[SYCL] Windows default build instructions without -DLLAMA_SYCL_F16 flag activated (#6767)

* Fix FP32/FP16 build instructions

* Fix typo

* Recommended build instruction

Co-authored-by: Neo Zhang Jianyu <jianyu.zhang@intel.com>

* Recommended build instruction

Co-authored-by: Neo Zhang Jianyu <jianyu.zhang@intel.com>

* Recommended build instruction

Co-authored-by: Neo Zhang Jianyu <jianyu.zhang@intel.com>

* Add comments in Intel GPU linux

---------

Co-authored-by: Anas Ahouzi <112881240+aahouzi-intel@users.noreply.github.com>
Co-authored-by: Neo Zhang Jianyu <jianyu.zhang@intel.com>

CMakeLists.txt

@@ -43,17 +43,11 @@ else()
     set(LLAMA_METAL_DEFAULT OFF)
 endif()
 
-# TODO: fix this for Android CI
-#       https://github.com/ggerganov/llama.cpp/pull/6716#issuecomment-2061509191
-#if (CMAKE_SYSTEM_NAME MATCHES "ANDROID")
-#    set(LLAMA_LLAMAFILE_DEFAULT OFF)
-#else()
-#    set(LLAMA_LLAMAFILE_DEFAULT ON)
-#endif()
-
-# TODO: temporary disable until MoE is fixed
-#       https://github.com/ggerganov/llama.cpp/pull/6716
-set(LLAMA_LLAMAFILE_DEFAULT OFF)
+if (CMAKE_SYSTEM_NAME MATCHES "ANDROID")
+    set(LLAMA_LLAMAFILE_DEFAULT OFF)
+else()
+    set(LLAMA_LLAMAFILE_DEFAULT ON)
+endif()
 
 # general
 option(BUILD_SHARED_LIBS                "build shared libraries"                                OFF)

Makefile

@@ -384,10 +384,6 @@ ifdef LLAMA_OPENBLAS
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
 endif # LLAMA_OPENBLAS
 
-# TODO: temporary disable until MoE is fixed
-#       https://github.com/ggerganov/llama.cpp/pull/6716
-LLAMA_NO_LLAMAFILE := 1
-
 ifndef LLAMA_NO_LLAMAFILE
 	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
 	OBJS        += sgemm.o

README-sycl.md

@@ -229,12 +229,12 @@ source /opt/intel/oneapi/setvars.sh
 # Build LLAMA with MKL BLAS acceleration for intel GPU
 mkdir -p build && cd build
 
-# Option 1: Use FP16 for better performance in long-prompt inference
-#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
-
-# Option 2: Use FP32 by default
+# Option 1: Use FP32 (recommended for better performance in most cases)
 cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
+# Option 2: Use FP16
+cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+
 #build all binary
 cmake --build . --config Release -j -v
 ```
@@ -250,12 +250,12 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
 # Build LLAMA with Nvidia BLAS acceleration through SYCL
 mkdir -p build && cd build
 
-# Option 1: Use FP16 for better performance in long-prompt  inference
-cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
-
-# Option 2: Use FP32 by default
+# Option 1: Use FP32 (recommended for better performance in most cases)
 cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
+# Option 2: Use FP16
+cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+
 #build all binary
 cmake --build . --config Release -j -v
 
@@ -416,6 +416,10 @@ mkdir -p build
 cd build
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
+# Option 1: Use FP32 (recommended for better performance in most cases)
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+
+# Option 2: Or FP16
 cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
 make -j

ggml.c

@@ -10825,7 +10825,7 @@ static void ggml_compute_forward_mul_mat(
 #endif
 
 #if GGML_USE_LLAMAFILE
-    if (nb10 == ggml_type_size(src1->type)) {
+    if (src1_cont) {
         for (int64_t i13 = 0; i13 < ne13; i13++)
             for (int64_t i12 = 0; i12 < ne12; i12++)
                 if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
@@ -10878,15 +10878,13 @@ UseGgmlGemm1:;
     const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
 #if GGML_USE_LLAMAFILE
-    if (nb10 == ggml_type_size(src1->type) || src1->type != vec_dot_type) {
+    if (src1->type != vec_dot_type) {
         for (int64_t i13 = 0; i13 < ne13; i13++)
             for (int64_t i12 = 0; i12 < ne12; i12++)
                 if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
                                      nb01/ggml_type_size(src0->type),
-                                     (const char *)wdata + ggml_row_size(vec_dot_type,
-                                         nb12/ggml_type_size(src1->type)*i12 +
-                                         nb13/ggml_type_size(src1->type)*i13),
+                                     (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
                                      row_size/ggml_type_size(vec_dot_type),
                                      (char *)dst->data + i12*nb2 + i13*nb3,
                                      nb1/ggml_type_size(dst->type),