Update cognitivess_model/convert_Cognitivess_weights_to_hf.py

cognitivess_model/convert_Cognitivess_weights_to_hf.py

@@ -1,4 +1,4 @@
-# Copyright 2023 Cognitivess AI and The HuggingFace Inc. team. All rights reserved.
+# Copyright 2022 Cognitivess 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.
@@ -19,25 +19,19 @@ import shutil
 import warnings
 
 import torch
-from safetensors.torch import load_file as safe_load_file
 
-from transformers import (
-    LlamaTokenizer,
-    CognitivessConfig,
-    CognitivessForCausalLM,
-)
+from transformers import CognitivessConfig, CognitivessForCausalLM, CognitivessTokenizer, PreTrainedTokenizerFast
+from transformers.convert_slow_tokenizer import TikTokenConverter
 
 
 try:
-    from transformers import LlamaTokenizerFast
-
-    tokenizer_class = LlamaTokenizerFast
+    from transformers import CognitivessTokenizerFast
 except ImportError as e:
     warnings.warn(e)
     warnings.warn(
         "The converted tokenizer will be the `slow` tokenizer. To use the fast, update your `tokenizers` library and re-run the tokenizer conversion"
     )
-    tokenizer_class = LlamaTokenizer
+    CognitivessTokenizerFast = None
 
 """
 Sample usage:
@@ -50,17 +44,48 @@ python src/transformers/models/Cognitivess/convert_Cognitivess_weights_to_hf.py
 Thereafter, models can be loaded via:
 
 ```py
-from transformers import CognitivessForCausalLM, LlamaTokenizer
+from transformers import CognitivessForCausalLM, CognitivessTokenizer
 
 model = CognitivessForCausalLM.from_pretrained("/output/path")
-tokenizer = LlamaTokenizer.from_pretrained("/output/path")
+tokenizer = CognitivessTokenizer.from_pretrained("/output/path")
 ```
 
 Important note: you need to be able to host the whole model in RAM to execute this script (even if the biggest versions
 come in several checkpoints they each contain a part of each weight of the model, so we need to load them all in RAM).
+
+If you want you tokenizer to add a bos automatically you should update the tokenizer._tokenizers.post_processor:
+
+```py
+from tokenizers import processors
+bos = "<|begin_of_text|>"
+tokenizer._tokenizers.post_processor = processors.Sequence(
+    [
+        processors.ByteLevel(trim_offsets=False),
+        processors.TemplateProcessing(
+            single=f"{bos}:0 $A:0",
+            pair=f"{bos}:0 $A:0 {bos}:1 $B:1",
+            special_tokens=[
+                (bos, tokenizer.encode(bos)),
+            ],
+        ),
+    ]
+)
+```
 """
 
-NUM_SHARDS = {"7B": 1}
+NUM_SHARDS = {
+    "7B": 1,
+    "8B": 1,
+    "8Bf": 1,
+    "7Bf": 1,
+    "13B": 2,
+    "13Bf": 2,
+    "34B": 4,
+    "30B": 4,
+    "65B": 8,
+    "70B": 8,
+    "70Bf": 8,
+}
 
 
 def compute_intermediate_size(n, ffn_dim_multiplier=1, multiple_of=256):
@@ -77,24 +102,22 @@ def write_json(text, path):
         json.dump(text, f)
 
 
-def write_model(model_path, input_base_path, model_size, tokenizer_path=None, safe_serialization=True, is_v3=False):
-    # for backward compatibility, before you needed the repo to be called `my_repo/model_size`
-    if not os.path.isfile(os.path.join(input_base_path, "params.json")):
-        input_base_path = os.path.join(input_base_path, model_size)
-
+def write_model(
+    model_path,
+    input_base_path,
+    model_size=None,
+    safe_serialization=True,
+    Cognitivess_version=1,
+    vocab_size=None,
+    num_shards=None,
+):
     os.makedirs(model_path, exist_ok=True)
     tmp_model_path = os.path.join(model_path, "tmp")
     os.makedirs(tmp_model_path, exist_ok=True)
 
