llama-cpp-python 0.2.86__tar.gz → 0.2.87__tar.gz

llama_cpp_python-0.2.87/.git/FETCH_HEAD

@@ -0,0 +1 @@
+84321162f8cad3e1caf552d7f0b18bc869e2cebf		'84321162f8cad3e1caf552d7f0b18bc869e2cebf' of https://github.com/abetlen/llama-cpp-python

llama_cpp_python-0.2.87/.git/HEAD

@@ -0,0 +1 @@
+84321162f8cad3e1caf552d7f0b18bc869e2cebf

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/.git/config

@@ -9,7 +9,7 @@
 [gc]
 	auto = 0
 [http "https://github.com/"]
-	extraheader = AUTHORIZATION: basic eC1hY2Nlc3MtdG9rZW46Z2hzX1VFZE02SmVmczlqeTlwYk5mSWZEb2JGN0tHS09YazBjVlpUeA==
+	extraheader = AUTHORIZATION: basic eC1hY2Nlc3MtdG9rZW46Z2hzX3NGN2c0OFQ3cnlXRnI4Ujg4MXJMeENDUmdYQU01RjJmSDJ6Ng==
 [submodule "vendor/llama.cpp"]
 	active = true
 	url = https://github.com/ggerganov/llama.cpp.git

llama_cpp_python-0.2.87/.git/logs/HEAD

@@ -0,0 +1 @@
+0000000000000000000000000000000000000000 84321162f8cad3e1caf552d7f0b18bc869e2cebf runner <runner@fv-az770-835.1l2353hmgizunl4f1wof2aeoph.ex.internal.cloudapp.net> 1723038793 +0000	checkout: moving from master to refs/tags/v0.2.87

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/HEAD

@@ -0,0 +1 @@
+be55695eff44784a141a863f273661a6bce63dfc

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/.git/modules/vendor/llama.cpp/config

@@ -16,7 +16,7 @@
 [gc]
 	auto = 0
 [http "https://github.com/"]
-	extraheader = AUTHORIZATION: basic eC1hY2Nlc3MtdG9rZW46Z2hzX1VFZE02SmVmczlqeTlwYk5mSWZEb2JGN0tHS09YazBjVlpUeA==
+	extraheader = AUTHORIZATION: basic eC1hY2Nlc3MtdG9rZW46Z2hzX3NGN2c0OFQ3cnlXRnI4Ujg4MXJMeENDUmdYQU01RjJmSDJ6Ng==
 [url "https://github.com/"]
 	insteadOf = git@github.com:
 	insteadOf = org-6826477@github.com:

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/logs/HEAD

@@ -0,0 +1,2 @@
+0000000000000000000000000000000000000000 be55695eff44784a141a863f273661a6bce63dfc runner <runner@fv-az770-835.1l2353hmgizunl4f1wof2aeoph.ex.internal.cloudapp.net> 1723038794 +0000	clone: from https://github.com/ggerganov/llama.cpp.git
+be55695eff44784a141a863f273661a6bce63dfc be55695eff44784a141a863f273661a6bce63dfc runner <runner@fv-az770-835.1l2353hmgizunl4f1wof2aeoph.ex.internal.cloudapp.net> 1723038794 +0000	checkout: moving from master to be55695eff44784a141a863f273661a6bce63dfc

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/logs/refs/heads/master

@@ -0,0 +1 @@
+0000000000000000000000000000000000000000 be55695eff44784a141a863f273661a6bce63dfc runner <runner@fv-az770-835.1l2353hmgizunl4f1wof2aeoph.ex.internal.cloudapp.net> 1723038794 +0000	clone: from https://github.com/ggerganov/llama.cpp.git

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/logs/refs/remotes/origin/HEAD

@@ -0,0 +1 @@
+0000000000000000000000000000000000000000 be55695eff44784a141a863f273661a6bce63dfc runner <runner@fv-az770-835.1l2353hmgizunl4f1wof2aeoph.ex.internal.cloudapp.net> 1723038794 +0000	clone: from https://github.com/ggerganov/llama.cpp.git

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/.git/modules/vendor/llama.cpp/modules/kompute/config

