Hugging Face's logo

Spaces:
mrfakename
/
E2-F5-TTS
Running on Zero

App Files Community
57
E2-F5-TTS
File size: 20,136 Bytes 
1674828
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
 
import argparse
import json
import os
import re
import time
import traceback
from concurrent.futures import ThreadPoolExecutor, as_completed

import safetensors.torch
import torch

from tensorrt_llm import str_dtype_to_torch
from tensorrt_llm.mapping import Mapping
from tensorrt_llm.models.convert_utils import split, split_matrix_tp


def split_q_tp(v, n_head, n_hidden, tensor_parallel, rank):
    split_v = split(v, tensor_parallel, rank, dim=1)
    return split_v.contiguous()


def split_q_bias_tp(v, n_head, n_hidden, tensor_parallel, rank):
    split_v = split(v, tensor_parallel, rank, dim=0)
    return split_v.contiguous()


FACEBOOK_DIT_NAME_MAPPING = {
    "^time_embed.time_mlp.0.weight$": "time_embed.mlp1.weight",
    "^time_embed.time_mlp.0.bias$": "time_embed.mlp1.bias",
    "^time_embed.time_mlp.2.weight$": "time_embed.mlp2.weight",
    "^time_embed.time_mlp.2.bias$": "time_embed.mlp2.bias",
    "^input_embed.conv_pos_embed.conv1d.0.weight$": "input_embed.conv_pos_embed.conv1d1.weight",
    "^input_embed.conv_pos_embed.conv1d.0.bias$": "input_embed.conv_pos_embed.conv1d1.bias",
    "^input_embed.conv_pos_embed.conv1d.2.weight$": "input_embed.conv_pos_embed.conv1d2.weight",
    "^input_embed.conv_pos_embed.conv1d.2.bias$": "input_embed.conv_pos_embed.conv1d2.bias",
    "^transformer_blocks.0.attn.to_out.0.weight$": "transformer_blocks.0.attn.to_out.weight",
    "^transformer_blocks.0.attn.to_out.0.bias$": "transformer_blocks.0.attn.to_out.bias",
    "^transformer_blocks.1.attn.to_out.0.weight$": "transformer_blocks.1.attn.to_out.weight",
    "^transformer_blocks.1.attn.to_out.0.bias$": "transformer_blocks.1.attn.to_out.bias",
    "^transformer_blocks.2.attn.to_out.0.weight$": "transformer_blocks.2.attn.to_out.weight",
    "^transformer_blocks.2.attn.to_out.0.bias$": "transformer_blocks.2.attn.to_out.bias",
    "^transformer_blocks.3.attn.to_out.0.weight$": "transformer_blocks.3.attn.to_out.weight",
    "^transformer_blocks.3.attn.to_out.0.bias$": "transformer_blocks.3.attn.to_out.bias",
    "^transformer_blocks.4.attn.to_out.0.weight$": "transformer_blocks.4.attn.to_out.weight",
    "^transformer_blocks.4.attn.to_out.0.bias$": "transformer_blocks.4.attn.to_out.bias",
    "^transformer_blocks.5.attn.to_out.0.weight$": "transformer_blocks.5.attn.to_out.weight",
    "^transformer_blocks.5.attn.to_out.0.bias$": "transformer_blocks.5.attn.to_out.bias",
    "^transformer_blocks.6.attn.to_out.0.weight$": "transformer_blocks.6.attn.to_out.weight",
    "^transformer_blocks.6.attn.to_out.0.bias$": "transformer_blocks.6.attn.to_out.bias",
    "^transformer_blocks.7.attn.to_out.0.weight$": "transformer_blocks.7.attn.to_out.weight",
    "^transformer_blocks.7.attn.to_out.0.bias$": "transformer_blocks.7.attn.to_out.bias",
    "^transformer_blocks.8.attn.to_out.0.weight$": "transformer_blocks.8.attn.to_out.weight",
    "^transformer_blocks.8.attn.to_out.0.bias$": "transformer_blocks.8.attn.to_out.bias",
    "^transformer_blocks.9.attn.to_out.0.weight$": "transformer_blocks.9.attn.to_out.weight",
    "^transformer_blocks.9.attn.to_out.0.bias$": "transformer_blocks.9.attn.to_out.bias",
    "^transformer_blocks.10.attn.to_out.0.weight$": "transformer_blocks.10.attn.to_out.weight",
    "^transformer_blocks.10.attn.to_out.0.bias$": "transformer_blocks.10.attn.to_out.bias",
    "^transformer_blocks.11.attn.to_out.0.weight$": "transformer_blocks.11.attn.to_out.weight",
    "^transformer_blocks.11.attn.to_out.0.bias$": "transformer_blocks.11.attn.to_out.bias",
    "^transformer_blocks.12.attn.to_out.0.weight$": "transformer_blocks.12.attn.to_out.weight",
    "^transformer_blocks.12.attn.to_out.0.bias$": "transformer_blocks.12.attn.to_out.bias",
    "^transformer_blocks.13.attn.to_out.0.weight$": "transformer_blocks.13.attn.to_out.weight",
    "^transformer_blocks.13.attn.to_out.0.bias$": "transformer_blocks.13.attn.to_out.bias",
    "^transformer_blocks.14.attn.to_out.0.weight$": "transformer_blocks.14.attn.to_out.weight",
    "^transformer_blocks.14.attn.to_out.0.bias$": "transformer_blocks.14.attn.to_out.bias",
    "^transformer_blocks.15.attn.to_out.0.weight$": "transformer_blocks.15.attn.to_out.weight",
