use newtype for logprobs count (-1 for all)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>

.buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml

@@ -21,6 +21,10 @@ steps:
     timeout_in_minutes: 30
     optional: true
     device: intel_gpu
+    agent_tags:
+      label: production
+      gpu: 2+
+      mem: 24+
     no_plugin: true
     env:
       REGISTRY: "public.ecr.aws/q9t5s3a7"
@@ -38,6 +42,10 @@ steps:
     timeout_in_minutes: 30
     optional: true
     device: intel_gpu
+    agent_tags:
+      label: production
+      gpu: 1+
+      mem: 16+
     no_plugin: true
     env:
       REGISTRY: "public.ecr.aws/q9t5s3a7"
@@ -55,6 +63,10 @@ steps:
     timeout_in_minutes: 30
     optional: true
     device: intel_gpu
+    agent_tags:
+      label: production
+      gpu: 1+
+      mem: 16+
     no_plugin: true
     env:
       REGISTRY: "public.ecr.aws/q9t5s3a7"

.buildkite/intel_jobs/basic_correctness.yaml

@@ -5,6 +5,10 @@ steps:
 - label: XPU Sleep Mode
   timeout_in_minutes: 30
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:

.buildkite/intel_jobs/engine_intel.yaml

@@ -5,6 +5,10 @@ steps:
 - label: Engine (1 GPU)
   timeout_in_minutes: 30
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:

.buildkite/intel_jobs/expert_parallelism_intel.yaml

@@ -6,6 +6,10 @@ steps:
   key: eplb-algorithm
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:

.buildkite/intel_jobs/kernels_intel.yaml

@@ -5,6 +5,10 @@ steps:
 - label: vLLM IR Tests
   timeout_in_minutes: 30
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:

.buildkite/intel_jobs/lora_intel.yaml

@@ -5,6 +5,10 @@ steps:
 - label: LoRA Runtime + Utils
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
   no_plugin: true
   working_dir: "."
   env:
@@ -34,6 +38,10 @@ steps:
 - label: LoRA Fused/MoE Kernels
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -54,6 +62,10 @@ steps:
 - label: LoRA Punica Kernels
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -74,6 +86,10 @@ steps:
 - label: LoRA Punica FP8/XPU Ops
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -94,6 +110,10 @@ steps:
 - label: LoRA Models
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 2+
+    mem: 24+
   no_plugin: true
   working_dir: "."
   env:
@@ -117,6 +137,10 @@ steps:
 - label: LoRA Multimodal
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:

.buildkite/intel_jobs/misc_intel.yaml

@@ -5,6 +5,10 @@ steps:
 - label: V1 Core + KV + Metrics
   timeout_in_minutes: 30
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -31,6 +35,10 @@ steps:
 - label: V1 Sample + Logits
   timeout_in_minutes: 30
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -71,6 +79,10 @@ steps:
 - label: XPU CPU Offload
   timeout_in_minutes: 60
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -95,6 +107,10 @@ steps:
   key: regression
   timeout_in_minutes: 30
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -126,6 +142,10 @@ steps:
   timeout_in_minutes: 30
   num_devices: 2
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 2+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -157,6 +177,10 @@ steps:
   key: async-engine-inputs-utils-worker
   timeout_in_minutes: 30
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
   no_plugin: true
   working_dir: "."
   env:

.buildkite/intel_jobs/model_runner_v2_intel.yaml

@@ -5,6 +5,10 @@ steps:
 - label: Model Runner V2 Core Tests (Intel)
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 2+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -30,6 +34,10 @@ steps:
 - label: Model Runner V2 Examples (Intel)
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
   no_plugin: true
   working_dir: "."
   env:

.buildkite/intel_jobs/models_multimodal_intel.yaml

@@ -6,6 +6,10 @@ steps:
   key: multi-modal-models-standard-1-qwen2
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -27,6 +31,10 @@ steps:
   key: multi-modal-models-standard-2-qwen3-gemma
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -47,6 +55,10 @@ steps:
   key: multi-modal-models-standard-3-llava-qwen2-vl
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
   no_plugin: true
   working_dir: "."
   env:
@@ -68,6 +80,10 @@ steps:
   key: multi-modal-models-standard-4-other-whisper
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:
@@ -88,6 +104,10 @@ steps:
   key: multi-modal-processor
   timeout_in_minutes: 45
   device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
   no_plugin: true
   working_dir: "."
   env:

.buildkite/intel_jobs/test-intel.yaml

@@ -19,6 +19,10 @@ steps:
       - image-build-xpu
     timeout_in_minutes: 30
     device: intel_gpu
+    agent_tags:
+      label: production
+      gpu: 2+
+      mem: 24+
     no_plugin: true
     env:
       REGISTRY: "public.ecr.aws/q9t5s3a7"
@@ -49,6 +53,10 @@ steps:
       - image-build-xpu
     timeout_in_minutes: 30
     device: intel_gpu
+    agent_tags:
+      label: production
+      gpu: 1+
+      mem: 16+
     no_plugin: true
     env:
       REGISTRY: "public.ecr.aws/q9t5s3a7"
@@ -74,6 +82,10 @@ steps:
       - image-build-xpu
     timeout_in_minutes: 30
     device: intel_gpu
+    agent_tags:
+      label: production
+      gpu: 1+
+      mem: 16+
     no_plugin: true
     env:
       REGISTRY: "public.ecr.aws/q9t5s3a7"
@@ -93,6 +105,10 @@ steps:
       - image-build-xpu
     timeout_in_minutes: 30
     device: intel_gpu
+    agent_tags:
+      label: production
+      gpu: 1+
+      mem: 16+
     no_plugin: true
     env:
       REGISTRY: "public.ecr.aws/q9t5s3a7"

.buildkite/scripts/install-kv-connectors.sh

@@ -4,6 +4,11 @@
 
 set -euo pipefail
 
+if python3 -c "import torch; raise SystemExit(0 if torch.version.hip is not None else 1)"; then
+    uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
+    exit 0
+fi
+
 REQUIREMENTS_FILE="${KV_CONNECTORS_REQUIREMENTS:-/vllm-workspace/requirements/kv_connectors.txt}"
 
 uv pip install --system -r "${REQUIREMENTS_FILE}"

.buildkite/test_areas/misc.yaml

@@ -105,6 +105,12 @@ steps:
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 60
+      depends_on:
+      - image-build-amd
 
 - label: V1 Others (CPU)
   key: v1-others-cpu

docs/contributing/model/basic.md

@@ -136,7 +136,7 @@ The model should also be added to the `MODELS_CONFIG_MAP` dictionary in [vllm/mo
 For case (2), we recommend using as a reference the implementation of [`JambaForCausalLM`](../../../vllm/model_executor/models/jamba.py) (for an example of a model that uses Mamba-1 and attention together) or [`NemotronHForCausalLM`](../../../vllm/model_executor/models/nemotron_h.py) (for an example of a model that uses Mamba-2 and attention together).
 These models should follow the same instructions as case (1), but they should inherit protocol `IsHybrid` (instead of `IsAttentionFree`) and it is *not* necessary to add them to the `MODELS_CONFIG_MAP` (their runtime defaults will be inferred from the protocol).
 
-For case (3), we recommend looking at the implementation of [`MiniMaxText01ForCausalLM`](../../../vllm/model_executor/models/minimax_text_01.py) or [`Lfm2ForCausalLM`](../../../vllm/model_executor/models/lfm2.py) as a reference, which use custom "mamba-like" layers `MiniMaxText01LinearAttention` and `ShortConv` respectively.
+For case (3), we recommend looking at the implementation of [`Lfm2ForCausalLM`](../../../vllm/model_executor/models/lfm2.py) as a reference, which uses a custom "mamba-like" layer `ShortConv`.
 Please follow the same guidelines as case (2) for implementing these models.
 We use "mamba-like" to refer to layers that possess a state that is updated in-place, rather than being appended-to (like KV cache for attention).
 For implementing new custom mamba-like layers, one should inherit from `MambaBase` and implement the methods `get_state_dtype`, `get_state_shape` to calculate the data types and state shapes at runtime, as well as `mamba_type` and `get_attn_backend`.
@@ -144,5 +144,5 @@ It is also necessary to implement the "attention meta-data" class which handles
 Please see [`LinearAttentionMetadata`](../../../vllm/v1/attention/backends/linear_attn.py) or [`ShortConvAttentionMetadata`](../../../vllm/v1/attention/backends/short_conv_attn.py) for examples of this.
 It is also worth noting that we should update `MambaAttentionBackendEnum` in [`registry.py`](../../../vllm/v1/attention/backends/registry.py) when adding a new mamba backend.
 Finally, if one wants to support torch compile and CUDA graphs, it necessary to wrap the call to the mamba-like layer inside a custom op and register it.
-Please see the calls to `direct_register_custom_op` in [vllm/model_executor/models/minimax_text_01.py](../../../vllm/model_executor/models/minimax_text_01.py) or [vllm/model_executor/layers/mamba/short_conv.py](../../../vllm/model_executor/layers/mamba/short_conv.py) for examples of this.
+Please see the calls to `direct_register_custom_op` in [vllm/model_executor/layers/mamba/linear/minimax_linear_attn.py](../../../vllm/model_executor/layers/mamba/linear/minimax_linear_attn.py) or [vllm/model_executor/layers/mamba/short_conv.py](../../../vllm/model_executor/layers/mamba/short_conv.py) for examples of this.
 The new custom op should then be added to the list `_attention_ops` in [vllm/config/compilation.py](../../../vllm/config/compilation.py) to ensure that piecewise CUDA graphs works as intended.

docs/design/attention_backends.md

@@ -170,8 +170,8 @@ Priority is **1 = highest** (tried first).
 | Backend | Version | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | MM Prefix | DCP | Attention Types | Compute Cap. |
 | ------- | ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | --------- | --- | --------------- | ------------ |
 | `CPU_ATTN` | | fp16, bf16, fp32 | `auto`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256, 512 | ❌ | ❌ | ❌ | ❌ | All | N/A |
-| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ❌ | ❌ | ❌ | ✅ | Decoder | 7.x-9.x |
-| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `nvfp4` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ✅ | ❌ | ❌ | ✅ | Decoder | 10.x |
+| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ❌ | ✅ | ❌ | ✅ | Decoder | 7.x-9.x |
+| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `nvfp4` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ✅ | ✅ | ❌ | ✅ | Decoder | 10.x |
 | `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ❌ | ✅ | All | ≥8.0 |
 | `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | 9.x |
 | `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | ≥10.0 |

docs/features/kv_offloading_usage.md

@@ -74,6 +74,7 @@ vllm serve <model> \
 | `max_tracker_size` | no | `64000` | single-tier | Max entries in the lookup tracker. |
 | `secondary_tiers` | no | `[]` | multi-tier | List of secondary tier configs (see below). |
 | `offload_prompt_only` | no | `true` | both | If `true`, only prompt (prefill) blocks are offloaded; decode blocks are skipped. |
+| `self_describing_kv_events` | no | `false` | single-tier | Opt-in. When `true` *and* KV cache events are enabled (`--kv-events-config` with `enable_kv_cache_events`), the connector emits self-describing block-granular `BlockStored`/`BlockRemoved` payloads (constituent block hashes, whole-chunk `token_ids`, per-block `block_size`, parent hash, LoRA + group/cache-spec metadata) instead of the placeholder fallback, so external KV-event consumers can index offloaded blocks. Inert unless events are enabled. Currently rejected by `TieringOffloadingSpec`. Full-attention groups only; sliding-window/SSM groups keep the placeholder fallback. In chunk mode (`block_size` > GPU block size), overlapping chunks re-announce shared per-block hashes, so consumers must reference-count (deduplicate) repeated store/remove announcements. |
 | `spec_module_path` | no | — | both | Python import path for a custom `OffloadingSpec` not in the built-in registry. Required only when `spec_name` is not built-in (advanced). |
 
 ## Secondary Tiers

docs/features/tool_calling.md

@@ -321,15 +321,6 @@ For Qwen2.5, the chat template in tokenizer_config.json has already included sup
 
 Flags: `--tool-call-parser hermes`
 
-### MiniMax Models (`minimax_m1`)
-
-Supported models:
-
-* `MiniMaxAi/MiniMax-M1-40k` (use with [examples/tool_chat_template_minimax_m1.jinja](../../examples/tool_chat_template_minimax_m1.jinja))
-* `MiniMaxAi/MiniMax-M1-80k` (use with [examples/tool_chat_template_minimax_m1.jinja](../../examples/tool_chat_template_minimax_m1.jinja))
-
-Flags: `--tool-call-parser minimax --chat-template examples/tool_chat_template_minimax_m1.jinja`
-
 ### DeepSeek-V3 Models (`deepseek_v3`)
 
 Supported models:

docs/models/pooling_models/token_embed.md

@@ -61,7 +61,7 @@ Models of any architecture can be converted into embedding models using `--conve
 | `ColModernVBertForRetrieval` | ColModernVBERT | T / I | `ModernVBERT/colmodernvbert-merged` | | |
 | `ColPaliForRetrieval` | ColPali | T / I | `vidore/colpali-v1.3-hf` | | |
 | `ColQwen3` | Qwen3-VL | T / I | `TomoroAI/tomoro-colqwen3-embed-4b`, `TomoroAI/tomoro-colqwen3-embed-8b` | | |
-| `ColQwen3_5` | ColQwen3.5 | T + I + V | `athrael-soju/colqwen3.5-4.5B-v3` | | |
+| `ColQwen3_5` | ColQwen3.5 | T + I + V | `athrael-soju/colqwen3.5-4.5B-v3`, `vultr/VultronRetrieverPrime-Qwen3.5-8B` | | |
 | `OpsColQwen3Model` | Qwen3-VL | T / I | `OpenSearch-AI/Ops-Colqwen3-4B`, `OpenSearch-AI/Ops-Colqwen3-8B` | | |
 | `Qwen3VLNemotronEmbedModel` | Qwen3-VL | T / I | `nvidia/nemotron-colembed-vl-4b-v2`, `nvidia/nemotron-colembed-vl-8b-v2` | ✅︎ | ✅︎ |
 | `*ForConditionalGeneration`<sup>C</sup>, `*ForCausalLM`<sup>C</sup>, etc. | Generative models | \* | N/A | \* | \* |

docs/models/supported_models.md

@@ -441,7 +441,6 @@ th {
 | `MiMoV2ForCausalLM` | MiMoV2Pro | `XiaomiMiMo/MiMo-V2.5-Pro`, etc. | | ✅︎ |
 | `MiniCPMForCausalLM` | MiniCPM | `openbmb/MiniCPM-2B-sft-bf16`, `openbmb/MiniCPM-2B-dpo-bf16`, `openbmb/MiniCPM-S-1B-sft`, etc. | ✅︎ | ✅︎ |
 | `MiniCPM3ForCausalLM` | MiniCPM3 | `openbmb/MiniCPM3-4B`, etc. | ✅︎ | ✅︎ |
-| `MiniMaxForCausalLM` | MiniMax-Text | `MiniMaxAI/MiniMax-Text-01-hf`, etc. | | |
 | `MiniMaxM2ForCausalLM` | MiniMax-M2, MiniMax-M2.1 | `MiniMaxAI/MiniMax-M2`, etc. | ✅︎ | ✅︎ |
 | `MistralForCausalLM` | Ministral-3, Mistral, Mistral-Instruct | `mistralai/Ministral-3-3B-Instruct-2512`, `mistralai/Mistral-7B-v0.1`, `mistralai/Mistral-7B-Instruct-v0.1`, etc. | ✅︎ | ✅︎ |
 | `MistralLarge3ForCausalLM` | Mistral-Large-3-675B-Base-2512, Mistral-Large-3-675B-Instruct-2512 | `mistralai/Mistral-Large-3-675B-Base-2512`, `mistralai/Mistral-Large-3-675B-Instruct-2512`, etc. | ✅︎ | ✅︎ |
@@ -487,8 +486,6 @@ th {
 | `TeleChat2ForCausalLM` | TeleChat2 | `Tele-AI/TeleChat2-3B`, `Tele-AI/TeleChat2-7B`, `Tele-AI/TeleChat2-35B`, etc. | ✅︎ | ✅︎ |
 | `TeleChat3ForCausalLM` | TeleChat3 | `Tele-AI/TeleChat3-36B-Thinking`, `Tele-AI/TeleChat3-Coder-36B-Thinking`, etc. | ✅︎ | ✅︎ |
 | `TeleFLMForCausalLM` | TeleFLM | `CofeAI/FLM-2-52B-Instruct-2407`, `CofeAI/Tele-FLM`, etc. | ✅︎ | ✅︎ |
-| `MiniMaxM1ForCausalLM` | MiniMax-Text | `MiniMaxAI/MiniMax-M1-40k`, `MiniMaxAI/MiniMax-M1-80k`, etc. | | |
-| `MiniMaxText01ForCausalLM` | MiniMax-Text | `MiniMaxAI/MiniMax-Text-01`, etc. | | |
 | `Zamba2ForCausalLM` | Zamba2 | `Zyphra/Zamba2-7B-instruct`, `Zyphra/Zamba2-2.7B-instruct`, `Zyphra/Zamba2-1.2B-instruct`, etc. | | |
 
 !!! note
@@ -595,6 +592,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `MiMoV2OmniForCausalLM` | MiMo-V2.5-Omni | T + I<sup>E+</sup> + V<sup>E+</sup> + A<sup>+</sup> | `XiaomiMiMo/MiMo-V2.5-Omni` | | ✅︎ |
 | `MiniCPMO` | MiniCPM-O | T + I<sup>E+</sup> + V<sup>E+</sup> + A<sup>E+</sup> | `openbmb/MiniCPM-o-2_6`, etc. | ✅︎ | ✅︎ |
 | `MiniCPMV` | MiniCPM-V | T + I<sup>E+</sup> + V<sup>E+</sup> | `openbmb/MiniCPM-V-2` (see note), `openbmb/MiniCPM-Llama3-V-2_5`, `openbmb/MiniCPM-V-2_6`, `openbmb/MiniCPM-V-4`, `openbmb/MiniCPM-V-4_5`, etc. | ✅︎ | |
+| `MiniMaxM3SparseForConditionalGeneration` | MiniMax-M3 | T + I<sup>+</sup> + V<sup>+</sup> | `MiniMaxAI/MiniMax-M3`, `MiniMaxAI/MiniMax-M3-MXFP8`, etc. | | |
 | `MiniMaxVL01ForConditionalGeneration` | MiniMax-VL | T + I<sup>E+</sup> | `MiniMaxAI/MiniMax-VL-01`, etc. | | ✅︎ |
 | `Mistral3ForConditionalGeneration` | Mistral3 (HF Transformers) | T + I<sup>+</sup> | `mistralai/Mistral-Small-3.1-24B-Instruct-2503`, etc. | ✅︎ | ✅︎ |
 | `MolmoForCausalLM` | Molmo | T + I<sup>+</sup> | `allenai/Molmo-7B-D-0924`, `allenai/Molmo-7B-O-0924`, etc. | ✅︎ | ✅︎ |

docs/usage/v1_guide.md

@@ -128,7 +128,7 @@ Models that use Mamba-2 and Mamba-1 layers (e.g., `Mamba2ForCausalLM`, `MambaFor
 Hybrid models that combine Mamba-2 and Mamba-1 layers with standard attention layers are also supported (e.g., `BambaForCausalLM`,
 `Zamba2ForCausalLM`, `NemotronHForCausalLM`, `FalconH1ForCausalLM` and `GraniteMoeHybridForCausalLM`, `JambaForCausalLM`, `Plamo2ForCausalLM`).
 