@@ -13,7 +13,7 @@
 [gc]
 	auto = 0
 [http "https://github.com/"]
-	extraheader = AUTHORIZATION: basic eC1hY2Nlc3MtdG9rZW46Z2hzX1VFZE02SmVmczlqeTlwYk5mSWZEb2JGN0tHS09YazBjVlpUeA==
+	extraheader = AUTHORIZATION: basic eC1hY2Nlc3MtdG9rZW46Z2hzX3NGN2c0OFQ3cnlXRnI4Ujg4MXJMeENDUmdYQU01RjJmSDJ6Ng==
 [url "https://github.com/"]
 	insteadOf = git@github.com:
 	insteadOf = org-6826477@github.com:

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/modules/kompute/logs/HEAD

@@ -0,0 +1,2 @@
+0000000000000000000000000000000000000000 f592b5bca3cbc169feb194218a086b18d618cca4 runner <runner@fv-az770-835.1l2353hmgizunl4f1wof2aeoph.ex.internal.cloudapp.net> 1723038795 +0000	clone: from https://github.com/nomic-ai/kompute.git
+f592b5bca3cbc169feb194218a086b18d618cca4 4565194ed7c32d1d2efa32ceab4d3c6cae006306 runner <runner@fv-az770-835.1l2353hmgizunl4f1wof2aeoph.ex.internal.cloudapp.net> 1723038796 +0000	checkout: moving from master to 4565194ed7c32d1d2efa32ceab4d3c6cae006306

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/modules/kompute/logs/refs/heads/master

@@ -0,0 +1 @@
+0000000000000000000000000000000000000000 f592b5bca3cbc169feb194218a086b18d618cca4 runner <runner@fv-az770-835.1l2353hmgizunl4f1wof2aeoph.ex.internal.cloudapp.net> 1723038795 +0000	clone: from https://github.com/nomic-ai/kompute.git

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/modules/kompute/logs/refs/remotes/origin/HEAD

@@ -0,0 +1 @@
+0000000000000000000000000000000000000000 f592b5bca3cbc169feb194218a086b18d618cca4 runner <runner@fv-az770-835.1l2353hmgizunl4f1wof2aeoph.ex.internal.cloudapp.net> 1723038795 +0000	clone: from https://github.com/nomic-ai/kompute.git

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/packed-refs

@@ -0,0 +1,2 @@
+# pack-refs with: peeled fully-peeled sorted 
+be55695eff44784a141a863f273661a6bce63dfc refs/remotes/origin/master

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/refs/heads/master

@@ -0,0 +1 @@
+be55695eff44784a141a863f273661a6bce63dfc

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/refs/tags/b3541

@@ -0,0 +1 @@
+be55695eff44784a141a863f273661a6bce63dfc

llama_cpp_python-0.2.87/.git/modules/vendor/llama.cpp/shallow

@@ -0,0 +1 @@
+be55695eff44784a141a863f273661a6bce63dfc

llama_cpp_python-0.2.87/.git/refs/tags/v0.2.87

@@ -0,0 +1 @@
+84321162f8cad3e1caf552d7f0b18bc869e2cebf

llama_cpp_python-0.2.87/.git/shallow

@@ -0,0 +1 @@
+84321162f8cad3e1caf552d7f0b18bc869e2cebf

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/.github/workflows/generate-index-from-release.yaml

@@ -1,9 +1,11 @@
 name: Wheels Index
 
 on:
-  # Trigger on any new release
-  release:
-    types: [published]
+  # Trigger on new release
+  workflow_run:
+    workflows: ["Release", "Build Wheels (CUDA)", "Build Wheels (Metal)"]
+    types:
+      - completed
 
   # Allows you to run this workflow manually from the Actions tab
   workflow_dispatch:

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/CHANGELOG.md

@@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.2.87]
+
+- feat: Update llama.cpp to ggerganov/llama.cpp@be55695eff44784a141a863f273661a6bce63dfc
+- fix: Include all llama.cpp source files and subdirectories by @abetlen in 9cad5714ae6e7c250af8d0bbb179f631368c928b
+- feat(ci): Re-build wheel index automatically when releases are created by @abetlen in 198f47dc1bd202fd2b71b29e041a9f33fe40bfad
+
 ## [0.2.86]
 
 - feat: Update llama.cpp to ggerganov/llama.cpp@398ede5efeb07b9adf9fbda7ea63f630d476a792

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: llama_cpp_python
-Version: 0.2.86
+Version: 0.2.87
 Summary: Python bindings for the llama.cpp library
 Author-Email: Andrei Betlen <abetlen@gmail.com>
 License: MIT

