updates for nm-magic-wand, nightly or release (#247) (#248)

SUMMARY:
* update GHA action `nm-build-vllm` to not install
`nm-magic-wand-nightly`
* update build script to not install `nm-magic-wand-nightly` (we might
consider getting rid of this script altogether, since we aren't really
using it)
* remove unused GHA action `nm-test-vllm`. this has been superseded by
`nm-install-test-whl`
* update GHA action `nm-install-test-whl` to get version of
`nm-magic-wand` if `nm-magic-wand-nightly` is not present
* update `setup.py` to default generate "nightly" package and add option
based on ENV to generate release package. this also includes managing
the dependency on `nm-magic-wand`.
* update `set-env` action to set ENV based on `wf_category` input
* update "release" workflow to include all supported python versions
* delete obsolete "gen-whl"

NOTES:
- "magic-wand" is only a runtime dependency, so no need to install it
during build phase.
- this PR makes it so that we by default generate a "nightly" package
with a "nightly" version number. if we want to generate a release
package we'll need to specify `wf_category` as `RELEASE`.

TEST PLAN:
runs on remote push. verifying that `wf_category` set to `RELEASE` will
generate appropriate package.

ran "build" workflow with `wf_category` set to `RELEASE` ... package
looks properly named and versioned ...
https://github.com/neuralmagic/nm-vllm/actions/runs/9129675592

the "remote push" defaulted to generating a "nightly" package ... please
see ... https://github.com/neuralmagic/nm-vllm/actions/runs/9129665988

Co-authored-by: andy-neuma <[email protected]>

Author: andy-neuma

.github/actions/nm-build-vllm/action.yml

@@ -32,7 +32,6 @@ runs:
       VENV="${{ inputs.venv }}-${COMMIT:0:7}"
       source $(pyenv root)/versions/${{ inputs.python }}/envs/${VENV}/bin/activate
       # TODO: adjust when we need a proper release. use nightly now.
-      pip3 install nm-magic-wand-nightly
       pip3 install -r requirements-cuda.txt -r requirements-build.txt
       # build
       SUCCESS=0

.github/actions/nm-install-test-whl/action.yml

@@ -52,6 +52,9 @@ runs:
         pip3 install ${WHL}[sparse]
         # report magic_wand version
         MAGIC_WAND=$(pip3 show nm-magic-wand-nightly | grep "Version" | cut -d' ' -f2)
+        if [ -z "${MAGIC_WAND}" ]; then
+          MAGIC_WAND=$(pip3 show nm-magic-wand | grep "Version" | cut -d' ' -f2)
+        fi
         echo "magic_wand=${MAGIC_WAND}" >> "$GITHUB_OUTPUT"
         # test and collect code coverage
         SUCCESS=0

.github/actions/nm-set-env/action.yml

@@ -1,6 +1,9 @@
 name: set neuralmagic env
 description: 'sets environment variables for neuralmagic'
 inputs:
+  wf_category:
+    description: "categories: REMOTE, NIGHTLY, RELEASE"
+    required: true
   hf_token:
     description: 'Hugging Face home'
     required: true
@@ -14,19 +17,27 @@ runs:
   using: composite
   steps:
   - run: |
+      # setup.py defaults to making 'nightly' package with 'nightly' version
+      if [[ "${{inputs.wf_category}}" == "RELEASE" ]]; then
+        echo "NM_RELEASE_TYPE=${{inputs.wf_category}}" >> $GITHUB_ENV
+      fi
+      # CUDA
       echo "TORCH_CUDA_ARCH_LIST=7.0 7.5 8.0 8.6 8.9 9.0+PTX" >> $GITHUB_ENV
+      echo "PATH=/usr/local/apps/pyenv/plugins/pyenv-virtualenv/shims:/usr/local/apps/pyenv/shims:/usr/local/apps/pyenv/bin:/usr/local/apps/nvm/versions/node/v19.9.0/bin:/usr/local/apps/nvm/versions/node/v16.20.2/bin:/usr/local/cuda-12.1/bin:/usr/local/cuda-12.1/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin:/home/${WHOAMI}/.local/bin:" >> $GITHUB_ENV
+      echo "LD_LIBRARY_PATH=/usr/local/cuda-12.1/lib64::/usr/local/cuda-12.1/lib64:" >> $GITHUB_ENV
+      # HF Cache
       echo "HF_TOKEN=${HF_TOKEN_SECRET}" >> $GITHUB_ENV
       echo "HF_HOME=/EFS/hf_home" >> $GITHUB_ENV
+      # build
       NUM_THREADS=$(./.github/scripts/determine-threading -G ${{ inputs.Gi_per_thread }})
       echo "MAX_JOBS=${NUM_THREADS}" >> $GITHUB_ENV
       echo "NVCC_THREADS=${{ inputs.nvcc_threads }}" >> $GITHUB_ENV
       echo "VLLM_INSTALL_PUNICA_KERNELS=1" >> $GITHUB_ENV
       echo "NCCL_IGNORE_DISABLED_P2P=1" >> $GITHUB_ENV
+      # pyenv
       echo "PYENV_ROOT=/usr/local/apps/pyenv" >> $GITHUB_ENV
+      # testmo
       echo "XDG_CONFIG_HOME=/usr/local/apps" >> $GITHUB_ENV
-      WHOAMI=$(whoami)
-      echo "PATH=/usr/local/apps/pyenv/plugins/pyenv-virtualenv/shims:/usr/local/apps/pyenv/shims:/usr/local/apps/pyenv/bin:/usr/local/apps/nvm/versions/node/v19.9.0/bin:/usr/local/apps/nvm/versions/node/v16.20.2/bin:/usr/local/cuda-12.1/bin:/usr/local/cuda-12.1/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin:/home/${WHOAMI}/.local/bin:" >> $GITHUB_ENV
-      echo "LD_LIBRARY_PATH=/usr/local/cuda-12.1/lib64::/usr/local/cuda-12.1/lib64:" >> $GITHUB_ENV
       echo "PROJECT_ID=12" >> $GITHUB_ENV
     env:
         HF_TOKEN_SECRET: ${{ inputs.hf_token }}

