The trick is to build psyco in ivm mode. The mode slower than the native mode(ivm mode involves a prolog generated-vm, you guess), however it's the only way to compile psyco on non-x86 system. And cross-compiling with distutils is not a trivial stuff, so I just compile it directly.
mipsel-linux-gcc -pthread -fno-strict-aliasing -DNDEBUG -fPIC -DALL_STATIC=1 -Ic/ivm -I/usr/include/python2.5 -c c/psyco.c -o build/temp.linux-mipsel-2.5/c/psyco.o
mipsel-linux-gcc -pthread -fno-strict-aliasing -DNDEBUG -fPIC -DALL_STATIC=1 -Ic/ivm -I/usr/include/python2.5 -c c/platform.c -o build/temp.linux-mipsel-2.5/c/platform.o
mipsel-linux-gcc -pthread -shared build/temp.linux-i686-2.5/c/psyco.o build/temp.linux-mipsel-2.5/c/platform.o -L/usr/lib -lpython2.5 -o build/lib.linux-i686-2.5/psyco/_psyco.so
This generates the C-module extension, and just copy all of py files to correct path and enjoy psyco speed-up! (well, not quite a case for me. It just use way too much memory on our embedded system. But I think maybe it'll useful in other project.)
ps: I should mentioned that I'm doing this on a uclibc linux system.