chore(inference): bump vLLM to 0.26.0

packages/prime-rl-configs/pyproject.toml

@@ -7,7 +7,7 @@ requires-python = "~=3.12.0"
 dependencies = [
     "pydantic>=1.10.13",
     "prime-pydantic-config>=0.4.2",
-    "renderers>=0.1.8",
+    "renderers>=0.1.9.dev10",
     "tomli>=2.2.1",
     "tomli-w>=1.2.0",
     "verifiers>=0.2.2.dev43",

pyproject.toml

@@ -18,7 +18,7 @@ dependencies = [
     "torchaudio",
     "torchdata>=0.11.0",
     "transformers==5.6.2",
-    "vllm>=0.24.0",
+    "vllm>=0.26.0",
     "mooncake-transfer-engine>=0.3.10.post2",
     "wandb>=0.26.1",
     "wandb-workspaces>=0.4.3",
@@ -200,8 +200,8 @@ vllm-router = [
     { url = "https://github.com/PrimeIntellect-ai/router/releases/download/v0.1.26/vllm_router-0.1.26-cp38-abi3-manylinux_2_28_aarch64.whl", marker = "platform_machine == 'aarch64'" },
 ]
 vllm = [
-    { url = "https://github.com/vllm-project/vllm/releases/download/v0.24.0/vllm-0.24.0+cu129-cp38-abi3-manylinux_2_28_x86_64.whl", marker = "platform_machine == 'x86_64'" },
-    { url = "https://github.com/vllm-project/vllm/releases/download/v0.24.0/vllm-0.24.0+cu129-cp38-abi3-manylinux_2_28_aarch64.whl", marker = "platform_machine == 'aarch64'" },
+    { url = "https://github.com/vllm-project/vllm/releases/download/v0.26.0/vllm-0.26.0+cu129-cp38-abi3-manylinux_2_28_x86_64.whl", marker = "platform_machine == 'x86_64'" },
+    { url = "https://github.com/vllm-project/vllm/releases/download/v0.26.0/vllm-0.26.0+cu129-cp38-abi3-manylinux_2_28_aarch64.whl", marker = "platform_machine == 'aarch64'" },
 ]
 deep-ep = [
     { url = "https://github.com/PrimeIntellect-ai/prime-rl/releases/download/v0.5.0/deep_ep-1.2.1+29d31c0-cp312-cp312-linux_x86_64.whl", marker = "platform_machine == 'x86_64'" },

src/prime_rl/inference/vllm/serving_tokens.py

@@ -1,7 +1,7 @@
 """Prime-RL extensions to vLLM's `/inference/v1/generate` handler.
 
-vLLM 0.22 ships a generic tokens-in / tokens-out handler at
-``vllm.entrypoints.serve.disagg.serving.ServingTokens`` that already covers
+vLLM ships a generic tokens-in / tokens-out handler at
+``vllm.entrypoints.scale_out.token_in_token_out.serving.ServingTokens`` that covers
 prefix-cache salting, lora dispatch, multimodal features, prompt logprobs,
 priority, ``data_parallel_rank`` header routing and server-side ``max_tokens``
 defaulting. We subclass it for the bits still missing from the upstream handler:
@@ -37,12 +37,12 @@
     RequestResponseMetadata,
     UsageInfo,
 )
-from vllm.entrypoints.serve.disagg.protocol import (
+from vllm.entrypoints.scale_out.token_in_token_out.protocol import (
     GenerateRequest,
     GenerateResponse,
     GenerateResponseChoice,
 )
-from vllm.entrypoints.serve.disagg.serving import ServingTokens
+from vllm.entrypoints.scale_out.token_in_token_out.serving import ServingTokens
 from vllm.entrypoints.serve.utils.api_utils import get_max_tokens
 from vllm.outputs import RequestOutput
 from vllm.sampling_params import RequestOutputKind, SamplingParams
@@ -176,7 +176,7 @@ async def serve_tokens(
         request: GenerateRequest,
         raw_request: Request | None = None,
     ) -> PrimeRlGenerateResponse | ErrorResponse | AsyncGenerator[str, None]:
-        # Mirrors upstream ``ServingTokens.serve_tokens`` (vllm 0.22). Diffs:
+        # Mirrors upstream ``ServingTokens.serve_tokens``. Diffs:
         # (a) inject ``data_parallel_rank`` from the inbound header into
         # ``engine_client.generate``; (b) default ``sampling_params.max_tokens``
         # to ``max_model_len - prompt_len`` when the caller didn't set it; and
@@ -200,10 +200,11 @@ async def serve_tokens(
         # Build the engine input — features-aware (MM) or text-only fallback.
         # Identical to upstream so we keep tracking it.
         if features := request.features:
-            from vllm.entrypoints.serve.disagg.mm_serde import decode_mm_kwargs_item
+            from vllm.entrypoints.scale_out.token_in_token_out.mm_serde import decode_mm_kwargs_item
             from vllm.inputs import mm_input
             from vllm.multimodal.inputs import (
                 MultiModalKwargsItem,
+                MultiModalKwargsItems,
                 PlaceholderRange,
             )
 