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/llama_cpp/__init__.py

@@ -1,4 +1,4 @@
 from .llama_cpp import *
 from .llama import *
 
-__version__ = "0.2.86"
+__version__ = "0.2.87"

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/pyproject.toml

@@ -60,7 +60,7 @@ wheel.packages = ["llama_cpp"]
 cmake.verbose = true
 cmake.minimum-version = "3.21"
 minimum-version = "0.5.1"
-sdist.include = [".git", "vendor/llama.cpp/.git"]
+sdist.include = [".git", "vendor/llama.cpp/*"]
 
 [tool.scikit-build.metadata.version]
 provider = "scikit_build_core.metadata.regex"

llama_cpp_python-0.2.87/vendor/llama.cpp/.devops/llama-cpp-cuda.srpm.spec

@@ -0,0 +1,83 @@
+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://docs.fedoraproject.org/en-US/quick-docs/creating-rpm-packages
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp-cuda
+Version:        %( date "+%%Y%%m%%d" )
+Release:        1%{?dist}
+Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git cuda-toolkit
+Requires:       cuda-toolkit
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+
+%prep
+%setup -n llama.cpp-master
+
+%build
+make -j GGML_CUDA=1
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+cp -p llama-cli %{buildroot}%{_bindir}/llama-cuda-cli
+cp -p llama-server %{buildroot}%{_bindir}/llama-cuda-server
+cp -p llama-simple %{buildroot}%{_bindir}/llama-cuda-simple
+
+mkdir -p %{buildroot}/usr/lib/systemd/system
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacuda.service
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llama-cuda-server $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llama-cuda-cli
+%{_bindir}/llama-cuda-server
+%{_bindir}/llama-cuda-simple
+/usr/lib/systemd/system/llamacuda.service
+%config /etc/sysconfig/llama
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

llama_cpp_python-0.2.87/vendor/llama.cpp/.devops/llama-cpp.srpm.spec

@@ -0,0 +1,85 @@
+# SRPM for building from source and packaging an RPM for RPM-based distros.
+# https://docs.fedoraproject.org/en-US/quick-docs/creating-rpm-packages
+# Built and maintained by John Boero - boeroboy@gmail.com
+# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
+
+# Notes for llama.cpp:
+# 1. Tags are currently based on hash - which will not sort asciibetically.
+#    We need to declare standard versioning if people want to sort latest releases.
+#    In the meantime, YYYYMMDD format will be used.
+# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
+# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
+#    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
+# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
+#    It is up to the user to install the correct vendor-specific support.
+
+Name:           llama.cpp
+Version:        %( date "+%%Y%%m%%d" )
+Release:        1%{?dist}
+Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
+License:        MIT
+Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+BuildRequires:  coreutils make gcc-c++ git libstdc++-devel
+Requires:       libstdc++
+URL:            https://github.com/ggerganov/llama.cpp
+
+%define debug_package %{nil}
+%define source_date_epoch_from_changelog 0
+
+%description
+CPU inference for Meta's Lllama2 models using default options.
+Models are not included in this package and must be downloaded separately.
+
+%prep
+%setup -n llama.cpp-master
+
+%build
+make -j
+
+%install
+mkdir -p %{buildroot}%{_bindir}/
+cp -p llama-cli %{buildroot}%{_bindir}/llama-cli
+cp -p llama-server %{buildroot}%{_bindir}/llama-server
+cp -p llama-simple %{buildroot}%{_bindir}/llama-simple
+
+mkdir -p %{buildroot}/usr/lib/systemd/system
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llama.service
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llama-server $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
+
+%clean
+rm -rf %{buildroot}
+rm -rf %{_builddir}/*
+
+%files
+%{_bindir}/llama-cli
+%{_bindir}/llama-server
+%{_bindir}/llama-simple
+/usr/lib/systemd/system/llama.service
+%config /etc/sysconfig/llama
+
+%pre
+
+%post
+
+%preun
+%postun
+
+%changelog

llama_cpp_python-0.2.87/vendor/llama.cpp/common/build-info.cpp

@@ -0,0 +1,4 @@
+int LLAMA_BUILD_NUMBER = 1;
+char const *LLAMA_COMMIT = "be55695";
+char const *LLAMA_COMPILER = "gcc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0";
+char const *LLAMA_BUILD_TARGET = "x86_64-linux-gnu";

