chore: bump version to 0.5.3 (#1909)

docker/Dockerfile

@@ -30,8 +30,8 @@ ARG GITHUB_ARTIFACTORY=github.com
 ARG FLASHINFER_VERSION=0.5.3
 
 # ktransformers wheel version (cu128torch28 for CUDA 12.8 + PyTorch 2.8)
-ARG KTRANSFORMERS_VERSION=0.4.2
-ARG KTRANSFORMERS_WHEEL=ktransformers-0.4.2+cu128torch28fancy-cp312-cp312-linux_x86_64.whl
+ARG KTRANSFORMERS_VERSION=0.5.3
+ARG KTRANSFORMERS_WHEEL=ktransformers-0.5.3+cu128torch28fancy-cp312-cp312-linux_x86_64.whl
 
 # flash_attn wheel for fine-tune env
 ARG FLASH_ATTN_WHEEL=flash_attn-2.8.3+cu12torch2.8cxx11abiTRUE-cp312-cp312-linux_x86_64.whl

docker/README-packaging.md

@@ -25,16 +25,16 @@ sglang-v{sglang版本}_ktransformers-v{ktransformers版本}_{cpu信息}_{gpu信
 
 **Tar file:**
 ```
-sglang-v0.5.6_ktransformers-v0.4.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022.tar
+sglang-v0.5.6_ktransformers-v0.5.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022.tar
 ```
 
 **DockerHub tags:**
 ```
 Full tag:
-kvcache/ktransformers:sglang-v0.5.6_ktransformers-v0.4.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022
+kvcache/ktransformers:sglang-v0.5.6_ktransformers-v0.5.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022
 
 Simplified tag:
-kvcache/ktransformers:v0.4.3-cu128
+kvcache/ktransformers:v0.5.3-cu128
 ```
 
 ### Name Components
@@ -42,7 +42,7 @@ kvcache/ktransformers:v0.4.3-cu128
 | Component | Description | Example |
 |-----------|-------------|---------|
 | sglang version | SGLang package version | `v0.5.6` |
-| ktransformers version | KTransformers version | `v0.4.3` |
+| ktransformers version | KTransformers version | `v0.5.3` |
 | cpu info | CPU instruction set support | `x86-intel-multi` (includes AMX/AVX512/AVX2) |
 | gpu info | CUDA version | `cu128` (CUDA 12.8) |
 | functionality | Feature mode | `sft_llamafactory-v0.9.3` or `infer` |
@@ -197,8 +197,8 @@ docker login
 ```
 
 This creates two tags:
-- Full: `kvcache/ktransformers:sglang-v0.5.6_ktransformers-v0.4.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022`
-- Simplified: `kvcache/ktransformers:v0.4.3-cu128`
+- Full: `kvcache/ktransformers:sglang-v0.5.6_ktransformers-v0.5.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022`
+- Simplified: `kvcache/ktransformers:v0.5.3-cu128`
 
 ### Example 4: Dry Run
 
@@ -225,12 +225,12 @@ Pass additional Docker build arguments:
 
 ```bash
 # Load the image
-docker load -i sglang-v0.5.6_ktransformers-v0.4.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022.tar
+docker load -i sglang-v0.5.6_ktransformers-v0.5.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022.tar
 
 # Run the container
 docker run -it --rm \
   --gpus all \
-  sglang-v0.5.6_ktransformers-v0.4.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022 \
+  sglang-v0.5.6_ktransformers-v0.5.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022 \
   /bin/bash
 ```
 
@@ -238,15 +238,15 @@ docker run -it --rm \
 
 ```bash
 # Pull with full tag
-docker pull kvcache/ktransformers:sglang-v0.5.6_ktransformers-v0.4.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022
+docker pull kvcache/ktransformers:sglang-v0.5.6_ktransformers-v0.5.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022
 
 # Or pull with simplified tag
-docker pull kvcache/ktransformers:v0.4.3-cu128
+docker pull kvcache/ktransformers:v0.5.3-cu128
 
 # Run the container
 docker run -it --rm \
   --gpus all \
-  kvcache/ktransformers:v0.4.3-cu128 \
+  kvcache/ktransformers:v0.5.3-cu128 \
   /bin/bash
 ```
 
@@ -301,7 +301,7 @@ cat /workspace/versions.env
 
 # Output:
 SGLANG_VERSION=0.5.6
-KTRANSFORMERS_VERSION=0.4.3
+KTRANSFORMERS_VERSION=0.5.3
 LLAMAFACTORY_VERSION=0.9.3
 ```
 

docker/docker-utils.sh

@@ -210,9 +210,9 @@ generate_image_name() {
 
 # Generate simplified tag for DockerHub
 # Input:
-#   $1: ktransformers_version (e.g., 0.4.3)
+#   $1: ktransformers_version (e.g., 0.5.3)
 #   $2: cuda_version (e.g., 12.8.1)
-# Output: Simplified tag (e.g., v0.4.3-cu128)
+# Output: Simplified tag (e.g., v0.5.3-cu128)
 generate_simplified_tag() {
     local ktrans_ver="$1"
     local cuda_version="$2"

docker/push-to-dockerhub.sh

@@ -137,13 +137,13 @@ OUTPUT:
       {registry}/{repository}:sglang-v{ver}_ktransformers-v{ver}_{cpu}_{gpu}_{func}_{timestamp}
 