-Hybrid models with mechanisms different to Mamba are also supported (e.g, `MiniMaxText01ForCausalLM`, `MiniMaxM1ForCausalLM`, `Lfm2ForCausalLM`).
+Hybrid models with mechanisms different to Mamba are also supported (e.g, `Lfm2ForCausalLM`).
 
 Please note that prefix caching is not yet supported for any of the above models.
 

examples/generate/multimodal/vision_language_offline.py

@@ -1481,39 +1481,6 @@ def run_minicpmv(questions: list[str], modality: str) -> ModelRequestData:
     return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-V-2_6")
 
 
-def run_minimax_vl_01(questions: list[str], modality: str) -> ModelRequestData:
-    assert modality == "image"
-
-    model_name = "MiniMaxAI/MiniMax-VL-01"
-
-    engine_args = EngineArgs(
-        model=model_name,
-        max_num_seqs=2,
-        limit_mm_per_prompt={modality: 1},
-        trust_remote_code=True,
-        tensor_parallel_size=8,
-    )
-
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-    messages = [
-        [
-            {
-                "role": "user",
-                "content": [{"type": "image"}, {"type": "text", "text": question}],
-            }
-        ]
-        for question in questions
-    ]
-    prompts = tokenizer.apply_chat_template(
-        messages, add_generation_prompt=True, tokenize=False
-    )
-
-    return ModelRequestData(
-        engine_args=engine_args,
-        prompts=prompts,
-    )
-
-
 # Mistral-3 HF-format
 def run_mistral3(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -2485,7 +2452,6 @@ model_example_map = {
     "mantis": run_mantis,
     "minicpmo": run_minicpmo,
     "minicpmv": run_minicpmv,
-    "minimax_vl_01": run_minimax_vl_01,
     "mistral3": run_mistral3,
     "molmo": run_molmo,
     "molmo2": run_molmo2,

examples/pooling/score/colqwen3_5_rerank_online.py

@@ -7,11 +7,27 @@ ColQwen3.5 is a multi-modal ColBERT-style model based on Qwen3.5.
 It produces per-token embeddings and uses MaxSim scoring for retrieval
 and reranking. Supports both text and image inputs.
 
+Works for any ColQwen3.5 checkpoint, e.g. `athrael-soju/colqwen3.5-4.5B-v3`
+or `vultr/VultronRetrieverPrime-Qwen3.5-8B`.
+
 Start the server with:
-    vllm serve athrael-soju/colqwen3.5-4.5B --max-model-len 4096
+    vllm serve athrael-soju/colqwen3.5-4.5B-v3 --max-model-len 4096 \
+        --mm-processor-kwargs '{"min_pixels": 65536, "max_pixels": 1835008}'
 
 Then run this script:
     python colqwen3_5_rerank_online.py
+
+Parity note (matching the native colpali ColQwen3_5Processor pipeline):
+  - Visual-token budget: ColQwen3_5Processor uses max_num_visual_tokens=1792,
+    i.e. max_pixels = 1792 * (patch_size*merge_size)^2 = 1792 * 32^2 = 1835008
+    (with min_pixels = shortest_edge = 65536). Pass these via --mm-processor-kwargs
+    as above; the default budget gives fewer visual tokens and lower retrieval ndcg.
+  - When you build prompts yourself (token_embed), reproduce the processor exactly:
+      image (document): wrap in the instruction template
+        "<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>"
+        "Describe the image.<|im_end|><|endoftext|>"
+      query: append the augmentation suffix  <text> + "<|endoftext|>" * 10
+    Omitting these reproduces a silent ~2.5 ndcg@10 drop vs the native pipeline.
 """
 
 import requests

examples/tool_chat_template_minimax_m1.jinja

@@ -1,91 +0,0 @@
-{{ '<begin_of_document>' -}}
-{%- if custom_tools is defined %}
-    {%- set tools = custom_tools %}
-{%- endif %}
-{%- if not tools is defined %}
-    {%- set tools = none %}
-{%- endif %}
-
-{#- Extract system message #}
-{% set ns = namespace(system_prompt='') -%}
-{%- if messages[0]['role'] == 'system' %}
-    {%- if messages[0]['content'] is string %}
-        {%- set ns.system_prompt = messages[0]['content']|trim %}
-    {%- else %}
-        {%- set ns.system_prompt = messages[0]['content'][0]['text']|trim %}
-    {%- endif %}
-    {%- set messages = messages[1:] %}
-{%- else %}
-    {%- if tools is not none %}
-        {%- set ns.system_prompt = "You are a helpful assistant created by Minimax based on MiniMax-M1 model." %}
-    {%- else %}
-        {%- set ns.system_prompt = "You are a helpful assistant created by Minimax based on MiniMax-M1 model." %}
-    {%- endif %}
-{%- endif %}
-
-{#- System message #}
-{%- if ns.system_prompt != '' %}
-{{ '<beginning_of_sentence>system ai_setting=assistant\n' + ns.system_prompt + '<end_of_sentence>\n' -}}
-{%- endif %}
-
-{#- Tools configuration #}
-{%- if tools is not none %}
-{{ '<beginning_of_sentence>system tool_setting=tools\nYou are provided with these tools:\n<tools>\n' -}}
-{%- for tool in tools %}
-{{ tool | tojson ~ '\n' -}}
-{%- endfor %}
-{{ '</tools>\n\nIf you need to call tools, please respond with <tool_calls></tool_calls> XML tags, and provide tool-name and json-object of arguments, following the format below:\n<tool_calls>\n{"name": <tool-name>, "arguments": <args-json-object>}\n...\n</tool_calls><end_of_sentence>\n' -}}
-{%- endif %}
-
-{#- Process messages #}
-{%- for message in messages %}
-    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}
-        {%- if message['role'] == 'user' %}
-{{ '<beginning_of_sentence>user name=user\n' -}}
-{%- if message['content'] is string %}
-{{ message['content']|trim -}}
-{%- else %}
-{%- for content in message['content'] %}
-{%- if content['type'] == 'text' %}
-{{ content['text']|trim -}}
-{%- endif %}
-{%- endfor %}
-{%- endif %}
-{{ '<end_of_sentence>\n' -}}
-        {%- elif message['role'] == 'assistant' %}
-{{ '<beginning_of_sentence>ai name=assistant\n' -}}
-{%- if message['content'] is string %}
-{{ message['content']|trim -}}
-{%- else %}
-{%- for content in message['content'] | selectattr('type', 'equalto', 'text') %}
-{{ content['text']|trim -}}
-{%- endfor %}
-{%- endif %}
-{{ '<end_of_sentence>\n' -}}
-        {%- endif %}
-    {%- elif 'tool_calls' in message %}
-{{ '<beginning_of_sentence>ai name=assistant\n<tool_calls>\n' -}}
-{%- for tool_call in message.tool_calls %}
-{{ '{"name": "' + tool_call.function.name + '", "arguments": ' + tool_call.function.arguments | tojson + '}\n' -}}
-{%- endfor %}
-{{ '</tool_calls><end_of_sentence>\n' -}}
-    {%- elif message.role == "tool" or message.role == "ipython" %}
-{{ '<beginning_of_sentence>tool name=tools\n' -}}
-{%- if message.content is string %}
-{{ 'tool result: ' + message.content + '\n\n' -}}
-{%- else %}
-{%- for content in message['content'] %}
-{%- if content['type'] == 'text' %}
-{{ 'tool result: ' + content['text'] + '\n\n' -}}
-{%- elif content.get('name') %}
-{{ 'tool name: ' + content['name'] + '\ntool result: ' + content['text'] + '\n\n' -}}
-{%- endif %}
-{%- endfor %}
-{%- endif %}
-{{ '<end_of_sentence>\n' -}}
-    {%- endif %}
-{%- endfor %}
-
-{%- if add_generation_prompt %}
-{{ '<beginning_of_sentence>ai name=assistant\n' -}}
-{%- endif %}

rust/src/chat/src/renderer/hf/format.rs

@@ -386,7 +386,6 @@ mod tests {
             tool_chat_template_llama3.2_pythonic.jinja         => String
             tool_chat_template_llama4_json.jinja               => OpenAi
             tool_chat_template_llama4_pythonic.jinja           => OpenAi
-            tool_chat_template_minimax_m1.jinja                => OpenAi
             tool_chat_template_mistral.jinja                   => String
             tool_chat_template_mistral3.jinja                  => OpenAi
             tool_chat_template_mistral_parallel.jinja          => String

rust/src/chat/tests/chat.rs

@@ -16,7 +16,8 @@ use vllm_engine_core_client::protocol::logprobs::{
     Logprobs, MaybeWireLogprobs, PositionLogprobs, TokenLogprob,
 };
 use vllm_engine_core_client::protocol::{
-    EngineCoreFinishReason, EngineCoreOutput, EngineCoreOutputs, EngineCoreRequest, StopReason,
+    EngineCoreFinishReason, EngineCoreOutput, EngineCoreOutputs, EngineCoreRequest, LogprobsCount,
+    StopReason,
 };
 use vllm_engine_core_client::test_utils::{IpcNamespace, spawn_mock_engine_task};
 use vllm_engine_core_client::{EngineCoreClient, EngineCoreClientConfig};
@@ -1387,8 +1388,8 @@ async fn chat_stream_and_collect_preserve_prompt_and_sample_logprobs() {
     .await;
 
     let mut request = sample_request("chat-logprobs");
-    request.sampling_params.logprobs = Some(1);
-    request.sampling_params.prompt_logprobs = Some(1);
+    request.sampling_params.logprobs = Some(LogprobsCount::Top(1));
+    request.sampling_params.prompt_logprobs = Some(LogprobsCount::Top(1));
 
     let mut stream = chat.chat(request.clone()).await.unwrap();
     match next_semantic(&mut stream).await.unwrap().unwrap() {

rust/src/chat/tests/templates/vllm_examples/tool_chat_template_minimax_m1.jinja

@@ -1,91 +0,0 @@
-{{ '<begin_of_document>' -}}
-{%- if custom_tools is defined %}
-    {%- set tools = custom_tools %}
-{%- endif %}
-{%- if not tools is defined %}
-    {%- set tools = none %}
-{%- endif %}
-
-{#- Extract system message #}
-{% set ns = namespace(system_prompt='') -%}
-{%- if messages[0]['role'] == 'system' %}
-    {%- if messages[0]['content'] is string %}
-        {%- set ns.system_prompt = messages[0]['content']|trim %}
-    {%- else %}
-        {%- set ns.system_prompt = messages[0]['content'][0]['text']|trim %}
-    {%- endif %}
-    {%- set messages = messages[1:] %}
-{%- else %}
-    {%- if tools is not none %}
-        {%- set ns.system_prompt = "You are a helpful assistant created by Minimax based on MiniMax-M1 model." %}
-    {%- else %}
-        {%- set ns.system_prompt = "You are a helpful assistant created by Minimax based on MiniMax-M1 model." %}
-    {%- endif %}
-{%- endif %}
-
-{#- System message #}
-{%- if ns.system_prompt != '' %}
-{{ '<beginning_of_sentence>system ai_setting=assistant\n' + ns.system_prompt + '<end_of_sentence>\n' -}}
-{%- endif %}
-
-{#- Tools configuration #}
-{%- if tools is not none %}
-{{ '<beginning_of_sentence>system tool_setting=tools\nYou are provided with these tools:\n<tools>\n' -}}
-{%- for tool in tools %}
-{{ tool | tojson ~ '\n' -}}
-{%- endfor %}
-{{ '</tools>\n\nIf you need to call tools, please respond with <tool_calls></tool_calls> XML tags, and provide tool-name and json-object of arguments, following the format below:\n<tool_calls>\n{"name": <tool-name>, "arguments": <args-json-object>}\n...\n</tool_calls><end_of_sentence>\n' -}}
-{%- endif %}
-
-{#- Process messages #}
-{%- for message in messages %}
-    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}
-        {%- if message['role'] == 'user' %}
-{{ '<beginning_of_sentence>user name=user\n' -}}
-{%- if message['content'] is string %}
-{{ message['content']|trim -}}
-{%- else %}
-{%- for content in message['content'] %}
-{%- if content['type'] == 'text' %}
-{{ content['text']|trim -}}
-{%- endif %}
-{%- endfor %}
-{%- endif %}
-{{ '<end_of_sentence>\n' -}}
-        {%- elif message['role'] == 'assistant' %}
-{{ '<beginning_of_sentence>ai name=assistant\n' -}}
-{%- if message['content'] is string %}
-{{ message['content']|trim -}}
-{%- else %}
-{%- for content in message['content'] | selectattr('type', 'equalto', 'text') %}
-{{ content['text']|trim -}}
-{%- endfor %}
-{%- endif %}
-{{ '<end_of_sentence>\n' -}}
-        {%- endif %}
-    {%- elif 'tool_calls' in message %}
-{{ '<beginning_of_sentence>ai name=assistant\n<tool_calls>\n' -}}
-{%- for tool_call in message.tool_calls %}
-{{ '{"name": "' + tool_call.function.name + '", "arguments": ' + tool_call.function.arguments | tojson + '}\n' -}}
-{%- endfor %}
-{{ '</tool_calls><end_of_sentence>\n' -}}
-    {%- elif message.role == "tool" or message.role == "ipython" %}
-{{ '<beginning_of_sentence>tool name=tools\n' -}}
-{%- if message.content is string %}
-{{ 'tool result: ' + message.content + '\n\n' -}}
-{%- else %}
-{%- for content in message['content'] %}
-{%- if content['type'] == 'text' %}
-{{ 'tool result: ' + content['text'] + '\n\n' -}}
-{%- elif content.get('name') %}
-{{ 'tool name: ' + content['name'] + '\ntool result: ' + content['text'] + '\n\n' -}}
-{%- endif %}
-{%- endfor %}
-{%- endif %}
-{{ '<end_of_sentence>\n' -}}
-    {%- endif %}
-{%- endfor %}
-
-{%- if add_generation_prompt %}
-{{ '<beginning_of_sentence>ai name=assistant\n' -}}
-{%- endif %}

rust/src/cmd/src/cli.rs

@@ -20,6 +20,7 @@ use serde_with::{DefaultOnNull, OneOrMany, serde_as};
 use thiserror_ext::AsReport as _;
 use uuid::Uuid;
 use vllm_engine_core_client::TransportMode;
+use vllm_engine_core_client::protocol::LogprobsCount;
 use vllm_managed_engine::ManagedEngineConfig;
 use vllm_managed_engine::cli::{ManagedEngineArgs, repartition_managed_engine_args};
 use vllm_server::{
@@ -136,9 +137,9 @@ pub struct SharedRuntimeArgs {
     pub max_model_len: Option<u32>,
     /// Maximum number of log probabilities to return when `logprobs` is
     /// specified in sampling parameters. `-1` means no cap.
-    #[arg(long, value_parser = clap::value_parser!(i32).range(-1..), allow_negative_numbers = true)]
+    #[arg(long, allow_negative_numbers = true)]
     #[serde(default)]
-    pub max_logprobs: Option<i32>,
+    pub max_logprobs: Option<LogprobsCount>,
     /// TCP port for the gRPC Generate service. When not set, no gRPC server is
     /// started.
     #[arg(long)]
@@ -529,7 +530,7 @@ impl ServeArgs {
         self.managed_engine.clone().into_config(
             self.runtime.model.clone(),
             self.runtime.max_model_len,
-            self.runtime.max_logprobs,
+            self.runtime.max_logprobs.map(managed_max_logprobs_to_i32),
             self.runtime.language_model_only,
             self.runtime.disable_log_stats,
             self.runtime.shutdown_timeout,
@@ -555,5 +556,9 @@ fn frontend_ipc_addresses() -> (String, String) {
     )
 }
 
