fix(norm,moe): restore cuDNN source tensor rank; correct MoE grouped-matmul graphs

.github/workflows/check-deserialize.yml

@@ -0,0 +1,84 @@
+# Copyright Advanced Micro Devices, Inc.
+# SPDX-License-Identifier: MIT
+
+name: check-deserialize
+
+on:
+  pull_request:
+    paths:
+      - 'Workloads/**'
+      - 'src/dnn_benchmarking/graph/**'
+      - 'tools/check_deserialize.py'
+      - '.github/workflows/check-deserialize.yml'
+  push:
+    branches: [main]
+    paths:
+      - 'Workloads/**'
+      - 'src/dnn_benchmarking/graph/**'
+      - 'tools/check_deserialize.py'
+      - '.github/workflows/check-deserialize.yml'
+
+permissions:
+  contents: read
+
+jobs:
+  check-deserialize:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 180
+    steps:
+      # rocm-sdk-devel unpacks to ~13 GB, same as the `rocm-build` job in
+      # setup.yml. No submodules: setup_env.py fetches prebuilt ROCm SDK
+      # wheels, it never builds against the rocm-libraries submodule source.
+      - name: Free disk space
+        run: |
+          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc \
+            /usr/local/.ghcup || true
+          docker system prune -af || true
+
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
+        with:
+          python-version: '3.12'
+
+      - name: Install DVC
+        run: pip install "dvc[s3]"
+
+      # .dvc/config sets allow_anonymous_login for the S3 remote, so this
+      # needs no credentials. Total tracked workload size is ~1 MB.
+      - name: Pull DVC-tracked workloads
+        run: dvc pull
+
+      - name: Extract workload tarballs
+        run: |
+          mkdir -p /tmp/workloads-extracted
+          find Workloads -iname '*.tar.gz' -print0 \
+            | xargs -0 -n1 -I{} tar xzf {} -C /tmp/workloads-extracted
+
+      # --level json only exercises dnn_benchmarking's pure-Python
+      # GraphLoader, which doesn't know hipDNN's real node contract (tensor
+      # rank, dtype, connected-components, per-node field constraints). It's
+      # a fast fail-fast gate for gross JSON/schema errors, not the safety
+      # net -- that's the opgraph-level check below, which calls real
+      # hipdnn_frontend and is the only level that has ever caught a graph
+      # bug in this repo (see the moe.tar.gz fix history).
+      - name: Fast pre-check (JSON schema only)
+        run: |
+          pip install -e .
+          python tools/check_deserialize.py --level json --src src \
+            /tmp/workloads-extracted
+
+      # No GPU required: from_json + validate + build_operation_graph only
+      # assemble and finalize the backend operation-graph descriptor, no
+      # plan build or kernel execution.
+      - name: Build hipDNN (no GPU)
+        env:
+          PIP_NO_CACHE_DIR: "1"
+        run: |
+          python setup_env.py --gpu-arch gfx90a --workspace .ci-workspace -y
+
+      - name: Check opgraph-level graph deserialization
+        run: |
+          .ci-workspace/.venv/bin/python tools/check_deserialize.py \
+            --level opgraph /tmp/workloads-extracted

Workloads/headline/moe.tar.gz.dvc

@@ -1,5 +1,5 @@
 outs:
-- md5: 59c000e95380ce937ac5c921879d7a69
-  size: 28183
+- md5: d9702d81b85dd61f968c41aa97a748ee
+  size: 28433
   hash: md5
   path: moe.tar.gz

Workloads/headline/norm.tar.gz.dvc

@@ -1,5 +1,5 @@
 outs:
-- md5: 266df8a828a6871f9857dd5b52e4bae7
-  size: 6629
+- md5: d12e33a357a93f4609299c978da7cab9
+  size: 8159
   hash: md5
   path: norm.tar.gz

Workloads/microbench/cudnn_bench_moe.tar.gz.dvc

@@ -1,5 +1,5 @@
 outs:
-- md5: 8f0a96aa10d5d9b4092158066fdfa752
-  size: 13204
+- md5: a20ac7876fbe01986fae74b009e3745d
+  size: 8855
   hash: md5
   path: cudnn_bench_moe.tar.gz

Workloads/microbench/cudnn_frontend.tar.gz.dvc

@@ -1,5 +1,5 @@
 outs:
-- md5: ecba7ded7a179bd4f8f64a41341218f8
-  size: 10071
+- md5: b004ab3259b8b5591cb33dc40b7db91e
+  size: 9932
   hash: md5
   path: cudnn_frontend.tar.gz

tools/check_deserialize.py

@@ -54,6 +54,14 @@ def main():
                 file=sys.stderr,
             )
             return 3
+        # This check only exercises from_json/validate/build_operation_graph,
+        # which assemble the backend graph descriptor from JSON and never
+        # touch an engine. Engine plugins (HIPBLASLT_ENGINE, MIOPEN_ENGINE,
+        # ASM_SDPA_ENGINE, ...) are otherwise loaded eagerly on Handle()
+        # construction and some initialize a real GPU context on load,
+        # which aborts hard on GPU-less machines/CI runners. Must be called
+        # before any Handle exists.
+        hipdnn.set_engine_plugin_paths([], mode=hipdnn.PluginLoadingMode.ABSOLUTE)
         handle = hipdnn.Handle()
 
     ok = fail = 0