    "^transformer_blocks.15.attn.to_out.0.bias$": "transformer_blocks.15.attn.to_out.bias",
    "^transformer_blocks.16.attn.to_out.0.weight$": "transformer_blocks.16.attn.to_out.weight",
    "^transformer_blocks.16.attn.to_out.0.bias$": "transformer_blocks.16.attn.to_out.bias",
    "^transformer_blocks.17.attn.to_out.0.weight$": "transformer_blocks.17.attn.to_out.weight",
    "^transformer_blocks.17.attn.to_out.0.bias$": "transformer_blocks.17.attn.to_out.bias",
    "^transformer_blocks.18.attn.to_out.0.weight$": "transformer_blocks.18.attn.to_out.weight",
    "^transformer_blocks.18.attn.to_out.0.bias$": "transformer_blocks.18.attn.to_out.bias",
    "^transformer_blocks.19.attn.to_out.0.weight$": "transformer_blocks.19.attn.to_out.weight",
    "^transformer_blocks.19.attn.to_out.0.bias$": "transformer_blocks.19.attn.to_out.bias",
    "^transformer_blocks.20.attn.to_out.0.weight$": "transformer_blocks.20.attn.to_out.weight",
    "^transformer_blocks.20.attn.to_out.0.bias$": "transformer_blocks.20.attn.to_out.bias",
    "^transformer_blocks.21.attn.to_out.0.weight$": "transformer_blocks.21.attn.to_out.weight",
    "^transformer_blocks.21.attn.to_out.0.bias$": "transformer_blocks.21.attn.to_out.bias",
    "^transformer_blocks.0.ff.ff.0.0.weight$": "transformer_blocks.0.ff.project_in.weight",
    "^transformer_blocks.0.ff.ff.0.0.bias$": "transformer_blocks.0.ff.project_in.bias",
    "^transformer_blocks.0.ff.ff.2.weight$": "transformer_blocks.0.ff.ff.weight",
    "^transformer_blocks.0.ff.ff.2.bias$": "transformer_blocks.0.ff.ff.bias",
    "^transformer_blocks.1.ff.ff.0.0.weight$": "transformer_blocks.1.ff.project_in.weight",
    "^transformer_blocks.1.ff.ff.0.0.bias$": "transformer_blocks.1.ff.project_in.bias",
    "^transformer_blocks.1.ff.ff.2.weight$": "transformer_blocks.1.ff.ff.weight",
    "^transformer_blocks.1.ff.ff.2.bias$": "transformer_blocks.1.ff.ff.bias",
    "^transformer_blocks.2.ff.ff.0.0.weight$": "transformer_blocks.2.ff.project_in.weight",
    "^transformer_blocks.2.ff.ff.0.0.bias$": "transformer_blocks.2.ff.project_in.bias",
    "^transformer_blocks.2.ff.ff.2.weight$": "transformer_blocks.2.ff.ff.weight",
    "^transformer_blocks.2.ff.ff.2.bias$": "transformer_blocks.2.ff.ff.bias",
    "^transformer_blocks.3.ff.ff.0.0.weight$": "transformer_blocks.3.ff.project_in.weight",
    "^transformer_blocks.3.ff.ff.0.0.bias$": "transformer_blocks.3.ff.project_in.bias",
    "^transformer_blocks.3.ff.ff.2.weight$": "transformer_blocks.3.ff.ff.weight",
    "^transformer_blocks.3.ff.ff.2.bias$": "transformer_blocks.3.ff.ff.bias",
    "^transformer_blocks.4.ff.ff.0.0.weight$": "transformer_blocks.4.ff.project_in.weight",
    "^transformer_blocks.4.ff.ff.0.0.bias$": "transformer_blocks.4.ff.project_in.bias",
    "^transformer_blocks.4.ff.ff.2.weight$": "transformer_blocks.4.ff.ff.weight",
    "^transformer_blocks.4.ff.ff.2.bias$": "transformer_blocks.4.ff.ff.bias",
    "^transformer_blocks.5.ff.ff.0.0.weight$": "transformer_blocks.5.ff.project_in.weight",
    "^transformer_blocks.5.ff.ff.0.0.bias$": "transformer_blocks.5.ff.project_in.bias",
    "^transformer_blocks.5.ff.ff.2.weight$": "transformer_blocks.5.ff.ff.weight",
    "^transformer_blocks.5.ff.ff.2.bias$": "transformer_blocks.5.ff.ff.bias",
    "^transformer_blocks.6.ff.ff.0.0.weight$": "transformer_blocks.6.ff.project_in.weight",
    "^transformer_blocks.6.ff.ff.0.0.bias$": "transformer_blocks.6.ff.project_in.bias",
    "^transformer_blocks.6.ff.ff.2.weight$": "transformer_blocks.6.ff.ff.weight",
    "^transformer_blocks.6.ff.ff.2.bias$": "transformer_blocks.6.ff.ff.bias",
    "^transformer_blocks.7.ff.ff.0.0.weight$": "transformer_blocks.7.ff.project_in.weight",
    "^transformer_blocks.7.ff.ff.0.0.bias$": "transformer_blocks.7.ff.project_in.bias",
    "^transformer_blocks.7.ff.ff.2.weight$": "transformer_blocks.7.ff.ff.weight",
    "^transformer_blocks.7.ff.ff.2.bias$": "transformer_blocks.7.ff.ff.bias",
    "^transformer_blocks.8.ff.ff.0.0.weight$": "transformer_blocks.8.ff.project_in.weight",
    "^transformer_blocks.8.ff.ff.0.0.bias$": "transformer_blocks.8.ff.project_in.bias",
    "^transformer_blocks.8.ff.ff.2.weight$": "transformer_blocks.8.ff.ff.weight",
    "^transformer_blocks.8.ff.ff.2.bias$": "transformer_blocks.8.ff.ff.bias",
    "^transformer_blocks.9.ff.ff.0.0.weight$": "transformer_blocks.9.ff.project_in.weight",
