chore(ci): fix download llama from upstream and drop cuda11.7 release

Author: zwpaper

.github/workflows/release.yml

@@ -33,7 +33,6 @@ jobs:
           - aarch64-apple-darwin
           - x86_64-manylinux_2_28
           - x86_64-manylinux_2_28-cuda123
-          - x86_64-windows-msvc
         include:
           - os: macos-latest
             target: aarch64-apple-darwin
@@ -49,11 +48,6 @@ jobs:
             binary: x86_64-manylinux_2_28-cuda123
             container: sameli/manylinux_2_28_x86_64_cuda_12.3@sha256:e12416bf249ab312f9dcfdebd7939b968dd6f1b6f810abbede818df875e86a7c
             build_args: --features binary,cuda
-          - os: windows-2022
-            target: x86_64-pc-windows-msvc
-            binary: x86_64-windows-msvc
-            build_args: --features binary
-            ext: .exe
 
     env:
       SCCACHE_GHA_ENABLED: true
@@ -235,13 +229,13 @@ jobs:
       - name: Display structure of downloaded files
         run: ls -R
 
-      - name: Package CUDA 11.7 for Windows
+      - name: Package CPU for Windows
         run: >
-          LLAMA_CPP_PLATFORM=win-cuda-cu11.7-x64 OUTPUT_NAME=tabby_x86_64-windows-msvc-cuda117 ./ci/package-from-upstream.sh
+          LLAMA_CPP_PLATFORM=win-cpu-x64 OUTPUT_NAME=tabby_x86_64-windows-msvc ./ci/package-from-upstream.sh
 
       - name: Package CUDA 12.4 for Windows
         run: >
-          LLAMA_CPP_PLATFORM=win-cuda-cu12.4-x64 OUTPUT_NAME=tabby_x86_64-windows-msvc-cuda124 ./ci/package-from-upstream.sh
+          LLAMA_CPP_PLATFORM=win-cuda-12.4-x64 OUTPUT_NAME=tabby_x86_64-windows-msvc-cuda124 ./ci/package-from-upstream.sh
 
       - name: Package Vulkan for Windows
         run: >

ci/package-from-upstream.sh

@@ -1,18 +1,25 @@
 #!/bin/bash
 
+set -e
+
 # get current bash file directory
 PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 LLAMA_CPP_PATH="${PROJECT_ROOT}/crates/llama-cpp-server/llama.cpp"
 
 # Input variables
 LLAMA_CPP_VERSION=${LLAMA_CPP_VERSION:-$(cd ${LLAMA_CPP_PATH} && git fetch --tags origin >/dev/null && git describe --tags --abbrev=0)}
 echo "LLAMA_CPP_VERSION=${LLAMA_CPP_VERSION}"
-LLAMA_CPP_PLATFORM=${LLAMA_CPP_PLATFORM:-win-cuda-cu11.7-x64}
-OUTPUT_NAME=${OUTPUT_NAME:-tabby_x86_64-windows-msvc-cuda117}
+LLAMA_CPP_PLATFORM=${LLAMA_CPP_PLATFORM:-win-cuda-cu12.4-x64}
 
-NAME=llama-${LLAMA_CPP_VERSION}-bin-${LLAMA_CPP_PLATFORM}
+if [[ $LLAMA_CPP_PLATFORM == *cuda* ]]; then
+    NAME=cudart-llama-bin-${LLAMA_CPP_PLATFORM}
+else
+    NAME=llama-${LLAMA_CPP_VERSION}-bin-${LLAMA_CPP_PLATFORM}
+fi
 ZIP_FILE=${NAME}.zip
 
+OUTPUT_NAME=${OUTPUT_NAME:-tabby_x86_64-windows-msvc-cuda124}
+
 if [[ ${LLAMA_CPP_PLATFORM} == win* ]]; then
     TABBY_BINARY=${TABBY_BINARY:-tabby_x86_64-windows-msvc.exe}
     TABBY_EXTENSION=.exe