+fn managed_max_logprobs_to_i32(count: LogprobsCount) -> i32 {
+    i32::try_from(count).expect("max_logprobs is parsed through i32")
+}
+
 #[cfg(test)]
 mod tests;

rust/src/cmd/src/cli/tests.rs

@@ -1,5 +1,6 @@
 use expect_test::expect;
 use vllm_engine_core_client::TransportMode;
+use vllm_engine_core_client::protocol::LogprobsCount;
 use vllm_server::{Config, HttpListenerMode, ParserSelection, RendererSelection};
 
 use super::{Cli, Command};
@@ -165,10 +166,10 @@ fn serve_args_forward_max_logprobs_to_frontend_and_managed_engine() {
     let Command::Serve(args) = cli.command else {
         panic!("expected serve args");
     };
-    assert_eq!(args.runtime.max_logprobs, Some(-1));
+    assert_eq!(args.runtime.max_logprobs, Some(LogprobsCount::All));
 
     let frontend_config = args.to_frontend_config("tcp://127.0.0.1:62100".to_string());
-    assert_eq!(frontend_config.max_logprobs, Some(-1));
+    assert_eq!(frontend_config.max_logprobs, Some(LogprobsCount::All));
 
     let engine_config = args.to_managed_engine_config(5555);
     assert_eq!(engine_config.python_args, vec!["--max-logprobs", "-1"]);
@@ -529,7 +530,7 @@ fn frontend_args_json_accepts_supported_non_default_fields() {
     assert_eq!(args.runtime.renderer, RendererSelection::DeepSeekV32);
     assert!(args.runtime.language_model_only);
     assert_eq!(args.runtime.max_model_len, Some(8192));
-    assert_eq!(args.runtime.max_logprobs, Some(-1));
+    assert_eq!(args.runtime.max_logprobs, Some(LogprobsCount::All));
     assert_eq!(args.runtime.shutdown_timeout, 3);
 }
 

rust/src/engine-core-client/examples/external_engine_logprobs.rs

@@ -6,7 +6,7 @@ use futures::StreamExt as _;
 use tokio::time::timeout;
 use tracing_subscriber::EnvFilter;
 use vllm_engine_core_client::protocol::{
-    EngineCoreFinishReason, EngineCoreRequest, EngineCoreSamplingParams,
+    EngineCoreFinishReason, EngineCoreRequest, EngineCoreSamplingParams, LogprobsCount,
 };
 use vllm_engine_core_client::{
     EngineCoreClient, EngineCoreClientConfig, EngineCoreStreamOutput, TransportMode,
@@ -33,10 +33,10 @@ struct Args {
     output_timeout_secs: u64,
     #[arg(long, default_value_t = 1)]
     max_tokens: u32,
-    #[arg(long, default_value_t = 2)]
-    logprobs: i32,
-    #[arg(long, default_value_t = 1)]
-    prompt_logprobs: i32,
+    #[arg(long, default_value_t = LogprobsCount::Top(2), allow_negative_numbers = true)]
+    logprobs: LogprobsCount,
+    #[arg(long, default_value_t = LogprobsCount::Top(1), allow_negative_numbers = true)]
+    prompt_logprobs: LogprobsCount,
     #[arg(long, default_value_t = 96)]
     prompt_repeats: usize,
 }
@@ -64,8 +64,8 @@ fn build_request(
     request_id: String,
     prompt_token_ids: Vec<u32>,
     max_tokens: u32,
-    logprobs: i32,
-    prompt_logprobs: i32,
+    logprobs: LogprobsCount,
+    prompt_logprobs: LogprobsCount,
     client_index: u32,
 ) -> EngineCoreRequest {
     EngineCoreRequest {

rust/src/engine-core-client/src/protocol/logprobs_count.rs

@@ -0,0 +1,135 @@
+use std::fmt;
+use std::str::FromStr;
+
+use serde::{Deserialize, Deserializer, Serialize, Serializer};
+
+/// Number of log probabilities requested for a token position.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum LogprobsCount {
+    /// Return the full model vocabulary.
+    All,
+    /// Return the top-N tokens by probability.
+    Top(u32),
+}
+
+impl LogprobsCount {
+    /// Expands the count to the actual number of logprobs to return, given the vocabulary size.
+    pub fn expanded(self, vocab_size: usize) -> usize {
+        match self {
+            Self::All => vocab_size,
+            Self::Top(count) => count as usize,
+        }
+    }
+}
+
+impl TryFrom<i32> for LogprobsCount {
+    type Error = String;
+
+    fn try_from(value: i32) -> Result<Self, Self::Error> {
+        match value {
+            -1 => Ok(Self::All),
+            value if value < -1 => Err(format!("must be non-negative or -1, got {value}")),
+            value => Ok(Self::Top(value as u32)),
+        }
+    }
+}
+
+impl TryFrom<LogprobsCount> for i32 {
+    type Error = String;
+
+    fn try_from(value: LogprobsCount) -> Result<Self, Self::Error> {
+        match value {
+            LogprobsCount::All => Ok(-1),
+            LogprobsCount::Top(count) => {
+                i32::try_from(count).map_err(|_| format!("must fit within i32, got {count}"))
+            }
+        }
+    }
+}
+
+impl FromStr for LogprobsCount {
+    type Err = String;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        let value = s
+            .parse::<i32>()
+            .map_err(|e| format!("must be an i32 integer, got {s:?}: {e}"))?;
+        Self::try_from(value)
+    }
+}
+
+impl fmt::Display for LogprobsCount {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            Self::All => (-1).fmt(f),
+            Self::Top(count) => count.fmt(f),
+        }
+    }
+}
+
+impl Serialize for LogprobsCount {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: Serializer,
+    {
+        let value: i32 = (*self).try_into().map_err(serde::ser::Error::custom)?;
+        value.serialize(serializer)
+    }
+}
+
+impl<'de> Deserialize<'de> for LogprobsCount {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        let value = i32::deserialize(deserializer)?;
+        Self::try_from(value).map_err(serde::de::Error::custom)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use rmpv::Value;
+
+    use super::*;
+    use crate::protocol::{decode_msgpack, encode_msgpack};
+
+    #[test]
+    fn logprobs_count_serializes_as_wire_integer() {
+        assert_eq!(serde_json::to_value(LogprobsCount::All).unwrap(), -1);
+        assert_eq!(serde_json::to_value(LogprobsCount::Top(3)).unwrap(), 3);
+    }
+
+    #[test]
+    fn logprobs_count_deserializes_wire_integer() {
+        assert_eq!(
+            serde_json::from_value::<LogprobsCount>(serde_json::json!(-1)).unwrap(),
+            LogprobsCount::All
+        );
+        assert_eq!(
+            serde_json::from_value::<LogprobsCount>(serde_json::json!(3)).unwrap(),
+            LogprobsCount::Top(3)
+        );
+        assert!(serde_json::from_value::<LogprobsCount>(serde_json::json!(-2)).is_err());
+        assert!(
+            serde_json::from_value::<LogprobsCount>(serde_json::json!(i64::from(i32::MAX) + 1))
+                .is_err()
+        );
+    }
+
+    #[test]
+    fn logprobs_count_decodes_msgpack_signed_and_unsigned() {
+        let mut encoded = Vec::new();
+        rmpv::encode::write_value(&mut encoded, &Value::from(-1)).unwrap();
+        assert_eq!(
+            decode_msgpack::<LogprobsCount>(&encoded).unwrap(),
+            LogprobsCount::All
+        );
+
+        let encoded = encode_msgpack(&LogprobsCount::Top(7)).unwrap();
+        assert_eq!(
+            decode_msgpack::<LogprobsCount>(&encoded).unwrap(),
+            LogprobsCount::Top(7)
+        );
+    }
+}

rust/src/engine-core-client/src/protocol/mod.rs

@@ -56,6 +56,7 @@ mod classified_outputs;
 pub mod dtype;
 pub mod handshake;
 pub mod logprobs;
+mod logprobs_count;
 pub mod lora;
 pub mod multimodal;
 pub mod stats;
@@ -66,6 +67,7 @@ pub use classified_outputs::{
 };
 pub use dtype::ModelDtype;
 pub use logprobs::decode_engine_core_outputs;
+pub use logprobs_count::LogprobsCount;
 
 /// Request types are encoded as single-byte protocol constants so they can be
 /// sent over the ZMQ socket without an extra encoding step.
@@ -277,14 +279,20 @@ pub struct EngineCoreSamplingParams {
     pub max_tokens: u32,
     /// Minimum number of tokens to generate before EOS or stop-token handling.
     pub min_tokens: u32,
+    /// Maximum number of reasoning ("thinking") tokens to emit before the
+    /// reasoning section is force-closed. `None` means unlimited; the
+    /// user-facing `-1` sentinel is normalized to `None` by the frontend before
+    /// reaching this DTO, so only non-negative values are sent. Enforced
+    /// engine-side (and only when a reasoning parser is configured).
+    pub thinking_token_budget: Option<u64>,
     /// Number of log probabilities to return per generated token.
     ///
-    /// `None` disables sample logprobs. `-1` requests the full vocabulary.
-    pub logprobs: Option<i32>,
+    /// `None` disables sample logprobs.
+    pub logprobs: Option<LogprobsCount>,
     /// Number of log probabilities to return per prompt token.
     ///
-    /// `None` disables prompt logprobs. `-1` requests the full vocabulary.
-    pub prompt_logprobs: Option<i32>,
+    /// `None` disables prompt logprobs.
+    pub prompt_logprobs: Option<LogprobsCount>,
     /// Minimum probability threshold for token sampling.
     pub min_p: f32,
     /// Frequency penalty applied by the sampler.
@@ -345,6 +353,7 @@ impl EngineCoreSamplingParams {
             seed: None,
             max_tokens: 65536,
             min_tokens: 0,
+            thinking_token_budget: None,
             logprobs: None,
             prompt_logprobs: None,
             min_p: 0.0,

rust/src/engine-core-client/src/tests/client.rs

@@ -150,6 +150,7 @@ fn sample_request_with_id(request_id: &str) -> EngineCoreRequest {
             top_k: 8,
             max_tokens: 32,
             min_tokens: 1,
+            thinking_token_budget: Some(256),
             stop_token_ids: vec![151643],
             eos_token_id: Some(151645),
             all_stop_token_ids: BTreeSet::from([151643, 151645]),
@@ -2502,6 +2503,7 @@ fn python_msgpack_fixtures_match_rust_encoding() {
             seed: None,
             max_tokens: 16,
             min_tokens: 0,
+            thinking_token_budget: None,
             logprobs: None,
             prompt_logprobs: None,
             min_p: 0.0,

rust/src/engine-core-client/src/tests/python_compat.py

@@ -39,6 +39,7 @@ class EngineCoreSamplingParams(msgspec.Struct, dict=True, omit_defaults=True):
     seed: int | None = None
     max_tokens: int = 16
     min_tokens: int = 0
+    thinking_token_budget: int | None = None
     min_p: float = 0.0
     frequency_penalty: float = 0.0
     presence_penalty: float = 0.0
@@ -122,6 +123,7 @@ request = EngineCoreRequest(
         seed=None,
         max_tokens=32,
         min_tokens=1,
+        thinking_token_budget=256,
         min_p=0.0,
         frequency_penalty=0.0,
         presence_penalty=0.0,

rust/src/server/src/config.rs

@@ -2,12 +2,13 @@ use std::collections::HashMap;
 use std::fmt;
 use std::time::Duration;
 
-use anyhow::{Result, bail};
+use anyhow::Result;
 use axum::http::{HeaderName, HeaderValue, Method};
 use educe::Educe;
 use serde::Serialize;
 use serde_json::Value;
 use vllm_chat::{ChatTemplateContentFormatOption, ParserSelection, RendererSelection};
+use vllm_engine_core_client::protocol::LogprobsCount;
 use vllm_engine_core_client::{CoordinatorMode as EngineCoreCoordinatorMode, TransportMode};
 
 /// How the HTTP server obtains its listening socket.
@@ -133,7 +134,7 @@ pub struct Config {
     pub chat_template_content_format: ChatTemplateContentFormatOption,
     /// Optional maximum number of top log probabilities accepted by the
     /// frontend. `None` delegates to the text layer default.
-    pub max_logprobs: Option<i32>,
+    pub max_logprobs: Option<LogprobsCount>,
     /// HTTP/API-server behavior switches.
     pub api_server_options: ApiServerOptions,
     /// CORS settings applied to every HTTP response.
@@ -158,15 +159,6 @@ impl Config {
     pub fn validate(&self) -> Result<()> {
         vllm_chat::validate_parser_overrides(&self.tool_call_parser, &self.reasoning_parser)?;
         self.cors.validate()?;
-        if let Some(max_logprobs) = self.max_logprobs
-            && max_logprobs < -1
-        {
-            bail!(
-                "max_logprobs must be non-negative or -1, got {}",
-                max_logprobs
-            );
-        }
-
         Ok(())
     }
 

rust/src/server/src/error.rs

@@ -103,6 +103,7 @@ fn is_request_validation_error(error: &vllm_text::Error) -> bool {
             | vllm_text::Error::EmptyPromptTokenIds { .. }
             | vllm_text::Error::Logprobs(_)
             | vllm_text::Error::OutOfVocab(_)
+            | vllm_text::Error::InvalidThinkingTokenBudget
             // An empty tokenized prompt detected later, at request prepare
             // time, surfaces through the transparent Llm wrapper.
             | vllm_text::Error::Llm(vllm_llm::Error::EmptyPromptTokenIds { .. })
@@ -127,6 +128,18 @@ mod tests {
         assert!(response.error.message.contains("9000"));
     }
 
+    #[test]
+    fn invalid_thinking_token_budget_maps_to_invalid_request() {
+        let api_error = text_submit_error(
+            "failed to submit completion request",
+            vllm_text::Error::InvalidThinkingTokenBudget,
+        );
+        assert_eq!(api_error.status_code(), StatusCode::BAD_REQUEST);
+        let response = api_error.to_error_response();
+        assert_eq!(response.error.error_type, "invalid_request_error");
+        assert!(response.error.message.contains("thinking_token_budget"));
+    }
+
     #[test]
     fn chat_wrapped_prompt_too_long_maps_to_invalid_request() {
         let error = vllm_chat::Error::Text(vllm_text::Error::PromptTooLong {

rust/src/server/src/grpc/convert.rs

@@ -3,7 +3,7 @@
 
 use tonic::Status;
 use uuid::Uuid;
-use vllm_engine_core_client::protocol::{StopReason, StructuredOutputsParams};
+use vllm_engine_core_client::protocol::{LogprobsCount, StopReason, StructuredOutputsParams};
 use vllm_text::{
     DecodedLogprobs, DecodedPromptLogprobs, FinishReason, Finished, Prompt, SamplingParams,
     TextDecodeOptions, TextRequest,
@@ -202,18 +202,22 @@ fn build_sampling_params(
 /// Map the proto `CandidateTokens` selector to a `(logprobs_count,
 /// logprob_token_ids)` pair.
 ///
-/// - `top_n(k)` → `(k, None)` — return top-k candidates by probability
-/// - `all` → `(-1, None)` — return the full vocabulary
+/// - `top_n(k)` → `(Top(k), None)` — return top-k candidates by probability
+/// - `all` → `(All, None)` — return the full vocabulary
 /// - `token_ids(n)` → `(1, Some(vec of n token ids))` — return logprobs for specific tokens (the
 ///   count `n` is stored in the proto as the number of token IDs that follow, but the actual IDs
 ///   are carried via `logprob_token_ids` on `SamplingParams`)
-/// - absent → `(1, None)` — just the sampled/scored token
-fn candidate_logprob_spec(candidates: Option<&pb::CandidateTokens>) -> (i32, Option<Vec<u32>>) {
+/// - absent → `(Top(1), None)` — just the sampled/scored token
+fn candidate_logprob_spec(
+    candidates: Option<&pb::CandidateTokens>,
+) -> (LogprobsCount, Option<Vec<u32>>) {
     match candidates.and_then(|c| c.select.as_ref()) {
-        Some(pb::candidate_tokens::Select::TopN(n)) => (*n as i32, None),
-        Some(pb::candidate_tokens::Select::All(true)) => (-1, None),
-        Some(pb::candidate_tokens::Select::TokenIds(ids)) => (1, Some(ids.ids.clone())),
-        _ => (1, None),
+        Some(pb::candidate_tokens::Select::TopN(n)) => (LogprobsCount::Top(*n), None),
+        Some(pb::candidate_tokens::Select::All(true)) => (LogprobsCount::All, None),
+        Some(pb::candidate_tokens::Select::TokenIds(ids)) => {
+            (LogprobsCount::Top(1), Some(ids.ids.clone()))
+        }
+        _ => (LogprobsCount::Top(1), None),
     }
 }
 

rust/src/server/src/routes/inference/generate/convert.rs

@@ -87,6 +87,7 @@ pub(super) fn prepare_generate_request(
 #[cfg(test)]
 mod tests {
     use serde_json::json;
+    use vllm_engine_core_client::protocol::LogprobsCount;
     use vllm_text::Prompt;
 
     use super::prepare_generate_request;
@@ -132,10 +133,13 @@ mod tests {
             Prompt::TokenIds(vec![11, 22, 33])
         );
         assert_eq!(prepared.text_request.sampling_params.max_tokens, Some(7));
-        assert_eq!(prepared.text_request.sampling_params.logprobs, Some(2));
+        assert_eq!(
+            prepared.text_request.sampling_params.logprobs,
+            Some(LogprobsCount::Top(2))
+        );
         assert_eq!(
             prepared.text_request.sampling_params.prompt_logprobs,
-            Some(1)
+            Some(LogprobsCount::Top(1))
         );
         assert!(prepared.text_request.sampling_params.ignore_eos);
         assert_eq!(prepared.text_request.priority, -3);
@@ -150,6 +154,33 @@ mod tests {
         );
     }
 