     Example:
-      docker.io/kvcache/ktransformers:sglang-v0.5.6_ktransformers-v0.4.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022
+      docker.io/kvcache/ktransformers:sglang-v0.5.6_ktransformers-v0.5.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022
 
     Simplified tag (if --also-push-simplified):
       {registry}/{repository}:v{ktransformers-ver}-{cuda}
 
     Example:
-      docker.io/kvcache/ktransformers:v0.4.3-cu128
+      docker.io/kvcache/ktransformers:v0.5.3-cu128
 
 EOF
     exit 0
@@ -372,7 +372,7 @@ generate_tags() {
         log_warning "DRY RUN: Using placeholder versions"
         # Use placeholder versions for dry run
         local versions="SGLANG_VERSION=0.5.6
-KTRANSFORMERS_VERSION=0.4.3
+KTRANSFORMERS_VERSION=0.5.3
 LLAMAFACTORY_VERSION=0.9.3"
     else
         # Extract versions from image
@@ -709,13 +709,13 @@ OUTPUT:
       {registry}/{repository}:sglang-v{ver}_ktransformers-v{ver}_{cpu}_{gpu}_{func}_{timestamp}
 
     Example:
-      docker.io/kvcache/ktransformers:sglang-v0.5.6_ktransformers-v0.4.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022
+      docker.io/kvcache/ktransformers:sglang-v0.5.6_ktransformers-v0.5.3_x86-intel-multi_cu128_sft_llamafactory-v0.9.3_20241212143022
 
     Simplified tag (if --also-push-simplified):
       {registry}/{repository}:v{ktransformers-ver}-{cuda}
 
     Example:
-      docker.io/kvcache/ktransformers:v0.4.3-cu128
+      docker.io/kvcache/ktransformers:v0.5.3-cu128
 
 EOF
     exit 0
@@ -944,7 +944,7 @@ generate_tags() {
         log_warning "DRY RUN: Using placeholder versions"
         # Use placeholder versions for dry run
         local versions="SGLANG_VERSION=0.5.6
-KTRANSFORMERS_VERSION=0.4.3
+KTRANSFORMERS_VERSION=0.5.3
 LLAMAFACTORY_VERSION=0.9.3"
     else
         # Extract versions from image

kt-kernel/CMakeLists.txt

@@ -27,7 +27,7 @@ option(KTRANSFORMERS_CPU_MOE_AMD "ktransformers: CPU use moe kernel for amd" OFF
 # LTO control
 option(CPUINFER_ENABLE_LTO "Enable link time optimization (IPO)" OFF)
 
-project(kt_kernel_ext VERSION 0.5.0)
+project(kt_kernel_ext VERSION 0.5.3)
 
 # Auto-detect CPU features early (unless building with LLAMA_NATIVE)
 if(NOT LLAMA_NATIVE AND NOT MSVC)
@@ -692,4 +692,3 @@ if(NUMA_LIBRARY)
 else()
     message(FATAL_ERROR "NUMA library not found, please install NUMA, sudo apt install libnuma-dev")
 endif()
-

kt-kernel/python/__init__.py

@@ -68,9 +68,9 @@ try:
             _version_ns = {}
             with open(_root_version_file, "r", encoding="utf-8") as f:
                 exec(f.read(), _version_ns)
-            __version__ = _version_ns.get("__version__", "0.4.3")
+            __version__ = _version_ns.get("__version__", "0.5.3")
         else:
-            __version__ = "0.4.3"
+            __version__ = "0.5.3"
 except ImportError:
     # Python < 3.8, fallback to pkg_resources or hardcoded version
     try:
@@ -79,8 +79,8 @@ except ImportError:
         try:
             __version__ = get_distribution("kt-kernel").version
         except DistributionNotFound:
-            __version__ = "0.4.3"
+            __version__ = "0.5.3"
     except ImportError:
-        __version__ = "0.4.3"
+        __version__ = "0.5.3"
 
 __all__ = ["KTMoEWrapper", "generate_gpu_experts_masks", "kt_kernel_ext", "__cpu_variant__", "__version__"]

kt-kernel/python/cli/__init__.py

@@ -5,4 +5,17 @@ This CLI provides a user-friendly interface to all KTransformers functionality,
 including model inference, fine-tuning, benchmarking, and more.
 """
 
-__version__ = "0.1.0"
+from importlib.metadata import PackageNotFoundError, version
+from pathlib import Path
+
+
+try:
+    __version__ = version("kt-kernel")
+except PackageNotFoundError:
+    _version_ns = {}
+    _root_version_file = Path(__file__).resolve().parents[3] / "version.py"
+    if _root_version_file.exists():
+        exec(_root_version_file.read_text(encoding="utf-8"), _version_ns)
+        __version__ = _version_ns.get("__version__", "0.5.3")
+    else:
+        __version__ = "0.5.3"

kt-kernel/setup.py

@@ -695,9 +695,9 @@ if _version_file.exists():
     _version_ns = {}
     with open(_version_file, "r", encoding="utf-8") as f:
         exec(f.read(), _version_ns)
-    _base_version = _version_ns.get("__version__", "0.5.0")
+    _base_version = _version_ns.get("__version__", "0.5.3")
 else:
-    _base_version = "0.5.0"
+    _base_version = "0.5.3"
 
 # Determine version
 if "CPUINFER_VERSION" in os.environ:

version.py

@@ -3,4 +3,4 @@ KTransformers version information.
 Shared across kt-kernel and kt-sft modules.
 """
 
-__version__ = "0.5.2.post1"
+__version__ = "0.5.3"