Refactor: Improve Proxy Handling and Secure Boot in GPU Install Script

.gitignore

@@ -11,3 +11,5 @@
 
 # Emacs
 *~
+
+gpu/install_gpu_driver.sh.d

gpu/README.md

@@ -28,8 +28,8 @@ CUDA | Full Version | Driver    | cuDNN     | NCCL   | Tested Dataproc Image Ver
 -----| ------------ | --------- | --------- | -------| ---------------------------
 11.8 | 11.8.0       | 525.147.05| 9.5.1.17  | 2.21.5 | 2.0, 2.1 (Debian/Ubuntu/Rocky); 2.2 (Ubuntu 22.04)
 12.0 | 12.0.1       | 525.147.05| 8.8.1.3   | 2.16.5 | 2.0, 2.1 (Debian/Ubuntu/Rocky); 2.2 (Rocky 9, Ubuntu 22.04)
-12.4 | 12.4.1       | 550.135   | 9.1.0.70  | 2.23.4 | 2.1 (Ubuntu 20.04, Rocky 8); Dataproc 2.2+
-12.6 | 12.6.3       | 550.142   | 9.6.0.74  | 2.23.4 | 2.1 (Ubuntu 20.04, Rocky 8); Dataproc 2.2+
+12.4 | 12.4.1       | 590.48.01| 9.1.0.70  | 2.23.4 | 2.1 (Ubuntu 20.04, Rocky 8); Dataproc 2.2+
+12.6 | 12.6.3       | 590.48.01| 9.6.0.74  | 2.23.4 | 2.1 (Ubuntu 20.04, Rocky 8); Dataproc 2.2+
 
 **Supported Operating Systems:**
 
@@ -289,6 +289,78 @@ handles metric creation and reporting.
     older versions of the `report_gpu_metrics.py` service. The current script
     and agent versions aim to mitigate this. If encountered, check agent logs.
 
+## Development and Testing
+
+If you are modifying this initialization action, you can use the provided test infrastructure to validate your changes locally before deploying them to production.
+
+### Local Integration Testing (Bazel / Podman)
+
+Before pushing any changes to GitHub, you **must** run the integration tests locally to validate your modifications against the full test matrix (`test_gpu.py`). These tests use `absl.testing.parameterized` and the `integration_tests.dataproc_test_case` framework to spin up ephemeral Dataproc clusters and validate GPU functionality (SINGLE, STANDARD, KERBEROS, MIG, etc.).
+
+We provide a Podman wrapper to execute the Bazel test suite locally, perfectly simulating the remote CI sandbox environment.
+
+1. **Credentials:** Ensure you have your Google Cloud Application Default Credentials (ADC) saved locally, typically at `~/.config/gcloud/application_default_credentials.json`, and copy it to `initialization-actions/key.json`.
+2. **Environment:** You must have a configured `env.json` in the `gpu/` directory.
+
+To run the full suite in the Podman container (Unfiltered):
+
+> ⚠️ **WARNING: HIGH RESOURCE CONSUMPTION**
+> An unfiltered run executes the entire test matrix (currently ~12 shards). Because the script is configured to run up to 10 jobs in parallel, this will concurrently provision up to 10 separate Dataproc clusters. This requires massive GCP quota (e.g., ~900 vCPUs and ~30 GPUs simultaneously if using `n1-standard-32` profiles) and will take 60-90 minutes.
+
+```bash
+cd initialization-actions
+# Test a specific Dataproc image version against the full suite
+./gpu/run-bazel-tests-with-podman.sh "2.2-ubuntu22"
+```
+
+To run a specific test filter to iterate quickly on a failure (Recommended):
+
+```bash
+cd initialization-actions
+
+# Filter by a specific test function
+./gpu/run-bazel-tests-with-podman.sh "2.2-ubuntu22" "--test_filter=test_gpu_allocation"
+
+# Filter by another specific test function
+./gpu/run-bazel-tests-with-podman.sh "2.2-ubuntu22" "--test_filter=test_install_gpu_cuda_nvidia_with_spark_job"
+
+# Filter by the entire class
+./gpu/run-bazel-tests-with-podman.sh "2.2-ubuntu22" "--test_filter=NvidiaGpuDriverTestCase"
+```
+
+### Manual Verification Scripts
+
+If you have already provisioned a Dataproc cluster (e.g., `my-cluster`) and want to verify its GPU configuration without running the full Bazel test suite, you can use the standalone verification scripts.
+
+```bash
+# Verify using the local Python script
+python3 gpu/verify_external_cluster.py \
+  --cluster=my-cluster \
+  --region=us-east4 \
+  --zone=us-east4-b \
+  --project=my-project \
+  --tests smi agent spark torch tf numa
+
+# Or using the bash equivalent
+export CLUSTER_NAME=my-cluster PROJECT_ID=my-project REGION=us-east4 ZONE=us-east4-b
+./gpu/verify_external_gpu_cluster.sh
+```
+
+### Advanced Spark / ML Validation
+
+For comprehensive validation of Spark RAPIDS, PyTorch, and TensorFlow on a running cluster, an external testing script is available in the associated `cloud-dataproc/gcloud` repository.
+
+```bash
+# Configure the gcloud test environment
+cd ../cloud-dataproc/gcloud
+source lib/env.sh  # Populates environment variables from env.json
+
+# Execute the comprehensive Spark GPU test suite against the configured cluster
+./t/spark-gpu-test.sh
+```
+
+This script will remotely execute SSH commands to validate NUMA configurations, run PyTorch/TensorFlow isolated in their Conda environments, verify NVCC/cuDNN, and submit `SparkPi` and `JavaIndexToStringExample` Spark jobs configured to use the RAPIDS accelerator plugin.
+
 ## Important notes
 