+    #[test]
+    fn prepare_generate_request_forwards_thinking_token_budget() {
+        let request: GenerateRequest = serde_json::from_value(json!({
+            "model": "Qwen/Qwen1.5-0.5B-Chat",
+            "token_ids": [11, 22, 33],
+            "sampling_params": {
+                "thinking_token_budget": 64
+            }
+        }))
+        .expect("parse request");
+
+        let prepared = prepare_generate_request(
+            request,
+            &served(&["Qwen/Qwen1.5-0.5B-Chat"]),
+            ResolvedRequestContext::default(),
+        )
+        .expect("prepare");
+
+        // The raw inference route shares `vllm_text::SamplingParams`, so the
+        // field is carried through to lowering exactly like the OpenAI routes
+        // (normalization/validation then happens in `lower_sampling_params`).
+        assert_eq!(
+            prepared.text_request.sampling_params.thinking_token_budget,
+            Some(64)
+        );
+    }
+
     #[test]
     fn prepare_generate_request_gates_continuous_usage_on_include_usage() {
         let request: GenerateRequest = serde_json::from_value(json!({

rust/src/server/src/routes/inference/generate/validate.rs

@@ -34,16 +34,6 @@ pub(super) fn validate_request_compat(
         );
     }
 
-    if let Some(prompt_logprobs) = request.sampling_params.prompt_logprobs
-        && prompt_logprobs < 0
-        && prompt_logprobs != -1
-    {
-        bail_invalid_request!(
-            param = "sampling_params",
-            "`prompt_logprobs` must be a non-negative value or -1."
-        );
-    }
-
     Ok(())
 }
 

rust/src/server/src/routes/openai/chat_completions/convert.rs

@@ -4,6 +4,7 @@ use vllm_chat::{
     ChatMessage as VllmChatMessage, ChatOptions, ChatRequest, ChatTool, ChatToolChoice,
     GenerationPromptMode, SamplingParams,
 };
+use vllm_engine_core_client::protocol::LogprobsCount;
 
 use super::types::ChatCompletionRequest;
 use super::validate;
@@ -94,7 +95,7 @@ pub(super) fn prepare_chat_request(
 
     // Auto-enable prompt logprobs for non-streaming echo, matching Python vLLM's
     // behavior.
-    let top_logprobs = request.top_logprobs.unwrap_or(0);
+    let top_logprobs = request.top_logprobs.unwrap_or(LogprobsCount::Top(0));
     let prompt_logprobs = request
         .prompt_logprobs
         .or((request.echo && !request.stream).then_some(top_logprobs));
@@ -115,6 +116,7 @@ pub(super) fn prepare_chat_request(
             seed: request.seed,
             max_tokens: request.max_completion_tokens,
             min_tokens: request.min_tokens,
+            thinking_token_budget: request.thinking_token_budget,
             logprobs: request.logprobs.then_some(top_logprobs),
             prompt_logprobs,
             min_p: request.min_p,
@@ -377,6 +379,7 @@ mod tests {
         ChatTool as VllmChatTool, ChatToolChoice, GenerationPromptMode,
         SamplingParams as VllmSamplingParams,
     };
+    use vllm_engine_core_client::protocol::LogprobsCount;
     use vllm_text::output::TextDecodeOptions;
 
     use super::prepare_chat_request;
@@ -613,6 +616,31 @@ mod tests {
         assert_eq!(prepared.chat_request.sampling_params, expected);
     }
 
+    #[test]
+    fn prepare_chat_request_passes_through_thinking_token_budget() {
+        let prepare = |budget: Option<i64>| {
+            prepare_chat_request(
+                ChatCompletionRequest {
+                    thinking_token_budget: budget,
+                    ..base_request()
+                },
+                &served(&["Qwen/Qwen1.5-0.5B-Chat"]),
+                ResolvedRequestContext::default(),
+            )
+            .expect("request is valid")
+            .chat_request
+            .sampling_params
+            .thinking_token_budget
+        };
+
+        // The convert layer forwards the raw value verbatim (including the `-1`
+        // "unlimited" sentinel); normalization/validation happens during
+        // lowering (see `vllm_text::lower`).
+        assert_eq!(prepare(Some(64)), Some(64));
+        assert_eq!(prepare(Some(-1)), Some(-1));
+        assert_eq!(prepare(None), None);
+    }
+
     #[test]
     fn prepare_chat_request_accepts_developer_messages() {
         let request = ChatCompletionRequest {
@@ -941,7 +969,7 @@ mod tests {
         let request = ChatCompletionRequest {
             stream: false,
             logprobs: true,
-            prompt_logprobs: Some(2),
+            prompt_logprobs: Some(LogprobsCount::Top(2)),
             ..base_request()
         };
 
@@ -954,10 +982,13 @@ mod tests {
 
         assert!(prepared.options.requested_logprobs);
         assert!(prepared.options.include_prompt_logprobs);
-        assert_eq!(prepared.chat_request.sampling_params.logprobs, Some(0));
+        assert_eq!(
+            prepared.chat_request.sampling_params.logprobs,
+            Some(LogprobsCount::Top(0))
+        );
         assert_eq!(
             prepared.chat_request.sampling_params.prompt_logprobs,
-            Some(2)
+            Some(LogprobsCount::Top(2))
         );
     }
 
@@ -965,7 +996,7 @@ mod tests {
     fn prepare_chat_request_keeps_prompt_logprobs_independent_from_echo() {
         let request = ChatCompletionRequest {
             logprobs: true,
-            top_logprobs: Some(3),
+            top_logprobs: Some(LogprobsCount::Top(3)),
             echo: true,
             ..base_request()
         };
@@ -977,7 +1008,10 @@ mod tests {
         )
         .expect("request is valid");
 
-        assert_eq!(prepared.chat_request.sampling_params.logprobs, Some(3));
+        assert_eq!(
+            prepared.chat_request.sampling_params.logprobs,
+            Some(LogprobsCount::Top(3))
+        );
         assert_eq!(prepared.chat_request.sampling_params.prompt_logprobs, None);
         assert!(!prepared.options.include_prompt_logprobs);
     }

rust/src/server/src/routes/openai/chat_completions/types.rs

@@ -6,6 +6,7 @@ use serde_json::Value;
 use serde_with::SerializeDisplay;
 use validator::Validate;
 use vllm_chat::ReasoningEffort;
+use vllm_engine_core_client::protocol::LogprobsCount;
 
 use crate::routes::openai::utils::structured_outputs::ResponseFormat;
 use crate::routes::openai::utils::types::{
@@ -44,10 +45,8 @@ pub struct ChatCompletionRequest {
     #[serde(default)]
     pub logprobs: bool,
 
-    /// An integer specifying the number of most likely tokens to return
-    /// -1 means return all
-    #[validate(range(min = -1))]
-    pub top_logprobs: Option<i32>,
+    /// Number of most likely tokens to return. `-1` means return full vocab.
+    pub top_logprobs: Option<LogprobsCount>,
 
     /// Deprecated: Replaced by max_completion_tokens
     #[deprecated(note = "Use max_completion_tokens instead")]
@@ -155,8 +154,8 @@ pub struct ChatCompletionRequest {
     /// Truncate prompt tokens to this length
     pub truncate_prompt_tokens: Option<i64>,
 
-    /// Number of prompt logprobs to return
-    pub prompt_logprobs: Option<i32>,
+    /// Number of prompt logprobs to return. `-1` means return full vocab.
+    pub prompt_logprobs: Option<LogprobsCount>,
 
     /// Restrict output to these token IDs only
     pub allowed_token_ids: Option<Vec<u32>>,
@@ -165,8 +164,10 @@ pub struct ChatCompletionRequest {
     pub bad_words: Option<Vec<String>>,
 
     // -------- Extra vLLM Parameters --------
-    /// Token budget for reasoning/thinking
-    pub thinking_token_budget: Option<u32>,
+    /// Token budget for reasoning/thinking. Accepts a non-negative integer, or
+    /// `-1` for unlimited (mirroring the Python frontend, which normalizes `-1`
+    /// to "no budget").
+    pub thinking_token_budget: Option<i64>,
 
     /// Whether to include reasoning content in the response
     #[serde(default = "default_true")]

rust/src/server/src/routes/openai/chat_completions/validate.rs

@@ -1,6 +1,7 @@
 use super::types::ChatCompletionRequest;
 use crate::error::{ApiError, bail_invalid_request};
 use crate::routes::openai::utils::types::{ChatMessage, Tool, ToolChoice, ToolChoiceValue};
+use vllm_engine_core_client::protocol::LogprobsCount;
 
 /// Enforce the minimal compatibility contract for the Rust OpenAI server.
 pub(super) fn validate_request_compat(
@@ -30,14 +31,12 @@ pub(super) fn validate_request_compat(
     }
 
     if let Some(prompt_logprobs) = request.prompt_logprobs {
-        if prompt_logprobs < 0 && prompt_logprobs != -1 {
-            bail_invalid_request!(
-                param = "prompt_logprobs",
-                "prompt_logprobs must be a non-negative value or -1."
-            );
-        }
-
-        if request.stream && (prompt_logprobs > 0 || prompt_logprobs == -1) {
+        if request.stream
+            && matches!(
+                prompt_logprobs,
+                LogprobsCount::All | LogprobsCount::Top(1..)
+            )
+        {
             bail_invalid_request!(
                 param = "prompt_logprobs",
                 "prompt_logprobs are not available when stream=true."
@@ -108,11 +107,6 @@ pub(super) fn validate_request_compat(
         "truncate_prompt_tokens",
         "truncate_prompt_tokens is not supported.",
     )?;
-    reject_non_default(
-        request.thinking_token_budget.as_ref(),
-        "thinking_token_budget",
-        "thinking_token_budget is not supported.",
-    )?;
     reject_non_default(
         request.media_io_kwargs.as_ref(),
         "media_io_kwargs",
@@ -159,6 +153,7 @@ mod tests {
 
     use serde_json::json;
     use vllm_chat::ReasoningEffort;
+    use vllm_engine_core_client::protocol::LogprobsCount;
 
     use super::validate_request_compat;
     use crate::routes::openai::chat_completions::types::ChatCompletionRequest;
@@ -304,7 +299,7 @@ mod tests {
     #[test]
     fn validate_request_compat_rejects_top_logprobs_without_logprobs() {
         let request = ChatCompletionRequest {
-            top_logprobs: Some(0),
+            top_logprobs: Some(LogprobsCount::Top(0)),
             ..base_request()
         };
         assert!(validate_request_compat(&request, &served(&["Qwen/Qwen1.5-0.5B-Chat"])).is_err());
@@ -313,26 +308,26 @@ mod tests {
     #[test]
     fn validate_request_compat_rejects_streaming_prompt_logprobs_requests() {
         let request = ChatCompletionRequest {
-            prompt_logprobs: Some(1),
+            prompt_logprobs: Some(LogprobsCount::Top(1)),
             ..base_request()
         };
         assert!(validate_request_compat(&request, &served(&["Qwen/Qwen1.5-0.5B-Chat"])).is_err());
 
         let request = ChatCompletionRequest {
-            prompt_logprobs: Some(-1),
+            prompt_logprobs: Some(LogprobsCount::All),
             ..base_request()
         };
         assert!(validate_request_compat(&request, &served(&["Qwen/Qwen1.5-0.5B-Chat"])).is_err());
     }
 
     #[test]
-    fn validate_request_compat_rejects_invalid_prompt_logprobs_value() {
-        let request = ChatCompletionRequest {
-            stream: false,
-            prompt_logprobs: Some(-2),
-            ..base_request()
-        };
-        assert!(validate_request_compat(&request, &served(&["Qwen/Qwen1.5-0.5B-Chat"])).is_err());
+    fn chat_request_deserialization_rejects_invalid_prompt_logprobs_value() {
+        let result = serde_json::from_value::<ChatCompletionRequest>(json!({
+            "model": "Qwen/Qwen1.5-0.5B-Chat",
+            "messages": [{"role": "user", "content": "hello"}],
+            "prompt_logprobs": -2
+        }));
+        assert!(result.is_err());
     }
 
     #[test]

rust/src/server/src/routes/openai/completions/convert.rs

@@ -1,3 +1,4 @@
+use vllm_engine_core_client::protocol::LogprobsCount;
 use vllm_text::{SamplingParams, TextDecodeOptions, TextRequest};
 
 use super::types::CompletionRequest;
@@ -61,15 +62,7 @@ pub(super) fn prepare_completion_request(
         .map(|request| request.lora_name.clone())
         .unwrap_or_else(|| lora_resolution.model_names.first().cloned().unwrap_or_default());
 
-    let logprobs = match request.logprobs {
-        Some(logprobs) => Some(i32::try_from(logprobs).map_err(|_| {
-            ApiError::invalid_request(
-                "`logprobs` must fit within a signed 32-bit integer.".to_string(),
-                Some("logprobs"),
-            )
-        })?),
-        None => None,
-    };
+    let logprobs = request.logprobs.map(LogprobsCount::Top);
     let prompt_only = request.echo && request.max_tokens == Some(0);
     let prompt_logprobs =
         request.prompt_logprobs.or(if request.echo && (!request.stream || prompt_only) {
@@ -108,6 +101,7 @@ pub(super) fn prepare_completion_request(
             seed: request.seed,
             max_tokens,
             min_tokens: request.min_tokens,
+            thinking_token_budget: request.thinking_token_budget,
             logprobs,
             prompt_logprobs,
             min_p: request.min_p,
@@ -162,6 +156,7 @@ pub(super) fn prepare_completion_request(
 mod tests {
     use axum::http::HeaderMap;
     use serde_json::json;
+    use vllm_engine_core_client::protocol::LogprobsCount;
     use vllm_text::Prompt;
 
     use super::prepare_completion_request;
@@ -246,7 +241,10 @@ mod tests {
             Prompt::TokenIds(vec![11, 22, 33])
         );
         assert_eq!(prepared.text_request.sampling_params.max_tokens, Some(7));
-        assert_eq!(prepared.text_request.sampling_params.logprobs, Some(2));
+        assert_eq!(
+            prepared.text_request.sampling_params.logprobs,
+            Some(LogprobsCount::Top(2))
+        );
         assert_eq!(prepared.text_request.sampling_params.top_p, Some(0.9));
         assert_eq!(prepared.text_request.sampling_params.top_k, Some(42));
         assert_eq!(prepared.text_request.sampling_params.min_p, Some(0.1));
@@ -266,6 +264,34 @@ mod tests {
         assert!(!prepared.text_request.decode_options.skip_special_tokens);
     }
 
+    #[test]
+    fn prepare_completion_request_passes_through_thinking_token_budget() {
+        let prepare = |budget: serde_json::Value| {
+            let request: CompletionRequest = serde_json::from_value(json!({
+                "model": "Qwen/Qwen1.5-0.5B-Chat",
+                "prompt": "hello",
+                "thinking_token_budget": budget,
+            }))
+            .expect("parse request");
+            prepare_completion_request(
+                request,
+                &served(&["Qwen/Qwen1.5-0.5B-Chat"]),
+                ResolvedRequestContext::default(),
+            )
+            .expect("prepare")
+            .text_request
+            .sampling_params
+            .thinking_token_budget
+        };
+
+        // The convert layer forwards the raw value verbatim (including the `-1`
+        // "unlimited" sentinel); normalization/validation happens during
+        // lowering (see `vllm_text::lower`).
+        assert_eq!(prepare(json!(64)), Some(64));
+        assert_eq!(prepare(json!(-1)), Some(-1));
+        assert_eq!(prepare(json!(null)), None);
+    }
+
     #[test]
     fn prepare_completion_request_maps_stream_usage_and_token_format_options() {
         let request: CompletionRequest = serde_json::from_value(json!({
@@ -381,10 +407,13 @@ mod tests {
         .expect("prepare");
 
         assert!(prepared.options.prompt_only);
-        assert_eq!(prepared.text_request.sampling_params.logprobs, Some(3));
+        assert_eq!(
+            prepared.text_request.sampling_params.logprobs,
+            Some(LogprobsCount::Top(3))
+        );
         assert_eq!(
             prepared.text_request.sampling_params.prompt_logprobs,
-            Some(3)
+            Some(LogprobsCount::Top(3))
         );
     }
 
@@ -406,10 +435,13 @@ mod tests {
         )
         .expect("prepare");
 
-        assert_eq!(prepared.text_request.sampling_params.logprobs, Some(3));
+        assert_eq!(
+            prepared.text_request.sampling_params.logprobs,
+            Some(LogprobsCount::Top(3))
+        );
         assert_eq!(
             prepared.text_request.sampling_params.prompt_logprobs,
-            Some(3)
+            Some(LogprobsCount::Top(3))
         );
     }
 
@@ -450,10 +482,13 @@ mod tests {
             ResolvedRequestContext::default(),
         )
         .expect("prepare");
-        assert_eq!(prepared.text_request.sampling_params.logprobs, Some(1));
+        assert_eq!(
+            prepared.text_request.sampling_params.logprobs,
+            Some(LogprobsCount::Top(1))
+        );
         assert_eq!(
             prepared.text_request.sampling_params.prompt_logprobs,
-            Some(2)
+            Some(LogprobsCount::Top(2))
         );
     }
 

rust/src/server/src/routes/openai/completions/types.rs

@@ -3,6 +3,7 @@ use std::collections::HashMap;
 use serde::{Deserialize, Serialize};
 use serde_json::{Map, Value};
 use validator::Validate;
+use vllm_engine_core_client::protocol::LogprobsCount;
 use vllm_text::Prompt;
 
 use crate::routes::openai::utils::types::{
@@ -131,8 +132,8 @@ pub struct CompletionRequest {
     /// Restrict output to these token IDs only
     pub allowed_token_ids: Option<Vec<u32>>,
 
-    /// Number of prompt logprobs to return
-    pub prompt_logprobs: Option<i32>,
+    /// Number of prompt logprobs to return. `-1` means return full vocab.
+    pub prompt_logprobs: Option<LogprobsCount>,
 
