move BLAS to a separate backend (#6210)

* move BLAS to a separate backend

* rename GGML_USE_OPENBLAS to GGML_USE_BLAS

* alloc : reuse same buffer when the same buffer type if used multiple times

* set number of threads automatically for openblas and blis

* sched : print assignments when GGML_SCHED_DEBUG env variable is set

* sched : allow ops with weights on an incompatible buffer type

This will cause the weight to be copied to a backend that supports the
op, which is very costly. The weight should have been stored in a buffer
of a backend that can run the op, but llama.cpp cannot do this
automatically at the moment.

---------

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

1 files changed, 10 insertions, 18 deletions

diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp
index e7d260bd..6f41ed27 100644
--- a/ggml-sycl.cpp
+++ b/ggml-sycl.cpp
@@ -16575,22 +16575,12 @@ GGML_CALL static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backen
     UNUSED(buft);
 }
 
-GGML_CALL static bool ggml_backend_sycl_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    if (!ggml_backend_is_sycl(backend)) {
-        return false;
-    }
-    ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
-    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
-    return buft_ctx->device == sycl_ctx->device;
-}
-
 static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_sycl_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_sycl_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_sycl_buffer_type_get_alignment,
     /* .get_max_size     = */ ggml_backend_sycl_buffer_type_get_max_size,
     /* .get_alloc_size   = */ ggml_backend_sycl_buffer_type_get_alloc_size,
-    /* .supports_backend = */ ggml_backend_sycl_buffer_type_supports_backend,
     /* .is_host          = */ nullptr,
 };
 
@@ -16942,12 +16932,6 @@ GGML_CALL static size_t ggml_backend_sycl_split_buffer_type_get_alloc_size(ggml_
     return total_size;
 }
 
-GGML_CALL static bool ggml_backend_sycl_split_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    return ggml_backend_is_sycl(backend);
-
-    UNUSED(buft);
-}
-
 GGML_CALL static bool ggml_backend_sycl_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
     return false;
 
@@ -16960,7 +16944,6 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_split_buffer_type_interface
     /* .get_alignment    = */ ggml_backend_sycl_split_buffer_type_get_alignment,
     /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_sycl_split_buffer_type_get_alloc_size,
-    /* .supports_backend = */ ggml_backend_sycl_split_buffer_type_supports_backend,
     /* .is_host          = */ ggml_backend_sycl_split_buffer_type_is_host,
 };
 
@@ -17046,7 +17029,6 @@ ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
             /* .get_max_size     = */ NULL, // TODO: return device.maxBufferLength
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
-            /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
         /* .context  = */ nullptr,
@@ -17311,6 +17293,14 @@ GGML_CALL static bool ggml_backend_sycl_offload_op(ggml_backend_t backend, const
     GGML_UNUSED(backend);
 }
 
+GGML_CALL static bool ggml_backend_sycl_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    if (buft->iface.get_name != ggml_backend_sycl_buffer_type_name) {
+        return false;
+    }
+    ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    return buft_ctx->device == sycl_ctx->device;
+}
 
 static ggml_backend_i ggml_backend_sycl_interface = {
     /* .get_name                = */ ggml_backend_sycl_name,
@@ -17322,9 +17312,11 @@ static ggml_backend_i ggml_backend_sycl_interface = {
     /* .synchronize             = */ ggml_backend_sycl_synchronize,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_sycl_graph_compute,
     /* .supports_op             = */ ggml_backend_sycl_supports_op,
+    /* .supports_buft           = */ ggml_backend_sycl_supports_buft,
     /* .offload_op              = */ ggml_backend_sycl_offload_op,
     /* .event_new               = */ NULL,
     /* .event_free              = */ NULL,