File size: 10,905 Bytes
9dc7665
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
# Copyright 2023 Mistral AI 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.
import argparse
import gc
import json
import os
import shutil
import warnings

import torch

from transformers import (
    LlamaTokenizer
)

from transformers.models.mistral.modeling_moe_mistral import MixtralForCausalLM
from transformers.models.mistral.configuration_dmoe_mistral import MixtralConfig

try:
    from transformers import LlamaTokenizerFast

    tokenizer_class = LlamaTokenizerFast
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

"""
Sample usage:

```
python src/transformers/models/mistral/convert_mistral_weights_to_hf.py \
    --input_dir /path/to/downloaded/mistral/weights --model_size 7B --output_dir /output/path
```

Thereafter, models can be loaded via:

```py
from transformers import MistralForCausalLM, LlamaTokenizer

model = MistralForCausalLM.from_pretrained("/output/path")
tokenizer = LlamaTokenizer.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).
"""

NUM_SHARDS = {"7B": 1}


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)


def read_json(path):
    with open(path, "r") as f:
        return json.load(f)


def write_json(text, path):
    with open(path, "w") as f:
        json.dump(text, f)


def write_model(model_path, input_base_path, model_size, tokenizer_path=None, safe_serialization=True):
    # 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)

    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]

    n_layers = params["n_layers"]
    n_heads = params["n_heads"]
    n_heads_per_shard = n_heads // num_shards
    dim = params["dim"]
    dims_per_head = dim // n_heads
    base = params.get("rope_theta", 100000.0)
    inv_freq = 1.0 / (base ** (torch.arange(0, dims_per_head, 2).float() / dims_per_head))
    max_position_embeddings = 4096 * 8
    num_experts_per_token = params["moe"]["num_experts_per_tok"]
    num_experts = params["moe"]["num_experts"]
    

    if tokenizer_path is not None:
        tokenizer = tokenizer_class(tokenizer_path)
        tokenizer.save_pretrained(model_path)
    vocab_size = tokenizer.vocab_size if tokenizer_path is not None else 32000

    if "n_kv_heads" in params:
        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
    else:  # compatibility with other checkpoints
        num_key_value_heads = n_heads
        num_local_key_value_heads = n_heads_per_shard
        key_value_dim = dim

    # permute for sliced rotary
    def permute(w, n_heads=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
    loaded = [
        torch.load(os.path.join(input_base_path, f"consolidated.{i:02d}.pth"), map_location="cpu")
        for i in range(num_shards)
    ]
    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(
                [
                    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
                    )
                    for i in range(num_shards)
                ],
                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
        )

        for expert in range(num_experts):
            state_dict[f"model.layers.{layer_i}.mlp.experts.{expert}.w1.weight"] = loaded[0][f"layers.{layer_i}.feed_forward.experts.{expert}.w1.weight"]
            state_dict[f"model.layers.{layer_i}.mlp.experts.{expert}.w2.weight"] = loaded[0][f"layers.{layer_i}.feed_forward.experts.{expert}.w2.weight"]
            state_dict[f"model.layers.{layer_i}.mlp.experts.{expert}.w3.weight"] = loaded[0][f"layers.{layer_i}.feed_forward.experts.{expert}.w3.weight"] 

        state_dict[f"model.layers.{layer_i}.mlp.gate.weight"] = loaded[0][f"layers.{layer_i}.feed_forward.gate.weight"]

        state_dict[f"model.layers.{layer_i}.self_attn.rotary_emb.inv_freq"] = inv_freq
        for k, v in state_dict.items():
            index_dict["weight_map"][k] = filename
            param_count += v.numel()
        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),
    }

    for k, v in state_dict.items():
        index_dict["weight_map"][k] = filename
        param_count += v.numel()
        print(param_count)
    torch.save(state_dict, os.path.join(tmp_model_path, filename))

    index_dict["metadata"] = {"total_size": param_count * 2}
    write_json(index_dict, os.path.join(tmp_model_path, "pytorch_model.bin.index.json"))
    config = MixtralConfig(
        hidden_size=dim,
        intermediate_size=params["hidden_dim"],
        num_attention_heads=params["n_heads"],
        num_hidden_layers=params["n_layers"],
        rms_norm_eps=params["norm_eps"],
        num_key_value_heads=num_key_value_heads,
        vocab_size=vocab_size,
        rope_theta=base,
        max_position_embeddings=max_position_embeddings,
        num_experts=num_experts,
        num_experts_per_token=num_experts_per_token
    )
    config.save_pretrained(tmp_model_path)

    del state_dict
    del loaded
    gc.collect()

    print("Loading the checkpoint in a Mistral model.")
    model = MixtralForCausalLM.from_pretrained(tmp_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
    # Avoid saving this as part of the config.
    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)


def write_tokenizer(tokenizer_path, input_tokenizer_path):
    # Initialize the tokenizer based on the `spm` model
    print(f"Saving a {tokenizer_class.__name__} to {tokenizer_path}.")
    tokenizer = tokenizer_class(input_tokenizer_path)
    tokenizer.save_pretrained(tokenizer_path)


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--input_dir",
        help="Location of Mistral weights, which contains tokenizer.model and model folders",
    )
    parser.add_argument(
        "--model_size",
        choices=["7B", "tokenizer_only"],
        help="'f' models correspond to the finetuned versions, and are specific to the Mistral2 official release. For more details on Mistral2, checkout the original repo: https://huggingface.co/meta-mistral",
    )
    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`.")
    args = parser.parse_args()
    spm_path = os.path.join(args.input_dir, "tokenizer.model")
    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,
        )
    else:
        write_tokenizer(args.output_dir, spm_path)


if __name__ == "__main__":
    main()