Add GPU-selectable cloud connections

.github/workflows/fal-deploy.yml

@@ -11,15 +11,6 @@ on:
           - main
           - prod
         default: 'main'
-      app:
-        description: 'Which app to deploy'
-        required: true
-        type: choice
-        options:
-          - both
-          - scope-app
-          - scope-livepeer
-        default: 'both'
   workflow_run:
     workflows: ["Build and Push Docker Image"]
     types:
@@ -51,10 +42,6 @@ jobs:
         run: pip install fal
 
       - name: Deploy Scope app to fal
-        if: >
-          github.event_name == 'workflow_run' ||
-          github.event.inputs.app == 'both' ||
-          github.event.inputs.app == 'scope-app'
         env:
           FAL_KEY: ${{ secrets.FAL_KEY }}
         run: |
@@ -68,11 +55,7 @@ jobs:
           echo "Deploying Scope app to $ENV"
           fal deploy --env $ENV --auth public src/scope/cloud/fal_app.py
 
-      - name: Deploy Livepeer runner to fal
-        if: >
-          github.event_name == 'workflow_run' ||
-          github.event.inputs.app == 'both' ||
-          github.event.inputs.app == 'scope-livepeer'
+      - name: Deploy default Livepeer runner to fal
         env:
           FAL_KEY: ${{ secrets.FAL_KEY }}
         run: |
@@ -84,8 +67,52 @@ jobs:
           fi
 
           echo "Deploying Livepeer runner to $ENV as scope-livepeer"
+          echo "Using default deploy GPU"
+
           fal deploy \
             --app-name scope-livepeer \
             --env $ENV \
             --auth private \
             src/scope/cloud/livepeer_fal_app.py
+
+      - name: Deploy RTX 4090 Livepeer runner to fal
+        env:
+          FAL_KEY: ${{ secrets.FAL_KEY }}
+          SCOPE_DEPLOY_GPU: GPU-RTX4090
+        run: |
+          # Use input environment for manual dispatch, default to main for workflow_run
+          if [ "${{ github.event_name }}" == "workflow_dispatch" ]; then
+            ENV="${{ github.event.inputs.environment }}"
+          else
+            ENV="main"
+          fi
+
+          echo "Deploying Livepeer runner to $ENV as scope-livepeer-rtx4090"
+          echo "Using deploy GPU override: ${SCOPE_DEPLOY_GPU}"
+
+          fal deploy \
+            --app-name scope-livepeer-rtx4090 \
+            --env $ENV \
+            --auth private \
+            src/scope/cloud/livepeer_fal_app.py
+
+      - name: Deploy RTX 5090 Livepeer runner to fal
+        env:
+          FAL_KEY: ${{ secrets.FAL_KEY }}
+          SCOPE_DEPLOY_GPU: GPU-RTX5090
+        run: |
+          # Use input environment for manual dispatch, default to main for workflow_run
+          if [ "${{ github.event_name }}" == "workflow_dispatch" ]; then
+            ENV="${{ github.event.inputs.environment }}"
+          else
+            ENV="main"
+          fi
+
+          echo "Deploying Livepeer runner to $ENV as scope-livepeer-rtx5090"
+          echo "Using deploy GPU override: ${SCOPE_DEPLOY_GPU}"
+
+          fal deploy \
+            --app-name scope-livepeer-rtx5090 \
+            --env $ENV \
+            --auth private \
+            src/scope/cloud/livepeer_fal_app.py

docs/livepeer.md

@@ -42,6 +42,23 @@ fal deploy --env main --auth public src/scope/cloud/livepeer_fal_app.py
 
 This starts the Livepeer runner as a subprocess inside the Fal container and proxies `/ws` traffic to it.
 
+Use `SCOPE_DEPLOY_GPU` to select the Fal machine type at deploy time:
+
+```bash
+# Default H100 deployment:
+fal deploy --app-name scope-livepeer --env prod --auth private src/scope/cloud/livepeer_fal_app.py
+
+# RTX 4090 deployment:
+SCOPE_DEPLOY_GPU=GPU-RTX4090 \
+fal deploy --app-name scope-livepeer-rtx4090 --env prod --auth private src/scope/cloud/livepeer_fal_app.py
+
+# RTX 5090 deployment:
+SCOPE_DEPLOY_GPU=GPU-RTX5090 \
+fal deploy --app-name scope-livepeer-rtx5090 --env prod --auth private src/scope/cloud/livepeer_fal_app.py
+```
+
+Supported deploy values are `GPU-RTX4090` and `GPU-RTX5090`. When `SCOPE_DEPLOY_GPU` is unset, the wrapper keeps the current default `GPU-H100`.
+
 ## Start the Scope Server
 
