Update yml file to run 8b tests on presubmit and 70b and 405b tests nightly

.github/workflows/ci-llama-large-tests.yaml

@@ -0,0 +1,90 @@
+# Copyright 2024 Advanced Micro Devices, Inc.
+#
+# Licensed under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+name: Llama Benchmarking Tests
+
+on:
+  workflow_dispatch:
+  schedule:
+    # Weekdays at 6:00 AM UTC = 11:00 PM PST.
+    - cron: "0 6 * * 1-5"
+
+concurrency:
+  # A PR number if a pull request and otherwise the commit hash. This cancels
+  # queued and in-progress runs for the same PR (presubmit) or commit
+  # (postsubmit). The workflow name is prepended to avoid conflicts between
+  # different workflows.
+  group: ${{ github.workflow }}-${{ github.event.number || github.sha }}
+  cancel-in-progress: true
+
+jobs:
+  test_llama_large:
+    name: "Llama Benchmarking Tests"
+    strategy:
+      matrix:
+        version: [3.11]
+      fail-fast: false
+    runs-on: llama-mi300x-1
+    defaults:
+      run:
+        shell: bash
+    env:
+      PIP_CACHE_DIR: "${{ github.workspace }}/.pip-cache"
+      VENV_DIR: ${{ github.workspace }}/.venv
+    steps:
+      - name: Get Current Date
+        id: date
+        run: echo "::set-output name=date::$(date +'%Y-%m-%d')"
+
+      - name: "Setting up Python"
+        id: setup_python
+        uses: actions/setup-python@v3
+        with:
+          python-version: ${{matrix.version}}
+
+      - name: "Checkout Code"
+        uses: actions/checkout@v3
+
+      - name: Cache Pip Packages
+        uses: actions/cache@v4
+        id: cache-pip
+        with:
+          path: ${{ env.PIP_CACHE_DIR }}
+          key: pip-${{ steps.setup_python.outputs.python-version }}-${{ hashFiles('*requirements.txt') }}
+
+      - name: Install pip deps
+        run: |
+          python -m pip install --no-compile --upgrade pip
+          # Note: We install in three steps in order to satisfy requirements
+          # from non default locations first. Installing the PyTorch CPU
+          # wheels saves multiple minutes and a lot of bandwidth on runner setup.
+          pip install --no-compile -r pytorch-cpu-requirements.txt
+          pip install --no-compile -f https://iree.dev/pip-release-links.html --src deps \
+            -e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
+          pip install --no-compile -r requirements.txt -r sharktank/requirements-tests.txt -e sharktank/
+
+          # Try with the latest nightly releases, not what iree-turbine pins.
+          # We could also pin to a known working or stable version.
+          # This should eventually stabilize. Do the best we can for now.
+          pip install -f https://iree.dev/pip-release-links.html --upgrade \
+            iree-compiler==20241104.1068 \
+            iree-runtime==20241104.1068 \
+            "numpy<2.0"
+
+      - name: Run llama tests
+        run: pytest sharktank/tests/models/llama/benchmark_amdgpu_test.py -v -s --run-all-llama --iree-hip-target=gfx942 --html=out/index.html
+
+      - name: Deploy to GitHub Pages
+        uses: peaceiris/actions-gh-pages@v3
+        with:
+          github_token: ${{ secrets.SHARK_PLATFORM_GH_TOKEN }}
+          publish_dir: ./out
+
+      - name: Upload llama executable files
+        uses: actions/upload-artifact@v4
+        with:
+          name: llama-files
+          path: ${{ github.workspace }}/${{ steps.date.outputs.date }}

.github/workflows/ci-llama.yaml → .github/workflows/ci-llama-quick-tests.yaml

@@ -4,7 +4,7 @@
 # See https://llvm.org/LICENSE.txt for license information.
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
-name: Llama Benchmarking Tests
+name: Llama Benchmarking 8B Tests
 
 on:
   workflow_dispatch:
@@ -22,8 +22,8 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  test_llama:
-    name: "Llama Benchmarking Tests"
+  test_llama_quick:
+    name: "Llama Benchmarking 8B Tests"
     strategy:
       matrix:
         version: [3.11]
@@ -71,18 +71,12 @@ jobs:
           # We could also pin to a known working or stable version.
           # This should eventually stabilize. Do the best we can for now.
           pip install -f https://iree.dev/pip-release-links.html --upgrade \
-            iree-compiler \
-            iree-runtime \
+            iree-compiler==20241104.1068 \
+            iree-runtime==20241104.1068 \
             "numpy<2.0"
 
-      - name: Run llama test
-        run: pytest sharktank/tests/models/llama/benchmark_amdgpu_test.py -v -s --longrun --iree-hip-target=gfx942 --html=out/index.html
-
-      # - name: Deploy to GitHub Pages
-      #   uses: peaceiris/actions-gh-pages@v3
-      #   with:
-      #     github_token: ${{ secrets.SHARK_PLATFORM_GH_TOKEN }}
-      #     publish_dir: ./out
+      - name: Run llama 8b tests
+        run: pytest sharktank/tests/models/llama/benchmark_amdgpu_test.py -v -s --iree-hip-target=gfx942 --run-8b-llama
 
