pnnx for torch exported program

.github/workflows/pnnx.yml

@@ -65,6 +65,42 @@ jobs:
         cd tools/pnnx
         cd build && ctest -C Release --output-on-failure -R test_nn_Conv
 
+  pt2-test:
+    name: pt2-tests
+    runs-on: ubuntu-latest
+    env:
+      PYTHONUSERBASE: ${{ github.workspace }}/pt2-python
+    steps:
+    - uses: actions/checkout@v7
+
+    - uses: actions/setup-python@v7
+      with:
+        python-version: '3.12'
+
+    - name: setup-pytorch
+      run: |
+        python3 -m pip config set global.break-system-packages true
+        pip3 install --user torch==2.13.0+cpu --index-url https://download.pytorch.org/whl/cpu
+        pip3 install --user numpy packaging
+
+    - name: python-ncnn
+      run: |
+        export PATH=${{ env.PYTHONUSERBASE }}/bin:$PATH
+        pip3 install --user . --verbose
+
+    - name: build-pnnx
+      run: |
+        cd tools/pnnx
+        mkdir build && cd build
+        cmake -DCMAKE_BUILD_TYPE=Release ..
+        cmake --build . --config Release -j 4
+
+    - name: pt2-tests
+      run: |
+        export PATH=${{ env.PYTHONUSERBASE }}/bin:$PATH
+        cd tools/pnnx/build
+        ctest -R 'test_ncnn_pt2_' --output-on-failure -j 1
+
   build:
     runs-on: [self-hosted, linux, ubuntu25]
 

tools/pnnx/local_diff.py

@@ -0,0 +1,57 @@
+"""本地诊断 sweep 的 DIFF：直接比对已落盘的 sw_*.ncnn.param 与 sw_*_ts.ncnn.param。
+复用 pt2_crosscheck.py 的 normalize_param，保证与 sweep 判定一致。无需跑 pnnx。"""
+import os
+import re
+import sys
+from collections import Counter
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, os.path.join(HERE, "tests", "ncnn"))
+from pt2_crosscheck import normalize_param  # noqa: E402
+
+D = HERE  # 与 sweep 的 CWD 一致（tools/pnnx/）
+
+pairs = []
+for f in sorted(os.listdir(D)):
+    if f.startswith("sw_") and f.endswith("_ts.ncnn.param"):
+        ts = f
+        pt2 = f.replace("_ts.ncnn.param", ".ncnn.param")
+        if os.path.exists(os.path.join(D, pt2)):
+            pairs.append((f[len("sw_"):-len("_ts.ncnn.param")], pt2, ts))
+
+print(f"# {len(pairs)} pairs\n")
+
+diffs = []
+for name, p2, pts in pairs:
+    a = normalize_param(os.path.join(D, p2))
+    b = normalize_param(os.path.join(D, pts))
+    if a == b:
+        continue
+    first = None
+    for i, (la, lb) in enumerate(zip(a, b)):
+        if la != lb:
+            first = (i, la, lb)
+            break
+    if first is None:
+        first = (None, f"len pt2={len(a)}", f"len pt={len(b)}")
+    diffs.append((name, first, len(a), len(b)))
+
+print(f"# DIFF count = {len(diffs)}\n")
+# 按首处差异的 op 类型归类
+kind = Counter()
+for name, (i, la, lb), _, _ in diffs:
+    toks = (la or "").split()
+    op = toks[0] if toks else "?"
+    if op == "7767517" or re.match(r"^\d+ \d+$", la or ""):
+        op = "HEADER(count/len)"
+    kind[op] += 1
+
+print("==== DIFF 首处差异所在 op 分布 ====")
+for op, c in kind.most_common():
+    print(f"  {op}: {c}")
+
+print("\n==== DIFF 明细（首处差异）====")
+for name, (i, la, lb), na, nb in diffs:
+    print(f"\n[{name}]  pt2={na}L pt={nb}L  line{i}")
+    print(f"   pt2: {(la or '')[:150]}")
+    print(f"   pt : {(lb or '')[:150]}")