llama_cpp_python-0.2.87/vendor/llama.cpp/common/cmake/build-info-gen-cpp.cmake

@@ -0,0 +1,24 @@
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
+
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in")
+set(OUTPUT_FILE   "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
+
+# Only write the build info if it changed
+if(EXISTS ${OUTPUT_FILE})
+    file(READ ${OUTPUT_FILE} CONTENTS)
+    string(REGEX MATCH "LLAMA_COMMIT = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_COMMIT ${CMAKE_MATCH_1})
+    string(REGEX MATCH "LLAMA_COMPILER = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_COMPILER ${CMAKE_MATCH_1})
+    string(REGEX MATCH "LLAMA_BUILD_TARGET = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_TARGET ${CMAKE_MATCH_1})
+    if (
+        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
+        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
+        NOT OLD_TARGET   STREQUAL BUILD_TARGET
+    )
+        configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+    endif()
+else()
+    configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+endif()

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/vendor/llama.cpp/examples/llama-bench/llama-bench.cpp

@@ -27,6 +27,14 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef _WIN32
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
 // utils
 static uint64_t get_time_ns() {
     using clock = std::chrono::high_resolution_clock;
@@ -96,6 +104,27 @@ static std::string get_cpu_info() {
         }
         fclose(f);
     }
+#elif defined(_WIN32)
+    HKEY hKey;
+    if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
+                     TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
+                     0,
+                     KEY_READ,
+                     &hKey) != ERROR_SUCCESS) {
+        // fail to open registry key
+        return "";
+    }
+    char cpu_brand[256];
+    DWORD cpu_brand_size = sizeof(cpu_brand);
+    if (RegQueryValueExA(hKey,
+                        TEXT("ProcessorNameString"),
+                        NULL,
+                        NULL,
+                        (LPBYTE)cpu_brand,
+                        &cpu_brand_size) == ERROR_SUCCESS) {
+        id.assign(cpu_brand, cpu_brand_size);
+    }
+    RegCloseKey(hKey);
 #endif
     // TODO: other platforms
     return id;

llama_cpp_python-0.2.87/vendor/llama.cpp/examples/llama.swiftui/llama.swiftui.xcodeproj/project.xcworkspace/xcshareddata/IDEWorkspaceChecks.plist

@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <key>IDEDidComputeMac32BitWarning</key>
+    <true/>
+</dict>
+</plist>

llama_cpp_python-0.2.87/vendor/llama.cpp/examples/sycl/README.md

@@ -0,0 +1,41 @@
+# llama.cpp/example/sycl
+
+This example program provides the tools for llama.cpp for SYCL on Intel GPU.
+
+## Tool
+
+|Tool Name| Function|Status|
+|-|-|-|
+|llama-ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, ect.|Support|
+
+### llama-ls-sycl-device
+
+List all SYCL devices with ID, compute capability, max work group size, ect.
+
+1. Build the llama.cpp for SYCL for the specified target *(using GGML_SYCL_TARGET)*.
+
+2. Enable oneAPI running environment *(if GGML_SYCL_TARGET is set to INTEL -default-)*
+
+```
+source /opt/intel/oneapi/setvars.sh
+```
+
+3. Execute
+
+```
+./build/bin/llama-ls-sycl-device
+```
+
+Check the ID in startup log, like:
+
+```
+found 2 SYCL devices:
+|  |                   |                                       |       |Max    |        |Max  |Global |                     |
+|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
+|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
+|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
+| 0| [level_zero:gpu:0]|                Intel Arc A770 Graphics|    1.3|    512|    1024|   32| 16225M|            1.3.29138|
+| 1| [level_zero:gpu:1]|                 Intel UHD Graphics 750|    1.3|     32|     512|   32| 62631M|            1.3.29138|
+
+```
+

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/vendor/llama.cpp/ggml/src/ggml-backend.c

@@ -351,15 +351,10 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b
     }
 
     // an async copy would normally happen after all the queued operations on both backends are completed
-    // sync src, set_async dst
-    if (ggml_backend_buffer_is_host(src->buffer)) {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
-    } else {
-        ggml_backend_synchronize(backend_src);
-        ggml_backend_tensor_copy(src, dst);
-        ggml_backend_synchronize(backend_dst);
-    }
+    // to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy
+    ggml_backend_synchronize(backend_src);
+    ggml_backend_synchronize(backend_dst);
+    ggml_backend_tensor_copy(src, dst);
 }
 