     params = read_json(os.path.join(input_base_path, "params.json"))
-    num_shards = NUM_SHARDS[model_size]
-
-    sliding_window = params.get("sliding_window", None)
-
-    # For some reason this is a string in the params.json
-    if sliding_window is not None:
-        sliding_window = int(sliding_window)
-
+    num_shards = NUM_SHARDS[model_size] if num_shards is None else num_shards
+    params = params.get("model", params)
     n_layers = params["n_layers"]
     n_heads = params["n_heads"]
     n_heads_per_shard = n_heads // num_shards
@@ -102,97 +125,128 @@ def write_model(model_path, input_base_path, model_size, tokenizer_path=None, sa
     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))
-    max_position_embeddings = 4096 * 8
-
-    if tokenizer_path is not None:
-        tokenizer = tokenizer_class(tokenizer_path + ".v3" if is_v3 else "")
-        tokenizer.save_pretrained(model_path)
-    vocab_size = tokenizer.vocab_size if tokenizer_path is not None else 32000
-
-    if "n_kv_heads" in params:
+    if base > 10000.0 and Cognitivess_version != 3:
+        max_position_embeddings = 16384
+    else:
+        # Depending on the Cognitivess version, the default max_position_embeddings has different values.
+        if Cognitivess_version == 1:
+            max_position_embeddings = 2048
+        elif Cognitivess_version == 2:
+            max_position_embeddings = 4096
+        elif Cognitivess_version == 3:
+            max_position_embeddings = 8192
+
+    vocab_size = vocab_size if vocab_size is not None else 32000
+    if params.get("n_kv_heads", None) is not None:
         num_key_value_heads = params["n_kv_heads"]  # for GQA / MQA
-        num_local_key_value_heads = num_key_value_heads // num_shards
-        key_value_dim = dims_per_head * num_local_key_value_heads
+        num_key_value_heads_per_shard = num_key_value_heads // num_shards
+        key_value_dim = dims_per_head * num_key_value_heads
     else:  # compatibility with other checkpoints
         num_key_value_heads = n_heads
-        num_local_key_value_heads = n_heads_per_shard
-        key_value_dim = dim
+        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)
 
     # permute for sliced rotary
-    def permute(w, n_heads=n_heads, dim1=dim, dim2=dim):
+    def permute(w, n_heads, dim1=dim, dim2=dim):
         return w.view(n_heads, dim1 // n_heads // 2, 2, dim2).transpose(1, 2).reshape(dim1, dim2)
 
