-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathgenerate_rana_model.py
182 lines (144 loc) · 8.45 KB
/
generate_rana_model.py
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
import argparse
import os
import pandas as pd
import torch
import torch
import pandas as pd
from tqdm import tqdm
from transformers import AutoTokenizer, AutoModelForCausalLM
from model_specific_utils import replace_qkv_layers_llama
from basis_learners import *
from input_samples import *
from input_samples import _get_input_tensors
from rank_adaptive_experts import *
from dataset import get_baseline_dataset_v2
from rank_adaptive import *
from model_specific_utils import get_mlps_llama
from basis_learners import _get_initial_lora_decomp
MODEL_REVISIONS = {
"meta-llama/Llama-2-7b-hf": "01c7f73d771dfac7d292323805ebc428287df4f9",
"google/gemma-2b": "68e273d91b1d6ea57c9e6024c4f887832f7b43fa",
}
def genereate_rana(PRUNE_RATIO, MODEL, DEVICE, DATASET_FILENAME, SAVE_FILENAME, PRE_LOAD_MODEL_STATE_DICT):
os.environ["CUDA_VISIBLE_DEVICES"] = f"{DEVICE}"
tokenizer = AutoTokenizer.from_pretrained(MODEL, revision=MODEL_REVISIONS[MODEL])
model = AutoModelForCausalLM.from_pretrained(MODEL, revision=MODEL_REVISIONS[MODEL])
if PRE_LOAD_MODEL_STATE_DICT is not None:
model.load_state_dict(torch.torch.load(PRE_LOAD_MODEL_STATE_DICT))
is_llama = MODEL == "meta-llama/Llama-2-7b-hf"
if is_llama:
model = model.half()
replace_qkv_layers_llama(model)
dataset = get_baseline_dataset_v2("togethercomputer/RedPajama-Data-1T-Sample", f"{DATASET_FILENAME}.pkl", train_size=100_000)
train_data, eval_data = dataset["train"], dataset["test"]
ttl_tokens_mult = 8
inp_sample_size = 32_000
seq_size = 1024
tokenized_data = get_tokenized_train_data(train_data.shuffle(seed=123), tokenizer, ttl_tokens=ttl_tokens_mult*inp_sample_size, seq_size=seq_size)
def custom_get_linears_llama(m):
linear_layers = []
for layer in model.model.layers:
attn = layer.self_attn
mlp = layer.mlp
linear_layers.append([attn.q_proj.stacked_qkv.qkv, mlp.up_proj, mlp.gate_proj])
return linear_layers
def custom_get_linears_gemma(m):
linear_layers = []
for layer in model.model.layers:
mlp = layer.mlp
linear_layers.append([mlp.up_proj, mlp.gate_proj])
return linear_layers
custom_get_linears = custom_get_linears_llama if is_llama else custom_get_linears_gemma
input_tensors = _get_input_tensors(model, tokenized_data, inp_sample_size=inp_sample_size,
ttl_tokens_mult=ttl_tokens_mult, seq_size=seq_size, mem_efficient=True, get_linears=custom_get_linears)
def _float_32_get_initial_lora_decomp(module, inp_tens, lora_decomp_func=get_lora_decomp):
torch.cuda.empty_cache()
usv = lora_decomp_func(module.weight.to(torch.float32), inp_tens.T.contiguous().to(torch.float32).cuda())
lora_decomp = LoRaDecomp(module, Q=usv[0].to(torch.float16), B=(usv[1]*usv[2]).to(torch.float16), X=inp_tens.T.contiguous())
torch.cuda.empty_cache()
return lora_decomp
get_initial_lora_decomp_func = _float_32_get_initial_lora_decomp if is_llama else _get_initial_lora_decomp
lora_decomps = get_initial_lora_decomps(model, input_tensors, custom_get_linears, get_initial_lora_decomp_func=get_initial_lora_decomp_func)
lora_decomps = [lora_decomps[i] for i in range(len(lora_decomps))]
mlps = get_mlps_llama(model)
if is_llama:
qkvs = [layer.self_attn.q_proj.stacked_qkv.qkv for layer in model.model.layers]
qkv_decomps = [lora_decomps[i][0] for i in range(len(lora_decomps))]
up_proj_decomps = [lora_decomps[i][1] for i in range(len(lora_decomps))]
gate_proj_decomps = [lora_decomps[i][2] for i in range(len(lora_decomps))]
up_proj_inps = [input_tensors[i][1] for i in range(len(input_tensors))]
qkv_inps = [input_tensors[i][0] for i in range(len(input_tensors))]
mlp_outs = []
qkv_outs = []
else:
up_proj_decomps = [lora_decomps[i][0] for i in range(len(lora_decomps))]
gate_proj_decomps = [lora_decomps[i][1] for i in range(len(lora_decomps))]
up_proj_inps = [input_tensors[i][0] for i in range(len(input_tensors))]
mlp_outs = []
model.eval()
with torch.no_grad():
if is_llama:
assert len(qkvs) == len(qkv_inps)
for qkv, qkv_inp in zip(qkvs, qkv_inps):
qkv_outs.append(qkv(qkv_inp.cuda()).detach().cpu())
assert len(mlps) == len(up_proj_inps)
for mlp, up_inp in zip(mlps, up_proj_inps):
mlp_outs.append(mlp(up_inp.cuda()).detach().cpu())
model.train()
import gc; gc.collect()
torch.cuda.empty_cache()
if is_llama:
print("Computing RaNA QKVs ...")