     // -------- Extra vLLM Parameters --------
     /// Whether to add special tokens (e.g. BOS) to the prompt
@@ -146,6 +147,11 @@ pub struct CompletionRequest {
     /// Additional kwargs for structured outputs
     pub structured_outputs: Option<Value>,
 
+    /// Token budget for reasoning/thinking. Accepts a non-negative integer, or
+    /// `-1` for unlimited (mirroring the Python frontend, which normalizes `-1`
+    /// to "no budget").
+    pub thinking_token_budget: Option<i64>,
+
     /// Request scheduling priority (lower means earlier; default 0)
     pub priority: Option<i32>,
 

rust/src/server/src/routes/openai/completions/validate.rs

@@ -1,3 +1,4 @@
+use vllm_engine_core_client::protocol::LogprobsCount;
 use vllm_text::Prompt;
 
 use super::types::CompletionRequest;
@@ -44,29 +45,18 @@ pub(super) fn validate_request_compat(
         bail_invalid_request!(param = "suffix", "suffix is not supported.");
     }
 
-    if let Some(logprobs) = request.logprobs
-        && logprobs > i32::MAX as u32
-    {
-        bail_invalid_request!(
-            param = "logprobs",
-            "`logprobs` must fit within a signed 32-bit integer."
-        );
-    }
-
     if let Some(prompt_logprobs) = request.prompt_logprobs {
-        if request.stream && (prompt_logprobs > 0 || prompt_logprobs == -1) {
+        if request.stream
+            && matches!(
+                prompt_logprobs,
+                LogprobsCount::All | LogprobsCount::Top(1..)
+            )
+        {
             bail_invalid_request!(
                 param = "prompt_logprobs",
                 "`prompt_logprobs` are not available when `stream=true`."
             );
         }
-
-        if prompt_logprobs < 0 && prompt_logprobs != -1 {
-            bail_invalid_request!(
-                param = "prompt_logprobs",
-                "`prompt_logprobs` must be a non-negative value or -1."
-            );
-        }
     }
 
     if request.use_beam_search {
@@ -101,6 +91,7 @@ pub(super) fn validate_request_compat(
 #[cfg(test)]
 mod tests {
     use serde_json::json;
+    use vllm_engine_core_client::protocol::LogprobsCount;
 
     use super::validate_request_compat;
     use crate::routes::openai::completions::types::CompletionRequest;
@@ -150,7 +141,7 @@ mod tests {
     #[test]
     fn validate_request_compat_rejects_streaming_prompt_logprobs() {
         let request = CompletionRequest {
-            prompt_logprobs: Some(1),
+            prompt_logprobs: Some(LogprobsCount::Top(1)),
             ..base_request()
         };
         assert!(
@@ -162,7 +153,7 @@ mod tests {
     fn validate_request_compat_accepts_non_stream_prompt_logprobs() {
         let request = CompletionRequest {
             stream: false,
-            prompt_logprobs: Some(-1),
+            prompt_logprobs: Some(LogprobsCount::All),
             ..base_request()
         };
         assert!(

rust/src/text/src/backend/mod.rs

@@ -2,6 +2,7 @@ pub mod hf;
 
 use std::sync::Arc;
 
+use vllm_engine_core_client::protocol::LogprobsCount;
 use vllm_tokenizer::DynTokenizer;
 
 use crate::error::Result;
@@ -26,9 +27,7 @@ pub struct SamplingLimits {
     /// Runtime context window size reported by the engine startup handshake.
     pub max_model_len: u32,
     /// Maximum number of top log probabilities accepted by this frontend.
-    ///
-    /// `-1` means allowing requests up to the model vocabulary size.
-    pub max_logprobs: i32,
+    pub max_logprobs: LogprobsCount,
 
     /// Model vocabulary size from the model config, used to bound generated
     /// token IDs and logits-domain sampling controls.
@@ -41,7 +40,7 @@ pub struct SamplingLimits {
 impl SamplingLimits {
     /// Original Python definition:
     /// <https://github.com/vllm-project/vllm/blob/b5adb027ad03c29b46181752ba3b1cb84eff1dd4/vllm/config/model.py#L216-L220>
-    pub const DEFAULT_MAX_LOGPROBS: i32 = 20;
+    pub const DEFAULT_MAX_LOGPROBS: LogprobsCount = LogprobsCount::Top(20);
     /// Original Python definition:
     /// <https://github.com/vllm-project/vllm/blob/b5adb027ad03c29b46181752ba3b1cb84eff1dd4/vllm/sampling_params.py#L30-L32>
     pub const MAX_LOGPROB_TOKEN_IDS: usize = 128;

rust/src/text/src/error.rs

@@ -20,6 +20,8 @@ pub enum Error {
     Logprobs(#[from] LogprobsError),
     #[error(transparent)]
     OutOfVocab(#[from] OutOfVocabError),
+    #[error("`thinking_token_budget` must be a non-negative integer or -1 for unlimited.")]
+    InvalidThinkingTokenBudget,
     #[error("text request stream `{request_id}` closed before terminal output")]
     StreamClosedBeforeTerminalOutput { request_id: String },
     #[error(transparent)]

rust/src/text/src/lib.rs

@@ -19,6 +19,7 @@ pub use output::{
 pub use request::{Prompt, SamplingParams, TextRequest};
 use trait_set::trait_set;
 use vllm_engine_core_client::EngineCoreClient;
+use vllm_engine_core_client::protocol::LogprobsCount;
 pub use vllm_llm::FinishReason;
 use vllm_llm::{GenerateOutputStream, Llm};
 use vllm_tokenizer::DynTokenizer;
@@ -48,7 +49,7 @@ pub struct TextLlm {
     /// Runtime context window size reported by the engine startup handshake.
     max_model_len: u32,
     /// Maximum number of top log probabilities accepted by this text facade.
-    max_logprobs: i32,
+    max_logprobs: LogprobsCount,
 }
 
 impl TextLlm {
@@ -68,7 +69,7 @@ impl TextLlm {
     }
 
     /// Override the maximum accepted logprobs count.
-    pub fn with_max_logprobs(mut self, max_logprobs: Option<i32>) -> Self {
+    pub fn with_max_logprobs(mut self, max_logprobs: Option<LogprobsCount>) -> Self {
         if let Some(max_logprobs) = max_logprobs {
             self.max_logprobs = max_logprobs;
         }

rust/src/text/src/lower.rs

@@ -87,6 +87,7 @@ pub fn lower_sampling_params(
         seed,
         max_tokens,
         min_tokens,
+        thinking_token_budget,
         logprobs,
         prompt_logprobs,
         min_p,
@@ -128,6 +129,7 @@ pub fn lower_sampling_params(
         prompt_len,
     )?;
     let min_tokens = min_tokens.unwrap_or(0);
+    let thinking_token_budget = normalize_thinking_token_budget(thinking_token_budget)?;
     let frequency_penalty = frequency_penalty.unwrap_or(0.0);
     let presence_penalty = presence_penalty.unwrap_or(0.0);
 
@@ -149,6 +151,7 @@ pub fn lower_sampling_params(
         seed,
         max_tokens,
         min_tokens,
+        thinking_token_budget,
         logprobs,
         prompt_logprobs,
         min_p,
@@ -170,6 +173,21 @@ pub fn lower_sampling_params(
     Ok(params)
 }
 
+/// Normalize the user-facing `thinking_token_budget` into the engine value.
+///
+/// Mirrors Python's `validate_thinking_token_budget`
+/// (<https://github.com/vllm-project/vllm/blob/ecf9d83520eb217401b47d8a5451a27c5231b8c2/vllm/sampling_params.py#L35-L55>):
+/// `None` and the `-1` "unlimited" sentinel both map to `None`; any other
+/// negative value is rejected; non-negative values pass through unchanged. Like
+/// Python's `int`, no upper bound is imposed.
+fn normalize_thinking_token_budget(value: Option<i64>) -> Result<Option<u64>> {
+    match value {
+        None | Some(-1) => Ok(None),
+        Some(budget) if budget >= 0 => Ok(Some(budget as u64)),
+        Some(_) => Err(Error::InvalidThinkingTokenBudget),
+    }
+}
+
 /// Convert bad-word strings into token-ID sequences, following the Python vLLM
 /// logic in `SamplingParams.update_from_tokenizer()`.
 ///
@@ -251,6 +269,7 @@ mod tests {
     use std::collections::{BTreeSet, HashMap};
 
     use serial_test::file_serial;
+    use vllm_engine_core_client::protocol::LogprobsCount;
 
     use super::*;
     use crate::backend::hf::HfTextBackend;
@@ -366,6 +385,36 @@ mod tests {
         )
     }
 
+    #[test]
+    fn lower_sampling_params_normalizes_thinking_token_budget() {
+        let lower = |budget: Option<i64>| {
+            lower_sampling_params_with_limits(
+                SamplingParams {
+                    thinking_token_budget: budget,
+                    ..SamplingParams::default()
+                },
+                sample_sampling_limits(),
+            )
+        };
+
+        // Non-negative budgets (including 0) pass through unchanged.
+        assert_eq!(lower(Some(256)).unwrap().thinking_token_budget, Some(256));
+        assert_eq!(lower(Some(0)).unwrap().thinking_token_budget, Some(0));
+        // `None` and the `-1` "unlimited" sentinel both disable the budget.
+        assert_eq!(lower(None).unwrap().thinking_token_budget, None);
+        assert_eq!(lower(Some(-1)).unwrap().thinking_token_budget, None);
+        // No upper bound is imposed, matching Python's `int`.
+        assert_eq!(
+            lower(Some(i64::from(u32::MAX) + 1)).unwrap().thinking_token_budget,
+            Some(u64::from(u32::MAX) + 1)
+        );
+        // Other negatives are rejected.
+        assert!(matches!(
+            lower(Some(-2)),
+            Err(Error::InvalidThinkingTokenBudget)
+        ));
+    }
+
     #[test]
     fn lower_text_request_applies_python_style_eos_hints() {
         let prepared = lower_text_request(
@@ -386,6 +435,7 @@ mod tests {
                 seed: None,
                 max_tokens: 999997,
                 min_tokens: 0,
+                thinking_token_budget: None,
                 logprobs: None,
                 prompt_logprobs: None,
                 min_p: 0.0,
@@ -437,6 +487,7 @@ mod tests {
                 seed: None,
                 max_tokens: 999997,
                 min_tokens: 0,
+                thinking_token_budget: None,
                 logprobs: None,
                 prompt_logprobs: None,
                 min_p: 0.0,
@@ -567,6 +618,7 @@ mod tests {
                 seed: None,
                 max_tokens: 40957,
                 min_tokens: 0,
+                thinking_token_budget: None,
                 logprobs: None,
                 prompt_logprobs: None,
                 min_p: 0.0,
@@ -628,6 +680,7 @@ mod tests {
                 seed: None,
                 max_tokens: 999997,
                 min_tokens: 0,
+                thinking_token_budget: None,
                 logprobs: None,
                 prompt_logprobs: None,
                 min_p: 0.0,
@@ -697,6 +750,7 @@ mod tests {
                 seed: None,
                 max_tokens: 32,
                 min_tokens: 2,
+                thinking_token_budget: None,
                 logprobs: None,
                 prompt_logprobs: None,
                 min_p: 0.1,
@@ -721,8 +775,8 @@ mod tests {
     #[test]
     fn lower_sampling_params_passes_logprobs_fields_through() {
         let sampling_params = SamplingParams {
-            logprobs: Some(3),
-            prompt_logprobs: Some(-1),
+            logprobs: Some(LogprobsCount::Top(3)),
+            prompt_logprobs: Some(LogprobsCount::All),
             ..Default::default()
         };
 
@@ -739,7 +793,7 @@ mod tests {
                 default_max_tokens: None,
             },
             SamplingLimits {
-                max_logprobs: -1,
+                max_logprobs: LogprobsCount::All,
                 ..sample_sampling_limits()
             },
             3,
@@ -747,15 +801,15 @@ mod tests {
         )
         .unwrap();
 
-        assert_eq!(params.logprobs, Some(3));
-        assert_eq!(params.prompt_logprobs, Some(-1));
+        assert_eq!(params.logprobs, Some(LogprobsCount::Top(3)));
+        assert_eq!(params.prompt_logprobs, Some(LogprobsCount::All));
     }
 
     #[test]
     fn lower_sampling_params_rejects_full_vocab_logprobs_over_default_cap() {
         let error = lower_sampling_params_with_limits(
             SamplingParams {
-                logprobs: Some(-1),
+                logprobs: Some(LogprobsCount::All),
                 ..Default::default()
             },
             sample_sampling_limits(),
@@ -776,24 +830,24 @@ mod tests {
     fn lower_sampling_params_expands_full_vocab_logprobs_from_model_vocab() {
         let params = lower_sampling_params_with_limits(
             SamplingParams {
-                logprobs: Some(-1),
+                logprobs: Some(LogprobsCount::All),
                 ..Default::default()
             },
             SamplingLimits {
-                max_logprobs: 1500,
+                max_logprobs: LogprobsCount::Top(1500),
                 ..sample_sampling_limits()
             },
         )
         .unwrap();
 
-        assert_eq!(params.logprobs, Some(-1));
+        assert_eq!(params.logprobs, Some(LogprobsCount::All));
     }
 
     #[test]
     fn lower_sampling_params_rejects_invalid_logprob_token_ids() {
         let error = lower_sampling_params_with_limits(
             SamplingParams {
-                logprobs: Some(1),
+                logprobs: Some(LogprobsCount::Top(1)),
                 logprob_token_ids: Some(vec![1000]),
                 ..Default::default()
             },
@@ -929,6 +983,7 @@ mod tests {
                 seed: None,
                 max_tokens: 128,
                 min_tokens: 0,
+                thinking_token_budget: None,
                 logprobs: None,
                 prompt_logprobs: None,
                 min_p: 0.1,

rust/src/text/src/lower/logprobs.rs

@@ -1,15 +1,14 @@
 //! Python-compatible validation for logprobs sampling params.
 //!
-//! `-1` is expanded only for bounds checks. The original request values are
+//! `All` is expanded only for bounds checks. The original request values are
 //! passed through to engine-core.
 
 use crate::backend::SamplingLimits;
 use thiserror::Error;
+use vllm_engine_core_client::protocol::LogprobsCount;
 
 #[derive(Debug, Error)]
 pub enum LogprobsError {
-    #[error("{parameter} must be non-negative or -1, got {value}")]
-    InvalidCount { parameter: &'static str, value: i32 },
     #[error(
         "requested {parameter} of {requested}, which is greater than max allowed: {max_allowed}"
     )]
@@ -30,19 +29,21 @@ pub enum LogprobsError {
         "when both logprobs and logprob_token_ids are set, logprobs must equal \
          len(logprob_token_ids). Got logprobs={logprobs}, len(logprob_token_ids)={num_token_ids}."
     )]
-    TokenIdsMismatch { logprobs: i32, num_token_ids: usize },
+    TokenIdsMismatch {
+        logprobs: LogprobsCount,
+        num_token_ids: usize,
+    },
 }
 
 /// Validate logprobs count sampling parameters.
 pub(super) fn validate_logprobs(
-    logprobs: Option<i32>,
-    prompt_logprobs: Option<i32>,
+    logprobs: Option<LogprobsCount>,
+    prompt_logprobs: Option<LogprobsCount>,
     logprob_token_ids: Option<&[u32]>,
     sampling_limits: SamplingLimits,
 ) -> Result<(), LogprobsError> {
     let vocab_size = sampling_limits.model_vocab_size;
-    let max_logprobs =
-        normalize_logprobs_count(sampling_limits.max_logprobs, vocab_size, "max_logprobs")?;
+    let max_logprobs = sampling_limits.max_logprobs.expanded(vocab_size);
 
     validate_logprobs_count(logprobs, max_logprobs, vocab_size, "logprobs")?;
     validate_logprobs_count(prompt_logprobs, max_logprobs, vocab_size, "prompt_logprobs")?;
@@ -50,7 +51,7 @@ pub(super) fn validate_logprobs(
 }
 
 fn validate_logprobs_count(
-    requested: Option<i32>,
+    requested: Option<LogprobsCount>,
     max_logprobs: usize,
     vocab_size: usize,
     parameter: &'static str,
@@ -59,7 +60,7 @@ fn validate_logprobs_count(
         return Ok(());
     };
 
-    let requested = normalize_logprobs_count(requested, vocab_size, parameter)?;
+    let requested = requested.expanded(vocab_size);
     if requested > max_logprobs {
         return Err(LogprobsError::TooManyCount {
             parameter,
@@ -72,7 +73,7 @@ fn validate_logprobs_count(
 }
 
 pub(super) fn validate_logprob_token_ids(
-    logprobs: Option<i32>,
+    logprobs: Option<LogprobsCount>,
     logprob_token_ids: Option<&[u32]>,
 ) -> Result<(), LogprobsError> {
     let Some(logprob_token_ids) = logprob_token_ids else {
@@ -88,7 +89,7 @@ pub(super) fn validate_logprob_token_ids(
     }
 
     if let Some(logprobs) = logprobs
-        && logprobs != n as i32
+        && logprobs != LogprobsCount::Top(n as u32)
     {
         return Err(LogprobsError::TokenIdsMismatch {
             logprobs,
@@ -98,15 +99,3 @@ pub(super) fn validate_logprob_token_ids(
 
     Ok(())
 }
-
-fn normalize_logprobs_count(
-    value: i32,
-    vocab_size: usize,
-    parameter: &'static str,
-) -> Result<usize, LogprobsError> {
-    match value {
-        -1 => Ok(vocab_size),
-        value if value < 0 => Err(LogprobsError::InvalidCount { parameter, value }),
-        value => Ok(value as usize),
-    }
-}

rust/src/text/src/output/decoded.rs

@@ -309,7 +309,7 @@ fn matches_stop_string(stops: &[String], output: &str, new_bytes: usize) -> Opti
         .find_map(|(ss_idx, (ss, len, start_off))| {
             output[start_off..]
                 .windows(len)
-                .rposition(|w| w == ss)
+                .position(|w| w == ss)
                 .map(|pos| (ss_idx, start_off + pos))
         })
 }
@@ -562,6 +562,13 @@ mod tests {
         assert_eq!(result, Some((0, 4)));
     }
 
