Segmentation fault (core dumped) in tensor creation APIs

[x0w3n]

Describe the bug

Inspired by problem #511, tensor multiplication can crash, and we found that the following tensor creation apis all have similar problems:

• jt.rand()
• jt.randn()
• jt.randint()
• jt.random()
• jt.empty()
• jt.full()
• jt.init.eye()

Here is the colab reference link:
https://colab.research.google.com/drive/1o4UTyHFKQvnGBvydJSQuEbXbbGkcN_bk?usp=sharing

Full Log

[i 1202 21:42:01.481311 48 compiler.py:956] Jittor(1.3.9.10) src: /home/miniconda3/envs/jittor/lib/python3.9/site-packages/jittor
[i 1202 21:42:01.491502 48 compiler.py:957] g++ at /usr/bin/g++(11.2.0)
[i 1202 21:42:01.491992 48 compiler.py:958] cache_path: /home/.cache/jittor/jt1.3.9/g++11.2.0/py3.9.12/Linux-5.15.0-1x30/INTELRXEONRGOLx51/ef26/default
[i 1202 21:42:01.561816 48 install_cuda.py:93] cuda_driver_version: [12, 2]
[i 1202 21:42:01.577170 48 __init__.py:412] Found /home/.cache/jittor/jtcuda/cuda12.2_cudnn8_linux/bin/nvcc(12.2.140) at /home/.cache/jittor/jtcuda/cuda12.2_cudnn8_linux/bin/nvcc.
[i 1202 21:42:01.729819 48 __init__.py:412] Found gdb(12.0.90) at /usr/bin/gdb.
[i 1202 21:42:01.732195 48 __init__.py:412] Found addr2line(2.38) at /usr/bin/addr2line.
[i 1202 21:42:01.955148 48 compiler.py:1013] cuda key:cu12.2.140_sm_89
[i 1202 21:42:02.983538 48 __init__.py:227] Total mem: 251.50GB, using 16 procs for compiling.
[i 1202 21:42:03.184558 48 jit_compiler.cc:28] Load cc_path: /usr/bin/g++
[i 1202 21:42:03.421654 48 init.cc:63] Found cuda archs: [89,]
Caught segfault at address 0x55dcf174e000, thread_name: '', flush log...
Segmentation fault (core dumped)

Minimal Reproduce

import jittor as jt
 
# a = jt.rand((300, 400))    # Segmentation fault (core dumped)
# a = jt.randn((300, 400)) #Segmentation fault (core dumped)
# a = jt.randint(300, shape=(300, 400))  #Segmentation fault (core dumped)
# a = jt.random((300, 400)) # Segmentation fault (core dumped)
# a = jt.empty((300, 400))  # Segmentation fault (core dumped)
# a = jt.full((300, 400),3.14)  # Segmentation fault (core dumped)
a = jt.init.eye((300,400))  # Segmentation fault (core dumped)

sumsum = 0
for i in range(300):
    for j in range(300):
        if i != j:
            sumsum += a[i] * a[j]

Expected behavior