Use vLLM triton kernels for MoE backend

docker/nemo/Dockerfile.nemo-megatron

@@ -15,8 +15,29 @@ RUN if [ "${INSTALL_FLASH_ATTN4}" = "1" ]; then \
         "flash-attn-4" "nvidia-cutlass-dsl"; \
     fi
 
+RUN set -eux; \
+    python3 - <<'PY'
+from pathlib import Path
+import shutil
+
+site = Path("/usr/local/lib/python3.12/dist-packages")
+src = site / "vllm/third_party/triton_kernels"
+dst = site / "triton_kernels"
+if not (src / "matmul_ogs.py").exists() or not (src / "tensor.py").exists():
+    raise SystemExit(f"missing vLLM triton_kernels source: {src}")
+if not dst.exists():
+    raise SystemExit(f"missing top-level triton_kernels package: {dst}")
+backup = site / "triton_kernels_nv26_backup"
+if backup.exists():
+    shutil.rmtree(backup)
+shutil.copytree(dst, backup, ignore=shutil.ignore_patterns("__pycache__"))
+shutil.rmtree(dst)
+shutil.copytree(src, dst, ignore=shutil.ignore_patterns("__pycache__"))
+PY
+
 RUN python3 - <<'PY'
 import importlib.metadata as metadata
+import importlib.util
 required = ["torch", "transformer-engine", "megatron-core", "megatron-bridge"]
 for package in required:
     print(package, metadata.version(package))
@@ -28,4 +49,8 @@ try:
     print("nvidia-cutlass-dsl", metadata.version("nvidia-cutlass-dsl"))
 except metadata.PackageNotFoundError:
     print("nvidia-cutlass-dsl MISSING")
+spec = importlib.util.find_spec("triton_kernels.matmul_ogs")
+if spec is None:
+    raise SystemExit("triton_kernels.matmul_ogs MISSING")
+print("triton_kernels.matmul_ogs", spec.origin)
 PY