 Set environment variables and launch the server:
@@ -71,7 +88,22 @@ LIVEPEER_WS_URL=wss://fal.run/<app-id>/ws \
 uv run daydream-scope
 ```
 
-To switch away from explicit runner overrides, unset both `LIVEPEER_WS_URL` and `SCOPE_CLOUD_APP_ID`. In that case the runner URL uses the default Livepeer flow.
+To switch away from explicit runner overrides, unset both `LIVEPEER_WS_URL` and `SCOPE_CLOUD_APP_ID`. In that case Scope uses the built-in default Livepeer app id.
+
+```bash
+# Default H100-backed Livepeer deployment:
+SCOPE_CLOUD_MODE=livepeer \
+LIVEPEER_TOKEN=<base64-json-token> \
+uv run daydream-scope
+
+# Select the RTX 4090 Livepeer deployment:
+SCOPE_CLOUD_MODE=livepeer \
+LIVEPEER_TOKEN=<base64-json-token> \
+SCOPE_CLOUD_GPU=rtx4090 \
+uv run daydream-scope
+```
+
+If `SCOPE_CLOUD_APP_ID` is set, Scope will use that app id as-is and will not rewrite it based on `SCOPE_CLOUD_GPU`.
 
 ### Environment Variables
 
@@ -82,6 +114,7 @@ To switch away from explicit runner overrides, unset both `LIVEPEER_WS_URL` and
 | `LIVEPEER_SIGNER`    | No       | Override signer URL used for Livepeer payments. To disable payments, set to a falsy value such as `"off"`. |
 | `LIVEPEER_WS_URL`    | No       | Explicit runner WebSocket URL (e.g. `ws://127.0.0.1:8001/ws`). |
 | `SCOPE_CLOUD_APP_ID` | No       | Fal app id used to construct `ws_url` as `wss://fal.run/<app-id>`. Must include `/ws` suffix. Used when `LIVEPEER_WS_URL` is not set. |
+| `SCOPE_CLOUD_GPU`    | No       | Livepeer GPU selector. Supported values: `h100`, `rtx4090`, `rtx5090`. Default `h100`. Ignored when `LIVEPEER_WS_URL` or `SCOPE_CLOUD_APP_ID` is set. |
 | `LIVEPEER_TOKEN`     | No       | Base64-encoded JSON token used to start the LV2V job. Can be used to override Livepeer orch / payments routing. |
 | `LIVEPEER_DEBUG`     | No       | Enables debug logging for the Livepeer Gateway SDK and local Livepeer modules. |
 | `LIVEPEER_DEV_MODE`  | No       | Used for developing against a local Livepeer orchestrator with self-signed certificates. |

src/scope/cloud/livepeer_fal_app.py

@@ -27,6 +27,7 @@
 RUNNER_MAX_FAILURES_PER_WINDOW = 20
 RUNNER_FAILURE_WINDOW_SECONDS = 60.0
 ASSETS_DIR_PATH = "/tmp/.daydream-scope/assets"
+SCOPE_DEPLOY_GPU_ENV = "SCOPE_DEPLOY_GPU"
 
 # Gates startup cleanup so only one cleanup run executes at a time.
 _cleanup_event: asyncio.Event | None = None
@@ -129,6 +130,21 @@ def _get_git_sha() -> str:
 )
 
 
+def _get_livepeer_machine_type() -> str:
+    """Return the Fal machine type selected by SCOPE_DEPLOY_GPU."""
+    deploy_gpu = os.getenv(SCOPE_DEPLOY_GPU_ENV, "").strip()
+    if not deploy_gpu:
+        return "GPU-H100"
+    if deploy_gpu in {"GPU-RTX4090", "GPU-RTX5090"}:
+        return deploy_gpu
+    raise ValueError(
+        "Invalid SCOPE_DEPLOY_GPU. Expected `GPU-RTX4090`, `GPU-RTX5090`, or unset."
+    )
+
+
+LIVEPEER_MACHINE_TYPE = _get_livepeer_machine_type()
+
+
 def _runner_is_ready() -> bool:
     """Return True when the local runner HTTP server responds."""
     import urllib.error
@@ -224,7 +240,7 @@ class LivepeerScopeApp(fal.App, keep_alive=300):
     """fal entrypoint that runs and proxies the existing Livepeer Scope runner."""
 
     image = custom_image
-    machine_type = "GPU-H100"
+    machine_type = LIVEPEER_MACHINE_TYPE
     requirements = [
         "websockets",
         "httpx",
@@ -235,6 +251,11 @@ def setup(self):
         import subprocess
 
         print(f"Starting Livepeer runner wrapper setup... (version: {GIT_SHA})")
+        print(
+            "Resolved Livepeer deploy GPU: "
+            f"{os.getenv(SCOPE_DEPLOY_GPU_ENV, '').strip() or 'default'} "
+            f"-> {self.machine_type}"
+        )
 
         try:
             result = subprocess.run(

src/scope/server/app.py

@@ -19,6 +19,7 @@
 from typing import TYPE_CHECKING
 
 import click
+import click.core
 import uvicorn
