Support BENCHMARK_JSON_PATH for provisioning-supplied benchmarks

start_server.sh clones pyworker into /workspace/vast-pyworker after the
provisioning phase has run, so a provisioning script that wants to ship
a custom benchmark workflow cannot write to misc/benchmark.json — that
path doesn't exist yet at provisioning time, and pre-creating it would
make the subsequent clone fail.

Allow provisioning to drop the workflow anywhere (e.g. /workspace) and
point the worker at it via the BENCHMARK_JSON_PATH env var. The in-tree
file still takes precedence (so forks with a baked-in benchmark keep
working unchanged); the env var is consulted only as a second choice,
and a misconfigured path logs a warning rather than silently degrading
to the SD1.5 fallback.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

workers/comfyui-json/README.md

@@ -104,13 +104,15 @@ Images will be saved locally AND uploaded to `s3://{bucket}/comfyui/{filename}`.
 
 ### Custom Benchmark Workflows
 
-You can provide a custom ComfyUI workflow for benchmarking by creating `workers/comfyui-json/misc/benchmark.json`. This allows you to test performance using your preferred models and workflow complexity.
+You can provide a custom ComfyUI workflow for benchmarking. This allows you to test performance using your preferred models and workflow complexity.
 
 **Ways to provide the benchmark file:**
-- Fork this repository and add your `benchmark.json` file
-- Write the file during worker provisioning (onstart script or setup phase)
+- **Fork this repository** and commit your workflow to `workers/comfyui-json/misc/benchmark.json`.
+- **Write the file during provisioning** to a path *outside* the pyworker tree (e.g. `/workspace/benchmark.json`) and export `BENCHMARK_JSON_PATH` so the worker can find it. The pyworker repo is cloned by `start_server.sh` *after* provisioning runs, so provisioning cannot write into `misc/` directly — the destination would be clobbered, or the clone would fail.
 
-An example file is provided in the repository. To ensure varied generations, use the placeholder `__RANDOM_INT__` in place of static seed values - it will be replaced with a random integer for each benchmark run.
+If both are present, the in-tree `misc/benchmark.json` wins; `BENCHMARK_JSON_PATH` is consulted only when no in-tree file exists. If the env var is set but points at a missing or unreadable file, the worker logs a warning and falls back to the default benchmark.
+
+An example workflow is provided at `workers/comfyui-json/misc/benchmark.json.example`. To ensure varied generations, use the placeholder `__RANDOM_INT__` in place of static seed values — it will be replaced with a random integer for each benchmark run.
 
 ### Default Benchmark (Fallback)
 
@@ -120,9 +122,10 @@ The default benchmark uses Stable Diffusion v1.5 with ComfyUI's standard text-to
 
 | Environment Variable | Default Value | Description |
 | -------------------- | ------------- | ----------- |
-| BENCHMARK_TEST_WIDTH | 512 | Image width (pixels) |
-| BENCHMARK_TEST_HEIGHT | 512 | Image height (pixels) |
-| BENCHMARK_TEST_STEPS | 20 | Number of denoising steps |
+| BENCHMARK_JSON_PATH | (unset) | Path to a custom workflow file outside the pyworker tree. Used only if `misc/benchmark.json` is absent. |
+| BENCHMARK_TEST_WIDTH | 512 | Fallback benchmark: image width (pixels) |
+| BENCHMARK_TEST_HEIGHT | 512 | Fallback benchmark: image height (pixels) |
+| BENCHMARK_TEST_STEPS | 20 | Fallback benchmark: number of denoising steps |
 
 Each benchmark run uses a random prompt from `misc/test_prompts.txt` and a random seed to ensure consistent GPU load patterns.
 

workers/comfyui-json/worker.py

@@ -2,10 +2,15 @@
 
 Each worker runs a benchmark on warm-up. The payload is selected as follows:
 
-  1. If ``misc/benchmark.json`` exists, it is used as a custom ComfyUI
-     workflow (recommended: match the workflow your endpoint will actually
-     serve, so the autoscaler's performance estimate is meaningful).
-  2. Otherwise an SD1.5 Text2Image fallback runs, parameterised by the
+  1. If ``misc/benchmark.json`` exists in the cloned worker tree, it is
+     used as a custom ComfyUI workflow. Use this if you fork the repo and
+     bake in your workflow.
+  2. Else, if ``$BENCHMARK_JSON_PATH`` is set and points at a readable
+     file, it is used. Use this from a provisioning script — provisioning
+     runs before pyworker is cloned, so it cannot write into ``misc/``,
+     but it can drop the workflow elsewhere (e.g. ``/workspace/``) and
+     export this env var.
+  3. Otherwise an SD1.5 Text2Image fallback runs, parameterised by the
      ``BENCHMARK_TEST_{WIDTH,HEIGHT,STEPS}`` env vars and a random prompt
      from ``misc/test_prompts.txt``.
 
@@ -53,17 +58,37 @@ TEST_PROMPTS   = MISC_DIR / "test_prompts.txt"
 log = logging.getLogger(__name__)
 
 
+def _resolve_benchmark_path() -> Path | None:
+    """Return the path to the custom benchmark workflow, or None if absent.
+
+    See module docstring for the precedence rule. ``$BENCHMARK_JSON_PATH``
+    is logged as a warning when set but missing, so a misconfigured
+    provisioning script doesn't silently degrade to the fallback benchmark.
+    """
+    if BENCHMARK_FILE.exists():
+        return BENCHMARK_FILE
+    env_path = os.getenv("BENCHMARK_JSON_PATH")
+    if not env_path:
+        return None
+    path = Path(env_path)
+    if not path.exists():
+        log.warning("BENCHMARK_JSON_PATH=%s does not exist; falling back to default benchmark", path)
+        return None
+    return path
+
+
 def _custom_workflow_payload() -> dict | None:
-    """Build a payload from ``misc/benchmark.json``, or None if unavailable."""
-    if not BENCHMARK_FILE.exists():
+    """Build a payload from a custom benchmark workflow JSON, or None if unavailable."""
+    path = _resolve_benchmark_path()
+    if path is None:
         return None
     try:
-        with open(BENCHMARK_FILE) as f:
+        with open(path) as f:
             workflow = json.load(f)
     except (json.JSONDecodeError, OSError) as e:
-        log.error("Failed to load %s: %s; falling back to default benchmark", BENCHMARK_FILE, e)
+        log.error("Failed to load %s: %s; falling back to default benchmark", path, e)
         return None
-    log.info("Using custom benchmark workflow from %s", BENCHMARK_FILE)
+    log.info("Using custom benchmark workflow from %s", path)
     return {
         "input": {
             "request_id": f"test-{random.randint(1000, 99999)}",