.github/actions/nm-test-vllm/action.yml

@@ -6,7 +6,6 @@ usage() {
     echo
     echo "usage: ${0} <options>"
     echo
-    echo "  -a    - pypi server address"
     echo "  -p    - python version"
     echo "  -v    - name for virtualenv"
     echo "  -h    - this list of options"
@@ -17,15 +16,12 @@ PYPI_IP=
 PYTHON=
 VENV=
 
-while getopts "ha:p:v:" OPT; do
+while getopts "hp:v:" OPT; do
     case "${OPT}" in
         h)
             usage
             exit 1
             ;;
-        a)
-            PYPI_IP="${OPTARG}"
-            ;;
         p)
             PYTHON="${OPTARG}"
             ;;
@@ -36,12 +32,6 @@ while getopts "ha:p:v:" OPT; do
 done
 
 # check if variables are valid
-if [ -z "${PYPI_IP}" ]; then
-    echo "please provide 'pypi' server address"
-    usage
-    exit 1
-fi
-
 if [ -z "${PYTHON}" ]; then
     echo "please provide python version, e.g. 3.10.12"
     usage
@@ -55,6 +45,5 @@ if [ -z "${VENV}" ]; then
 fi
 
 source $(pyenv root)/versions/${PYTHON}/envs/${VENV}/bin/activate
-pip3 install --index-url http://${PYPI_IP}:8080/ --trusted-host ${PYPI_IP} nm-magic-wand-nightly
 pip3 install -r requirements-cuda.txt -r requirements-build.txt
 pip3 install -e .

.github/scripts/build

@@ -116,6 +116,7 @@ jobs:
     BUILD:
         uses: ./.github/workflows/build.yml
         with:
+            wf_category: ${{ inputs.wf_category }}
             build_label: ${{ inputs.build_label }}
             timeout: ${{ inputs.build_timeout }}
             gitref: ${{ github.ref }}

.github/workflows/build-test.yml

@@ -3,6 +3,10 @@ on:
   # makes workflow reusable
   workflow_call:
     inputs:
+      wf_category:
+        description: "categories: REMOTE, NIGHTLY, RELEASE"
+        type: string
+        default: "REMOTE"
       build_label:
         description: "requested runner label (specifies instance)"
         type: string
@@ -31,6 +35,10 @@ on:
   # makes workflow manually callable
   workflow_dispatch:
     inputs:
+      wf_category:
+        description: "categories: REMOTE, NIGHTLY, RELEASE"
+        type: string
+        default: "REMOTE"
       build_label:
         description: "requested runner label (specifies instance)"
         type: string
@@ -83,6 +91,7 @@ jobs:
               id: setenv
               uses: ./.github/actions/nm-set-env/
               with:
+                wf_category: ${{ inputs.wf_category }}
                 hf_token: ${{ secrets.NM_HF_TOKEN }}
                 Gi_per_thread: ${{ inputs.Gi_per_thread }}
                 nvcc_threads: ${{ inputs.nvcc_threads }}

.github/workflows/build.yml

@@ -17,7 +17,43 @@ concurrency:
 
 jobs:
 