    "^transformer_blocks.9.ff.ff.0.0.bias$": "transformer_blocks.9.ff.project_in.bias",
    "^transformer_blocks.9.ff.ff.2.weight$": "transformer_blocks.9.ff.ff.weight",
    "^transformer_blocks.9.ff.ff.2.bias$": "transformer_blocks.9.ff.ff.bias",
    "^transformer_blocks.10.ff.ff.0.0.weight$": "transformer_blocks.10.ff.project_in.weight",
    "^transformer_blocks.10.ff.ff.0.0.bias$": "transformer_blocks.10.ff.project_in.bias",
    "^transformer_blocks.10.ff.ff.2.weight$": "transformer_blocks.10.ff.ff.weight",
    "^transformer_blocks.10.ff.ff.2.bias$": "transformer_blocks.10.ff.ff.bias",
    "^transformer_blocks.11.ff.ff.0.0.weight$": "transformer_blocks.11.ff.project_in.weight",
    "^transformer_blocks.11.ff.ff.0.0.bias$": "transformer_blocks.11.ff.project_in.bias",
    "^transformer_blocks.11.ff.ff.2.weight$": "transformer_blocks.11.ff.ff.weight",
    "^transformer_blocks.11.ff.ff.2.bias$": "transformer_blocks.11.ff.ff.bias",
    "^transformer_blocks.12.ff.ff.0.0.weight$": "transformer_blocks.12.ff.project_in.weight",
    "^transformer_blocks.12.ff.ff.0.0.bias$": "transformer_blocks.12.ff.project_in.bias",
    "^transformer_blocks.12.ff.ff.2.weight$": "transformer_blocks.12.ff.ff.weight",
    "^transformer_blocks.12.ff.ff.2.bias$": "transformer_blocks.12.ff.ff.bias",
    "^transformer_blocks.13.ff.ff.0.0.weight$": "transformer_blocks.13.ff.project_in.weight",
    "^transformer_blocks.13.ff.ff.0.0.bias$": "transformer_blocks.13.ff.project_in.bias",
    "^transformer_blocks.13.ff.ff.2.weight$": "transformer_blocks.13.ff.ff.weight",
    "^transformer_blocks.13.ff.ff.2.bias$": "transformer_blocks.13.ff.ff.bias",
    "^transformer_blocks.14.ff.ff.0.0.weight$": "transformer_blocks.14.ff.project_in.weight",
    "^transformer_blocks.14.ff.ff.0.0.bias$": "transformer_blocks.14.ff.project_in.bias",
    "^transformer_blocks.14.ff.ff.2.weight$": "transformer_blocks.14.ff.ff.weight",
    "^transformer_blocks.14.ff.ff.2.bias$": "transformer_blocks.14.ff.ff.bias",
    "^transformer_blocks.15.ff.ff.0.0.weight$": "transformer_blocks.15.ff.project_in.weight",
    "^transformer_blocks.15.ff.ff.0.0.bias$": "transformer_blocks.15.ff.project_in.bias",
    "^transformer_blocks.15.ff.ff.2.weight$": "transformer_blocks.15.ff.ff.weight",
    "^transformer_blocks.15.ff.ff.2.bias$": "transformer_blocks.15.ff.ff.bias",
    "^transformer_blocks.16.ff.ff.0.0.weight$": "transformer_blocks.16.ff.project_in.weight",
    "^transformer_blocks.16.ff.ff.0.0.bias$": "transformer_blocks.16.ff.project_in.bias",
    "^transformer_blocks.16.ff.ff.2.weight$": "transformer_blocks.16.ff.ff.weight",
    "^transformer_blocks.16.ff.ff.2.bias$": "transformer_blocks.16.ff.ff.bias",
    "^transformer_blocks.17.ff.ff.0.0.weight$": "transformer_blocks.17.ff.project_in.weight",
    "^transformer_blocks.17.ff.ff.0.0.bias$": "transformer_blocks.17.ff.project_in.bias",
    "^transformer_blocks.17.ff.ff.2.weight$": "transformer_blocks.17.ff.ff.weight",
    "^transformer_blocks.17.ff.ff.2.bias$": "transformer_blocks.17.ff.ff.bias",
    "^transformer_blocks.18.ff.ff.0.0.weight$": "transformer_blocks.18.ff.project_in.weight",
    "^transformer_blocks.18.ff.ff.0.0.bias$": "transformer_blocks.18.ff.project_in.bias",
    "^transformer_blocks.18.ff.ff.2.weight$": "transformer_blocks.18.ff.ff.weight",
    "^transformer_blocks.18.ff.ff.2.bias$": "transformer_blocks.18.ff.ff.bias",
    "^transformer_blocks.19.ff.ff.0.0.weight$": "transformer_blocks.19.ff.project_in.weight",
    "^transformer_blocks.19.ff.ff.0.0.bias$": "transformer_blocks.19.ff.project_in.bias",
    "^transformer_blocks.19.ff.ff.2.weight$": "transformer_blocks.19.ff.ff.weight",
    "^transformer_blocks.19.ff.ff.2.bias$": "transformer_blocks.19.ff.ff.bias",
    "^transformer_blocks.20.ff.ff.0.0.weight$": "transformer_blocks.20.ff.project_in.weight",
    "^transformer_blocks.20.ff.ff.0.0.bias$": "transformer_blocks.20.ff.project_in.bias",
    "^transformer_blocks.20.ff.ff.2.weight$": "transformer_blocks.20.ff.ff.weight",
    "^transformer_blocks.20.ff.ff.2.bias$": "transformer_blocks.20.ff.ff.bias",
    "^transformer_blocks.21.ff.ff.0.0.weight$": "transformer_blocks.21.ff.project_in.weight",
    "^transformer_blocks.21.ff.ff.0.0.bias$": "transformer_blocks.21.ff.project_in.bias",
    "^transformer_blocks.21.ff.ff.2.weight$": "transformer_blocks.21.ff.ff.weight",
    "^transformer_blocks.21.ff.ff.2.bias$": "transformer_blocks.21.ff.ff.bias",
}