+    #[test]
+    fn stop_string_matches_leftmost_with_multiple_new_bytes() {
+        let stops = vec!["\n".to_string()];
+        let result = matches_stop_string(&stops, "Answer\n\n", 2);
+        assert_eq!(result, Some((0, 6)));
+    }
+
     #[test]
     fn stop_string_matches_at_beginning() {
         let stops = vec!["say".to_string()];

rust/src/text/src/request.rs

@@ -3,9 +3,9 @@ use std::collections::HashMap;
 use enum_as_inner::EnumAsInner;
 use serde::{Deserialize, Serialize};
 use serde_json::Value;
-use vllm_engine_core_client::protocol::StructuredOutputsParams;
 use vllm_engine_core_client::protocol::lora::LoraRequest;
 use vllm_engine_core_client::protocol::multimodal::MmFeatures;
+use vllm_engine_core_client::protocol::{LogprobsCount, StructuredOutputsParams};
 
 use crate::error::{Error, Result};
 use crate::output::TextDecodeOptions;
@@ -56,14 +56,20 @@ pub struct SamplingParams {
     pub max_tokens: Option<u32>,
     /// Minimum number of tokens to generate before EOS or stop-token handling.
     pub min_tokens: Option<u32>,
+    /// Maximum number of reasoning ("thinking") tokens to emit before the
+    /// reasoning section is force-closed. `None` or the user-facing `-1`
+    /// "unlimited" sentinel both disable the budget. The raw value is carried
+    /// here; `-1` is normalized to `None` (and other negatives rejected) during
+    /// lowering (see `lower_sampling_params`).
+    pub thinking_token_budget: Option<i64>,
     /// Number of log probabilities to return per generated token.
     ///
-    /// `None` disables sample logprobs. `-1` requests the full vocabulary.
-    pub logprobs: Option<i32>,
+    /// `None` disables sample logprobs.
+    pub logprobs: Option<LogprobsCount>,
     /// Number of log probabilities to return per prompt token.
     ///
-    /// `None` disables prompt logprobs. `-1` requests the full vocabulary.
-    pub prompt_logprobs: Option<i32>,
+    /// `None` disables prompt logprobs.
+    pub prompt_logprobs: Option<LogprobsCount>,
     /// Minimum probability threshold for token sampling. `None` means no
     /// explicit user override.
     pub min_p: Option<f32>,
@@ -116,6 +122,7 @@ impl Default for SamplingParams {
             seed: None,
             max_tokens: None,
             min_tokens: None,
+            thinking_token_budget: None,
             logprobs: None,
             prompt_logprobs: None,
             min_p: None,

tests/entrypoints/serve/tokenize/test_serving_tokenization.py

@@ -15,7 +15,7 @@ from vllm.entrypoints.serve.tokenize.protocol import (
     TokenizeChatRequest,
     TokenizeCompletionRequest,
 )
-from vllm.entrypoints.serve.tokenize.serving import OpenAIServingTokenization
+from vllm.entrypoints.serve.tokenize.serving import ServingTokenization
 from vllm.v1.engine.async_llm import AsyncLLM
 
 MODEL_NAME = "openai-community/gpt2"
@@ -58,7 +58,7 @@ class MockModelConfig:
         return self.diff_sampling_param or {}
 
 
-def _build_serving_tokenization(engine: AsyncLLM) -> OpenAIServingTokenization:
+def _build_serving_tokenization(engine: AsyncLLM) -> ServingTokenization:
     models = OpenAIServingModels(
         engine_client=engine,
         base_model_paths=BASE_MODEL_PATHS,
@@ -71,8 +71,7 @@ def _build_serving_tokenization(engine: AsyncLLM) -> OpenAIServingTokenization:
         chat_template=None,
         chat_template_content_format="auto",
     )
-    return OpenAIServingTokenization(
-        engine,
+    return ServingTokenization(
         models,
         openai_serving_render=serving_render,
         request_logger=None,

tests/kernels/attention/test_attention_selector.py

@@ -435,7 +435,7 @@ def test_per_head_quant_scales_backend_selection(
     ]
     + (
         [
-            ("FLASHINFER", True, False),  # FlashInfer does not support non-causal
+            ("FLASHINFER", True, True),  # FlashInfer supports non-causal
             ("FLASHINFER", False, True),  # FlashInfer works with causal
         ]
         if CudaPlatform is not None

tests/kernels/test_mhc_kernels.py

@@ -8,8 +8,8 @@ from vllm.model_executor.kernels.mhc.tilelang import (
     _tilelang_hc_prenorm_gemm,
     _torch_hc_prenorm_gemm,
 )
+from vllm.model_executor.layers.mhc import HAS_TILELANG_MHC
 from vllm.platforms import current_platform
-from vllm.utils.import_utils import has_tilelang
 from vllm.utils.torch_utils import set_random_seed
 
 DEVICE = current_platform.device_type
@@ -97,8 +97,8 @@ def hc_head_ref(
 
 
 @pytest.mark.skipif(
-    not (current_platform.is_cuda_alike() and has_tilelang()),
-    reason="CUDA or ROCm and tilelang required",
+    not HAS_TILELANG_MHC,
+    reason="TileLang MHC support required",
 )
 @pytest.mark.parametrize("num_tokens", [1, 4, 8, 128])
 @pytest.mark.parametrize("hidden_size", [4096, 7168])
@@ -150,8 +150,8 @@ def test_mhc_pre_tilelang(num_tokens, hidden_size, hc_mult):
 
 
 @pytest.mark.skipif(
-    not (current_platform.is_cuda_alike() and has_tilelang()),
-    reason="CUDA or ROCm and tilelang required",
+    not HAS_TILELANG_MHC,
+    reason="TileLang MHC support required",
 )
 @pytest.mark.parametrize(
     ("num_tokens", "hidden_size"),
@@ -190,8 +190,8 @@ def test_hc_prenorm_gemm_tilelang(num_tokens, hidden_size):
 
 
 @pytest.mark.skipif(
-    not (current_platform.is_cuda_alike() and has_tilelang()),
-    reason="CUDA or ROCm and tilelang required",
+    not HAS_TILELANG_MHC,
+    reason="TileLang MHC support required",
 )
 @pytest.mark.parametrize("num_tokens", [1, 4, 8, 128])
 @pytest.mark.parametrize("hidden_size", [4096, 7168])
@@ -217,8 +217,8 @@ def test_mhc_post_tilelang(num_tokens, hidden_size, hc_mult):
 
 
 @pytest.mark.skipif(
-    not (current_platform.is_cuda_alike() and has_tilelang()),
-    reason="CUDA or ROCm and tilelang required",
+    not HAS_TILELANG_MHC,
+    reason="TileLang MHC support required",
 )
 @pytest.mark.parametrize("num_tokens", [1, 4, 8, 128])
 @pytest.mark.parametrize("hidden_size", [4096, 7168])
@@ -324,8 +324,8 @@ def test_hc_head_triton(num_tokens, hidden_size, hc_mult):
 
 
 @pytest.mark.skipif(
-    not (current_platform.is_cuda_alike() and has_tilelang()),
-    reason="CUDA or ROCm and tilelang required",
+    not HAS_TILELANG_MHC,
+    reason="TileLang MHC support required",
 )
 @pytest.mark.parametrize("num_tokens", [1, 4, 8, 128])
 @pytest.mark.parametrize("hidden_size", [4096, 7168])

tests/models/multimodal/generation/test_common.py

@@ -810,29 +810,6 @@ VLM_TEST_SETTINGS = {
         hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
         patch_hf_runner=model_utils.minicpmv_26_patch_hf_runner,
     ),
-    "minimax_vl_01": VLMTestInfo(
-        models=["MiniMaxAI/MiniMax-VL-01"],
-        prompt_formatter=lambda img_prompt: f"<beginning_of_sentence>user: {img_prompt} assistant:<end_of_sentence>",  # noqa: E501
-        img_idx_to_prompt=lambda _: "<image>",
-        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
-        max_model_len=8192,
-        max_num_seqs=4,
-        dtype="bfloat16",
-        hf_output_post_proc=model_utils.minimax_vl_01_hf_output,
-        patch_hf_runner=model_utils.minimax_vl_01_patch_hf_runner,
-        auto_cls=AutoModelForImageTextToText,
-        marks=[
-            large_gpu_mark(min_gb=80),
-            # TODO: [ROCm] Fix pickle issue with ROCm spawn and tp>1
-            pytest.mark.skipif(
-                current_platform.is_rocm(),
-                reason=(
-                    "ROCm: Model too large for single GPU; "
-                    "multi-GPU blocked by HF _LazyConfigMapping pickle issue with spawn"
-                ),
-            ),
-        ],
-    ),
     "molmo": VLMTestInfo(
         models=["allenai/Molmo-7B-D-0924"],
         test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),

tests/models/multimodal/generation/vlm_utils/model_utils.py

@@ -245,13 +245,6 @@ def minicpmv_trunc_hf_output(hf_output: RunnerOutput, model: str) -> RunnerOutpu
     return output_ids, output_str, out_logprobs
 
 
-def minimax_vl_01_hf_output(hf_output: RunnerOutput, model: str) -> RunnerOutput:
-    output_ids, output_str, out_logprobs = hf_output
-    if output_str.endswith("<end_of_sentence>"):
-        output_str = output_str.split("<end_of_sentence>")[0]
-    return output_ids, output_str, out_logprobs
-
-
 def ultravox_trunc_hf_output(hf_output: RunnerOutput, model: str) -> RunnerOutput:
     output_ids, output_str, out_logprobs = hf_output
 
@@ -1023,17 +1016,6 @@ def minicpmv_26_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
     return hf_model
 
 
-def minimax_vl_01_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
-    orig_generate = hf_model.model.generate
-
-    def _generate(self, *args, image_sizes=None, **kwargs):
-        return orig_generate(*args, decode_text=False, **kwargs)
-
-    hf_model.model.generate = types.MethodType(_generate, hf_model.model)
-
-    return hf_model
-
-
 def molmo_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
     """Patches and returns an instance of the HfRunner to use for Molmo."""
     hf_processor = hf_model.processor

tests/models/multimodal/pooling/test_colqwen3_5.py

@@ -152,3 +152,21 @@ def test_colqwen3_5_relevance_ordering(
     dtype: str,
 ) -> None:
     _run_relevance_test(vllm_runner, model, dtype=dtype)
+
+
+def test_colqwen3_5_config_enables_bidirectional_attention() -> None:
+    """ColQwen3.5 retrieval must be served BIDIRECTIONAL (is_causal=False) so the
+    full_attention layers build with AttentionType.ENCODER_ONLY. This guards the
+    silent-causal regression (no GPU / model load needed)."""
+    from types import SimpleNamespace
+
+    from vllm.model_executor.models.config import (
+        MODELS_CONFIG_MAP,
+        ColQwen3_5Config,
+    )
+
+    assert MODELS_CONFIG_MAP["ColQwen3_5"] is ColQwen3_5Config
+
+    model_config = SimpleNamespace(hf_config=SimpleNamespace())
+    ColQwen3_5Config.verify_and_update_model_config(model_config)
+    assert model_config.hf_config.is_causal is False

tests/models/multimodal/processing/test_minimax_vl_01.py

@@ -1,113 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import pytest
-from PIL import Image
-
-from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.parse import ImageSize
-from vllm.multimodal.processing import BaseMultiModalProcessor
-
-from ....conftest import ImageTestAssets
-from ...utils import build_model_context
-
-
-@pytest.mark.parametrize("model_id", ["MiniMaxAI/MiniMax-VL-01"])
-@pytest.mark.parametrize("num_imgs", [1, 2])
-def test_processor_override(
-    image_assets: ImageTestAssets,
-    model_id: str,
-    num_imgs: int,
-):
-    ctx = build_model_context(
-        model_id,
-        mm_processor_kwargs=None,
-        limit_mm_per_prompt={"image": num_imgs},
-    )
-    processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
-    prompt = "<image>" * num_imgs
-    image = Image.new("RGB", size=(364, 364))
-    mm_data = {"image": [image] * num_imgs}
-
-    processed_inputs = processor(
-        prompt,
-        mm_items=processor.info.parse_mm_data(mm_data),
-        hf_processor_mm_kwargs={},
-    )
-    image_placeholders = processed_inputs["mm_placeholders"]["image"]
-
-    assert len(image_placeholders) == num_imgs
-
-
-def _validate_image_prompt_replacements_one(
-    processor: BaseMultiModalProcessor,
-    num_imgs: int,
-    failed_size_excs: list[tuple[ImageSize, Exception]],
-    image_size: ImageSize,
-) -> None:
-    prompt = "<image>" * num_imgs
-    image = Image.new("RGB", size=image_size)
-    mm_data = {"image": [image] * num_imgs}
-
-    try:
-        processed_inputs = processor(
-            prompt,
-            mm_items=processor.info.parse_mm_data(mm_data),
-            hf_processor_mm_kwargs={},
-        )
-
-        image_placeholders = processed_inputs["mm_placeholders"]["image"]
-        assert len(image_placeholders) == num_imgs
-
-    except Exception as exc:
-        failed_size_excs.append((image_size, exc))
-
-
-def _test_image_prompt_replacements(
-    processor,
-    *,
-    num_imgs: int,
-    image_sizes: list[ImageSize],
-) -> None:
-    failed_size_excs = list[tuple[ImageSize, Exception]]()
-
-    for size in image_sizes:
-        _validate_image_prompt_replacements_one(
-            processor, num_imgs, failed_size_excs, size
-        )
-
-    if failed_size_excs:
-        msg = "Found failing image sizes:" + "\n========\n".join(
-            f"[{size}]\n{exc}" for size, exc in failed_size_excs
-        )
-        raise AssertionError(msg)
-
-
-@pytest.mark.parametrize("model_id", ["MiniMaxAI/MiniMax-VL-01"])
-@pytest.mark.parametrize("num_imgs", [1, 2])
-def test_processor_prompt_replacements_regression(model_id, num_imgs):
-    ctx = build_model_context(
-        model_id,
-        mm_processor_kwargs=None,
-        limit_mm_per_prompt={"image": num_imgs},
-    )
-    processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
-
-    image_ratios = [
-        (171, 152),
-        (184, 161),
-        (198, 176),
-        (333, 296),
-        (369, 328),
-        (488, 183),
-        (2560, 1669),
-    ]
-    image_sizes = [
-        size for w, h in image_ratios for size in [ImageSize(w, h), ImageSize(h, w)]
-    ]
-
-    _test_image_prompt_replacements(
-        processor,
-        num_imgs=num_imgs,
-        image_sizes=image_sizes,
-    )

tests/models/registry.py

@@ -421,15 +421,6 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
         },
         trust_remote_code=True,
     ),
-    "MiniMaxForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-Text-01-hf"),
-    "MiniMaxText01ForCausalLM": _HfExamplesInfo(
-        "MiniMaxAI/MiniMax-Text-01",
-        trust_remote_code=True,
-        revision="a59aa9cbc53b9fb8742ca4e9e1531b9802b6fdc3",
-    ),
-    "MiniMaxM1ForCausalLM": _HfExamplesInfo(
-        "MiniMaxAI/MiniMax-M1-40k", trust_remote_code=True
-    ),
     "MiniMaxM2ForCausalLM": _HfExamplesInfo(
         "MiniMaxAI/MiniMax-M2",
         trust_remote_code=True,
@@ -1113,10 +1104,6 @@ _MULTIMODAL_EXAMPLE_MODELS = {
         "openbmb/MiniCPM-V-4_6",
         min_transformers_version="5.7.0",
     ),
-    "MiniMaxVL01ForConditionalGeneration": _HfExamplesInfo(
-        "MiniMaxAI/MiniMax-VL-01",
-        trust_remote_code=True,
-    ),
     "MiniMaxM3SparseForConditionalGeneration": _HfExamplesInfo(
         "MiniMaxAI/MiniMax-M3",
         trust_remote_code=True,

tests/models/test_initialization.py

@@ -98,11 +98,6 @@ def can_initialize(
         vllm_config.validate_block_size()
         return scheduler_kv_cache_config
 
