File size: 2,548 Bytes
c25e25e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2023, Tri Dao, Albert Gu.

import argparse
import time
import json

import torch
import torch.nn.functional as F

from einops import rearrange

from transformers import AutoTokenizer, AutoModelForCausalLM, LlamaTokenizer

from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel


parser = argparse.ArgumentParser(description="Generation benchmarking")
parser.add_argument("--model-name", type=str, default="state-spaces/mamba-130m")
parser.add_argument("--prompt", type=str, default=None)
parser.add_argument("--promptlen", type=int, default=100)
parser.add_argument("--genlen", type=int, default=100)
parser.add_argument("--temperature", type=float, default=1.0)
parser.add_argument("--topk", type=int, default=1)
parser.add_argument("--topp", type=float, default=1.0)
parser.add_argument("--repetition-penalty", type=float, default=1.0)
parser.add_argument("--batch", type=int, default=1)
args = parser.parse_args()

repeats = 3
device = "cuda"
dtype = torch.float16

print(f"Loading model {args.model_name}")
tokenizer = LlamaTokenizer.from_pretrained(args.model_name)
model = MambaLMHeadModel.from_pretrained(args.model_name, device=device, dtype=dtype)
model.eval()
print(f"Number of parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}")

torch.random.manual_seed(0)
if args.prompt is None:
    input_ids = torch.randint(1, 1000, (args.batch, args.promptlen), dtype=torch.long, device="cuda")
    attn_mask = torch.ones_like(input_ids, dtype=torch.long, device="cuda")
else:
    args.prompt = args.prompt.replace('\\n', '\n')
    tokens = tokenizer(args.prompt, return_tensors="pt")
    input_ids = tokens.input_ids.to(device=device)
    attn_mask = tokens.attention_mask.to(device=device)
max_length = input_ids.shape[1] + args.genlen

fn = lambda: model.generate(
    input_ids=input_ids,
    max_length=max_length,
    cg=True,
    return_dict_in_generate=True,
    output_scores=True,
    enable_timing=False,
    temperature=args.temperature,
    top_k=args.topk,
    top_p=args.topp,
    repetition_penalty=args.repetition_penalty,
)
out = fn()
if args.prompt is not None:
    l = tokenizer.batch_decode(out.sequences.tolist())
    text = ''.join(l)
    print(text)

torch.cuda.synchronize()
start = time.time()
for _ in range(repeats):
    fn()
torch.cuda.synchronize()
print(f"Prompt length: {len(input_ids[0])}, generation length: {len(out.sequences[0]) - len(input_ids[0])}")
print(f"{args.model_name} prompt processing + decoding time: {(time.time() - start) / repeats * 1000:.0f}ms")