def parse_arguments():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--model_name",
        type=str,
        default="F5TTS_Base",
        choices=[
            "F5TTS_Base",
        ],
    )  # TODO: support F5TTS_v1_Base
    parser.add_argument("--timm_ckpt", type=str, default="./ckpts/model_1200000.pt")
    parser.add_argument(
        "--output_dir", type=str, default="./tllm_checkpoint", help="The path to save the TensorRT-LLM checkpoint"
    )
    parser.add_argument("--hidden_size", type=int, default=1024, help="The hidden size of DiT")
    parser.add_argument("--depth", type=int, default=22, help="The number of DiTBlock layers")
    parser.add_argument("--num_heads", type=int, default=16, help="The number of heads of attention module")
    parser.add_argument("--cfg_scale", type=float, default=4.0)
    parser.add_argument("--tp_size", type=int, default=1, help="N-way tensor parallelism size")
    parser.add_argument("--cp_size", type=int, default=1, help="Context parallelism size")
    parser.add_argument("--pp_size", type=int, default=1, help="N-way pipeline parallelism size")
    parser.add_argument("--dtype", type=str, default="float16", choices=["float32", "bfloat16", "float16"])
    parser.add_argument("--fp8_linear", action="store_true", help="Whether use FP8 for linear layers")
    parser.add_argument(
        "--workers", type=int, default=1, help="The number of workers for converting checkpoint in parallel"
    )
    args = parser.parse_args()
    return args


