rwkv-x 0.1.8

The RWKV-X Language Model

The RWKV-X Language Model

https://github.com/howard-hou/RWKV-X

# !!! set these before import RWKV !!!
import os

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
# fp32 = good for CPU
# bf16 = 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_x.model import RWKV_X
from rwkv_x.utils import PIPELINE, PIPELINE_ARGS

# download models: https://huggingface.co/howard-hou/RWKV-X/tree/main
model = RWKV_X(model_path='RWKV-X-0.2B-64k-Base.pth', strategy='cpu fp32')

pipeline = PIPELINE(model) # for "world" models

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 = [], # 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')

# !!! model.forward(tokens, state) will modify state in-place !!!

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')