       - name: Upload llama executable files
         uses: actions/upload-artifact@v4

sharktank/conftest.py

@@ -69,7 +69,23 @@ def pytest_addoption(parser):
         action="store_true",
         dest="longrun",
         default=False,
-        help="Enable long and slow tests",
+        help="Enable long tests",
+    )
+
+    parser.addoption(
+        "--run-8b-llama",
+        action="store_true",
+        dest="run-8b-llama",
+        default=False,
+        help="Enable llama 8b benchmarking tests",
+    )
+
+    parser.addoption(
+        "--run-all-llama",
+        action="store_true",
+        dest="run-all-llama",
+        default=False,
+        help="Enable all llama benchmarking tests",
     )
 
     # TODO: Remove all hardcoded paths in CI tests

sharktank/sharktank/examples/sharding/shard_llm_dataset.py

@@ -30,6 +30,14 @@ def main(raw_args=None):
     args = cli.parse(parser, args=raw_args)
     dataset = cli.get_input_dataset(args)
 
+    if args.output_irpa_file is None:
+        raise RuntimeError(f"Need file destination for IRPA file")
+
+    if args.tensor_parallelism_size < 2:
+        raise RuntimeError(
+            f"Expect sharding greater than 1 found {args.tensor_parallelism_size}"
+        )
+
     hp = LlamaHParams.from_gguf_props(dataset.properties)
     llama_config = LlamaModelConfig(
         hp, tensor_parallelism_size=args.tensor_parallelism_size

sharktank/sharktank/utils/export_artifacts.py

@@ -123,17 +123,18 @@ def wrapper(*args, **kwargs):
     def shard_irpa_file(
         self,
         *,
-        output_file: str,
+        gguf_file: str,
+        output_irpa: str,
     ):
         shard_irpa_args = [
             "python3",
             "-m",
-            "sharktank.models.llama.tools.shard_llama",
-            "--irpa-file",
-            self.irpa_path,
-            "--output-file",
-            output_file,
-            "--shard_count",
+            "sharktank.examples.sharding.shard_llm_dataset",
+            "--gguf-file",
+            gguf_file,
+            "--output-irpa-file",
+            output_irpa,
+            "--tensor-parallelism-size",
             str(self.tensor_parallelism_size),
         ]
 
@@ -145,7 +146,7 @@ def shard_irpa_file(
         proc = subprocess.run(cmd, shell=True, capture_output=True, cwd=cwd, text=True)
         if proc.returncode != 0:
             logger.error(
-                f"Error sharding irpa file with shard_llama.py\n"
+                f"Error sharding irpa file with shard_llm_dataset.py\n"
                 f"{proc.stdout+proc.stderr}"
             )
         else:
@@ -203,6 +204,12 @@ def compile_to_vmfb(
             f"--iree-hal-target-backends={self.iree_hal_target_backends}",
             f"-o={vmfb_path}",
         ]
+        if self.tensor_parallelism_size > 1:
+            iree_hal_target_devices = [
+                f"--iree-hal-target-device=hip[{i}]"
+                for i in range(self.tensor_parallelism_size)
+            ]
+            compile_args += iree_hal_target_devices
         if hal_dump_path:
             compile_args += [
                 f"--iree-hal-dump-executable-files-to={hal_dump_path}/files"
@@ -234,16 +241,34 @@ def iree_benchmark_vmfb(
             compile_cmd: Command used to compile the program, for inclusion in error messages.
         Raises Exception if running fails for some reason.
         """
-        benchmark_args = [
-            f"ROCR_VISIBLE_DEVICES={hip_device_id}",
+        benchmark_args = []
+        if self.tensor_parallelism_size > 1:
+            base_irpa_path, _ = os.path.splitext(irpa_path)
+            rocr_visible_devices = [
+                f"ROCR_VISIBLE_DEVICES={','.join(str(i) for i in range(self.tensor_parallelism_size))}"
+            ]
+            params = [f"--parameters=model={base_irpa_path}.irpa"]
+            params += [
+                f"--parameters=model={base_irpa_path}.rank{i}.irpa"
+                for i in range(self.tensor_parallelism_size)
+            ]
+            devices = [
+                f"--device=hip://{i}" for i in range(self.tensor_parallelism_size)
+            ]
+        else:
+            rocr_visible_devices = [f"ROCR_VISIBLE_DEVICES={hip_device_id}"]
+            params = [f"--parameters=model={irpa_path}"]
+            devices = [f"--device=hip://{hip_device_id}"]
+        benchmark_args += rocr_visible_devices
+        benchmark_args += [
             "iree-benchmark-module",
-            f"--device=hip://{hip_device_id}",
             "--hip_use_streams=true",
             "--hip_allow_inline_execution=true",
             "--device_allocator=caching",
             f"--module={vmfb_name}",
-            f"--parameters=model={irpa_path}",
         ]
+        benchmark_args += params
+        benchmark_args += devices
         benchmark_args += args
         cmd = subprocess.list2cmdline(benchmark_args)
         logging.getLogger().info(f"Launching run command:\n" f"cd {cwd} && {cmd}")