Created
August 30, 2022 00:54
-
-
Save moyix/b325c0938a09156e7a7157e844a7ab51 to your computer and use it in GitHub Desktop.
Demo of extending a rotary position embedding model to a longer context than it was trained on
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
(sfcodegen) moyix@isabella:~$ python load_codegen_with_longer_context.py | |
vocab_file vocab.json | |
merges_file merges.txt | |
tokenizer_file tokenizer.json | |
added_tokens_file added_tokens.json | |
special_tokens_map_file special_tokens_map.json | |
tokenizer_config_file tokenizer_config.json | |
Partial prompt from /usr/include/stdlib.h: | |
[...] restrict __nptr, | |
char **__restrict __endptr) | |
__THROW __nonnull ((1)); | |
#[Generated text start]endif | |
#if __HAVE_FLOAT128 && __GLIBC_USE (IEC_60559_TYPES_EXT) | |
extern _Float128 strtold (const char *__restrict __nptr, | |
char **__restrict __endptr) | |
__THROW __nonnull ((1)); | |
#endif | |
#if __HAVE_FLOAT128X && __GLIBC_USE (IEC_60559_TYPES_EXT) | |
extern _Float128 strtoldx (const char *__restrict __nptr, | |
char **__restrict __endptr) | |
__THROW __nonnull ((1)); | |
#endif | |
#if __HAVE_FLOAT128 && __GLIBC_USE (IEC_60559_TYPES_EXT) | |
extern _Float128 strtodx (const char *__restrict __nptr, | |
char **__restrict __endptr) | |
__THROW __nonnull ((1)); | |
#endif | |
#if __HAVE_FLOAT128X && __GLIBC_USE (IEC_60559_TYPES_EXT) | |
extern _Float128x strtodx (const char *__restrict __nptr, | |
char **__restrict __endptr) | |
__THROW __nonnull ((1)); | |
#endif | |
#if __HAVE_FLOAT128X && __GLIBC_USE (IEC_60559_TYPES_EXT) | |
extern _Float128x strtodx (const char *__restrict __nptr, | |
char **__restrict __endptr) | |
__THROW __nonnull ((1)); | |
#endif | |
#if __HAVE_CONFIG_H | |
#include <config.h> | |
#endif | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h> | |
#include <sys/types.h | |
#include <sys/types.h | |
#endif | |
#include <sys/types.h | |
#include <sys/types.h> | |
#endif | |
#include <sys/types.h> | |
#include <sys/types.h | |
#include <sys/sys | |
#include <sys/types.h> | |
#include <sys/sys | |
#endif | |
#include <sys | |
#include <sys | |
# | |
# | |
# | |
#include | |
#endif | |
# | |
# | |
# | |
# | |
# | |
#include | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# Ifl | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# H | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
*/ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#c | |
# | |
#l_C_c_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# Copyrighted | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#c | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# /* | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#_ | |
# | |
# | |
#) | |
# | |
# | |
# | |
#_p_ | |
# | |
# | |
#c_ | |
# | |
# | |
# | |
# _ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#adm_ | |
# | |
# | |
# | |
# | |
#_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#include | |
#include | |
#_ | |
# | |
# | |
#edl_D_t | |
#include | |
#include | |
#include | |
#include | |
#include | |
#include | |
#include | |
#include_H_Hd_DRSF_H_H_H | |
#end_ | |
# | |
#include_c_ | |
# | |
#include | |
#include | |
#crun_ | |
# | |
#C_ | |
# | |
#r_c | |
# | |
#include | |
# | |
# | |
# | |
#lib_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#include _ | |
#r_t | |
# | |
#include | |
#include | |
# | |
#include | |
#cubd_start_m_D_un_start_un_D_start_H_m_L4_m_ | |
# | |
# | |
# | |
# | |
_r_in_ | |
#c_ | |
# | |
#ifn_H_L_h_ | |
#c_L_# | |
#r_C_ | |
#c_ | |
#include | |
#c_ | |
# | |
#c_L_ | |
# | |
#r1 | |
#r_ | |
#r_# | |
#include | |
# | |
# | |
# | |
#c_ | |
# | |
# | |
# | |
#if_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#pr_ | |
#if_free_free_p | |
#if_C_ | |
# | |
#c_c_c_ | |
# | |
#pr_C_INTRACL_L_ | |
# | |
# | |
# | |
#r_H_P_r_r_H_ | |
# | |
#if_in_H_legal_ | |
# | |
#if_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#r_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#include | |
# | |
#include | |
# | |
# | |
#include | |
# | |
# | |
#legal | |
# | |
# | |
# | |
# | |
# | |
#pr_legal_locp | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#pr_ | |
#_ | |
#p_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#_ | |
#_ | |
#_ | |
#_) | |
#_ | |
# | |
#_ | |
#_ | |
# | |
# | |
# | |
#_ | |
#_except_H_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#_ | |
# | |
# | |
# | |
# | |
# | |
#_ | |
#_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
#_ | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
# | |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
#!/usr/bin/env python | |
import sys | |
import torch | |
from transformers import CodeGenConfig, CodeGenForCausalLM, AutoTokenizer | |
from transformers.utils.hub import cached_file | |
COLOR_BOLD = '\033[1m' | |
COLOR_RESET = '\033[0m' | |
NEW_SIZE = 4096 | |
cg_config = CodeGenConfig.from_pretrained('Salesforce/codegen-350M-multi') | |
cg_config.n_ctx = NEW_SIZE | |
cg_config.n_positions = NEW_SIZE | |
weights_file = cached_file('Salesforce/codegen-350M-multi', 'pytorch_model.bin') | |
state_dict = torch.load(weights_file) | |
# Expand the causal mask to the new size. Unclear why this is in the saved | |
# model weights, because it's a constant... | |
for k in list(state_dict.keys()): | |
if k.endswith('causal_mask'): | |
# This is copied from CodeGen's __init__() method. | |
state_dict[k] = torch.tril(torch.ones((NEW_SIZE, NEW_SIZE), dtype=torch.uint8)).view( | |
1, 1, NEW_SIZE, NEW_SIZE | |
) | |
model = CodeGenForCausalLM.from_pretrained(None, config=cg_config, state_dict=state_dict, torch_dtype='auto') | |
model.to('cuda') | |
model.eval() | |
# Try to generate something | |
filename = '/usr/include/stdlib.h' | |
text = open(filename).read() | |
tokenizer = AutoTokenizer.from_pretrained('Salesforce/codegen-350M-multi') | |
tokenizer.model_max_length = sys.maxsize | |
enc = tokenizer.encode(text) | |
enc = enc[:2048] | |
enc = torch.tensor(enc, dtype=torch.long).unsqueeze(0) | |
enc = enc.to('cuda') | |
with torch.no_grad(): | |
out = model.generate( | |
input_ids=enc, | |
do_sample=True, | |
temperature=0.2, | |
top_p=1.0, | |
max_new_tokens=2048, | |
num_return_sequences=1, | |
pad_token_id=tokenizer.eos_token_id, | |
eos_token_id=tokenizer.eos_token_id, | |
) | |
print(f"Generated {len(out[0])-2048} tokens.") | |
print(f"Partial prompt from {filename}:\n") | |
print(f"{COLOR_BOLD}[...]{COLOR_RESET} ", end='') | |
print(tokenizer.decode(out[0][2048-32:2048].cpu().tolist()), end='') | |
print(f"{COLOR_BOLD}[Generated text start]{COLOR_RESET}", end='') | |
print(tokenizer.decode(out[0][2048:].cpu().tolist())) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment