fix memcpy() crash, add missed cmd in guide, fix softmax

[NeoZhangJianyu]

1. When install latest Intel GPU driver, memcpy() from host to device will crash.
   fix it by use host buf.
2. Add missed cmd in guide.
3. Fix softmax error in UT.

[NeoZhangJianyu]

@slaren I have tried a solution to return support_mmap by backend in callback function.
But the caller is too earlier that the backend is not initial at that time.
So, I define a static function in ggml-backend.c

[slaren]

This is not ok, if you need a workaround add a memcpy to a temporary buffer in the set_tensor function.

[NeoZhangJianyu]

@slaren
Sorry, this is misunderstanding.
memcpy() usage is OK.
due to the new driver issue, the memcpy() will crash.
The workaround is disable mmap() used in sycl backend.

[slaren]

If the driver crashes when it receives an address from mmap, then what you can do is memcpy the buffer to a temporary buffer allocated in a conventional mean, eg. with malloc or std::vector, and pass that to the driver.

[NeoZhangJianyu]

driver is not crash.
The source ptr provided by mmap is unavailable. That lead to memcpy() crash.
If disable mmap() and read the gguf file into host buf. The memcpy() is from host buf to device. It's OK.
If memcpy from ptr of mmap() to device, it will crash.

[slaren]

memcpy the ptr you get in the tensor_set function to a temporary buffer. host -> host. Then memcpy host -> device using the temporary buffer.

[NeoZhangJianyu]

host -> host -> device. this solution should be OK.
Let me try.

[NeoZhangJianyu]

@slaren Yes, it works well.
I will update the PR as soon.
Thank you!

[NeoZhangJianyu]

@slaren @airMeng I update with the new solution. Please review again!

[airMeng]

approved after this is removed

[NeoZhangJianyu]

yes, remove host_buffer. use malloc/free for thread safe and better performance.

[jeffzhou2000]

driver is not crash. The source ptr provided by mmap is unavailable. That lead to memcpy() crash. If disable mmap() and read the gguf file into host buf. The memcpy() is from host buf to device. It's OK. If memcpy from ptr of mmap() to device, it will crash.

why the source ptr provided by mmap is unavailable? I'd like to reproduce this issue in my local dev, could you help provide a quick guide for this: whether I only need to read the docs/backend/SYCL.md?

[jeffzhou2000]

your explanation in #12734 is very clear, thanks so much.