-    if model_arch == "MiniMaxVL01ForConditionalGeneration":
-        pytest.skip(
-            "pickle error when loading `transformers.models.auto.CONFIG_MAPPING`"
-        )
-
     if model_arch == "MoonshotKimiaForCausalLM":
         pytest.skip(
             "Kimi-Audio requires SpeechToTextConfig "

tests/models/utils.py

@@ -507,7 +507,13 @@ def dummy_hf_overrides(
     # Only set MoE related config when the model has MoE layers.
     # Otherwise all models detected as MoE by _get_transformers_backend_cls.
     if model_arch_config.num_experts > 0:
-        num_experts_per_tok = 1 if model_arch == "Llama4ForConditionalGeneration" else 2
+        num_experts_per_tok = 2
+        if model_arch in (
+            "Llama4ForConditionalGeneration",
+            "Llama4ForCausalLM",
+            "EagleLlama4ForCausalLM",
+        ):
+            num_experts_per_tok = 1
         update_dict.update(
             {
                 "num_experts": num_experts,

tests/v1/attention/test_attention_backends_selection.py

@@ -6,10 +6,12 @@ from types import SimpleNamespace
 
 import pytest
 
+from vllm.model_executor.layers.mamba.linear.minimax_linear_attn import (
+    MiniMaxText01LinearAttention,
+)
 from vllm.model_executor.layers.mamba.mamba_mixer import MambaMixer
 from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
 from vllm.model_executor.layers.mamba.short_conv import ShortConv
-from vllm.model_executor.models.minimax_text_01 import MiniMaxText01LinearAttention
 from vllm.v1.attention.backends.linear_attn import LinearAttentionBackend
 from vllm.v1.attention.backends.mamba1_attn import Mamba1AttentionBackend
 from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionBackend

tests/v1/attention/test_gdn_metadata_builder.py

@@ -16,6 +16,7 @@ from tests.v1.attention.utils import (
     create_vllm_config,
 )
 from vllm.config import SpeculativeConfig
+from vllm.config.compilation import CUDAGraphMode
 from vllm.v1.attention.backends.gdn_attn import (
     GDNAttentionMetadata,
     GDNAttentionMetadataBuilder,
@@ -123,9 +124,15 @@ GDN_BUILD_TEST_CASES = {
 
 def _create_gdn_builder(
     num_speculative_tokens: int = 0,
+    full_cuda_graph: bool = False,
 ) -> GDNAttentionMetadataBuilder:
     """Create a GDNAttentionMetadataBuilder with minimal config."""
-    vllm_config = create_vllm_config(block_size=BLOCK_SIZE)
+    vllm_config = create_vllm_config(
+        model_name="Qwen/Qwen3.5-0.8B",
+        block_size=BLOCK_SIZE,
+    )
+    if full_cuda_graph:
+        vllm_config.compilation_config.cudagraph_mode = CUDAGraphMode.FULL_AND_PIECEWISE
     if num_speculative_tokens > 0:
         vllm_config.speculative_config = SpeculativeConfig(
             method="ngram",
@@ -189,3 +196,28 @@ def test_has_initial_state_after_reclassification():
     assert meta.has_initial_state is not None
     # req0 has context_lens = 65 - 1 = 64 > 0, so has_initial_state[0] = True
     assert meta.has_initial_state[0].item() is True
+
+
+def test_full_cudagraph_spec_metadata_uses_request_count():
+    """FULL cudagraph token padding must not pad request-indexed metadata."""
+    num_speculative_tokens = 3
+    builder = _create_gdn_builder(
+        num_speculative_tokens=num_speculative_tokens,
+        full_cuda_graph=True,
+    )
+    batch = BatchSpec(seq_lens=[80, 96], query_lens=[4, 4])
+    meta = _build(builder, batch, num_decode_draft_tokens=[3, 3])
+
+    assert meta.num_spec_decodes == batch.batch_size
+    assert meta.num_spec_decode_tokens == batch.compute_num_tokens()
+    assert meta.spec_state_indices_tensor is not None
+    assert meta.spec_state_indices_tensor.shape == (
+        batch.batch_size,
+        num_speculative_tokens + 1,
+    )
+    assert meta.spec_sequence_masks is not None
+    assert meta.spec_sequence_masks.shape == (batch.batch_size,)
+    assert meta.spec_query_start_loc is not None
+    assert meta.spec_query_start_loc.shape == (batch.batch_size + 1,)
+    assert meta.num_accepted_tokens is not None
+    assert meta.num_accepted_tokens.shape == (batch.batch_size,)

tests/v1/core/prefix_cache/test_partial_prefix_cache_primitives.py

@@ -0,0 +1,460 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from collections.abc import Callable
+
+import pytest
+
+import vllm.v1.core.kv_cache_utils as kv_cache_utils
+from vllm.distributed.kv_events import BlockRemoved, BlockStored
+from vllm.sampling_params import SamplingParams
+from vllm.utils.hashing import sha256
+from vllm.v1.core.block_pool import BlockPool
+from vllm.v1.core.kv_cache_utils import (
+    BlockHash,
+    BlockHashListWithBlockSize,
+    KVCacheBlock,
+    get_request_block_hasher,
+    hash_block_tokens,
+    init_none_hash,
+)
+from vllm.v1.request import Request
+
+pytestmark = pytest.mark.cpu_test
+
+
+@pytest.fixture(autouse=True)
+def _auto_init_hash_fn():
+    init_none_hash(sha256)
+
+
+def make_request(
+    request_id: str,
+    prompt_token_ids: list[int],
+    hash_block_size: int,
+    hash_fn: Callable,
+) -> Request:
+    sampling_params = SamplingParams(max_tokens=17)
+    sampling_params.update_from_generation_config({}, eos_token_id=100)
+    return Request(
+        request_id=request_id,
+        prompt_token_ids=prompt_token_ids,
+        sampling_params=sampling_params,
+        pooling_params=None,
+        block_hasher=get_request_block_hasher(hash_block_size, hash_fn),
+    )
+
+
+def boundary_hash(req: Request, hash_block_size: int, num_tokens: int) -> BlockHash:
+    # Every boundary at a hash_block_size multiple is just the fine-grained
+    # chain hash ending there.
+    return req.block_hashes[num_tokens // hash_block_size - 1]
+
+
+def cache_full_block_and_partial_tail(
+    token_ids: list[int],
+    *,
+    enable_kv_cache_events: bool = False,
+) -> tuple[BlockPool, Request, list[KVCacheBlock], BlockHash]:
+    hash_block_size = 2
+    block_size = 6
+    kv_cache_group_id = 0
+    req = make_request("0", token_ids, hash_block_size, sha256)
+    pool = BlockPool(
+        num_gpu_blocks=3,
+        enable_caching=True,
+        hash_block_size=hash_block_size,
+        enable_kv_cache_events=enable_kv_cache_events,
+    )
+    blocks = pool.get_new_blocks(2)
+
+    pool.cache_full_blocks(
+        request=req,
+        blocks=blocks,
+        num_cached_blocks=0,
+        num_full_blocks=1,
+        block_size=block_size,
+        kv_cache_group_id=kv_cache_group_id,
+    )
+    partial_hash = boundary_hash(req, hash_block_size, len(token_ids))
+    assert pool.cache_partial_block(
+        request=req,
+        block=blocks[1],
+        num_tokens=len(token_ids),
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+    return pool, req, blocks, partial_hash
+
+
+def test_boundary_hashes_reuse_fine_grained_chain():
+    hash_block_size = 2
+    block_size = 6
+    token_ids = [0, 0, 1, 1, 2, 2, 3, 3, 4, 4]
+    req = make_request("0", token_ids, hash_block_size, sha256)
+
+    coarse = BlockHashListWithBlockSize(req.block_hashes, hash_block_size, block_size)
+    # The block_size=6 full-block hash is the fine hash at the 6-token boundary,
+    # not a concatenation of the three fine hashes inside the block.
+    assert coarse[0] == req.block_hashes[6 // hash_block_size - 1]
+    assert coarse[0] != BlockHash(
+        req.block_hashes[0] + req.block_hashes[1] + req.block_hashes[2]
+    )
+    # A partial tail at 10 tokens is the fine hash at the 10-token boundary,
+    # which chains over the entire prefix.
+    tail_hash = boundary_hash(req, hash_block_size, 10)
+    assert tail_hash == req.block_hashes[4]
+    assert tail_hash == hash_block_tokens(sha256, req.block_hashes[3], token_ids[8:10])
+
+
+def test_cache_partial_block_kv_cache_events():
+    hash_block_size = 4
+    block_size = 12
+    kv_cache_group_id = 2
+
+    pool = BlockPool(
+        num_gpu_blocks=2,
+        enable_caching=True,
+        hash_block_size=hash_block_size,
+        enable_kv_cache_events=True,
+    )
+    req = make_request(
+        "req_partial_events",
+        prompt_token_ids=list(range(hash_block_size * 2)),
+        hash_block_size=hash_block_size,
+        hash_fn=sha256,
+    )
+
+    block = pool.get_new_blocks(1)[0]
+    partial_entry_hash = pool.cache_partial_block(
+        request=req,
+        block=block,
+        num_tokens=hash_block_size * 2,
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+
+    events = pool.take_events()
+    assert len(events) == 1
+    stored_event = events[0]
+    assert isinstance(stored_event, BlockStored)
+    assert partial_entry_hash is not None
+    assert stored_event.block_hashes == [
+        kv_cache_utils.maybe_convert_block_hash(req.block_hashes[1])
+    ]
+    assert stored_event.parent_block_hash == kv_cache_utils.maybe_convert_block_hash(
+        req.block_hashes[0]
+    )
+    assert stored_event.token_ids == req.all_token_ids[hash_block_size:]
+    assert stored_event.block_size == 4
+    assert stored_event.group_idx == kv_cache_group_id
+
+    duplicate_entry_hash = pool.cache_partial_block(
+        request=req,
+        block=block,
+        num_tokens=hash_block_size * 2,
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+    assert duplicate_entry_hash == partial_entry_hash
+    assert pool.take_events() == []
+
+    pool.free_blocks([block])
+    pool.get_new_blocks(1)
+    events = pool.take_events()
+    assert len(events) == 1
+    removed_event = events[0]
+    assert isinstance(removed_event, BlockRemoved)
+    assert removed_event.block_hashes == stored_event.block_hashes
+    assert removed_event.group_idx == kv_cache_group_id
+
+
+def test_partial_block_replacement_emits_remove_then_store_events():
+    hash_block_size = 2
+    block_size = 6
+    kv_cache_group_id = 0
+    req = make_request("0", [0, 0, 1, 1, 2, 2, 3, 3], hash_block_size, sha256)
+    pool = BlockPool(
+        num_gpu_blocks=3,
+        enable_caching=True,
+        hash_block_size=hash_block_size,
+        enable_kv_cache_events=True,
+    )
+    blocks = pool.get_new_blocks(2)
+
+    pool.cache_full_blocks(
+        request=req,
+        blocks=blocks,
+        num_cached_blocks=0,
+        num_full_blocks=1,
+        block_size=block_size,
+        kv_cache_group_id=kv_cache_group_id,
+    )
+    partial_hash_8 = boundary_hash(req, hash_block_size, 8)
+    assert pool.cache_partial_block(
+        request=req,
+        block=blocks[1],
+        num_tokens=8,
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+    assert pool.get_cached_block(partial_hash_8, [kv_cache_group_id]) == [blocks[1]]
+    pool.take_events()
+
+    req.append_output_token_ids([4, 4])
+    partial_hash_10 = boundary_hash(req, hash_block_size, 10)
+    assert pool.cache_partial_block(
+        request=req,
+        block=blocks[1],
+        num_tokens=10,
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+    events = pool.take_events()
+
+    assert len(events) == 2
+    removed_event, stored_event = events
+    assert isinstance(removed_event, BlockRemoved)
+    assert removed_event.block_hashes == [
+        kv_cache_utils.maybe_convert_block_hash(partial_hash_8)
+    ]
+    assert removed_event.group_idx == kv_cache_group_id
+    assert isinstance(stored_event, BlockStored)
+    assert stored_event.block_hashes == [
+        kv_cache_utils.maybe_convert_block_hash(partial_hash_10)
+    ]
+    assert stored_event.parent_block_hash == kv_cache_utils.maybe_convert_block_hash(
+        boundary_hash(req, hash_block_size, 8)
+    )
+    assert stored_event.token_ids == req.all_token_ids[8:10]
+    assert stored_event.block_size == hash_block_size
+    assert stored_event.group_idx == kv_cache_group_id
+    assert pool.get_cached_block(partial_hash_8, [kv_cache_group_id]) is None
+    assert pool.get_cached_block(partial_hash_10, [kv_cache_group_id]) == [blocks[1]]
+
+
+def test_later_request_hits_cached_partial_tail():
+    hash_block_size = 2
+    block_size = 6
+    kv_cache_group_id = 0
+    cached_token_ids = [0, 0, 1, 1, 2, 2, 3, 3, 4, 4]
+    req = make_request("0", cached_token_ids, hash_block_size, sha256)
+    pool = BlockPool(
+        num_gpu_blocks=3,
+        enable_caching=True,
+        hash_block_size=hash_block_size,
+    )
+    blocks = pool.get_new_blocks(2)
+
+    pool.cache_full_blocks(
+        request=req,
+        blocks=blocks,
+        num_cached_blocks=0,
+        num_full_blocks=1,
+        block_size=block_size,
+        kv_cache_group_id=kv_cache_group_id,
+    )
+    partial_hash_10 = boundary_hash(req, hash_block_size, 10)
+    assert pool.cache_partial_block(
+        request=req,
+        block=blocks[1],
+        num_tokens=10,
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+
+    replay = make_request("1", cached_token_ids, hash_block_size, sha256)
+    replay_hash_10 = boundary_hash(replay, hash_block_size, 10)
+    assert replay_hash_10 == partial_hash_10
+    assert pool.get_cached_block(replay_hash_10, [kv_cache_group_id]) == [blocks[1]]
+
+    extended = make_request("2", cached_token_ids + [10], hash_block_size, sha256)
+    extended_hash_10 = boundary_hash(extended, hash_block_size, 10)
+    assert extended_hash_10 == partial_hash_10
+    assert pool.get_cached_block(extended_hash_10, [kv_cache_group_id]) == [blocks[1]]
+
+
+def test_cache_partial_block_uses_fine_grained_boundary_hash():
+    hash_block_size = 2
+    block_size = 6
+    kv_cache_group_id = 0
+    token_ids = [0, 0, 1, 1, 2, 2, 3, 3, 4, 4]
+    req = make_request("0", token_ids, hash_block_size, sha256)
+    pool = BlockPool(
+        num_gpu_blocks=3,
+        enable_caching=True,
+        hash_block_size=hash_block_size,
+    )
+    blocks = pool.get_new_blocks(2)
+
+    pool.cache_full_blocks(
+        request=req,
+        blocks=blocks,
+        num_cached_blocks=0,
+        num_full_blocks=1,
+        block_size=block_size,
+        kv_cache_group_id=kv_cache_group_id,
+    )
+
+    partial_entry_hash = pool.cache_partial_block(
+        request=req,
+        block=blocks[1],
+        num_tokens=10,
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+    # The partial entry is keyed by the fine-grained hash at the 10-token
+    # boundary, regardless of the owning group's block_size.
+    expected = boundary_hash(req, hash_block_size, 10)
+    assert partial_entry_hash == kv_cache_utils.make_block_hash_with_group_id(
+        expected, kv_cache_group_id
+    )
+    assert pool.get_cached_block(expected, [kv_cache_group_id]) == [blocks[1]]
+
+
+def test_cache_partial_block_requires_hash_boundary():
+    hash_block_size = 2
+    block_size = 4
+    req = make_request("0", [0, 0, 1, 1], hash_block_size, sha256)
+    pool = BlockPool(
+        num_gpu_blocks=2,
+        enable_caching=True,
+        hash_block_size=hash_block_size,
+    )
+    block = pool.get_new_blocks(1)[0]
+
+    with pytest.raises(AssertionError):
+        pool.cache_partial_block(
+            request=req,
+            block=block,
+            num_tokens=3,
+            kv_cache_group_id=0,
+            block_size=block_size,
+        )
+
+
+def test_cache_partial_block_duplicate_checks_all_blocks_for_hash():
+    hash_block_size = 2
+    block_size = 4
+    kv_cache_group_id = 0
+    req = make_request("0", [0, 0, 1, 1], hash_block_size, sha256)
+    pool = BlockPool(
+        num_gpu_blocks=4,
+        enable_caching=True,
+        hash_block_size=hash_block_size,
+    )
+    blocks = pool.get_new_blocks(2)
+
+    first_entry_hash = pool.cache_partial_block(
+        request=req,
+        block=blocks[0],
+        num_tokens=2,
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+    second_entry_hash = pool.cache_partial_block(
+        request=req,
+        block=blocks[1],
+        num_tokens=2,
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+    assert first_entry_hash == second_entry_hash
+
+    duplicate_entry_hash = pool.cache_partial_block(
+        request=req,
+        block=blocks[1],
+        num_tokens=2,
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+    assert duplicate_entry_hash == second_entry_hash
+    assert pool.cached_block_hashes_by_block == {}
+
+
+def test_reset_prefix_cache_clears_partial_entry_metadata():
+    pool, req, blocks, partial_hash_10 = cache_full_block_and_partial_tail(
+        [0, 0, 1, 1, 2, 2, 3, 3, 4, 4]
+    )
+    full_hash = BlockHashListWithBlockSize(req.block_hashes, 2, 6)[0]
+
+    assert pool.get_cached_block(full_hash, [0]) == [blocks[0]]
+    assert pool.get_cached_block(partial_hash_10, [0]) == [blocks[1]]
+
+    pool.free_blocks(blocks)
+    assert pool.reset_prefix_cache()
+
+    assert pool.get_cached_block(full_hash, [0]) is None
+    assert pool.get_cached_block(partial_hash_10, [0]) is None
+    assert pool.cached_block_hashes_by_block == {}
+
+
+def test_evict_cached_block_removes_full_hash_and_partial_entry():
+    pool, req, blocks, partial_hash_10 = cache_full_block_and_partial_tail(
+        [0, 0, 1, 1, 2, 2, 3, 3, 4, 4]
+    )
+    full_hash = BlockHashListWithBlockSize(req.block_hashes, 2, 6)[0]
+
+    assert pool.get_cached_block(full_hash, [0]) == [blocks[0]]
+    assert pool.get_cached_block(partial_hash_10, [0]) == [blocks[1]]
+
+    pool.evict_blocks({blocks[0].block_id, blocks[1].block_id})
+
+    assert pool.get_cached_block(full_hash, [0]) is None
+    assert pool.get_cached_block(partial_hash_10, [0]) is None
+    assert pool.cached_block_hashes_by_block == {}
+
+
+def test_partial_block_promotes_to_direct_full_block_hash():
+    hash_block_size = 2
+    block_size = 6
+    kv_cache_group_id = 0
+    token_ids = [0, 0, 1, 1, 2, 2, 3, 3, 4, 4]
+    req = make_request("0", token_ids, hash_block_size, sha256)
+    pool = BlockPool(
+        num_gpu_blocks=3,
+        enable_caching=True,
+        hash_block_size=hash_block_size,
+    )
+    blocks = pool.get_new_blocks(2)
+
+    pool.cache_full_blocks(
+        request=req,
+        blocks=blocks,
+        num_cached_blocks=0,
+        num_full_blocks=1,
+        block_size=block_size,
+        kv_cache_group_id=kv_cache_group_id,
+    )
+    partial_hash_10 = boundary_hash(req, hash_block_size, 10)
+    assert pool.cache_partial_block(
+        request=req,
+        block=blocks[1],
+        num_tokens=10,
+        kv_cache_group_id=kv_cache_group_id,
+        block_size=block_size,
+    )
+    assert pool.get_cached_block(partial_hash_10, [kv_cache_group_id]) == [blocks[1]]
+
+    req.append_output_token_ids([5, 5])
+    full_hashes = BlockHashListWithBlockSize(
+        req.block_hashes, hash_block_size, block_size
+    )
+    promoted_full_hash = full_hashes[1]
+    # The promoted full-block hash is the fine hash at the 12-token boundary,
+    # not a concatenation of the fine hashes inside the block.
+    assert promoted_full_hash == req.block_hashes[12 // hash_block_size - 1]
+    assert promoted_full_hash != BlockHash(
+        req.block_hashes[3] + req.block_hashes[4] + req.block_hashes[5]
+    )
+
+    pool.cache_full_blocks(
+        request=req,
+        blocks=blocks,
+        num_cached_blocks=1,
+        num_full_blocks=2,
+        block_size=block_size,
+        kv_cache_group_id=kv_cache_group_id,
+    )
+    assert pool.get_cached_block(promoted_full_hash, [kv_cache_group_id]) == [blocks[1]]
+    assert pool.get_cached_block(partial_hash_10, [kv_cache_group_id]) is None

tests/v1/core/test_kv_cache_utils.py

@@ -225,7 +225,7 @@ def test_kv_cache_block():
 