 from fastapi import BackgroundTasks, Depends, FastAPI, HTTPException, Query, Request
 from fastapi.middleware.cors import CORSMiddleware
@@ -3115,7 +3116,7 @@ async def connect_to_cloud(
         # Use request body credentials if provided, otherwise fall back to CLI/env
         app_id = request.app_id or os.environ.get("SCOPE_CLOUD_APP_ID")
         api_key = request.api_key or os.environ.get("SCOPE_CLOUD_API_KEY")
-        if not app_id:
+        if not app_id and not is_livepeer_enabled():
             raise HTTPException(
                 status_code=400,
                 detail="cloud credentials not configured. Use --cloud-app-id and --cloud-api-key CLI args, "
@@ -3425,7 +3426,7 @@ def run_server(reload: bool, host: str, port: int, no_browser: bool):
 )
 @click.option(
     "--cloud-app-id",
-    default="Daydream/scope-app--prod/ws",
+    default=None,
     envvar="SCOPE_CLOUD_APP_ID",
     help="Cloud app ID for cloud mode (e.g., 'username/scope-app')",
 )
@@ -3460,8 +3461,6 @@ def main(
 
     # MCP mode: run the MCP stdio server instead of the HTTP server
     if mcp:
-        import click.core
-
         from .mcp_server import run_mcp_server
 
         # Only pre-connect if --port was explicitly provided on the command line

src/scope/server/livepeer.py

@@ -34,7 +34,7 @@
 
 def is_livepeer_enabled() -> bool:
     """Check if Livepeer mode is enabled via environment variables."""
-    return os.getenv("SCOPE_CLOUD_MODE", "").lower() == "livepeer"
+    return os.getenv("SCOPE_CLOUD_MODE", "livepeer").lower() == "livepeer"
 
 
 class LivepeerConnection:

src/scope/server/livepeer_client.py

@@ -57,6 +57,31 @@
 TASK_DRAIN_TIMEOUT_S = 0.25
 RUNNER_RESTART_TIMEOUT_S = 30.0
 PAYMENT_SEND_INTERVAL_S = 10.0
+SCOPE_CLOUD_GPU_ENV = "SCOPE_CLOUD_GPU"
+DEFAULT_LIVEPEER_APP_ID = "daydream/scope-livepeer--prod/ws"
+
+
+def _normalize_optional_string(value: str | None) -> str | None:
+    if value is None:
+        return None
+    normalized = value.strip()
+    return normalized or None
+
+
+def _resolve_livepeer_app_id(
+    app_id: str | None,
+) -> str | None:
+    normalized_app_id = _normalize_optional_string(app_id)
+    if normalized_app_id is not None:
+        return normalized_app_id.strip("/")
+
+    gpu = _normalize_optional_string(os.getenv(SCOPE_CLOUD_GPU_ENV))
+    if gpu not in {None, "h100", "rtx4090", "rtx5090"}:
+        raise ValueError(
+            "Invalid SCOPE_CLOUD_GPU. Expected `h100`, `rtx4090`, `rtx5090`, or unset."
+        )
+    gpu_suffix = "" if gpu in {None, "h100"} else f"-{gpu}"
+    return f"daydream/scope-livepeer{gpu_suffix}--prod/ws"
 
 
 @dataclass(slots=True)
@@ -312,14 +337,11 @@ def _normalize_ws_url(value: str | None) -> str | None:
 
     @staticmethod
     def _ws_url_from_app_id(value: str | None) -> str | None:
-        # HACK: Ignore the default app_id to use the orchestrator's own config
-        if not value or value == "Daydream/scope-app--prod/ws":
+        resolved_app_id = _resolve_livepeer_app_id(value)
+        if not resolved_app_id:
             return None
         try:
-            trimmed = value.strip()
-            if not trimmed:
-                raise ValueError
-            app_id = trimmed.strip("/")
+            app_id = resolved_app_id.strip("/")
             if not app_id.endswith("/ws"):
                 raise ValueError
             ws_url = f"wss://fal.run/{app_id}"
@@ -329,7 +351,7 @@ def _ws_url_from_app_id(value: str | None) -> str | None:
         except Exception:
             raise ValueError(
                 "Invalid cloud app id. Expected a non-empty app id ending in "
-                "`/ws` (for example `daydream/scope-app/ws`)."
+                "`/ws` (for example `daydream/custom-runner--prod/ws`)."
             ) from None
         if parsed.scheme not in {"ws", "wss"}:
             raise ValueError("Invalid ws_url. Expected a valid ws:// or wss:// URL.")