From d9cd3d9667d83f68040378dc2834a49962e08244 Mon Sep 17 00:00:00 2001
From: Daniel Hiltgen <daniel@ollama.com>
Date: Wed, 20 Dec 2023 14:46:15 -0800
Subject: [PATCH] Revive windows build

The windows native setup still needs some more work, but this gets it building
again and if you set the PATH properly, you can run the resulting exe on a cuda system.
---
 llm/ext_server.go             |  4 +--
 llm/llama.cpp/gen_windows.ps1 | 61 ++++++++++++++++++++++++++++-------
 2 files changed, 51 insertions(+), 14 deletions(-)

diff --git a/llm/ext_server.go b/llm/ext_server.go
index 048b1a65..45251cc5 100644
--- a/llm/ext_server.go
+++ b/llm/ext_server.go
@@ -22,8 +22,8 @@ package llm
 #cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cpu/libllama.a
 #cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cpu/libggml_static.a
 #cgo linux LDFLAGS: -lrt -lpthread -ldl -lstdc++ -lm
-#cgo windows LDFLAGS: -L${SRCDIR}/llama.cpp/gguf/build/wincuda/dist/bin
-#cgo windows LDFLAGS: -lext_server_shared -lpthread
+#cgo windows LDFLAGS: -L${SRCDIR}/llama.cpp/gguf/build/wincpu/dist/lib
+#cgo windows LDFLAGS: -lcpu_server -lpthread
 
 #include <stdlib.h>
 #include "server.h"
diff --git a/llm/llama.cpp/gen_windows.ps1 b/llm/llama.cpp/gen_windows.ps1
index f85f1a45..2c77d4ab 100644
--- a/llm/llama.cpp/gen_windows.ps1
+++ b/llm/llama.cpp/gen_windows.ps1
@@ -3,52 +3,89 @@
 $ErrorActionPreference = "Stop"
 
 function init_vars {
-    $script:buildDir="gguf/build/wincuda"
-    $script:installDir="gguf/build/wincuda/dist"
     $script:patches = @("0001-Expose-callable-API-for-server.patch")
-    $script:cmakeDefs = @("-DLLAMA_NATIVE=off", "-DLLAMA_F16C=off", "-DLLAMA_FMA=off", "-DLLAMA_AVX512=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX=on", "-DLLAMA_K_QUANTS=on", "-DLLAMA_ACCELERATE=on", "-DLLAMA_CUBLAS=ON","-DCMAKE_VERBOSE_MAKEFILE=ON","-DBUILD_SHARED_LIBS=on","-A","x64")
+    $script:cmakeDefs = @("-DBUILD_SHARED_LIBS=on", "-DLLAMA_NATIVE=off", "-DLLAMA_F16C=off", "-DLLAMA_FMA=off", "-DLLAMA_AVX512=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX=on", "-DLLAMA_K_QUANTS=on", "-DLLAMA_ACCELERATE=on", "-A","x64")
 
     if ($env:CGO_CFLAGS -contains "-g") {
         $script:cmakeDefs += @("-DCMAKE_VERBOSE_MAKEFILE=on")
-        $script:config += "RelWithDebInfo"
+        $script:config = "RelWithDebInfo"
     } else {
-        $script:config += "Release"
+        $script:config = "Release"
     }
 }
 
 function git_module_setup {
     # TODO add flags to skip the init/patch logic to make it easier to mod llama.cpp code in-repo
     & git submodule init
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     & git submodule update --force gguf
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
 }
 
 function apply_patches {
     rm -erroraction ignore -path "gguf/examples/server/server.h"
-    foreach ($patch in $patches) {
+    foreach ($patch in $script:patches) {
         write-host "Applying patch $patch"
         & git -C gguf apply ../patches/$patch
+        if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
     }
 }
 
 function build {
-    write-host "generating config with: cmake -S gguf -B $buildDir $cmakeDefs"
+    write-host "generating config with: cmake -S gguf -B $script:buildDir $script:cmakeDefs"
     & cmake --version
-    & cmake -S gguf -B $buildDir $cmakeDefs
-    write-host "building with: cmake --build $buildDir --config $config"
-    & cmake --build $buildDir --config $config
+    & cmake -S gguf -B $script:buildDir $script:cmakeDefs
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+    write-host "building with: cmake --build $script:buildDir --config $script:config"
+    & cmake --build $script:buildDir --config $script:config
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
 }
 
 function install {
-    rm -erroraction ignore -recurse -force -path $installDir
-    & cmake --install $buildDir --prefix $installDir --config $config
+    rm -erroraction ignore -recurse -force -path $script:installDir
+    & cmake --install $script:buildDir --prefix $script:installDir --config $script:config
+    if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
 
 }
 
 init_vars
 git_module_setup
 apply_patches
+
+# first build CPU based
+$script:buildDir="gguf/build/wincpu"
+$script:installDir="gguf/build/wincpu/dist"
+
+build
+# install
+
+md gguf/build/lib -ea 0
+md gguf/build/wincpu/dist/lib -ea 0
+mv gguf/build/wincpu/bin/$script:config/ext_server_shared.dll gguf/build/wincpu/dist/lib/cpu_server.dll
+
+
+# Nope, this barfs on lots of symbol problems
+#mv gguf/build/wincpu/examples/server/$script:config/ext_server_shared.dll gguf/build/wincpu/dist/lib/cpu_server.lib
+# Nope: this needs lots of include paths to pull in things like msvcprt.lib and other deps
+# & cl.exe `
+#     gguf/build/wincpu/examples/server/$script:config/ext_server.lib `
+#     gguf/build/wincpu/common/$script:config/common.lib `
+#     gguf/build/wincpu/$script:config/llama.lib `
+#     gguf/build/wincpu/$script:config/ggml_static.lib `
+#     /link /DLL /DEF:cpu_server.def /NOENTRY /MACHINE:X64  /OUT:gguf/build/wincpu/dist/lib/cpu_server.dll
+# if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
+
+# Then build cuda as a dynamically loaded library
+init_vars
+$script:buildDir="gguf/build/wincuda"
+$script:installDir="gguf/build/wincuda/dist"
+$script:cmakeDefs += @("-DLLAMA_CUBLAS=ON", "-DBUILD_SHARED_LIBS=on")
 build
 install
+cp gguf/build/wincuda/dist/bin/ext_server_shared.dll gguf/build/lib/cuda_server.dll
+
+# TODO - more to do here to create a usable dll
+
 
 # TODO - implement ROCm support on windows
 md gguf/build/winrocm/lib -ea 0