   * This initialization script will install NVIDIA GPU drivers in all nodes in

gpu/TESTING.md

@@ -0,0 +1,110 @@
+# Testing the GPU Initialization Script
+
+This document details the recommended iterative development and testing process for the `install_gpu_driver.sh` script, bypassing the slow integration runner when developing and ensuring comprehensive testing when complete.
+
+## Fast Iterative Development (SSH/Manual)
+
+This initialization action is designed to be **idempotent**, meaning it can be run multiple times on the same node without breaking the environment. It achieves this by writing "completion sentinels" to `/opt/install-dpgce/complete/` after successfully finishing each phase (e.g., `build-dependencies`, `nccl`, `cuda`).
+
+To facilitate rapid iteration, we use the tooling provided in the companion `cloud-dataproc/gcloud` repository. This repo contains the test infrastructure, environment configuration (`env.json`), and lifecycle management scripts (`recreate-dpgce`, `ssh-m`, `scp-m`) necessary to provision and interact with test clusters efficiently.
+
+When making structural or execution logic changes, you want to avoid destroying and recreating the entire Dataproc cluster during each test cycle. Instead, follow this incremental workflow:
+
+### 1. Provision a "Bare" GPU Cluster
+First, configure your target OS and versions in `cloud-dataproc/gcloud/env.json`. Then, use the `--no-init-action` flag on the recreation script to provision a cluster with GPUs attached, but *without* running any initialization actions during boot.
+
+```bash
+cd cloud-dataproc/gcloud
+./bin/recreate-dpgce --gpu --no-init-action
+```
+
+### 2. Compile and Stage the Script
+The `install_gpu_driver.sh` script is built from fragments. First, compile the fragments, then use the optimized `scp-m` command to transfer your local changes to the -m node. This script stages the file in the GCS temp bucket and pulls it down to `/tmp/install_gpu_driver.sh` over SSH.
+
+```bash
+cd initialization-actions
+cat gpu/install_gpu_driver.sh.d/*.sh > gpu/install_gpu_driver.sh
+cd ../cloud-dataproc/gcloud
+./bin/scp-m ../../initialization-actions/gpu/install_gpu_driver.sh
+```
+
+### 3. Execute and Monitor (Incremental Testing)
+Execute the script manually over SSH as root. Pumping the output through `tee` captures the logs identically to how Dataproc normally records initialization scripts.
+
+**Crucially, when re-running the script to test a specific fix, you must purge the relevant completion sentinels** (and partial build directories like `nccl`) so the script doesn't skip the phase you are trying to test.
+
+*   To run the *entire* script from scratch: `sudo rm -rf /opt/install-dpgce/complete`
+*   To re-test only the NCCL build: `sudo rm -f /opt/install-dpgce/complete/nccl && sudo rm -rf /opt/install-dpgce/nccl`
+
+```bash
+cd cloud-dataproc/gcloud
+./bin/ssh-m 'sudo rm -rf /opt/install-dpgce/complete' # Example: clear everything
+cd ../../initialization-actions
+./gpu/install-in-screen.sh
+```
+
+If your SSH connection drops, simply run `./gpu/install-in-screen.sh` again to instantly re-attach to the running session without losing context or interrupting the installation.
+
