char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Sigbjørn Skjæretci : switch from pyright to ty (#20826)29b28a982

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
18.7 KiB450 linesraw
1#!/usr/bin/env python3
2from __future__ import annotations
3
4import logging
5import argparse
6import os
7import struct
8import sys
9from enum import IntEnum
10from pathlib import Path
11
12import numpy as np
13
14if 'NO_LOCAL_GGUF' not in os.environ:
15    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
16import gguf
17
18logger = logging.getLogger("ggml-to-gguf")
19
20
21class GGMLFormat(IntEnum):
22    GGML = 0
23    GGMF = 1
24    GGJT = 2
25
26
27class GGMLFType(IntEnum):
28    ALL_F32              = 0
29    MOSTLY_F16           = 1
30    MOSTLY_Q4_0          = 2
31    MOSTLY_Q4_1          = 3
32    MOSTLY_Q4_1_SOME_F16 = 4
33    MOSTLY_Q8_0          = 7
34    MOSTLY_Q5_0          = 8
35    MOSTLY_Q5_1          = 9
36    MOSTLY_Q2_K          = 10
37    MOSTLY_Q3_K_S        = 11
38    MOSTLY_Q3_K_M        = 12
39    MOSTLY_Q3_K_L        = 13
40    MOSTLY_Q4_K_S        = 14
41    MOSTLY_Q4_K_M        = 15
42    MOSTLY_Q5_K_S        = 16
43    MOSTLY_Q5_K_M        = 17
44    MOSTLY_Q6_K          = 18
45
46
47class Hyperparameters:
48    def __init__(self):
49        self.n_vocab = self.n_embd = self.n_mult = self.n_head = 0
50        self.n_layer = self.n_rot = self.n_ff = 0
51        self.ftype = GGMLFType.ALL_F32
52
53    def set_n_ff(self, model):
54        ff_tensor_idx = model.tensor_map.get(b'layers.0.feed_forward.w1.weight')
55        assert ff_tensor_idx is not None, 'Missing layer 0 FF tensor'
56        ff_tensor = model.tensors[ff_tensor_idx]
57        self.n_ff = ff_tensor.dims[1]
58
59    def load(self, data, offset):
60        (
61            self.n_vocab,
62            self.n_embd,
63            self.n_mult,
64            self.n_head,
65            self.n_layer,
66            self.n_rot,
67            ftype,
68        ) = struct.unpack('<7I', data[offset:offset + (4 * 7)])
69        try:
70            self.ftype = GGMLFType(ftype)
71        except ValueError:
72            raise ValueError(f'Invalid ftype {ftype}')
73        return 4 * 7
74
75    def __str__(self):
76        return f'<Hyperparameters: n_vocab={self.n_vocab}, n_embd={self.n_embd}, n_mult={self.n_mult}, n_head={self.n_head}, n_layer={self.n_layer}, n_rot={self.n_rot}, n_ff={self.n_ff}, ftype={self.ftype.name}>'
77
78
79class Vocab:
80    def __init__(self, load_scores = True):
81        self.items = []
82        self.load_scores = load_scores
83
84    def load(self, data, offset, n_vocab):
85        orig_offset = offset
86        for _ in range(n_vocab):
87            itemlen = struct.unpack('<I', data[offset:offset + 4])[0]
88            assert itemlen < 4096, 'Absurd vocab item length'
89            offset += 4
90            item_text = bytes(data[offset:offset + itemlen])
91            offset += itemlen
92            if self.load_scores:
93                item_score = struct.unpack('<f', data[offset:offset + 4])[0]
94                offset += 4
95            else:
96                item_score = 0.0
97            self.items.append((item_text, item_score))
98        return offset - orig_offset
99
100
101class Tensor:
102    def __init__(self, use_padding = True):
103        self.name = None
104        self.dims: tuple[int, ...] = ()
105        self.dtype = None
106        self.start_offset = 0
107        self.len_bytes = np.int64(0)
108        self.use_padding = use_padding
109
110    def load(self, data, offset):
111        orig_offset = offset
112        (n_dims, name_len, dtype) = struct.unpack('<3I', data[offset:offset + 12])
113        assert n_dims >= 0 and n_dims <= 4, f'Invalid tensor dimensions {n_dims}'
114        assert name_len < 4096, 'Absurd tensor name length'
115        self.dtype = gguf.GGMLQuantizationType(dtype)
116        quant = gguf.GGML_QUANT_SIZES.get(self.dtype)
117        assert quant is not None, 'Unknown tensor type'
118        (blksize, tysize) = quant
119        offset += 12
120        self.dims = struct.unpack(f'<{n_dims}I', data[offset:offset + (4 * n_dims)])
121        offset += 4 * n_dims
122        self.name = bytes(data[offset:offset + name_len])
123        offset += name_len
124        pad = ((offset + 31) & ~31) - offset if self.use_padding else 0
125        offset += pad
126        n_elems = np.prod(self.dims)
127        n_bytes = np.int64(np.int64(n_elems) * np.int64(tysize)) // np.int64(blksize)
128        self.start_offset = offset
129        self.len_bytes = n_bytes
130        offset += n_bytes
131        return offset - orig_offset
132
133
134class GGMLModel:
135
136    file_format: GGMLFormat
137    format_version: int
138
139    def __init__(self):
140        self.hyperparameters = None
141        self.vocab = None
142        self.tensor_map = {}
143        self.tensors = []
144
145    def validate_header(self, data, offset):
146        magic = bytes(data[offset:offset + 4])
147        if magic == b'GGUF':
148            raise ValueError('File is already in GGUF format.')
