Merge branch 'main' into include-sharktank

Author: marbre

.github/workflows/ci-libshortfin.yml

@@ -9,15 +9,9 @@ name: CI - shortfin
 on:
   workflow_dispatch:
   pull_request:
-    paths:
-      - '.github/workflows/ci-libshortfin.yml'
-      - 'shortfin/**'
   push:
     branches:
       - main
-    paths:
-      - '.github/workflows/ci-libshortfin.yml'
-      - 'shortfin/**'
 
 permissions:
   contents: read
@@ -152,3 +146,22 @@ jobs:
       run: |
         ctest --timeout 30 --output-on-failure --test-dir build
         pytest -s --durations=10
+
+  # Depends on all other jobs to provide an aggregate job status.
+  ci_libshortfin_summary:
+    if: always()
+    runs-on: ubuntu-24.04
+    needs:
+      - build-and-test
+    steps:
+      - name: Getting failed jobs
+        run: |
+          echo '${{ toJson(needs) }}'
+          FAILED_JOBS="$(echo '${{ toJson(needs) }}' \
+            | jq --raw-output \
+            'map_values(select(.result!="success" and .result!="skipped")) | keys | join(",")' \
+          )"
+          if [[ "${FAILED_JOBS}" != "" ]]; then
+            echo "The following jobs failed: ${FAILED_JOBS}"
+            exit 1
+          fi

.github/workflows/ci-sharktank.yml

@@ -191,3 +191,23 @@ jobs:
         run: |
           pytest -v sharktank/ -m punet_quick \
             --durations=0
+
+  # Depends on other jobs to provide an aggregate job status.
+  # TODO(#584): move test_with_data and test_integration to a pkgci integration test workflow?
+  ci_sharktank_summary:
+    if: always()
+    runs-on: ubuntu-24.04
+    needs:
+      - test
+    steps:
+      - name: Getting failed jobs
+        run: |
+          echo '${{ toJson(needs) }}'
+          FAILED_JOBS="$(echo '${{ toJson(needs) }}' \
+            | jq --raw-output \
+            'map_values(select(.result!="success" and .result!="skipped")) | keys | join(",")' \
+          )"
+          if [[ "${FAILED_JOBS}" != "" ]]; then
+            echo "The following jobs failed: ${FAILED_JOBS}"
+            exit 1
+          fi

sharktank/sharktank/utils/export_artifacts.py

@@ -92,6 +92,7 @@ def __init__(
         iree_hal_target_backends: str,
         attention_kernel: str,
         tensor_parallelism_size: int,
+        block_seq_stride: Optional[int] = None,
     ):
         self.sharktank_dir = str(
             Path(os.path.dirname(os.path.abspath(__file__))).parent.parent.parent
@@ -102,6 +103,7 @@ def __init__(
         self.iree_hal_target_backends = iree_hal_target_backends
         self.attention_kernel = attention_kernel
         self.tensor_parallelism_size = tensor_parallelism_size
+        self.block_seq_stride = block_seq_stride
 
     def timeit(func):
         def wrapper(*args, **kwargs):
@@ -184,6 +186,8 @@ def export_to_mlir(
         if self.attention_kernel in ["decomposed", "torch"]:
             export_args.append("--attention-kernel")
             export_args.append(self.attention_kernel)
+        if self.block_seq_stride:
+            export_args.append(f"--block-seq-stride={self.block_seq_stride}")
 
         cwd = self.sharktank_dir
         cmd = subprocess.list2cmdline(export_args)
@@ -280,7 +284,6 @@ def iree_benchmark_vmfb(
         benchmark_args += [
             "iree-benchmark-module",
             "--hip_use_streams=true",
-            "--hip_allow_inline_execution=true",
             "--device_allocator=caching",
             f"--module={vmfb_name}",
         ]