-  BUILD-TEST:
+  PYTHON-3-8:
+    uses: ./.github/workflows/build-test.yml
+    with:
+      wf_category: 'RELEASE'
+      python: 3.8.17
+      gitref: ${{ github.ref }}
+
+      test_label_solo: aws-avx2-32G-a10g-24G
+      test_label_multi: aws-avx2-192G-4-a10g-96G
+      test_timeout: 480
+      test_skip_list: neuralmagic/tests/skip-for-release.txt
+
+      benchmark_label: aws-avx2-32G-a10g-24G
+      benchmark_config_list_file: ./.github/data/nm_benchmark_nightly_configs_list.txt
+      benchmark_timeout: 180
+      push_benchmark_results_to_gh_pages: ${{ inputs.push_benchmark_results_to_gh_pages }}
+    secrets: inherit
+
+  PYTHON-3-9:
+    uses: ./.github/workflows/build-test.yml
+    with:
+      wf_category: 'RELEASE'
+      python: 3.9.17
+      gitref: ${{ github.ref }}
+
+      test_label_solo: aws-avx2-32G-a10g-24G
+      test_label_multi: aws-avx2-192G-4-a10g-96G
+      test_timeout: 480
+      test_skip_list: neuralmagic/tests/skip-for-release.txt
+
+      benchmark_label: aws-avx2-32G-a10g-24G
+      benchmark_config_list_file: ./.github/data/nm_benchmark_nightly_configs_list.txt
+      benchmark_timeout: 180
+      push_benchmark_results_to_gh_pages: ${{ inputs.push_benchmark_results_to_gh_pages }}
+    secrets: inherit
+
+  PYTHON-3-10:
     uses: ./.github/workflows/build-test.yml
     with:
       wf_category: 'RELEASE'
@@ -34,3 +70,21 @@ jobs:
       benchmark_timeout: 180
       push_benchmark_results_to_gh_pages: ${{ inputs.push_benchmark_results_to_gh_pages }}
     secrets: inherit
+
+  PYTHON-3-11:
+    uses: ./.github/workflows/build-test.yml
+    with:
+      wf_category: 'RELEASE'
+      python: 3.11.4
+      gitref: ${{ github.ref }}
+
+      test_label_solo: aws-avx2-32G-a10g-24G
+      test_label_multi: aws-avx2-192G-4-a10g-96G
+      test_timeout: 480
+      test_skip_list: neuralmagic/tests/skip-for-release.txt
+
+      benchmark_label: aws-avx2-32G-a10g-24G
+      benchmark_config_list_file: ./.github/data/nm_benchmark_nightly_configs_list.txt
+      benchmark_timeout: 180
+      push_benchmark_results_to_gh_pages: ${{ inputs.push_benchmark_results_to_gh_pages }}
+    secrets: inherit

.github/workflows/gen-whl.yml

@@ -2,6 +2,7 @@
 # UPSTREAM SYNC: noqa is required for passing ruff.
 # This file has been modified by Neural Magic
 
+import datetime
 import importlib.util
 import io
 import logging
@@ -306,9 +307,28 @@ def find_version(filepath: str) -> str:
         raise RuntimeError("Unable to find version string.")
 
 
+# Neuralmagic packaging ENV's
+NM_RELEASE_TYPE = 'NM_RELEASE_TYPE'
+
+
+def get_nm_vllm_package_name() -> str:
+    nm_release_type = os.getenv(NM_RELEASE_TYPE)
+    package_name = None
+    if nm_release_type == 'RELEASE':
+        package_name = 'nm-vllm'
+    else:
+        package_name = 'nm-vllm-nightly'
+    return package_name
+
+
 def get_vllm_version() -> str:
     version = find_version(get_path("vllm", "__init__.py"))
 
+    nm_release_type = os.getenv(NM_RELEASE_TYPE)
+    if nm_release_type != 'RELEASE':
+        date = datetime.date.today().strftime("%Y%m%d")
+        version += f'.{date}'
+
     if _is_cuda():
         cuda_version = str(get_nvcc_cuda_version())
         if cuda_version != MAIN_CUDA_VERSION:
@@ -393,6 +413,9 @@ def _read_requirements(filename: str) -> List[str]:
 
 # UPSTREAM SYNC: needed for sparsity
 _sparsity_deps = ["nm-magic-wand-nightly"]
+nm_release_type = os.getenv(NM_RELEASE_TYPE)
+if nm_release_type == 'RELEASE':
+    _sparsity_deps = ["nm-magic-wand"]
 
 package_data = {
     "vllm": ["py.typed", "model_executor/layers/fused_moe/configs/*.json"]
@@ -402,7 +425,7 @@ def _read_requirements(filename: str) -> List[str]:
     package_data["vllm"].append("*.so")
 
 setup(
-    name="nm-vllm",
+    name=get_nm_vllm_package_name(),
     version=get_vllm_version(),
     author="vLLM Team, Neural Magic",
     author_email="[email protected]",