149        if magic == b'lmgg':
150            self.file_format = GGMLFormat.GGML
151            self.format_version = 1
152            return 4
153        version = struct.unpack('<I', data[offset + 4:offset + 8])[0]
154        if magic == b'fmgg':
155            if version != 1:
156                raise ValueError(f'Cannot handle unexpected GGMF file version {version}')
157            self.file_format = GGMLFormat.GGMF
158            self.format_version = version
159            return 8
160        if magic == b'tjgg':
161            if version < 1 or version > 3:
162                raise ValueError(f'Cannot handle unexpected GGJT file version {version}')
163            self.file_format = GGMLFormat.GGJT
164            self.format_version = version
165            return 8
166        raise ValueError(f"Unexpected file magic {magic!r}! This doesn't look like a GGML format file.")
167
168    def validate_conversion(self, ftype):
169        err = ''
170        if (self.file_format < GGMLFormat.GGJT or self.format_version < 2):
171            if ftype not in (GGMLFType.ALL_F32, GGMLFType.MOSTLY_F16):
172                err = 'Quantizations changed in GGJTv2. Can only convert unquantized GGML files older than GGJTv2.'
173        elif (self.file_format == GGMLFormat.GGJT and self.format_version == 2):
174            if ftype in (GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1,
175                         GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0):
176                err = 'Q4 and Q8 quantizations changed in GGJTv3.'
177        if len(err) > 0:
178            raise ValueError(f'{err} Sorry, your {self.file_format.name}v{self.format_version} file of type {ftype.name} is not eligible for conversion.')
179
180    def load(self, data, offset):
181        offset += self.validate_header(data, offset)
182        hp = Hyperparameters()
183        offset += hp.load(data, offset)
184        logger.info(f'* File format: {self.file_format.name}v{self.format_version} with ftype {hp.ftype.name}')
185        self.validate_conversion(hp.ftype)
186        vocab = Vocab(load_scores = self.file_format > GGMLFormat.GGML)
187        offset += vocab.load(data, offset, hp.n_vocab)
188        tensors: list[Tensor] = []
189        tensor_map = {}
190        while offset < len(data):
191            tensor = Tensor(use_padding = self.file_format > GGMLFormat.GGMF)
192            offset += tensor.load(data, offset)
193            tensor_map[tensor.name] = len(tensors)
194            tensors.append(tensor)
195        self.hyperparameters = hp
196        self.vocab = vocab
197        self.tensors = tensors
198        self.tensor_map = tensor_map
199        hp.set_n_ff(self)
200        return offset
201
202
203class GGMLToGGUF:
204    def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None):
205        hp = ggml_model.hyperparameters
206        self.model = ggml_model
207        self.data = data
208        self.cfg = cfg
209        self.params_override = params_override
210        self.vocab_override = vocab_override
211        self.special_vocab = special_vocab
212        if params_override is not None:
213            n_kv_head = params_override.n_head_kv
214        else:
215            if cfg.gqa == 1:
216                n_kv_head = hp.n_head
217            else:
218                gqa = float(cfg.gqa)
219                n_kv_head = None
220                for x in range(1, 256):
221                    if float(hp.n_head) / float(x) == gqa:
222                        n_kv_head = x
223                assert n_kv_head is not None, "Couldn't determine n_kv_head from GQA param"
224                logger.info(f'- Guessed n_kv_head = {n_kv_head} based on GQA {cfg.gqa}')
225        self.n_kv_head = n_kv_head
226        self.name_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.LLAMA, ggml_model.hyperparameters.n_layer)
227
228    def save(self):
229        logger.info('* Preparing to save GGUF file')
230        gguf_writer = gguf.GGUFWriter(
231            self.cfg.output,
232            gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA],
233            use_temp_file = False)
234        self.add_params(gguf_writer)
235        self.add_vocab(gguf_writer)
236        if self.special_vocab is not None:
237            self.special_vocab.add_to_gguf(gguf_writer)
238        self.add_tensors(gguf_writer)
239        logger.info("    gguf: write header")
240        gguf_writer.write_header_to_file()
241        logger.info("    gguf: write metadata")
242        gguf_writer.write_kv_data_to_file()
243        logger.info("    gguf: write tensors")
244        gguf_writer.write_tensors_to_file()
245        gguf_writer.close()
246
247    def add_params(self, gguf_writer):
248        hp = self.model.hyperparameters
249        cfg = self.cfg
250        if cfg.desc is not None:
251            desc = cfg.desc
252        else:
253            desc = f'converted from legacy {self.model.file_format.name}v{self.model.format_version} {hp.ftype.name} format'
254        try:
255            # Filenames aren't necessarily valid UTF8.
