Optimize GPTQ-Pro kernels for Ampere decode and prefill

.github/workflows/gptq-pro-cuda-compile.yml

@@ -0,0 +1,139 @@
+name: GPTQ-Pro Kernel CI
+
+on:
+  pull_request:
+    paths:
+      - "gptqmodel_ext/gptq_pro/**"
+      - "gptqmodel/utils/gptq_pro.py"
+      - "gptqmodel/nn_modules/qlinear/gptq_pro.py"
+      - "scripts/benchmark_gptq_pro_kernel.py"
+      - "tests/kernels/test_gptq_pro_ampere_pipeline.py"
+      - ".github/workflows/gptq-pro-cuda-compile.yml"
+  push:
+    branches:
+      - perf/ampere-kernel-pipeline
+    paths:
+      - "gptqmodel_ext/gptq_pro/**"
+      - "gptqmodel/utils/gptq_pro.py"
+      - "gptqmodel/nn_modules/qlinear/gptq_pro.py"
+      - "scripts/benchmark_gptq_pro_kernel.py"
+      - "tests/kernels/test_gptq_pro_ampere_pipeline.py"
+      - ".github/workflows/gptq-pro-cuda-compile.yml"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  python-style:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.13"
+      - name: Install Ruff
+        run: python -m pip install ruff==0.13.0
+      - name: Check modified Python sources
+        run: |
+          ruff check \
+            gptqmodel/utils/gptq_pro.py \
+            gptqmodel/nn_modules/qlinear/gptq_pro.py \
+            scripts/benchmark_gptq_pro_kernel.py \
+            tests/kernels/test_gptq_pro_ampere_pipeline.py
+          ruff format --check \
+            gptqmodel/utils/gptq_pro.py \
+            gptqmodel/nn_modules/qlinear/gptq_pro.py \
+            scripts/benchmark_gptq_pro_kernel.py \
+            tests/kernels/test_gptq_pro_ampere_pipeline.py
+          python -m py_compile \
+            gptqmodel/utils/gptq_pro.py \
+            gptqmodel/nn_modules/qlinear/gptq_pro.py \
+            scripts/benchmark_gptq_pro_kernel.py \
+            tests/kernels/test_gptq_pro_ampere_pipeline.py
+
+  cuda-compile:
+    runs-on: ubuntu-24.04
+    container:
+      image: nvidia/cuda:12.6.3-devel-ubuntu22.04
+    steps:
+      - uses: actions/checkout@v6
+      - name: Install host compiler
+        run: |
+          apt-get update
+          apt-get install -y --no-install-recommends build-essential
+          rm -rf /var/lib/apt/lists/*
+      - name: Compile standalone validator for Ampere targets
+        working-directory: gptqmodel_ext/gptq_pro
+        run: |
+          nvcc -O3 -std=c++17 -lineinfo \
+            -U__CUDA_NO_HALF_OPERATORS__ \
+            -U__CUDA_NO_HALF_CONVERSIONS__ \
+            -gencode arch=compute_80,code=sm_80 \
+            -gencode arch=compute_86,code=sm_86 \
+            -gencode arch=compute_87,code=sm_87 \
+            -gencode arch=compute_87,code=compute_87 \
+            gptq_pro_validate.cu gptq_pro_kernel.cu \
+            -o /tmp/gptq_pro_validate
+      - name: Record binary metadata
+        run: |
+          test -x /tmp/gptq_pro_validate
+          ls -lh /tmp/gptq_pro_validate
+
+  pytorch-extension-compile:
+    runs-on: ubuntu-24.04
+    container:
+      image: nvidia/cuda:12.6.3-devel-ubuntu22.04
+    env:
+      TORCH_CUDA_ARCH_LIST: "8.0;8.6;8.7+PTX"
+    steps:
+      - uses: actions/checkout@v6
+      - name: Install isolated Python build environment
+        run: |
+          apt-get update
+          apt-get install -y --no-install-recommends \
+            build-essential python3 python3-pip python3-dev python3-venv ninja-build
+          rm -rf /var/lib/apt/lists/*
+          python3 -m venv /tmp/gptq-pro-ci
+          /tmp/gptq-pro-ci/bin/python -m pip install --upgrade pip setuptools wheel ninja
+          /tmp/gptq-pro-ci/bin/python -m pip install \
+            --index-url https://download.pytorch.org/whl/cu126 \
+            torch==2.8.0
+          /tmp/gptq-pro-ci/bin/python -m pip install pytest
+      - name: Compile and import PyTorch extension
+        run: |
+          mkdir -p /tmp/gptq_pro_extension_ci
+          /tmp/gptq-pro-ci/bin/python - <<'PY'
+          from pathlib import Path
+          from torch.utils.cpp_extension import load
+
+          root = Path.cwd()
+          extension = load(
+              name="gptqmodel_gptq_pro_ci",
+              sources=[
+                  str(root / "gptqmodel_ext/gptq_pro/gptq_pro_torch.cpp"),
+                  str(root / "gptqmodel_ext/gptq_pro/gptq_pro_kernel.cu"),
+              ],
+              extra_cflags=["-O3", "-std=c++17"],
+              extra_cuda_cflags=[
+                  "-O3",
+                  "-std=c++17",
+                  "-lineinfo",
+                  "-U__CUDA_NO_HALF_OPERATORS__",
+                  "-U__CUDA_NO_HALF_CONVERSIONS__",
+                  "-gencode=arch=compute_80,code=sm_80",
+                  "-gencode=arch=compute_86,code=sm_86",
+                  "-gencode=arch=compute_87,code=sm_87",
+                  "-gencode=arch=compute_87,code=compute_87",
+              ],
+              build_directory="/tmp/gptq_pro_extension_ci",
+              verbose=True,
+          )
+          assert hasattr(extension, "gptq_pro_gemm")
+          print(extension.gptq_pro_gemm.__doc__)
+          PY
+      - name: Run CPU packing and dispatch contracts
+        run: |
+          /tmp/gptq-pro-ci/bin/python -m pytest -q \
+            --confcutdir=tests/kernels \
+            tests/kernels/test_gptq_pro_ampere_pipeline.py