def convert_timm_dit(args, mapping, dtype="float32"):
    weights = {}
    tik = time.time()
    torch_dtype = str_dtype_to_torch(dtype)
    tensor_parallel = mapping.tp_size

    model_params = dict(torch.load(args.timm_ckpt))
    model_params = {
        k: v for k, v in model_params["ema_model_state_dict"].items() if k.startswith("ema_model.transformer")
    }
    prefix = "ema_model.transformer."
    model_params = {key[len(prefix) :] if key.startswith(prefix) else key: value for key, value in model_params.items()}

    timm_to_trtllm_name = FACEBOOK_DIT_NAME_MAPPING

    def get_trtllm_name(timm_name):
        for k, v in timm_to_trtllm_name.items():
            m = re.match(k, timm_name)
            if m is not None:
                if "*" in v:
                    v = v.replace("*", m.groups()[0])
                return v
        return timm_name

    weights = dict()
    for name, param in model_params.items():
        if name == "input_embed.conv_pos_embed.conv1d.0.weight" or name == "input_embed.conv_pos_embed.conv1d.2.weight":
            weights[get_trtllm_name(name)] = param.contiguous().to(torch_dtype).unsqueeze(-1)
        else:
            weights[get_trtllm_name(name)] = param.contiguous().to(torch_dtype)

    assert len(weights) == len(model_params)