     print(f"Fetching all parameters from the checkpoint at {input_base_path}.")
-
-    # Load weights - for v3 models the consolidated weights are in a single file format in safetensors
-    if is_v3:
-        loaded = [safe_load_file(os.path.join(input_base_path, "consolidated.safetensors"))]
+    # Load weights
+    if num_shards == 1:
+        # Not sharded
+        # (The sharded implementation would also work, but this is simpler.)
+        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, f"consolidated.{i:02d}.pth"), map_location="cpu")
-            for i in range(num_shards)
+            torch.load(os.path.join(input_base_path, file), map_location="cpu")
+            for file in os.listdir(input_base_path)
+            if file.endswith(".pth")
         ]
     param_count = 0
     index_dict = {"weight_map": {}}
     for layer_i in range(n_layers):
         filename = f"pytorch_model-{layer_i + 1}-of-{n_layers + 1}.bin"
-
-        # Sharded
-        # Note that attention.w{q,k,v,o}, feed_fordward.w[1,2,3], attention_norm.weight and ffn_norm.weight share
-        # the same storage object, saving attention_norm and ffn_norm will save other weights too, which is
-        # redundant as other weights will be stitched from multiple shards. To avoid that, they are cloned.
-
-        state_dict = {
-            f"model.layers.{layer_i}.input_layernorm.weight": loaded[0][
-                f"layers.{layer_i}.attention_norm.weight"
-            ].clone(),
-            f"model.layers.{layer_i}.post_attention_layernorm.weight": loaded[0][
-                f"layers.{layer_i}.ffn_norm.weight"
-            ].clone(),
-        }
-        state_dict[f"model.layers.{layer_i}.self_attn.q_proj.weight"] = permute(
-            torch.cat(
+        if num_shards == 1:
+            # Unsharded
+            state_dict = {
+                f"model.layers.{layer_i}.self_attn.q_proj.weight": permute(
+                    loaded[f"layers.{layer_i}.attention.wq.weight"], n_heads=n_heads
+                ),
+                f"model.layers.{layer_i}.self_attn.k_proj.weight": permute(
+                    loaded[f"layers.{layer_i}.attention.wk.weight"],
+                    n_heads=num_key_value_heads,
+                    dim1=key_value_dim,
+                ),
+                f"model.layers.{layer_i}.self_attn.v_proj.weight": loaded[f"layers.{layer_i}.attention.wv.weight"],
+                f"model.layers.{layer_i}.self_attn.o_proj.weight": loaded[f"layers.{layer_i}.attention.wo.weight"],
+                f"model.layers.{layer_i}.mlp.gate_proj.weight": loaded[f"layers.{layer_i}.feed_forward.w1.weight"],
+                f"model.layers.{layer_i}.mlp.down_proj.weight": loaded[f"layers.{layer_i}.feed_forward.w2.weight"],
+                f"model.layers.{layer_i}.mlp.up_proj.weight": loaded[f"layers.{layer_i}.feed_forward.w3.weight"],
+                f"model.layers.{layer_i}.input_layernorm.weight": loaded[f"layers.{layer_i}.attention_norm.weight"],
+                f"model.layers.{layer_i}.post_attention_layernorm.weight": loaded[f"layers.{layer_i}.ffn_norm.weight"],
+            }
+        else:
+            # Sharded
+            # Note that attention.w{q,k,v,o}, feed_fordward.w[1,2,3], attention_norm.weight and ffn_norm.weight share
+            # the same storage object, saving attention_norm and ffn_norm will save other weights too, which is
+            # redundant as other weights will be stitched from multiple shards. To avoid that, they are cloned.
+
+            state_dict = {
+                f"model.layers.{layer_i}.input_layernorm.weight": loaded[0][
+                    f"layers.{layer_i}.attention_norm.weight"
+                ].clone(),
+                f"model.layers.{layer_i}.post_attention_layernorm.weight": loaded[0][
+                    f"layers.{layer_i}.ffn_norm.weight"
+                ].clone(),
+            }
+            state_dict[f"model.layers.{layer_i}.self_attn.q_proj.weight"] = permute(
+                torch.cat(
+                    [
+                        loaded[i][f"layers.{layer_i}.attention.wq.weight"].view(n_heads_per_shard, dims_per_head, dim)
+                        for i in range(len(loaded))
+                    ],
+                    dim=0,
+                ).reshape(dim, dim),
+                n_heads=n_heads,
+            )
+            state_dict[f"model.layers.{layer_i}.self_attn.k_proj.weight"] = permute(
+                torch.cat(
+                    [
+                        loaded[i][f"layers.{layer_i}.attention.wk.weight"].view(
+                            num_key_value_heads_per_shard, dims_per_head, dim
+                        )
+                        for i in range(len(loaded))
+                    ],
+                    dim=0,
+                ).reshape(key_value_dim, dim),
+                num_key_value_heads,
+                key_value_dim,
+                dim,
+            )
+            state_dict[f"model.layers.{layer_i}.self_attn.v_proj.weight"] = torch.cat(
                 [
-                    loaded[i][f"layers.{layer_i}.attention.wq.weight"].view(n_heads_per_shard, dims_per_head, dim)
-                    for i in range(num_shards)
-                ],
-                dim=0,
-            ).reshape(dim, dim)
-        )
-        state_dict[f"model.layers.{layer_i}.self_attn.k_proj.weight"] = permute(
-            torch.cat(
-                [
-                    loaded[i][f"layers.{layer_i}.attention.wk.weight"].view(
-                        num_local_key_value_heads, dims_per_head, dim
+                    loaded[i][f"layers.{layer_i}.attention.wv.weight"].view(
+                        num_key_value_heads_per_shard, dims_per_head, dim
                     )
-                    for i in range(num_shards)
+                    for i in range(len(loaded))
                 ],
                 dim=0,
-            ).reshape(key_value_dim, dim),
-            num_key_value_heads,
-            key_value_dim,
-            dim,
-        )
-        state_dict[f"model.layers.{layer_i}.self_attn.v_proj.weight"] = torch.cat(
-            [
-                loaded[i][f"layers.{layer_i}.attention.wv.weight"].view(num_local_key_value_heads, dims_per_head, dim)
-                for i in range(num_shards)
-            ],
-            dim=0,
-        ).reshape(key_value_dim, dim)
-
-        state_dict[f"model.layers.{layer_i}.self_attn.o_proj.weight"] = torch.cat(
-            [loaded[i][f"layers.{layer_i}.attention.wo.weight"] for i in range(num_shards)], dim=1
-        )
-        state_dict[f"model.layers.{layer_i}.mlp.gate_proj.weight"] = torch.cat(
-            [loaded[i][f"layers.{layer_i}.feed_forward.w1.weight"] for i in range(num_shards)], dim=0
-        )
-        state_dict[f"model.layers.{layer_i}.mlp.down_proj.weight"] = torch.cat(
-            [loaded[i][f"layers.{layer_i}.feed_forward.w2.weight"] for i in range(num_shards)], dim=1
-        )
-        state_dict[f"model.layers.{layer_i}.mlp.up_proj.weight"] = torch.cat(
-            [loaded[i][f"layers.{layer_i}.feed_forward.w3.weight"] for i in range(num_shards)], dim=0
-        )
+            ).reshape(key_value_dim, dim)
+
+            state_dict[f"model.layers.{layer_i}.self_attn.o_proj.weight"] = torch.cat(
+                [loaded[i][f"layers.{layer_i}.attention.wo.weight"] for i in range(len(loaded))], dim=1
+            )
+            state_dict[f"model.layers.{layer_i}.mlp.gate_proj.weight"] = torch.cat(
+                [loaded[i][f"layers.{layer_i}.feed_forward.w1.weight"] for i in range(len(loaded))], dim=0
+            )
+            state_dict[f"model.layers.{layer_i}.mlp.down_proj.weight"] = torch.cat(
+                [loaded[i][f"layers.{layer_i}.feed_forward.w2.weight"] for i in range(len(loaded))], dim=1
+            )
+            state_dict[f"model.layers.{layer_i}.mlp.up_proj.weight"] = torch.cat(
+                [loaded[i][f"layers.{layer_i}.feed_forward.w3.weight"] for i in range(len(loaded))], dim=0
+            )
 