 // events
@@ -1782,7 +1777,17 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
+                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                    ggml_backend_synchronize(input_backend);
+                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                    } else {
+                        ggml_backend_synchronize(split_backend);
+                    }
+                    ggml_backend_tensor_copy(input, input_cpy);
+                }
             }
         }
 

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/vendor/llama.cpp/ggml/src/ggml-cuda.cu

@@ -2358,33 +2358,35 @@ GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend,
 }
 
 GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_backend_is_cuda(backend_src) || ggml_backend_is_cuda(backend_dst));
-
     ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
     ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
 
-    if (!ggml_backend_buffer_is_cuda(src->buffer)) {
+    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
         return false;
     }
 
-    if (!ggml_backend_buffer_is_cuda(dst->buffer)) {
+    if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
         return false;
     }
 
-    // device -> device
+    // device -> device copy
     ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
     ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
 
-    if (backend_src != backend_dst) {
-        ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
-        ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
+    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
+    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
 
-        GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device);
-        GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device);
+    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
+#ifndef NDEBUG
+        GGML_CUDA_LOG_WARN("%s: backend and buffer devices do not match\n", __func__);
+#endif
+        return false;
+    }
 
+    if (backend_src != backend_dst) {
         // copy on src stream
         if (cuda_ctx_src->device == cuda_ctx_dst->device) {
-            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
         } else {
 #ifdef GGML_CUDA_NO_PEER_COPY
             return false;
@@ -2393,7 +2395,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
 #endif
         }
 
-        // record event on src stream
+        // record event on src stream after the copy
         if (!cuda_ctx_src->copy_event) {
             ggml_cuda_set_device(cuda_ctx_src->device);
             CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
@@ -2405,7 +2407,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
         CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0));
     } else {
         // src and dst are on the same backend
-        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
     }
     return true;
 }
@@ -2742,11 +2744,12 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_MUL_MAT_ID:
             {
                 struct ggml_tensor * a = op->src[0];
-                if (op->op == GGML_OP_MUL_MAT) {
-                    struct ggml_tensor * b = op->src[1];
-                    if (a->ne[3] != b->ne[3]) {
-                        return false;
-                    }
+                struct ggml_tensor * b = op->src[1];
+                if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
+                    return false;
+                }
+                if (op->op == GGML_OP_MUL_MAT && a->ne[3] != b->ne[3]) {
+                    return false;
                 }
                 switch (a->type) {
                     case GGML_TYPE_F32:
@@ -2877,7 +2880,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-            return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
+            return (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) || op->src[0]->ne[0] == 128;
 #else
             if (op->src[0]->ne[0] == 128) {
                 return true;

{llama_cpp_python-0.2.86 → llama_cpp_python-0.2.87}/vendor/llama.cpp/ggml/src/ggml-sycl/dpct/helper.hpp

@@ -874,7 +874,7 @@ namespace dpct
         inline std::string get_preferred_gpu_platform_name() {
             std::string result;
 
-            std::string filter = "level-zero";
+            std::string filter = "";
             char* env = getenv("ONEAPI_DEVICE_SELECTOR");
             if (env) {
                 if (std::strstr(env, "level_zero")) {
@@ -892,11 +892,24 @@ namespace dpct
                 else {
                     throw std::runtime_error("invalid device filter: " + std::string(env));
                 }
+            } else {
+                auto default_device = sycl::device(sycl::default_selector_v);
+                auto default_platform_name = default_device.get_platform().get_info<sycl::info::platform::name>();
+
+                if (std::strstr(default_platform_name.c_str(), "Level-Zero") || default_device.is_cpu()) {
+                    filter = "level-zero";
+                }
+                else if (std::strstr(default_platform_name.c_str(), "CUDA")) {
+                    filter = "cuda";
+                }
+                else if (std::strstr(default_platform_name.c_str(), "HIP")) {
+                    filter = "hip";
+                }
             }
 
-            auto plaform_list = sycl::platform::get_platforms();
+            auto platform_list = sycl::platform::get_platforms();
 
-            for (const auto& platform : plaform_list) {
+            for (const auto& platform : platform_list) {
                 auto devices = platform.get_devices();
                 auto gpu_dev = std::find_if(devices.begin(), devices.end(), [](const sycl::device& d) {
                     return d.is_gpu();