The RWKV Language Model
Project description
The RWKV Language Model
https://github.com/BlinkDL/ChatRWKV
https://github.com/BlinkDL/RWKV-LM
# set these before import RWKV
os.environ['RWKV_JIT_ON'] = '1'
os.environ["RWKV_CUDA_ON"] = '0' # '1' to compile CUDA kernel (10x faster), requires c++ compiler & cuda libraries
########################################################################################################
#
# Use '/' in model path, instead of '\'. Use ctx4096 models if you need long ctx.
#
# fp16 = good for GPU (!!! DOES NOT support CPU !!!)
# fp32 = good for CPU
# bf16 = worse accuracy, supports CPU
# xxxi8 (example: fp16i8, fp32i8) = xxx with int8 quantization to save 50% VRAM/RAM, slower, slightly less accuracy
#
# We consider [ln_out+head] to be an extra layer, so L12-D768 (169M) has "13" layers, L24-D2048 (1.5B) has "25" layers, etc.
# Strategy Examples: (device = cpu/cuda/cuda:0/cuda:1/...)
# 'cpu fp32' = all layers cpu fp32
# 'cuda fp16' = all layers cuda fp16
# 'cuda fp16i8' = all layers cuda fp16 with int8 quantization
# 'cuda fp16i8 *10 -> cpu fp32' = first 10 layers cuda fp16i8, then cpu fp32 (increase 10 for better speed)
# 'cuda:0 fp16 *10 -> cuda:1 fp16 *8 -> cpu fp32' = first 10 layers cuda:0 fp16, then 8 layers cuda:1 fp16, then cpu fp32
#
# Basic Strategy Guide: (fp16i8 works for any GPU)
# 100% VRAM = 'cuda fp16' # all layers cuda fp16
# 98% VRAM = 'cuda fp16i8 *1 -> cuda fp16' # first 1 layer cuda fp16i8, then cuda fp16
# 96% VRAM = 'cuda fp16i8 *2 -> cuda fp16' # first 2 layers cuda fp16i8, then cuda fp16
# 94% VRAM = 'cuda fp16i8 *3 -> cuda fp16' # first 3 layers cuda fp16i8, then cuda fp16
# ...
# 50% VRAM = 'cuda fp16i8' # all layers cuda fp16i8
# 48% VRAM = 'cuda fp16i8 -> cpu fp32 *1' # most layers cuda fp16i8, last 1 layer cpu fp32
# 46% VRAM = 'cuda fp16i8 -> cpu fp32 *2' # most layers cuda fp16i8, last 2 layers cpu fp32
# 44% VRAM = 'cuda fp16i8 -> cpu fp32 *3' # most layers cuda fp16i8, last 3 layers cpu fp32
# ...
# 0% VRAM = 'cpu fp32' # all layers cpu fp32
#
# Use '+' for STREAM mode, which can save VRAM too, and it is sometimes faster
# 'cuda fp16i8 *10+' = first 10 layers cuda fp16i8, then fp16i8 stream the rest to it (increase 10 for better speed)
#
# Extreme STREAM: 3G VRAM is enough to run RWKV 14B (slow. will be faster in future)
# 'cuda fp16i8 *0+ -> cpu fp32 *1' = stream all layers cuda fp16i8, last 1 layer [ln_out+head] cpu fp32
#
# ########################################################################################################
from rwkv.model import RWKV
from rwkv.utils import PIPELINE, PIPELINE_ARGS
# download models: https://huggingface.co/BlinkDL
model = RWKV(model='/fsx/BlinkDL/HF-MODEL/rwkv-4-pile-169m/RWKV-4-Pile-169M-20220807-8023', strategy='cpu fp32')
pipeline = PIPELINE(model, "20B_tokenizer.json") # 20B_tokenizer.json is in https://github.com/BlinkDL/ChatRWKV
ctx = "\nIn a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley, in Tibet. Even more surprising to the researchers was the fact that the dragons spoke perfect Chinese."
print(ctx, end='')
def my_print(s):
print(s, end='', flush=True)
# For alpha_frequency and alpha_presence, see "Frequency and presence penalties":
# https://platform.openai.com/docs/api-reference/parameter-details
args = PIPELINE_ARGS(temperature = 1.0, top_p = 0.7, top_k = 100, # top_k = 0 then ignore
alpha_frequency = 0.25,
alpha_presence = 0.25,
alpha_decay = 0.996, # gradually decay the penalty
token_ban = [0], # ban the generation of some tokens
token_stop = [], # stop generation whenever you see any token here
chunk_len = 256) # split input into chunks to save VRAM (shorter -> slower)
pipeline.generate(ctx, token_count=200, args=args, callback=my_print)
print('\n')
out, state = model.forward([187, 510, 1563, 310, 247], None)
print(out.detach().cpu().numpy()) # get logits
out, state = model.forward([187, 510], None)
out, state = model.forward([1563], state) # RNN has state (use deepcopy to clone states)
out, state = model.forward([310, 247], state)
print(out.detach().cpu().numpy()) # same result as above
print('\n')
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
rwkv-beta-0.8.5.tar.gz
(404.1 kB
view hashes)
Built Distribution
rwkv_beta-0.8.5-py3-none-any.whl
(406.2 kB
view hashes)
Close
Hashes for rwkv_beta-0.8.5-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 3c9c1e8d0fa3832899b7a75038d71ccc23f579151844c62279f0d52bc326bc6e |
|
MD5 | b4b2008a2f0b28e1890121c0d56c1cd1 |
|
BLAKE2b-256 | 39a62005dd58bc0eb8423da7b033a9906de1c9e40528ae44797ab70b43d4ee0c |