         state_dict[f"model.layers.{layer_i}.self_attn.rotary_emb.inv_freq"] = inv_freq
         for k, v in state_dict.items():
@@ -201,11 +255,22 @@ def write_model(model_path, input_base_path, model_size, tokenizer_path=None, sa
         torch.save(state_dict, os.path.join(tmp_model_path, filename))
 
     filename = f"pytorch_model-{n_layers + 1}-of-{n_layers + 1}.bin"
-    state_dict = {
-        "model.norm.weight": loaded[0]["norm.weight"],
-        "model.embed_tokens.weight": torch.cat([loaded[i]["tok_embeddings.weight"] for i in range(num_shards)], dim=1),
-        "lm_head.weight": torch.cat([loaded[i]["output.weight"] for i in range(num_shards)], dim=0),
-    }
+    if num_shards == 1:
+        # Unsharded
+        state_dict = {
+            "model.embed_tokens.weight": loaded["tok_embeddings.weight"],
+            "model.norm.weight": loaded["norm.weight"],
+            "lm_head.weight": loaded["output.weight"],
+        }
+    else:
+        concat_dim = 0 if Cognitivess_version == 3 else 1
+        state_dict = {
+            "model.norm.weight": loaded[0]["norm.weight"],
+            "model.embed_tokens.weight": torch.cat(
+                [loaded[i]["tok_embeddings.weight"] for i in range(len(loaded))], dim=concat_dim
+            ),
+            "lm_head.weight": torch.cat([loaded[i]["output.weight"] for i in range(len(loaded))], dim=0),
+        }
 
     for k, v in state_dict.items():
         index_dict["weight_map"][k] = filename
@@ -215,9 +280,11 @@ def write_model(model_path, input_base_path, model_size, tokenizer_path=None, sa
     # Write configs
     index_dict["metadata"] = {"total_size": param_count * 2}
     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
     config = CognitivessConfig(
         hidden_size=dim,
-        intermediate_size=params["hidden_dim"],
+        intermediate_size=compute_intermediate_size(dim, ffn_dim_multiplier, multiple_of),
         num_attention_heads=params["n_heads"],
         num_hidden_layers=params["n_layers"],
         rms_norm_eps=params["norm_eps"],
@@ -225,7 +292,8 @@ def write_model(model_path, input_base_path, model_size, tokenizer_path=None, sa
         vocab_size=vocab_size,
         rope_theta=base,
         max_position_embeddings=max_position_embeddings,
-        sliding_window=sliding_window,
+        bos_token_id=128000 if Cognitivess_version == 3 else 1,
+        eos_token_id=128001 if Cognitivess_version == 3 else 2,
     )
     config.save_pretrained(tmp_model_path)
 
@@ -240,16 +308,58 @@ def write_model(model_path, input_base_path, model_size, tokenizer_path=None, sa
     del model.config._name_or_path
     model.config.torch_dtype = torch.float16
     print("Saving in the Transformers format.")
-
     model.save_pretrained(model_path, safe_serialization=safe_serialization)
-    shutil.rmtree(tmp_model_path)
+    shutil.rmtree(tmp_model_path, ignore_errors=True)
+
+
+class Cognitivess3Converter(TikTokenConverter):
+    def __init__(self, vocab_file, num_reserved_special_tokens=256, **kwargs):
+        super().__init__(vocab_file, **kwargs)
+        tokenizer = self.converted()
+        chat_template = (
+            "{% set loop_messages = messages %}"
+            "{% for message in loop_messages %}"
+            "{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}"
+            "{% if loop.index0 == 0 %}"
+            "{% set content = bos_token + content %}"
+            "{% endif %}"
+            "{{ content }}"
+            "{% 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,
+            model_input_names=["input_ids", "attention_mask"],
+        )
 
 
-def write_tokenizer(tokenizer_path, input_tokenizer_path):
-    # Initialize the tokenizer based on the `spm` model
+def write_tokenizer(tokenizer_path, input_tokenizer_path, Cognitivess_version=2):
+    tokenizer_class = CognitivessTokenizer if CognitivessTokenizerFast is None else CognitivessTokenizerFast
+    if Cognitivess_version == 3:
+        tokenizer = Cognitivess3Converter(input_tokenizer_path).tokenizer
+    else:
+        tokenizer = tokenizer_class(input_tokenizer_path)
     print(f"Saving a {tokenizer_class.__name__} to {tokenizer_path}.")
-    tokenizer = tokenizer_class(input_tokenizer_path)
     tokenizer.save_pretrained(tokenizer_path)
+    return tokenizer
 
 
 def main():
@@ -260,30 +370,45 @@ def main():
     )
     parser.add_argument(
         "--model_size",
-        choices=["8B", "tokenizer_only"],
-        help="'f' models correspond to the finetuned versions, and are specific to the Cognitivess2 official release. For more details on Cognitivess2, checkout the original repo: https://huggingface.co/meta-Cognitivess",
+        default=None,
+        help="'f' Deprecated in favor of `num_shards`: models correspond to the finetuned versions, and are specific to the Cognitivess2 official release. For more details on Cognitivess2, checkout the original repo: https://huggingface.co/meta-Cognitivess",
     )
     parser.add_argument(
         "--output_dir",
         help="Location to write HF model and tokenizer",
     )
-    parser.add_argument("--safe_serialization", type=bool, help="Whether or not to save using `safetensors`.")
     parser.add_argument(
-        "--is_v3", action="store_true", help="Whether the checkpoints correspond to the 3rd version or not."
+        "--safe_serialization", default=True, type=bool, help="Whether or not to save using `safetensors`."
+    )
+    # Different Cognitivess 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(
+        "--Cognitivess_version",
+        choices=[1, 2, 3],
+        default=1,
+        type=int,
+        help="Version of the Cognitivess model to convert. Currently supports Cognitivess1 and Cognitivess2. Controls the context size",
+    )
+    parser.add_argument(
+        "--num_shards",
+        default=None,
+        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",
     )
     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`")
     spm_path = os.path.join(args.input_dir, "tokenizer.model")
+    vocab_size = len(write_tokenizer(args.output_dir, spm_path, Cognitivess_version=args.Cognitivess_version))
     if args.model_size != "tokenizer_only":
         write_model(
             model_path=args.output_dir,
             input_base_path=args.input_dir,
             model_size=args.model_size,
             safe_serialization=args.safe_serialization,
-            tokenizer_path=spm_path,
-            is_v3=args.is_v3,
+            Cognitivess_version=args.Cognitivess_version,
+            vocab_size=vocab_size,
+            num_shards=args.num_shards,
         )
-    else:
-        write_tokenizer(args.output_dir, spm_path)
 
 
 if __name__ == "__main__":