[GITHUB ACTION][NIXL]update install_nixl.py script (#543)

Follow upstream fix to fix our install_nixl script to work with nixl
0.7.0

Signed-off-by: Chendi Xue <chendi.xue@intel.com>

Author: xuechendi

.github/workflows/create-release-branch.yaml

@@ -315,7 +315,7 @@ jobs:
             -e HF_HOME=/workspace/hf_cache \
             -e HF_TOKEN=${{ secrets.HF_TOKEN }} \
             -v /mnt/hf_cache:/workspace/hf_cache \
-            -v /mnt/wheels_cache:/workspace/wheels_cache \
+            -v /mnt/wheels_cache:/tmp/wheels_cache \
             hpu-plugin-v1-${{ needs.prepare-release-branch.outputs.tag_name }} \
             /bin/bash -c "
               pip install lm-eval[api] &&

.github/workflows/hourly-ci.yaml

@@ -258,7 +258,7 @@ jobs:
             -e HF_HOME=/workspace/hf_cache \
             -e HF_TOKEN=${{ secrets.HF_TOKEN }} \
             -v /mnt/hf_cache:/workspace/hf_cache \
-            -v /mnt/wheels_cache:/workspace/wheels_cache \
+            -v /mnt/wheels_cache:/tmp/wheels_cache \
             hpu-plugin-v1-test-env-hourly-ci \
             /bin/bash -c "
               pip install lm-eval[api] &&

.github/workflows/pre-merge.yaml

@@ -379,7 +379,7 @@ jobs:
             -e HF_HOME=/workspace/hf_cache \
             -e HF_TOKEN=${{ secrets.HF_TOKEN }} \
             -v /mnt/hf_cache:/workspace/hf_cache \
-            -v /mnt/wheels_cache:/workspace/wheels_cache \
+            -v /mnt/wheels_cache:/tmp/wheels_cache \
             hpu-plugin-v1-test-env-pre-merge-${{ github.event.pull_request.head.sha }} \
             /bin/bash -c "
                 pip install lm-eval[api] &&

install_nixl.py

@@ -4,9 +4,11 @@
 import sys
 import argparse
 import glob
+import json
+import urllib
 
 # --- Configuration ---
-WHEELS_CACHE_HOME = os.environ.get("WHEELS_CACHE_HOME", "/workspace/wheels_cache")
+WHEELS_CACHE_HOME = os.environ.get("WHEELS_CACHE_HOME", "/tmp/wheels_cache")
 ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
 UCX_DIR = os.path.join('/tmp', 'ucx_source')
 NIXL_DIR = os.path.join('/tmp', 'nixl_source')
@@ -16,6 +18,20 @@
 
 
 # --- Helper Functions ---
+def get_latest_nixl_version():
+    """Helper function to get latest release version of NIXL"""
+    try:
+        nixl_release_url = "https://api.github.com/repos/ai-dynamo/nixl/releases/latest"
+        with urllib.request.urlopen(nixl_release_url) as response:
+            data = json.load(response)
+            return data.get("tag_name", "0.7.0")
+    except Exception:
+        return "0.7.0"
+
+
+NIXL_VERSION = os.environ.get("NIXL_VERSION", get_latest_nixl_version())
+
+
 def run_command(command, cwd='.', env=None):
     """Helper function to run a shell command and check for errors."""
     print(f"--> Running command: {' '.join(command)} in '{cwd}'", flush=True)
@@ -33,7 +49,7 @@ def is_pip_package_installed(package_name):
 def find_nixl_wheel_in_cache(cache_dir):
     """Finds a nixl wheel file in the specified cache directory."""
     # The repaired wheel will have a 'manylinux' tag, but this glob still works.
-    search_pattern = os.path.join(cache_dir, "nixl-*.whl")
+    search_pattern = os.path.join(cache_dir, f"nixl*{NIXL_VERSION}*.whl")
     wheels = glob.glob(search_pattern)
     if wheels:
         # Sort to get the most recent/highest version if multiple exist
@@ -125,6 +141,10 @@ def build_and_install_prerequisites(args):
     print("\n[2/3] Building NIXL wheel from source...", flush=True)
     if not os.path.exists(NIXL_DIR):
         run_command(['git', 'clone', NIXL_REPO_URL, NIXL_DIR])
+    else:
+        run_command(["git", "fetch", "--tags"], cwd=NIXL_DIR)
+    run_command(["git", "checkout", NIXL_VERSION], cwd=NIXL_DIR)
+    print(f"--> Checked out NIXL version: {NIXL_VERSION}", flush=True)
 
     build_env = os.environ.copy()
     build_env['PKG_CONFIG_PATH'] = os.path.join(ucx_install_path, 'lib', 'pkgconfig')
@@ -169,7 +189,14 @@ def build_and_install_prerequisites(args):
 
     print(f"--> Successfully built self-contained wheel: {os.path.basename(newly_built_wheel)}. Now installing...",
           flush=True)
-    install_command = [sys.executable, '-m', 'pip', 'install', newly_built_wheel]
+    install_command = [
+        sys.executable,
+        "-m",
+        "pip",
+        "install",
+        "--no-deps",  # w/o "no-deps", it will install cuda-torch
+        newly_built_wheel,
+    ]
     if args.force_reinstall:
         install_command.insert(-1, '--force-reinstall')