File size: 5,487 Bytes
76b1ec5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3

from .promptops import *

from .aux import CmdlineArgs, log
from .data import get_data_loader
from .trainllm import env_stuff, load_model, load_tokenizer


import sys
import torch
import json
import torch.distributed as dist

from accelerate import Accelerator

from datetime import datetime

"""
This currently assumes the batch size to be 1. With larger batches the padding tokens went
into the decoder. Right-padding as a solution?
"""
def llm_generate(model, tokenizer, tok_batch, debug=False, max_len=2000):
    tok_batch['input_ids'] = tok_batch['input_ids'].to(model.device)
    tok_batch['attention_mask'] = tok_batch['attention_mask'].to(model.device)
    start_time = datetime.now()

    if debug:
        log(f"Tokenized input: {tok_batch['input_ids']}")

    raw_output_toks = model.generate(**tok_batch, tokenizer=tokenizer,
                                 do_sample=False, num_beams=4, max_length=max_len, top_p=None, temperature=None,
                                 eos_token_id=[tokenizer.eos_token_id,
                                               tokenizer.convert_tokens_to_ids("<|reserved_special_token_14|>")])

    #clean_output_toks = remove_prompt_from_output(tok_batch['attention_mask'], raw_output_toks, filler_id)
    assert len(raw_output_toks) == 1, "Only batch size=1 supported %-("
    gen_idx = len(tok_batch['attention_mask'][0])

    if debug:
        log(f"Full tokenized output: {raw_output_toks[0]}")
        log(f"Full tokens: {tokenizer.convert_ids_to_tokens(raw_output_toks[0])}")
        full_out = tokenizer.batch_decode([raw_output_toks[0]], skip_special_tokens=True)
        log(f"Full text: {full_out[0]}")

    clean_output_toks = raw_output_toks[0][gen_idx:]
    clean_outputs = tokenizer.batch_decode([clean_output_toks], skip_special_tokens=True)

    if debug:
        log(f"Pruned tokenized output: {clean_output_toks}")
        log(f"Pruned tokens: {tokenizer.convert_ids_to_tokens(clean_output_toks)}")
        log(f"Cleaned output: {clean_outputs[0]}")

        end_time = datetime.now()
        log(f"This took: {end_time - start_time}")

    return clean_outputs


def reassemble_multi(list_of_lists):
    result = []

    for gen_idx in range(len(list_of_lists[0])):
        for i in range(len(list_of_lists)):
            if gen_idx < len(list_of_lists[i]):
                result.append(list_of_lists[i][gen_idx])

    return result


def predict(model, tokenizer, data_loader, accel, multi=False, debug=False, max_len=2000):
    outs_final = []

    with torch.no_grad():
        for idx, batch in enumerate(data_loader):
            if idx % accel.num_processes == accel.process_index:
                start_time = datetime.now()
                outputs = llm_generate(model, tokenizer, batch, debug=debug, max_len=max_len)
                end_time = datetime.now()
                log(f"Generated for {idx} in proc {accel.process_index} in {end_time - start_time}")
                outs_final += outputs

    if multi:
        accel.wait_for_everyone()

        rank0_buffer = [None] * accel.num_processes if accel.is_main_process else None
        dist.gather_object(outs_final, rank0_buffer, dst=0)
        if accel.is_main_process:
            outs_final = reassemble_multi(rank0_buffer)
        else:
            outs_final = None

    return outs_final


def _cmdline_args():
    inputs = sys.argv[1:]

    description = """Predict output for an input via prompting"""

    pos_args = ["mdl_id"]

    #post-process the arguments
    args = CmdlineArgs(description, pos_args, input_args=inputs,
                       kw_arg_dict={"debug": False,
                                    "input_file": "none",
                                    "output_file": "none",
                                    "multiproc": False,
                                    "max_len": 2000,
                                    "prompt_format": PF_ALPACA})

    if args.input_file == "none":
        args.input_file = None
    if args.output_file == "none":
        args.output_file = None

    log(f"Launched as {args}")

    return args


def save_all(outputs, args, acc):
    if acc.is_main_process:
        if args.output_file is None:
            log("Writing to STDOUT")
            out_fh = sys.stdout
        else:
            out_fh = open(args.output_file, "w")

        if args.prompt_format in {PF_RAW, PF_RAWLINES}:
            for line in outputs:
                out_fh.write(line + "\n")
        else:
            json.dump(outputs, out_fh)


def and_i_called_this_function_do_main_too():
    args = _cmdline_args()

    if args.multiproc:
        env_stuff()

    acc = Accelerator()
    device = acc.device

    log(f"Device: {device}.", accelerator=acc)

    if not args.multiproc and not acc.is_main_process:
        log("Not launched in multi-processing mode, exiting non-main process.")
        sys.exit(0)

    tokenizer = load_tokenizer(args.mdl_id, acc)

    data_loader = get_data_loader(args.input_file, args.prompt_format, tokenizer, debug=args.debug)

    model = load_model(args.mdl_id, device, acc, attention="eager")
    model.eval()

    log(f"Device: {model.device}.", accelerator=acc)

    log("Model loaded, starting to generate")
    outputs = predict(model, tokenizer, data_loader, acc, multi=args.multiproc, debug=args.debug, max_len=args.max_len)

    save_all(outputs, args, acc)

    log("Done")


if __name__ == "__main__":
    and_i_called_this_function_do_main_too()