From 232921ca0ce34092c91581aab6ef378cbf8a657d Mon Sep 17 00:00:00 2001
From: Rick Staa <rick.staa@outlook.com>
Date: Wed, 1 Jul 2026 15:15:41 +0200
Subject: [PATCH 1/2] feat(vllm): single-shot mode + no-gateway offchain call

Switch the vLLM runner to single-shot with label routing, and document
calling it directly through the orchestrator (/apps/vllm/app/...) with
plain curl -- no gateway.py, no SDK. Offchain-only until single-shot
on-chain payment lands (go-livepeer#3955), so the on-chain section is a
pending note for now.

Relates to #4, #5.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 vllm/README.md    | 26 ++++++++++++++------------
 vllm/runners.json |  3 ++-
 2 files changed, 16 insertions(+), 13 deletions(-)

diff --git a/vllm/README.md b/vllm/README.md
index f07a6b5..2c4f89a 100644
--- a/vllm/README.md
+++ b/vllm/README.md
@@ -5,7 +5,7 @@ Runs an OpenAI-compatible LLM on the Livepeer network and consumes it with the *
 |              |                                            |
 | ------------ | ------------------------------------------ |
 | App id       | `vllm/qwen2.5-0.5b-instruct`               |
-| Runner mode  | persistent (single-shot by nature)         |
+| Runner mode  | single-shot                                |
 | Registration | static (orchestrator config + health poll) |
 | Transport    | HTTP + SSE (OpenAI `/v1/chat/completions`) |
 | Port         | 8000 (vLLM), 8080 (gateway)                |
@@ -40,6 +40,17 @@ curl http://localhost:8080/v1/chat/completions \
   -d '{"model":"Qwen/Qwen2.5-0.5B-Instruct","messages":[{"role":"user","content":"hi"}]}'
 ```
 
+### Call it directly (no gateway)
+
+As a **single-shot** runner with `"routing": "label"`, vLLM is reachable straight through the orchestrator's proxy (`/apps/<label>/app/<path>`) — no `gateway.py`, no SDK, not even the OpenAI client. This path skips discovery and payment, so it is **offchain-only**:
+
+```sh
+curl -Nk https://localhost:8935/apps/vllm/app/v1/chat/completions \
+  -d '{"model":"Qwen/Qwen2.5-0.5B-Instruct","stream":true,"messages":[{"role":"user","content":"hi"}]}'
+```
+
+It also shows the orchestrator reverse-proxy is streaming-transparent: the runner's `text/event-stream` passes straight through, no buffering.
+
 ### Streaming (SSE)
 
 > [!IMPORTANT]
@@ -61,14 +72,5 @@ curl -N http://localhost:8080/v1/chat/completions \
 
 ## Run on-chain (paid)
 
-Layer `docker-compose.onchain.yml` to add a remote signer and run the orchestrator on-chain, so vLLM advertises the price from `runners.json` and the gateway pays per call. Needs an Ethereum RPC, a funded signer wallet (deposit + reserve), and an orchestrator wallet — see [On-chain (paid) setup](../README.md#on-chain-paid-setup) in the repo README.
-
-```sh
-cp .env.example .env   # fill in RPC, network, keystore paths, accounts
-docker compose -f docker-compose.yml -f docker-compose.onchain.yml up -d
-uv run gateway.py --signer http://localhost:7936 &     # gateway pays per request
-uv run client.py --prompt "In one sentence, what is Livepeer?"
-kill %1; docker compose -f docker-compose.yml -f docker-compose.onchain.yml down
-```
-
-The client is **unchanged** — only the gateway gets `--signer`; it pays per call through the remote signer, so the consumer never sees discovery or payment. The price is set in `runners.json` (see the comments in `.env.example`).
+> [!IMPORTANT]
+> Single-shot on-chain payment is not implemented yet ([go-livepeer#3955](https://github.com/livepeer/go-livepeer/issues/3955)). Until it lands, vLLM is **offchain-only**, and the direct route above skips payment. This section — the paid path and the per-call billing model — is finalized once single-shot billing ships.
diff --git a/vllm/runners.json b/vllm/runners.json
index 48a8739..896c0d8 100644
--- a/vllm/runners.json
+++ b/vllm/runners.json
@@ -5,7 +5,8 @@
       "app": "vllm/qwen2.5-0.5b-instruct",
       "runner_url": "http://vllm:8000",
       "health_url": "/health",
-      "mode": "persistent",
+      "routing": "label",
+      "mode": "single-shot",
       "capacity": 1,
       "price_info": { "price_per_unit": 1, "pixels_per_unit": 1000, "unit": "USD" }
     }

From f4dfe6caec4ad18e4d5b46af191509a53075852d Mon Sep 17 00:00:00 2001
From: Rick Staa <rick.staa@outlook.com>
Date: Wed, 1 Jul 2026 15:26:24 +0200
Subject: [PATCH 2/2] docs(vllm): make on-chain section a pre-merge TODO, not a
 shipped caveat

---
 vllm/README.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/vllm/README.md b/vllm/README.md
index 2c4f89a..2bbc5e6 100644
--- a/vllm/README.md
+++ b/vllm/README.md
@@ -72,5 +72,5 @@ curl -N http://localhost:8080/v1/chat/completions \
 
 ## Run on-chain (paid)
 
-> [!IMPORTANT]
-> Single-shot on-chain payment is not implemented yet ([go-livepeer#3955](https://github.com/livepeer/go-livepeer/issues/3955)). Until it lands, vLLM is **offchain-only**, and the direct route above skips payment. This section — the paid path and the per-call billing model — is finalized once single-shot billing ships.
+<!-- TODO(go-livepeer#3955): document the single-shot on-chain paid path (per-call billing, how the direct route / gateway pays) once single-shot billing ships. This PR merges only after #3955 lands, so replace this with the real instructions and undraft — do not merge with this comment in place. -->
+_Paid single-shot path — pending [go-livepeer#3955](https://github.com/livepeer/go-livepeer/issues/3955)._