256            name = cfg.name if cfg.name is not None else cfg.input.name
257        except UnicodeDecodeError:
258            name = None
259        logger.info('* Adding model parameters and KV items')
260        if name is not None:
261            gguf_writer.add_name(name)
262        gguf_writer.add_description(desc)
263        gguf_writer.add_file_type(int(hp.ftype))
264        if self.params_override is not None:
265            po = self.params_override
266            assert po.n_embd == hp.n_embd, 'Model hyperparams mismatch'
267            assert po.n_layer == hp.n_layer, 'Model hyperparams mismatch'
268            assert po.n_head == hp.n_head, 'Model hyperparams mismatch'
269            gguf_writer.add_context_length      (po.n_ctx)
270            gguf_writer.add_embedding_length    (po.n_embd)
271            gguf_writer.add_block_count         (po.n_layer)
272            gguf_writer.add_feed_forward_length (po.n_ff)
273            gguf_writer.add_rope_dimension_count(po.n_embd // po.n_head)
274            gguf_writer.add_head_count          (po.n_head)
275            gguf_writer.add_head_count_kv       (po.n_head_kv)
276            gguf_writer.add_layer_norm_rms_eps  (po.f_norm_eps)
277            return
278        gguf_writer.add_context_length(cfg.context_length)
279        gguf_writer.add_embedding_length(hp.n_embd)
280        gguf_writer.add_block_count(hp.n_layer)
281        gguf_writer.add_feed_forward_length(hp.n_ff)
282        gguf_writer.add_rope_dimension_count(hp.n_embd // hp.n_head)
283        gguf_writer.add_head_count(hp.n_head)
284        gguf_writer.add_head_count_kv(self.n_kv_head)
285        gguf_writer.add_layer_norm_rms_eps(float(cfg.eps))
286
287    def add_vocab(self, gguf_writer):
288        hp = self.model.hyperparameters
289        gguf_writer.add_tokenizer_model('llama')
290        gguf_writer.add_tokenizer_pre('default')
291        tokens = []
292        scores = []
293        toktypes = []
294        if self.vocab_override is not None:
295            vo = self.vocab_override
296            logger.info('* Adding vocab item(s)')
297            for (_, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
298                tokens.append(vbytes)
299                scores.append(score)
300                toktypes.append(ttype)
301            assert len(tokens) == hp.n_vocab, \
302                f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}'
303            gguf_writer.add_token_list(tokens)
304            gguf_writer.add_token_scores(scores)
305            if len(toktypes) > 0:
306                gguf_writer.add_token_types(toktypes)
307            return
308        logger.info(f'* Adding {hp.n_vocab} vocab item(s)')
309        assert len(self.model.vocab.items) >= 3, 'Cannot handle unexpectedly short model vocab'
310        for (tokid, (vbytes, vscore)) in enumerate(self.model.vocab.items):
311            tt = 1 # Normal
312            # Special handling for UNK, BOS, EOS tokens.
313            if tokid <= 2:
314                if tokid == 0:
315                    vbytes = b'<unk>'
316                    tt = 2
317                elif tokid == 1:
318                    vbytes = b'<s>'
319                    tt = 3
320                else:
321                    vbytes = b'</s>'
322                    tt = 3
323            elif len(vbytes) == 0:
324                tt = 3 # Control
325            elif tokid >= 3 and tokid <= 258 and len(vbytes) == 1:
326                vbytes = bytes(f'<0x{vbytes[0]:02X}>', encoding = 'UTF-8')
327                tt = 6 # Byte
328            else:
329                vbytes = vbytes.replace(b' ', b'\xe2\x96\x81')
330            toktypes.append(tt)
331            tokens.append(vbytes)
332            scores.append(vscore)
333        gguf_writer.add_token_list(tokens)
334        gguf_writer.add_token_scores(scores)
335        gguf_writer.add_token_types(toktypes)
336        gguf_writer.add_unk_token_id(0)
337        gguf_writer.add_bos_token_id(1)
338        gguf_writer.add_eos_token_id(2)
339
340    def add_tensors(self, gguf_writer):
341        tensor_map = self.name_map
342        data = self.data
343        logger.info(f'* Adding {len(self.model.tensors)} tensor(s)')
344        for tensor in self.model.tensors:
345            name = str(tensor.name, 'UTF-8')
346            mapped_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
347            assert mapped_name is not None, f'Bad name {name}'
348            tempdims = list(tensor.dims[:])
349            if len(tempdims) > 1:
350                temp = tempdims[1]
351                tempdims[1] = tempdims[0]
352                tempdims[0] = temp
353            gguf_writer.add_tensor(
354                mapped_name,
355                data[tensor.start_offset:tensor.start_offset + tensor.len_bytes],
356                raw_shape = tempdims,
357                raw_dtype = tensor.dtype)
358
359
360def handle_metadata(cfg, hp):
361    import examples.convert_legacy_llama as convert
362
363    assert cfg.model_metadata_dir.is_dir(), 'Metadata dir is not a directory'
364    hf_config_path   = cfg.model_metadata_dir / "config.json"
365    orig_config_path = cfg.model_metadata_dir / "params.json"
366    # We pass a fake model here. "original" mode will check the shapes of some
367    # tensors if information is missing in the .json file: other than that, the
368    # model data isn't used so this should be safe (at least for now).
