我正在尝试通过JittorLLMs使用chatglm。
当我执行“python3 cli_demo.py chatglm”时，报错“nvcc fatal : Value ‘compute_30’ is not defined for option ‘gpu-architecture’”。
root@fjuliusd2023:~/JittorLLMs# python3 cli_demo.py chatglm
[i 0418 17:00:29.895090 04 compiler.py:955] Jittor(1.3.7.13) src: /usr/local/lib/python3.8/dist-packages/jittor
[i 0418 17:00:29.897524 04 compiler.py:956] g++ at /usr/bin/g++(9.4.0)
[i 0418 17:00:29.897646 04 compiler.py:957] cache_path: /root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default
[i 0418 17:00:29.902873 04 install_cuda.py:93] cuda_driver_version: [11, 4]
Downloading https://cg.cs.tsinghua.edu.cn/jittor/assets/cuda11.2_cudnn8_linux.tgz to /root/.cache/jittor/jtcuda/cuda11.2_cudnn8_linux.tgz
1.62GB [01:58, 14.8MB/s]
[i 0418 17:02:50.584458 04 __init__.py:411] Found /root/.cache/jittor/jtcuda/cuda11.2_cudnn8_linux/bin/nvcc(11.2.152) at /root/.cache/jittor/jtcuda/cuda11.2_cudnn8_linux/bin/nvcc.
[i 0418 17:02:50.590444 04 __init__.py:411] Found addr2line(2.34) at /usr/bin/addr2line.
[i 0418 17:02:50.703961 04 compiler.py:1010] cuda key:cu11.2.152_sm_30
[i 0418 17:02:50.725154 04 compiler.py:34] Create cache dir: /root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default/cu11.2.152_sm_30
[i 0418 17:02:50.725447 04 compiler.py:34] Create cache dir: /root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default/cu11.2.152_sm_30/jit
[i 0418 17:02:50.725600 04 compiler.py:34] Create cache dir: /root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default/cu11.2.152_sm_30/obj_files
[i 0418 17:02:50.725781 04 compiler.py:34] Create cache dir: /root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default/cu11.2.152_sm_30/gen
[i 0418 17:02:50.725923 04 compiler.py:34] Create cache dir: /root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default/cu11.2.152_sm_30/tmp
[i 0418 17:02:50.726062 04 compiler.py:34] Create cache dir: /root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default/cu11.2.152_sm_30/checkpoints
[i 0418 17:03:12.281893 56 __init__.py:227] Total mem: 31.30GB, using 10 procs for compiling.
Compiling jittor_core(151/151) used: 52.121s eta: 0.000ss
[i 0418 17:04:04.723650 56 jit_compiler.cc:28] Load cc_path: /usr/bin/g++
[i 0418 17:04:04.770611 56 init.cc:62] Found cuda archs: [30,]
[i 0418 17:04:06.258919 56 compile_extern.py:522] mpicc not found, distribution disabled.
[i 0418 17:04:06.259268 56 compiler.py:34] Create cache dir: /root/.cache/jittor/cutt
[i 0418 17:04:06.259480 56 compile_extern.py:339] Downloading cutt...
Downloading https://codeload.github.com/Jittor/cutt/zip/v1.2 to /root/.cache/jittor/cutt/cutt-1.2.zip
296kB [00:01, 265kB/s]
[i 0418 17:04:07.415762 56 compile_extern.py:352] installing cutt...
nvcc fatal : Value 'compute_30' is not defined for option 'gpu-architecture'
nvcc fatal : Value 'compute_30' is not defined for option 'gpu-architecture'
nvcc fatal : Value 'compute_30' is not defined for option 'gpu-architecture'
nvcc fatal : Value 'compute_30' is not defined for option 'gpu-architecture'
nvcc fatal : Value 'compute_30' is not defined for option 'gpu-architecture'
multiprocessing.pool.RemoteTraceback:
"""
Traceback (most recent call last):
File "/usr/lib/python3.8/multiprocessing/pool.py", line 125, in worker
result = (True, func(*args, **kwds))
File "/usr/local/lib/python3.8/dist-packages/jittor_utils/__init__.py", line 197, in do_compile
return cc.cache_compile(cmd, cache_path, jittor_path)