@@ -220,13 +221,13 @@ async def serve_tokens(
                     mm_kwargs[modality] = [None] * len(hashes)
             engine_input = mm_input(
                 prompt_token_ids=request.token_ids,
-                mm_kwargs=mm_kwargs,  # type: ignore[arg-type]
+                mm_kwargs=MultiModalKwargsItems(mm_kwargs),
                 mm_hashes=features.mm_hashes,
                 mm_placeholders=mm_placeholders,
                 cache_salt=request.cache_salt,
             )
         else:
-            (engine_input,) = await self.openai_serving_render.preprocess_completion(
+            (engine_input,) = await self.online_renderer.preprocess_completion(
                 request,
                 prompt_input=request.token_ids,
                 prompt_embeds=None,

src/prime_rl/utils/client.py

@@ -608,7 +608,7 @@ async def prefill_logprobs(openai: AsyncOpenAI, model: str, token_ids: list[int]
     + ``prompt_logprobs`` (the prime-rl server-side extension in
     ``inference/vllm/serving_tokens.py``). Returns one logprob per token (0.0 for
     the leading token, which has no preceding context)."""
-    from vllm.entrypoints.serve.disagg.protocol import GenerateResponse
+    from vllm.entrypoints.scale_out.token_in_token_out.protocol import GenerateResponse
 
     # `/inference/v1/generate` is mounted at server root, not under `/v1`: pass an
     # absolute URL so the SDK skips the base-url merge. vLLM's `GenerateResponse`

tests/unit/inference/test_serving_tokens.py

@@ -1,6 +1,6 @@
 """Sanity tests for the prime-RL ``ServingTokens`` subclass.
 
-The full happy-path is owned upstream by vLLM 0.20's
+The full happy-path is owned upstream by vLLM's
 ``vllm/entrypoints/serve/disagg`` test suite. We only cover the prime-RL
 deltas here:
     * ``serialize_routed_experts`` round-trips a compact raw-byte payload.
@@ -15,7 +15,7 @@
 import numpy as np
 import pybase64
 from vllm.entrypoints.openai.engine.protocol import UsageInfo
-from vllm.entrypoints.serve.disagg.protocol import GenerateResponse, GenerateResponseChoice
+from vllm.entrypoints.scale_out.token_in_token_out.protocol import GenerateResponse, GenerateResponseChoice
 
 from prime_rl.inference.vllm.routed_experts import serialize_routed_experts
 from prime_rl.inference.vllm.serving_tokens import (

tests/unit/orchestrator/test_qwen3_vl_e2e.py

@@ -96,8 +96,8 @@ def test_renderer_client_qwen3_vl_e2e_features_payload_roundtrips_through_vllm()
         ClientConfig,
         UserMessage,
     )
-    from vllm.entrypoints.serve.disagg.mm_serde import decode_mm_kwargs_item
-    from vllm.entrypoints.serve.disagg.protocol import GenerateRequest
+    from vllm.entrypoints.scale_out.token_in_token_out.mm_serde import decode_mm_kwargs_item
+    from vllm.entrypoints.scale_out.token_in_token_out.protocol import GenerateRequest
 
     # ── Build a real Qwen3VLRenderer with a real processor. ─────────────
     tokenizer = load_tokenizer(_MODEL)