369    fakemodel = {
370        'tok_embeddings.weight': convert.LazyTensor.__new__(convert.LazyTensor),
371        'layers.0.feed_forward.w1.weight': convert.LazyTensor.__new__(convert.LazyTensor),
372    }
373    fakemodel['tok_embeddings.weight'].shape = [hp.n_vocab]
374    fakemodel['layers.0.feed_forward.w1.weight'].shape = [hp.n_ff]
375    if hf_config_path.exists():
376        params = convert.Params.loadHFTransformerJson(fakemodel, hf_config_path)
377    elif orig_config_path.exists():
378        params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path)
379    else:
380        raise ValueError('Unable to load metadata')
381    vocab_path = Path(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir)
382    vocab_factory = convert.VocabFactory(vocab_path)
383    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype.split(","), cfg.model_metadata_dir)
384    convert.check_vocab_size(params, vocab)
385    return params, vocab, special_vocab
386
387
388def handle_args():
389    parser = argparse.ArgumentParser(description = 'Convert GGML models to GGUF')
390    parser.add_argument('--input', '-i', type = Path, required = True,
391                        help = 'Input GGMLv3 filename')
392    parser.add_argument('--output', '-o', type = Path, required = True,
393                        help ='Output GGUF filename')
394    parser.add_argument('--name',
395                        help = 'Set model name')
396    parser.add_argument('--desc',
397                        help = 'Set model description')
398    parser.add_argument('--gqa', type = int, default = 1,
399                        help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
400    parser.add_argument('--eps', default = '5.0e-06',
401                        help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
402    parser.add_argument('--context-length', '-c', type=int, default = 2048,
403                        help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
404    parser.add_argument('--model-metadata-dir', '-m', type = Path,
405                        help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
406    parser.add_argument("--vocab-dir", type=Path,
407                        help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
408    parser.add_argument("--vocabtype", default="spm,hfft",
409                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm,hfft)")
410    parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
411    return parser.parse_args()
412
413
414def main():
415    cfg = handle_args()
416    logging.basicConfig(level=logging.DEBUG if cfg.verbose else logging.INFO)
417    logger.info(f'* Using config: {cfg}')
418    logger.warning('=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===')
419    if cfg.model_metadata_dir is None and (cfg.gqa == 1 or cfg.eps == '5.0e-06'):
420        logger.info('- Note: If converting LLaMA2, specifying "--eps 1e-5" is required. 70B models also need "--gqa 8".')
421    data = np.memmap(cfg.input, mode = 'r')
422    model = GGMLModel()
423    logger.info('* Scanning GGML input file')
424    offset = model.load(data, 0)  # noqa
425    logger.info(f'* GGML model hyperparameters: {model.hyperparameters}')
426    vocab_override = None
427    params_override = None
428    special_vocab = None
429    if cfg.model_metadata_dir is not None:
430        (params_override, vocab_override, special_vocab) = handle_metadata(cfg, model.hyperparameters)
431        logger.info('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.')
432        logger.info(f'* Overriding params: {params_override}')
433        logger.info(f'* Overriding vocab: {vocab_override}')
434        logger.info(f'* Special vocab: {special_vocab}')
435    else:
436        logger.warning('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
437        if model.file_format == GGMLFormat.GGML:
438            logger.info('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!')
439    converter = GGMLToGGUF(
440        model, data, cfg,
441        params_override = params_override,
442        vocab_override = vocab_override,
443        special_vocab = special_vocab
444    )
445    converter.save()
446    logger.info(f'* Successful completion. Output saved to: {cfg.output}')
447
448
449if __name__ == '__main__':
450    main()