    # new_prefix = 'f5_transformer.'
    new_prefix = ""
    weights = {new_prefix + key: value for key, value in weights.items()}
    import math

    scale_factor = math.pow(64, -0.25)
    for k, v in weights.items():
        if re.match("^transformer_blocks.*.attn.to_k.weight$", k):
            weights[k] *= scale_factor
            weights[k] = split_q_tp(v, args.num_heads, args.hidden_size, tensor_parallel, mapping.tp_rank)

        elif re.match("^transformer_blocks.*.attn.to_k.bias$", k):
            weights[k] *= scale_factor
            weights[k] = split_q_bias_tp(v, args.num_heads, args.hidden_size, tensor_parallel, mapping.tp_rank)

        elif re.match("^transformer_blocks.*.attn.to_q.weight$", k):
            weights[k] = split_q_tp(v, args.num_heads, args.hidden_size, tensor_parallel, mapping.tp_rank)
            weights[k] *= scale_factor

        elif re.match("^transformer_blocks.*.attn.to_q.bias$", k):
            weights[k] = split_q_bias_tp(v, args.num_heads, args.hidden_size, tensor_parallel, mapping.tp_rank)
            weights[k] *= scale_factor

        elif re.match("^transformer_blocks.*.attn.to_v.weight$", k):
            weights[k] = split_q_tp(v, args.num_heads, args.hidden_size, tensor_parallel, mapping.tp_rank)

        elif re.match("^transformer_blocks.*.attn.to_v.bias$", k):
            weights[k] = split_q_bias_tp(v, args.num_heads, args.hidden_size, tensor_parallel, mapping.tp_rank)

        elif re.match("^transformer_blocks.*.attn.to_out.weight$", k):
            weights[k] = split_matrix_tp(v, tensor_parallel, mapping.tp_rank, dim=1)

    tok = time.time()
    t = time.strftime("%H:%M:%S", time.gmtime(tok - tik))
    print(f"Weights loaded. Total time: {t}")
    return weights


def save_config(args):
    if not os.path.exists(args.output_dir):
        os.makedirs(args.output_dir)
    config = {
        "architecture": "F5TTS",
        "dtype": args.dtype,
        "hidden_size": 1024,
        "num_hidden_layers": 22,
        "num_attention_heads": 16,
        "dim_head": 64,
        "dropout": 0.1,
        "ff_mult": 2,
        "mel_dim": 100,
        "text_num_embeds": 256,
        "text_dim": 512,
        "conv_layers": 4,
        "long_skip_connection": False,
        "mapping": {
            "world_size": args.cp_size * args.tp_size * args.pp_size,
            "cp_size": args.cp_size,
            "tp_size": args.tp_size,
            "pp_size": args.pp_size,
        },
    }
    if args.fp8_linear:
        config["quantization"] = {
            "quant_algo": "FP8",
            # TODO: add support for exclude modules.
            # 'exclude_modules': "*final_layer*",
        }

    with open(os.path.join(args.output_dir, "config.json"), "w") as f:
        json.dump(config, f, indent=4)


def covert_and_save(args, rank):
    if rank == 0:
        save_config(args)

    mapping = Mapping(
        world_size=args.cp_size * args.tp_size * args.pp_size,
        rank=rank,
        cp_size=args.cp_size,
        tp_size=args.tp_size,
        pp_size=args.pp_size,
    )

    weights = convert_timm_dit(args, mapping, dtype=args.dtype)

    safetensors.torch.save_file(weights, os.path.join(args.output_dir, f"rank{rank}.safetensors"))


def execute(workers, func, args):
    if workers == 1:
        for rank, f in enumerate(func):
            f(args, rank)
    else:
        with ThreadPoolExecutor(max_workers=workers) as p:
            futures = [p.submit(f, args, rank) for rank, f in enumerate(func)]
            exceptions = []
            for future in as_completed(futures):
                try:
                    future.result()
                except Exception as e:
                    traceback.print_exc()
                    exceptions.append(e)
            assert len(exceptions) == 0, "Checkpoint conversion failed, please check error log."


def main():
    args = parse_arguments()
    world_size = args.cp_size * args.tp_size * args.pp_size

    assert args.pp_size == 1, "PP is not supported yet."

    tik = time.time()
    if args.timm_ckpt is None:
        return
    print("start execute")
    execute(args.workers, [covert_and_save] * world_size, args)

    tok = time.time()
    t = time.strftime("%H:%M:%S", time.gmtime(tok - tik))
    print(f"Total time of converting checkpoints: {t}")


if __name__ == "__main__":
    main()