[SYCL] enhance the api to support peer-to-peer copy

[arthw]

Overview

It maybe fix the issue: #27198.

Additional information

There is missed to call the API to enable peer-to-peer memory copy.
This PR adds it to make the code following the guide and stable.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO.

[stolk]

Applied your patch:

commit 7e4ccad92c66c272c6b0ed8f2ebc5e80109323c0 (HEAD -> master)
Author: zhang.jianyu <zhang.jianyu@outlook.com>
Date:   Sat Aug 22 23:57:57 2026 +0800

    enhance the api to support peer-to-peer copy

commit b21e4de74567f5eef213765c9476a843c2e43f0d (tag: b10586, origin/master, origin/HEAD)
Author: Xuan-Son Nguyen <son@huggingface.co>
Date:   Sat Aug 22 16:33:47 2026 +0200

    mtmd: use ggml_rope_set_offset (#27521)
    
    * mtmd: use ggml_rope_set_offset
    
    * add comment

Rebuild my tree with SYCL back end enabled, took out the export GGML_SYCL_DEV2DEV_MEMCPY=2 and launced llama.cpp

No improvement: still crashes on first prompt input, in:

__syscall_cancel_arch () at ../sysdeps/unix/sysv/linux/x86_64/syscall_cancel.S:56
⚠️ warning: 56	../sysdeps/unix/sysv/linux/x86_64/syscall_cancel.S: No such file or directory
#0  __syscall_cancel_arch () at ../sysdeps/unix/sysv/linux/x86_64/syscall_cancel.S:56
56	in ../sysdeps/unix/sysv/linux/x86_64/syscall_cancel.S
#1  0x000079d228ea061c in __internal_syscall_cancel (a1=49577, a2=0, a3=0, a4=0, a5=0, a6=0, nr=61) at ./nptl/cancellation.c:49
⚠️ warning: 49	./nptl/cancellation.c: No such file or directory
#2  __syscall_cancel (a1=a1@entry=49577, a2=a2@entry=0, a3=a3@entry=0, a4=a4@entry=0, a5=a5@entry=0, a6=a6@entry=0, nr=61) at ./nptl/cancellation.c:75
75	in ./nptl/cancellation.c
#3  0x000079d228f1cd4f in __GI___wait4 (pid=pid@entry=49577, stat_loc=stat_loc@entry=0x0, options=options@entry=0, usage=usage@entry=0x0) at ../sysdeps/unix/sysv/linux/wait4.c:30
⚠️ warning: 30	../sysdeps/unix/sysv/linux/wait4.c: No such file or directory
#4  0x000079d228f1cd9b in __GI___waitpid (pid=pid@entry=49577, stat_loc=stat_loc@entry=0x0, options=options@entry=0) at ./posix/waitpid.c:38
⚠️ warning: 38	./posix/waitpid.c: No such file or directory
#5  0x000079d23f67abb6 in ggml_print_backtrace () at /usr/src/llama.cpp/ggml/src/ggml.c:234
234	        waitpid(child_pid, NULL, 0);
#6  0x000079d23f679d89 in ggml_abort (file=0x79d22aec1f9b "/usr/src/llama.cpp/ggml/src/ggml-sycl/../ggml-sycl/common.hpp", line=155, fmt=0x79d22aec1fd9 "SYCL error") at /usr/src/llama.cpp/ggml/src/ggml.c:268
268	        ggml_print_backtrace();
#7  0x000079d22a0256a8 in ggml_sycl_error (stmt=<optimized out>, func=0x79d22aec213e "dev2dev_memcpy", file=0x79d22aebfc82 "/usr/src/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp", line=line@entry=714, msg=<optimized out>) at /usr/src/llama.cpp/ggml/src/ggml-sycl/../ggml-sycl/common.hpp:155
155	  GGML_ABORT("SYCL error");
#8  0x000079d22a02d55b in dev2dev_memcpy (device_dst=<optimized out>, q_dst=..., device_src=<optimized out>, q_src=..., ptr_dst=ptr_dst@entry=0x79bb78166000, ptr_src=ptr_src@entry=0x79b377c00000, size=3563520) at /usr/src/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp:714
714	            SYCL_CHECK(CHECK_TRY_ERROR(q_dst.memcpy(ptr_dst, ptr_src, size).wait()));
#9  0x000079d22a02cc7c in ggml_backend_sycl_comm_allreduce_tensor (comm_ctx_v=<optimized out>, tensors=<optimized out>) at /usr/src/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp:6686
6686	    dev2dev_memcpy(ctx0->device, *q0, ctx1->device, *q1, inbox0, outbox1, bf16_bytes);
#10 0x000079d23f6a2fdc in ggml_backend_meta_graph_compute (backend=<optimized out>, cgraph=<optimized out>) at /usr/src/llama.cpp/ggml/src/ggml-backend-meta.cpp:2215
2215	                backend_allreduce_success = backend_ctx->comm_allreduce(backend_ctx->comm_ctx, nodes.data());
#11 0x000079d23f69d2c6 in ggml_backend_graph_compute_async (backend=0x360eba20, cgraph=0x0) at /usr/src/llama.cpp/ggml/src/ggml-backend.cpp:452
452	    return backend->iface.graph_compute(backend, cgraph);
#12 ggml_backend_sched_compute_splits (sched=0x350b5c70) at /usr/src/llama.cpp/ggml/src/ggml-backend.cpp:1743
1743	            enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
#13 ggml_backend_sched_graph_compute_async (sched=0x350b5c70, graph=<optimized out>) at /usr/src/llama.cpp/ggml/src/ggml-backend.cpp:1973
1973	    return ggml_backend_sched_compute_splits(sched);
...
#30 0x000079d23f08c07f in llama_server (argc=39, argv=0x7ffdb8a2d5f8) at /usr/src/llama.cpp/tools/server/server.cpp:112
112	    return llama_server(params, argc, argv);
#31 0x000079d228e2a601 in __libc_start_call_main (main=main@entry=0x4064e0 <main(int, char**)>, argc=argc@entry=39, argv=argv@entry=0x7ffdb8a2d5f8) at ../sysdeps/nptl/libc_start_call_main.h:59
⚠️ warning: 59	../sysdeps/nptl/libc_start_call_main.h: No such file or directory
#32 0x000079d228e2a718 in __libc_start_main_impl (main=0x4064e0 <main(int, char**)>, argc=39, argv=0x7ffdb8a2d5f8, init=<optimized out>, fini=<optimized out>, rtld_fini=<optimized out>, stack_end=0x7ffdb8a2d5e8) at ../csu/libc-start.c:360
⚠️ warning: 360	../csu/libc-start.c: No such file or directory
#33 0x0000000000406415 in _start ()

I double checked: your commit is properly applied, as I see my line nr shift by one and matches your change.