rwkv-14b / config.py
gururise's picture
update to latest rwkvstic
f9ef5de
from rwkvstic.agnostic.backends import TORCH, TORCH_QUANT
import torch
quantized = {
"mode": TORCH_QUANT,
"runtimedtype": torch.bfloat16,
"useGPU": torch.cuda.is_available(),
"chunksize": 32, # larger = more accurate, but more memory (and slower)
"target": 15 # your gpu max size, excess vram offloaded to cpu
}
# UNCOMMENT TO SELECT OPTIONS
# Not full list of options, see https://pypi.org/project/rwkvstic/ and https://huggingface.co/BlinkDL/ for more models/modes
# RWKV 1B5 instruct test 2 model
# Approximate
# [Vram usage: 6.0GB]
# [File size: 3.0GB]
# config = {
# "path": "https://huggingface.co/BlinkDL/rwkv-4-pile-1b5/resolve/main/RWKV-4-Pile-1B5-Instruct-test2-20230209.pth",
# "mode": TORCH,
# "runtimedtype": torch.float32,
# "useGPU": torch.cuda.is_available(),
# "dtype": torch.float32
# }
# title = "RWKV-4 (1.5b Instruct Test 2)"
# RWKV 1B5 instruct model quantized
# Approximate
# [Vram usage: 1.3GB]
# [File size: 3.0GB]
# config = {
# "path": "https://huggingface.co/BlinkDL/rwkv-4-pile-1b5/resolve/main/RWKV-4-Pile-1B5-Instruct-test1-20230124.pth",
# **quantized
# }
# title = "RWKV-4 (1.5b Instruct Quantized)"
# RWKV 7B instruct pre-quantized (settings baked into model)
# Approximate
# [Vram usage: 7.0GB]
# [File size: 8.0GB]
# config = {
# "path": "https://huggingface.co/Hazzzardous/RWKV-8Bit/resolve/main/RWKV-4-Pile-7B-Instruct.pqth"
# }
# title = "RWKV-4 (7b Instruct Quantized)"
# RWKV 14B quantized (latest as of feb 9)
# Approximate
# [Vram usage: 15.0GB]
# [File size: 28.0GB]
# config = {
# "path": "https://huggingface.co/BlinkDL/rwkv-4-pile-14b/resolve/main/RWKV-4-Pile-14B-20230204-7324.pth",
# **quantized
# }
# title = "RWKV-4 (14b Quantized)"
# RWKV 14B quantized (latest as of feb 13)
# Approximate
# [Vram usage: 15.0GB]
# [File size: 14.4GB]
config = {
# "path": "https://huggingface.co/Hazzzardous/RWKV-8Bit/resolve/main/RWKV-4-Pile-14B-20230204-7324.pqth"
"path": "https://huggingface.co/yahma/RWKV-14b_quant/resolve/main/RWKV-4-Pile-14B-20230213-8019.pqth"
}
title = "RWKV-4 (14b Quantized - Feb 13)"
# RWKV 14B (latest as of feb 9)
# Approximate
# [Vram usage: 27.0GB]
# [File size: 28.4GB]
# config = {
# "path": "https://huggingface.co/BlinkDL/rwkv-4-pile-14b/resolve/main/RWKV-4-Pile-14B-20230204-7324.pth",
# "mode": TORCH,
# "runtimedtype": torch.bfloat16,
# "useGPU": torch.cuda.is_available(),
# "dtype": torch.bfloat16
# }
# title = "RWKV-4 (14b Feb 4 Snapshot)"