fix(gpu_test): isolate concurrent test sessions

.github/workflows/ci.yml

@@ -57,3 +57,6 @@ jobs:
 
       - name: Ruff check
         run: uv run ruff check gpu_test/ && uv run ruff format --check gpu_test/
+
+      - name: Test GPU infrastructure
+        run: uv run pytest -q -m "not gpu"

README.md

@@ -128,3 +128,20 @@ cmake --build build --target check-warpforth
 # Run end-to-end GPU tests (requires Vast.ai API key)
 VASTAI_API_KEY=xxx uv run pytest -v -m gpu
 ```
+
+Each GPU test process rents its own uniquely labeled Vast.ai instance, so GPU
+tests can run concurrently from separate checkouts or Amp threads using the
+same Vast.ai account. Each process logs existing instances whose labels start
+with `warpforth-test-` but never destroys instances owned by another run.
+Running multiple processes therefore increases Vast.ai usage and cost;
+pytest-xdist also rents one instance per worker.
+
+Each process generates an ephemeral SSH key and attaches only its public key to
+the instance it owns. The private key remains in a local temporary directory
+until teardown, so ad hoc machines and CI require only `VASTAI_API_KEY`; no
+account-level Vast.ai SSH key configuration is required.
+
+Normal teardown verifies that the process's instance was removed. A hard kill
+or machine failure can still bypass cleanup. If startup logs report a possible
+orphan, inspect the instance ID with `uv run vastai show instances` and remove
+only the confirmed orphan with `uv run vastai destroy instance <id>`.