+### 4. Verify with the Test Suite
+Once the installation script completes without errors, run the external testing suite to ensure all Conda environments (PyTorch, TensorFlow, RAPIDS) and Spark services correctly bind to the GPU.
+
+```bash
+cd cloud-dataproc/gcloud
+bash t/spark-gpu-test.sh
+```
+
+## Continuous Integration Testing (Bazel/Podman)
+
+Once the manual tests pass, you **must** verify the script behaves correctly within the isolated Python `absl` test harness (`test_gpu.py`) before pushing your changes to GitHub. This validates the full matrix of installation scenarios (SINGLE, STANDARD, KERBEROS, MIG, etc.).
+
+We use a Podman wrapper to execute the Bazel test suite locally, perfectly simulating the remote CI environment.
+
+1. **Credentials:** Ensure your Google Cloud Application Default Credentials (ADC) are saved locally (typically `~/.config/gcloud/application_default_credentials.json`). Copy them to the root of the repository:
+   ```bash
+   cp ~/.config/gcloud/application_default_credentials.json ./key.json
+   ```
+
+2. **Execute Full Suite (Unfiltered):** To execute the entire parameterized test matrix, run the wrapper script without a test filter. 
+
+   > ⚠️ **WARNING: HIGH RESOURCE CONSUMPTION**
+   > An unfiltered run executes all ~12 active parameterized shards. Because the script runs with `--jobs=10`, this will concurrently provision up to 10 separate Dataproc clusters. This requires massive GCP quota (roughly ~900 vCPUs and ~30 GPUs simultaneously if using `n1-standard-32` profiles) and will take approximately 60 to 90 minutes to complete. Do not run this unless you are finalizing a major PR.
+
+   ```bash
+   cd initialization-actions
+   ./gpu/run-bazel-tests-with-podman.sh "2.2-ubuntu22"
+   ```
+
+3. **Execute Specific Tests (Recommended for Iteration):** When iterating on a specific feature or failure, always pass Bazel arguments to filter the test execution. This saves significant time and quota. You can filter by test function name or class.
+
+   *Filter by a specific test function:*
+   ```bash
+   cd initialization-actions
+   ./gpu/run-bazel-tests-with-podman.sh "2.2-ubuntu22" "--test_filter=test_gpu_allocation"
+   ```
+
+   *Filter by a specific test function that executes spark jobs:*
+   ```bash
+   cd initialization-actions
+   ./gpu/run-bazel-tests-with-podman.sh "2.2-ubuntu22" "--test_filter=test_install_gpu_cuda_nvidia_with_spark_job"
+   ```
+
+   *Filter by test class (runs all tests in the class):*
+   ```bash
+   cd initialization-actions
+   ./gpu/run-bazel-tests-with-podman.sh "2.2-ubuntu22" "--test_filter=NvidiaGpuDriverTestCase"
+   ```
+
+## Compiling the AST Splitter Tool (`split.go`)
+
+If you need to re-split `install_gpu_driver.sh` into its `.d/` fragments (e.g. if the main script was modified instead of the fragments), we use a Go-based AST parsing tool (`split.go`) to accurately chunk the bash script.
+
+To compile the tool locally:
+
+```bash
+cd initialization-actions/gpu
+go mod init split
+go get mvdan.cc/sh/v3/syntax
+go build -o split_ast split.go
+```
+
+Once compiled, executing `./split_ast install_gpu_driver.sh` will parse the script and populate the `install_gpu_driver.sh.d/` directory with the chunked components.