
     parser.add_argument(
@@ -394,11 +433,34 @@ def main():
         type=int,
         help="The number of individual shards used for the model. Does not have to be the same as the number of consolidated_xx.pth",
     )
+    parser.add_argument(
+        "--special_tokens",
+        default=None,
+        type=List[str],
+        help="The list of special tokens that should be added to the model.",
+    )
+    parser.add_argument(
+        "--instruct",
+        default=False,
+        type=bool,
+        help="Whether the model is an instruct model or not. Will affect special tokens for llama 3.1.",
+    )
     args = parser.parse_args()
     if args.model_size is None and args.num_shards is None:
         raise ValueError("You have to set at least `num_shards` if you are not giving the `model_size`")
+    if args.special_tokens is None:
+        args.special_tokens = DEFAULT_LLAMA_SPECIAL_TOKENS[str(args.llama_version)]
+
     spm_path = os.path.join(args.input_dir, "tokenizer.model")
-    vocab_size = len(write_tokenizer(args.output_dir, spm_path, llama_version=args.llama_version))
+    vocab_size = len(
+        write_tokenizer(
+            args.output_dir,
+            spm_path,
+            llama_version=args.llama_version,
+            special_tokens=args.special_tokens,
+            instruct=args.instruct,
+        )
+    )
     if args.model_size != "tokenizer_only":
         write_model(
             model_path=args.output_dir,
@@ -408,6 +470,7 @@ def main():
             llama_version=args.llama_version,
             vocab_size=vocab_size,
             num_shards=args.num_shards,
+            instruct=args.instruct,
         )