final touches and enable ci

.github/workflows/pkgci_shark_ai.yml

@@ -66,6 +66,54 @@ jobs:
           name: smoke-test-${{ matrix.name }}
           path: smoke-test-${{ matrix.name }}.xml
 
+
+  direct_to_batcher_test:
+    name: "Direct to Batcher Test (${{ matrix.name }})"
+    runs-on: ${{ matrix.runs-on }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - name: cpu
+            runs-on: azure-cpubuilder-linux-scale
+            test_device: cpu
+            python-version: 3.11
+          - name: amdgpu_rocm_mi300_gfx942
+            runs-on: linux-mi300-1gpu-ossci
+            test_device: gfx942
+            python-version: 3.11
+    defaults:
+      run:
+        shell: bash
+    env:
+      PACKAGE_DOWNLOAD_DIR: ${{ github.workspace }}/.packages
+      VENV_DIR: ${{ github.workspace }}/.venv
+    steps:
+      - name: Run rocminfo
+        if: contains(matrix.test_device, 'gfx')
+        run: rocminfo
+      - name: "Checkout Code"
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      - name: "Set up environment and install PkgCI Artifacts"
+        uses: ./.github/actions/pkgci-setup
+        with:
+          python-version: ${{matrix.python-version}}
+          artifact-run-id: ${{ inputs.artifact_run_id }}
+      - name: Run Direct-to-batcher Test
+        run: |
+          source ${VENV_DIR}/bin/activate
+          pytest -v -s --test_device=${{ matrix.test_device }} \
+            --junitxml=direct-to-batcher-test-${{ matrix.name }}.xml \
+            app_tests/integration_tests/llm/shortfin/direct_to_batcher_test.py \
+            --log-cli-level=INFO
+      - name: Upload Test Results
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: direct-to-batcher-test-${{ matrix.name }}
+          path: direct-to-batcher-test-${{ matrix.name }}.xml
+
+
   integration_test:
     name: "Integration Test (${{ matrix.name }})"
     runs-on: ${{ matrix.runs-on }}

app_tests/integration_tests/llm/shortfin/conftest.py

@@ -78,6 +78,23 @@ def server(model_artifacts, request):
     process.wait()
 
 
+@pytest.fixture(scope="module")
+def generate_service(model_artifacts, request):
+    """Starts and manages the test server."""
+    model_config = model_artifacts.model_config
+
+    server_config = ServerConfig(
+        artifacts=model_artifacts,
+        device_settings=model_config.device_settings,
+        prefix_sharing_algorithm=request.param.get("prefix_sharing", "none"),
+    )
+
+    server_instance = ServerInstance(server_config)
+    server_instance.port = 0
+    with server_instance.start_service_only() as gs:
+        yield gs
+
+
 @pytest.fixture(scope="module")
 def encoded_prompt(model_artifacts: ModelArtifacts, request) -> list[int]:
     tokenizer = Tokenizer.from_file(str(model_artifacts.tokenizer_path))

app_tests/integration_tests/llm/shortfin/direct_to_batcher_test.py

@@ -12,29 +12,13 @@
 from shortfin_apps.llm.components.messages import InferencePhase, InferenceExecRequest
 
 
-@pytest.fixture
-def processor():
-    return ModelProcessor(base_dir="/tmp/model_management")
-
-
-@pytest.fixture
-def model_config():
-    config = TEST_MODELS["tinystories_llama2_25m"]
-    config.device_settings = CPU
-    return config
-
-
-@pytest.fixture
-def server_instance(processor, model_config):
-    artifacts = processor.process_model(model_config)
-    sconf = ServerConfig(
-        artifacts=artifacts,
-        device_settings=CPU,
-        prefix_sharing_algorithm="none",
-    )
-    sinst = ServerInstance(sconf)
-    sinst.port = 0
-    return sinst
+pytestmark = pytest.mark.parametrize(
+    "model_artifacts,generate_service",
+    [
+        ["tinystories_llama2_25m", {"prefix_sharing": "none"}],
+    ],
+    indirect=True,
+)
 
 
 class BatchConsistencyTestProcess(sf.Process):
@@ -97,7 +81,7 @@ async def run(self):
             ), f"Inconsistent results between batch sizes {self.batch_sizes[0]} and {batch_size}"
 
 
-def test_batch_and_nobatch_consistency(server_instance):
+def test_batch_and_nobatch_consistency(model_artifacts, generate_service):
     """
     Test that requests produce identical results regardless of batch size.
 
@@ -107,12 +91,11 @@ def test_batch_and_nobatch_consistency(server_instance):
     - improper seq_len / current_position handling in service.py
     - improper masking in sharktank
     """
-    with server_instance.start_service_only() as generate_service:
-        # Create and run the test process
-        test_process = BatchConsistencyTestProcess(
-            generate_service,
-            input_tokens=[1, 2, 3, 4],
-            batch_sizes=[1, 2, 3, 4],
-            max_response_length=3,
-        )
-        test_process.launch()
+    # Create and run the test process
+    test_process = BatchConsistencyTestProcess(
+        generate_service,
+        input_tokens=[1, 2, 3, 4],
+        batch_sizes=[1, 2, 3, 4],
+        max_response_length=3,
+    )
+    test_process.launch()