RuntimeError: [f 0418 17:04:07.433644 56 log.cc:608] Check failed ret(256) == 0(0) Run cmd failed: "/root/.cache/jittor/jtcuda/cuda11.2_cudnn8_linux/bin/nvcc" -arch=compute_30 -code=sm_30 "/root/.cache/jittor/cutt/cutt-1.2/src/cuttkernel.cu" -std=c++14 -Xcompiler -fPIC -Xcompiler -march=native -Xcompiler -fdiagnostics-color=always -I"/usr/local/lib/python3.8/dist-packages/jittor/src" -I/usr/include/python3.8 -I/usr/include/python3.8 -DHAS_CUDA -DIS_CUDA -I"/root/.cache/jittor/jtcuda/cuda11.2_cudnn8_linux/include" -I"/usr/local/lib/python3.8/dist-packages/jittor/extern/cuda/inc" -I"/root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default/cu11.2.152_sm_30" -O2 -I"/root/.cache/jittor/cutt/cutt-1.2/include" -I"/root/.cache/jittor/cutt/cutt-1.2/src" -c -o "/root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default/cu11.2.152_sm_30/obj_files/cuttkernel.cu.o" -x cu --cudart=shared -ccbin="/usr/bin/g++" --use_fast_math -w -I"/usr/local/lib/python3.8/dist-packages/jittor/extern/cuda/inc"
"""
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "cli_demo.py", line 8, in <module>
model = models.get_model(args)
File "/root/JittorLLMs/models/__init__.py", line 38, in get_model
globals()[f"get_{model_name}"]()
File "/root/JittorLLMs/models/util.py", line 51, in get_chatglm
new_path.append(download_fromhub(f"jittorhub://{f}", tdir="chat-glm"))
File "/root/JittorLLMs/models/util.py", line 5, in download_fromhub
import jittor as jt
File "/usr/local/lib/python3.8/dist-packages/jittor/__init__.py", line 25, in <module>
from . import compile_extern
File "/usr/local/lib/python3.8/dist-packages/jittor/compile_extern.py", line 589, in <module>
setup_cutt()
File "/usr/local/lib/python3.8/dist-packages/jittor/compile_extern.py", line 387, in setup_cutt
install_cutt(cutt_path)
File "/usr/local/lib/python3.8/dist-packages/jittor/compile_extern.py", line 367, in install_cutt
compile(cc_path, cutt_flags, files2, cache_path+"/libcutt"+so, cuda_flags=arch_flag)
File "/usr/local/lib/python3.8/dist-packages/jittor/compiler.py", line 151, in compile
jit_utils.run_cmds(cmds, cache_path, jittor_path, "Compiling "+base_output)
File "/usr/local/lib/python3.8/dist-packages/jittor_utils/__init__.py", line 251, in run_cmds
for i,_ in enumerate(p.imap_unordered(do_compile, cmds)):
File "/usr/lib/python3.8/multiprocessing/pool.py", line 868, in next
raise value
RuntimeError: [f 0418 17:04:07.433644 56 log.cc:608] Check failed ret(256) == 0(0) Run cmd failed: "/root/.cache/jittor/jtcuda/cuda11.2_cudnn8_linux/bin/nvcc" -arch=compute_30 -code=sm_30 "/root/.cache/jittor/cutt/cutt-1.2/src/cuttkernel.cu" -std=c++14 -Xcompiler -fPIC -Xcompiler -march=native -Xcompiler -fdiagnostics-color=always -I"/usr/local/lib/python3.8/dist-packages/jittor/src" -I/usr/include/python3.8 -I/usr/include/python3.8 -DHAS_CUDA -DIS_CUDA -I"/root/.cache/jittor/jtcuda/cuda11.2_cudnn8_linux/include" -I"/usr/local/lib/python3.8/dist-packages/jittor/extern/cuda/inc" -I"/root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default/cu11.2.152_sm_30" -O2 -I"/root/.cache/jittor/cutt/cutt-1.2/include" -I"/root/.cache/jittor/cutt/cutt-1.2/src" -c -o "/root/.cache/jittor/jt1.3.7/g++9.4.0/py3.8.10/Linux-5.19.0-3x55/IntelRCoreTMi7xea/default/cu11.2.152_sm_30/obj_files/cuttkernel.cu.o" -x cu --cudart=shared -ccbin="/usr/bin/g++" --use_fast_math -w -I"/usr/local/lib/python3.8/dist-packages/jittor/extern/cuda/inc"
terminate called without an active exception
Aborted (core dumped)
我本地是ubuntu22.04的系统，显卡GTX760，显卡驱动470.57.02对应CUDA11.4但没有在本地安装CUDA。
由于本地的gcc版本太高，对应CUDA超过了显卡支持的版本，所以我选择在docker中通过ubuntu20.04来使用JittorLLMs。
本地的docker版本是23.0.4，nvidia-container-toolkit版本是1.13.0，可以在容器内正确执行nvidia-smi。
无论是我手动在容器内安装了CUDA11.4并设置好nvcc_path的值，还是另外创建容器让脚本自动安装CUDA11.2，都会遇到报错“nvcc fatal : Value ‘compute_30’ is not defined for option ‘gpu-architecture’”。
我在NVIDIA官网查询我的显卡对应的架构版本的确是30，这是否意味着容器内其实没有正确识别我的显卡？对了，我创建容器的语句是“docker create -it --name jittor --network host --runtime=nvidia --gpus all ubuntu:20.04”。
请各位大佬不吝赐教，谢谢！