assert len(qkvs) == len(qkv_outs) == len(qkv_decomps)
rank_adaptive_qkvs = []
for orig_qkv, qkv_out, qkv_proj_decomp in zip(qkvs, qkv_outs, qkv_decomps):
orig_qkv_32 = orig_qkv.float()
qkv_out_32 = qkv_out.to(torch.float32)
qkv_proj_decomp_32 = LoRaDecomp(qkv_proj_decomp.module.float(), qkv_proj_decomp.Q.to(torch.float32), qkv_proj_decomp.B.to(torch.float32), qkv_proj_decomp.X.to(torch.float32), out=(None,))
rank_adaptive_qkvs.append(get_best_rank_adaptive_qkv(orig_qkv_32, qkv_proj_decomp_32, PRUNE_RATIO))
orig_qkv_32.half()
qkv_proj_decomp_32.module.half()
print("Computing RaNA MLPs ...")
assert len(mlps) == len(mlp_outs) == len(up_proj_decomps) == len(gate_proj_decomps)
rank_adaptive_mlps = []
for orig_mlp, mlp_out, up_proj_decomp, gate_proj_decomp in zip(mlps, mlp_outs, up_proj_decomps, gate_proj_decomps):
orig_mlp_32 = orig_mlp.float()
mlp_out_32 = mlp_out.to(torch.float32)
up_proj_decomp_32 = LoRaDecomp(up_proj_decomp.module.float(), up_proj_decomp.Q.to(torch.float32), up_proj_decomp.B.to(torch.float32), up_proj_decomp.X.to(torch.float32), out=(None,))
gate_proj_decomp_32 = LoRaDecomp(gate_proj_decomp.module.float(), gate_proj_decomp.Q.to(torch.float32), gate_proj_decomp.B.to(torch.float32), gate_proj_decomp.X.to(torch.float32), out=(None,))
rank_adaptive_mlps.append(get_best_rank_adaptive_mlp_up_down_gate(orig_mlp_32, up_proj_decomp_32, gate_proj_decomp_32, PRUNE_RATIO, mlp_out_32, ret_descr_ratio=True))
orig_mlp_32.half()
up_proj_decomp_32.module.half()
gate_proj_decomp_32.module.half()
else:
print("Computing RaNA MLPs ...")
assert len(mlps) == len(mlp_outs) == len(up_proj_decomps) == len(gate_proj_decomps)
rank_adaptive_mlps = []
for orig_mlp, mlp_out, up_proj_decomp, gate_proj_decomp in zip(mlps, mlp_outs, up_proj_decomps, gate_proj_decomps):
rank_adaptive_mlps.append(get_best_rank_adaptive_mlp_up_down_gate(orig_mlp, up_proj_decomp, gate_proj_decomp, PRUNE_RATIO, mlp_out, ret_descr_ratio=True))
for mlp in rank_adaptive_mlps:
mlp[0].down_proj = None
os.makedirs("./rana_models/", exist_ok=True)
pd.to_pickle(rank_adaptive_mlps, f"./rana_models/{SAVE_FILENAME}_{PRUNE_RATIO}_mlps.pkl")
if is_llama:
pd.to_pickle(rank_adaptive_qkvs, f"./rana_models/{SAVE_FILENAME}_{PRUNE_RATIO}_qkvs.pkl")
print("DONE GENERATING RANA")
def main():
parser = argparse.ArgumentParser(description="generate_rana_model")
parser.add_argument("--model", default="meta-llama/Llama-2-7b-hf", type=str, help="Model Name")
parser.add_argument("--prune_ratio", type=float, help="Prune Ratio", required=True)
parser.add_argument("--device", default=0, type=int, help="CUDA Device ID")
parser.add_argument("--dataset_filename", default="red_pajama_dataset", type=str, help="dataset file name (where to save / where to load from the dataset")
parser.add_argument("--save_filename", default="rana_non_ft", type=str, help="model file name (where to save the rana model)")
parser.add_argument("--pre-load-model-state-dict", default=None, type=str, help="State-dict to load into model before generating rana adapters")
args = parser.parse_args()
model = args.model
prune_ratio = args.prune_ratio
device = args.device
dataset_filename = args.dataset_filename
save_filename = args.save_filename
pre_load_model_state_dict = args.pre_load_model_state_dict
assert model in MODEL_REVISIONS, f"MODEL: {model} not supported"
genereate_rana(prune_ratio, model, device, dataset_filename, save_filename, pre_load_model_state_dict)
if __name__ == "__main__":
main()