fix the issue in mixed different model GPUs in FA (llama/29071)

* fix mixed different model GPUs issue

* Update ggml/src/ggml-sycl/ggml-sycl.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

ggml/src/ggml-sycl/fattn-vec.hpp

@@ -589,10 +589,10 @@ void ggml_sycl_flash_attn_ext_vec_case_impl(ggml_backend_sycl_context & ctx, ggm
     const bool need_f16_V = type_V == GGML_TYPE_F16;
     constexpr size_t nbytes_shared = 0;
 
+    const auto arch = ggml_sycl_info().devices[ggml_sycl_get_device()].hw_info.arch;
     // D=512 does not fit the default register file; it spills up to 343 bytes per thread, against at most 57 for D <= 256. This kernel is decode only, so thread occupancy is not the limit. It is 1.9x faster at every KV depth on Battlemage.
     constexpr bool use_large_grf = D >= 512;
 
-    const auto arch = ggml_sycl_info().devices[ctx.device].hw_info.arch;
     const int nthreads = ggml_sycl_fattn_vec_get_nthreads_device(arch);
     if constexpr (D <= 256) {
         if (nthreads == 256) {

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -792,6 +792,14 @@ static bool ggml_sycl_is_l0_discrete_gpu(int device) {
 }
 #endif
 
+static void memcpy_host_forward(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst,
+                         const void *ptr_src, size_t size) {
+    char *host_buf = (char *)malloc(size);
+    q_src.memcpy(host_buf, (const char *)ptr_src, size).wait();
+    q_dst.memcpy((char *)ptr_dst, host_buf, size).wait();
+    free(host_buf);
+}
+
 static void dev2dev_memcpy(int device_dst, sycl::queue &q_dst, int device_src, sycl::queue &q_src, void *ptr_dst,
                     const void *ptr_src, size_t size) {
 
@@ -835,10 +843,7 @@ static void dev2dev_memcpy(int device_dst, sycl::queue &q_dst, int device_src, s
     } else {
         GGML_SYCL_DEBUG("[SYCL] dev2dev memcpy by host forward for SYCL/L0 fallback\n");
     }
-    char *host_buf = (char *)malloc(size);
+    memcpy_host_forward(q_dst, q_src, ptr_dst, ptr_src, size);
-    q_src.memcpy(host_buf, (const char *)ptr_src, size).wait();
-    q_dst.memcpy((char *)ptr_dst, host_buf, size).wait();
-    free(host_buf);
 }
 
 static bool