     # Test block hash setting and resetting
     block_hash = make_block_hash_with_group_id(BlockHash(b"abc"), 0)
-    block.block_hash = block_hash
+    block.set_block_hash(block_hash)
     assert block.block_hash == block_hash
 
     block.reset_hash()

tests/v1/core/test_prefix_caching.py

@@ -2003,7 +2003,7 @@ def test_maybe_evict_cached_block():
     assert len(pool.blocks) == len(block_hashes)
     # Manually add all blocks to cached_blocks
     for block, block_hash in zip(pool.blocks, block_hashes):
-        block.block_hash = block_hash
+        block.set_block_hash(block_hash)
         pool.cached_block_hash_to_block.insert(block_hash, block)
 
     block0, block1, block2, block3 = pool.blocks

tests/v1/kv_connector/unit/offloading_connector/test_events.py

@@ -0,0 +1,355 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from unittest.mock import MagicMock
+
+import pytest
+import torch
+
+from tests.v1.kv_connector.unit.utils import create_vllm_config
+from vllm.config import KVEventsConfig, KVTransferConfig
+from vllm.distributed.kv_events import BlockRemoved, BlockStored
+from vllm.distributed.kv_transfer.kv_connector.v1.offloading.events import (
+    OffloadingEventGroupSpec,
+    OffloadingEventsTracker,
+)
+from vllm.distributed.kv_transfer.kv_connector.v1.offloading.scheduler import (
+    GroupOffloadConfig,
+)
+from vllm.v1.core.kv_cache_utils import BlockHash, maybe_convert_block_hash
+from vllm.v1.kv_cache_interface import (
+    FullAttentionSpec,
+    KVCacheConfig,
+    KVCacheGroupSpec,
+    KVCacheSpecKind,
+)
+from vllm.v1.kv_offload.base import (
+    OffloadingEvent,
+    OffloadingKVEventsConfig,
+    OffloadKey,
+    make_offload_key,
+)
+from vllm.v1.kv_offload.cpu.common import CPULoadStoreSpec
+from vllm.v1.kv_offload.tiering.spec import TieringOffloadingSpec
+
+_CPU_MEDIUM = CPULoadStoreSpec.medium()
+_FULL_ATTENTION_EVENT_SPEC = OffloadingEventGroupSpec(
+    kv_cache_spec_kind=KVCacheSpecKind.FULL_ATTENTION.value,
+    kv_cache_spec_sliding_window=None,
+)
+
+
+def _tracker(
+    *,
+    enable_kv_cache_events: bool = True,
+    self_describing_kv_events: bool = True,
+) -> OffloadingEventsTracker:
+    return OffloadingEventsTracker(
+        OffloadingKVEventsConfig(
+            enable_kv_cache_events=enable_kv_cache_events,
+            self_describing_kv_events=self_describing_kv_events,
+        )
+    )
+
+
+def _hash(i: int) -> BlockHash:
+    return BlockHash(str(i).encode())
+
+
+def _wire_hash(block_hash: BlockHash):
+    return maybe_convert_block_hash(block_hash)
+
+
+def _request(*, block_hashes: list[BlockHash], token_count: int):
+    req = MagicMock()
+    req.block_hashes = block_hashes
+    req.all_token_ids = list(range(1, token_count + 1))
+    req.lora_request = None
+    return req
+
+
+def _group_config(
+    *,
+    group_idx: int = 0,
+    block_size: int = 4,
+    block_size_factor: int = 1,
+    sliding_window_size_in_blocks: int | None = None,
+) -> GroupOffloadConfig:
+    return GroupOffloadConfig(
+        group_idx=group_idx,
+        gpu_block_size=block_size,
+        offloaded_block_size=block_size * block_size_factor,
+        hash_block_size_factor=block_size_factor,
+        sliding_window_size_in_blocks=sliding_window_size_in_blocks,
+        kv_event_group_spec=_FULL_ATTENTION_EVENT_SPEC,
+    )
+
+
+def _record_chunks(
+    tracker: OffloadingEventsTracker,
+    req,
+    group_config: GroupOffloadConfig,
+    num_chunks: int,
+) -> list[OffloadKey]:
+    keys: list[OffloadKey] = []
+    hbf = group_config.hash_block_size_factor
+    for chunk_idx in range(num_chunks):
+        tail_hash = req.block_hashes[(chunk_idx + 1) * hbf - 1]
+        assert tail_hash is not None
+        key = make_offload_key(tail_hash, group_config.group_idx)
+        tracker.record_store(req, group_config, chunk_idx, key)
+        keys.append(key)
+    return keys
+
+
+def _stored_event(keys: list[OffloadKey]) -> OffloadingEvent:
+    return OffloadingEvent(keys=keys, medium=_CPU_MEDIUM, removed=False)
+
+
+def _removed_event(keys: list[OffloadKey]) -> OffloadingEvent:
+    return OffloadingEvent(keys=keys, medium=_CPU_MEDIUM, removed=True)
+
+
+def test_take_events_publishes_routable_block_stored():
+    block_size = 4
+    tracker = _tracker()
+    group_config = _group_config(block_size=block_size)
+    req = _request(
+        block_hashes=[_hash(i) for i in range(6)],
+        token_count=block_size * 6,
+    )
+    keys = _record_chunks(tracker, req, group_config, num_chunks=6)
+
+    batch1 = list(tracker.take_events([_stored_event(keys[:3])]))
+    assert len(batch1) == 3
+
+    for i, event in enumerate(batch1):
+        assert isinstance(event, BlockStored)
+        assert event.medium == _CPU_MEDIUM
+        assert event.block_hashes == [_wire_hash(_hash(i))]
+        assert event.block_size == block_size
+        assert event.token_ids == list(
+            range(i * block_size + 1, (i + 1) * block_size + 1)
+        )
+        if i == 0:
+            assert event.parent_block_hash is None
+        else:
+            assert event.parent_block_hash == _wire_hash(_hash(i - 1))
+        assert event.lora_id is None
+        assert event.lora_name is None
+        assert event.extra_keys is None
+        assert event.group_idx == 0
+        assert event.kv_cache_spec_kind == KVCacheSpecKind.FULL_ATTENTION.value
+        assert event.kv_cache_spec_sliding_window is None
+
+    batch2 = list(tracker.take_events([_stored_event(keys[3:])]))
+    assert len(batch2) == 3
+    assert batch2[0].parent_block_hash == batch1[-1].block_hashes[-1]
+
+    assert len(tracker._pending_event_metadata) == 6
+
+
+def test_take_events_factor_gt_1_chunk_store_and_remove():
+    block_size = 4
+    block_size_factor = 3
+    tracker = _tracker()
+    group_config = _group_config(
+        block_size=block_size, block_size_factor=block_size_factor
+    )
+    req = _request(
+        block_hashes=[_hash(i) for i in range(6)],
+        token_count=block_size * block_size_factor * 2,
+    )
+    keys = _record_chunks(tracker, req, group_config, num_chunks=2)
+
+    stored = list(tracker.take_events([_stored_event(keys)]))
+    assert len(stored) == 2
+
+    expected_hashes = []
+    for chunk_idx, event in enumerate(stored):
+        assert isinstance(event, BlockStored)
+        expected_chunk_hashes = [
+            _wire_hash(_hash(i))
+            for i in range(
+                chunk_idx * block_size_factor,
+                (chunk_idx + 1) * block_size_factor,
+            )
+        ]
+        assert event.block_hashes == expected_chunk_hashes
+        assert event.block_size == block_size
+        assert len(event.token_ids) == block_size * block_size_factor
+        if chunk_idx == 0:
+            assert event.parent_block_hash is None
+        else:
+            assert event.parent_block_hash == _wire_hash(_hash(block_size_factor - 1))
+        expected_hashes.extend(expected_chunk_hashes)
+
+    assert len(tracker._pending_event_metadata) == 2
+
+    removed = list(tracker.take_events([_removed_event(keys)]))
+    assert len(removed) == 1
+    assert isinstance(removed[0], BlockRemoved)
+    assert removed[0].block_hashes == expected_hashes
+    assert removed[0].medium == _CPU_MEDIUM
+    assert removed[0].group_idx == 0
+    assert not tracker._pending_event_metadata
+
+
+def test_take_events_factor_gt_1_store_is_order_independent():
+    block_size_factor = 3
+    tracker = _tracker()
+    group_config = _group_config(block_size_factor=block_size_factor)
+    req = _request(
+        block_hashes=[_hash(i) for i in range(6)],
+        token_count=4 * block_size_factor * 2,
+    )
+    keys = _record_chunks(tracker, req, group_config, num_chunks=2)
+    unknown_key = make_offload_key(_hash(12345), 0)
+
+    events = list(tracker.take_events([_stored_event([keys[1], unknown_key, keys[0]])]))
+
+    assert len(events) == 3
+    chunk1, placeholder, chunk0 = events
+    assert [len(event.block_hashes) for event in events] == [3, 1, 3]
+    assert placeholder.block_size == 0
+    assert placeholder.token_ids == []
+    assert chunk0.parent_block_hash is None
+    assert chunk1.parent_block_hash == chunk0.block_hashes[-1]
+
+
+def test_take_events_opt_out_keeps_placeholders():
+    tracker = _tracker(self_describing_kv_events=False)
+    group_config = _group_config()
+    req = _request(block_hashes=[_hash(i) for i in range(3)], token_count=12)
+    keys = _record_chunks(tracker, req, group_config, num_chunks=3)
+
+    assert not tracker.self_describing_enabled
+    assert not tracker._pending_event_metadata
+
+    events = list(
+        tracker.take_events(
+            [
+                _stored_event(keys),
+                _removed_event(keys),
+            ]
+        )
+    )
+    assert len(events) == 4
+    for event in events[:3]:
+        assert isinstance(event, BlockStored)
+        assert event.block_size == 0
+        assert event.token_ids == []
+        assert event.parent_block_hash is None
+    assert isinstance(events[3], BlockRemoved)
+    assert len(events[3].block_hashes) == 3
+
+
+def test_record_store_skips_sliding_window_group():
+    tracker = _tracker()
+    group_config = _group_config(sliding_window_size_in_blocks=2)
+    req = _request(block_hashes=[_hash(i) for i in range(3)], token_count=12)
+    keys = _record_chunks(tracker, req, group_config, num_chunks=3)
+
+    assert not tracker._pending_event_metadata
+
+    events = list(tracker.take_events([_stored_event(keys[:1])]))
+    assert len(events) == 1
+    assert isinstance(events[0], BlockStored)
+    assert events[0].block_size == 0
+
+
+def test_take_events_groups_removed_hashes_by_kv_group():
+    tracker = _tracker()
+    group0_config = _group_config(group_idx=0, block_size_factor=2)
+    group1_config = _group_config(group_idx=1, block_size_factor=2)
+    req0 = _request(block_hashes=[_hash(0), _hash(1)], token_count=8)
+    req1 = _request(block_hashes=[_hash(10), _hash(11)], token_count=8)
+    key0 = _record_chunks(tracker, req0, group0_config, num_chunks=1)[0]
+    key1 = _record_chunks(tracker, req1, group1_config, num_chunks=1)[0]
+
+    removed = list(tracker.take_events([_removed_event([key0, key1])]))
+
+    assert len(removed) == 2
+    by_group = {event.group_idx: event.block_hashes for event in removed}
+    assert by_group == {
+        0: [_wire_hash(_hash(0)), _wire_hash(_hash(1))],
+        1: [_wire_hash(_hash(10)), _wire_hash(_hash(11))],
+    }
+
+
+def test_take_events_supports_restore_after_eviction():
+    block_size = 4
+    tracker = _tracker()
+    group_config = _group_config(block_size=block_size)
+    req = _request(block_hashes=[_hash(0)], token_count=block_size)
+    key = _record_chunks(tracker, req, group_config, num_chunks=1)[0]
+
+    first_store = list(tracker.take_events([_stored_event([key])]))
+    assert len(first_store) == 1
+    assert isinstance(first_store[0], BlockStored)
+    assert first_store[0].token_ids == [1, 2, 3, 4]
+
+    removed = list(tracker.take_events([_removed_event([key])]))
+    assert len(removed) == 1
+    assert isinstance(removed[0], BlockRemoved)
+    assert not tracker._pending_event_metadata
+
+    req.all_token_ids = [5, 6, 7, 8]
+    tracker.record_store(req, group_config, offload_block_idx=0, offload_key=key)
+
+    second_store = list(tracker.take_events([_stored_event([key])]))
+    assert len(second_store) == 1
+    assert isinstance(second_store[0], BlockStored)
+    assert second_store[0].token_ids == [5, 6, 7, 8]
+
+
+def test_reset_cache_clears_side_table():
+    tracker = _tracker()
+    group_config = _group_config()
+    req = _request(block_hashes=[_hash(i) for i in range(3)], token_count=12)
+    _record_chunks(tracker, req, group_config, num_chunks=3)
+
+    assert tracker._pending_event_metadata
+
+    tracker.reset()
+
+    assert not tracker._pending_event_metadata
+
+
+def test_tiering_rejects_self_describing_kv_events():
+    vllm_config = create_vllm_config(
+        block_size=4,
+        max_num_batched_tokens=16,
+        disable_hybrid_kv_cache_manager=False,
+    )
+    vllm_config.kv_transfer_config = KVTransferConfig(
+        kv_connector="OffloadingConnector",
+        kv_role="kv_both",
+        kv_connector_extra_config={
+            "spec_name": "TieringOffloadingSpec",
+            "cpu_bytes_to_use": 1 << 20,
+            "self_describing_kv_events": True,
+            "secondary_tiers": [{"type": "example"}],
+        },
+    )
+    vllm_config.kv_events_config = KVEventsConfig(
+        enable_kv_cache_events=True,
+        publisher="null",
+    )
+    kv_cache_config = KVCacheConfig(
+        num_blocks=0,
+        kv_cache_tensors=[],
+        kv_cache_groups=[
+            KVCacheGroupSpec(
+                ["layer"],
+                FullAttentionSpec(
+                    block_size=4,
+                    num_kv_heads=1,
+                    head_size=1,
+                    dtype=torch.float32,
+                ),
+            )
+        ],
+    )
+
+    with pytest.raises(ValueError, match="TieringOffloadingSpec"):
+        TieringOffloadingSpec(vllm_config, kv_cache_config)

tests/v1/kv_connector/unit/offloading_connector/test_scheduler.py

@@ -1,6 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from collections.abc import Iterable
 from types import SimpleNamespace
 from unittest.mock import MagicMock
 
@@ -12,19 +11,16 @@ from tests.v1.kv_connector.unit.offloading_connector.utils import (
     to_keys,
 )
 from tests.v1.kv_connector.unit.utils import EOS_TOKEN_ID
-from vllm.distributed.kv_events import BlockRemoved, BlockStored
 from vllm.distributed.kv_transfer.kv_connector.v1.offloading.scheduler import (
     OffloadingConnectorScheduler,
     RequestOffloadState,
 )
-from vllm.v1.core.kv_cache_utils import BlockHash
 from vllm.v1.kv_cache_interface import (
     FullAttentionSpec,
     KVCacheGroupSpec,
     SlidingWindowSpec,
 )
 from vllm.v1.kv_offload.base import (
-    OffloadingEvent,
     OffloadingManager,
     OffloadPolicy,
     ReqContext,
@@ -146,31 +142,6 @@ def test_offloading_connector(request_runner, async_scheduling: bool):
     runner.connector_scheduler._maximal_prefix_lookup = lambda key, req_context: 1
     runner.run(decoded_tokens=[EOS_TOKEN_ID], expected_loaded=(3, 4, 5))
 
-    # test take_events
-    def to_hashes(int_hashes: list[int]) -> list[BlockHash]:
-        return [BlockHash(str(i).encode()) for i in int_hashes]
-
-    def take_events() -> Iterable[OffloadingEvent]:
-        yield OffloadingEvent(keys=to_keys([1, 2, 3]), medium="A", removed=False)
-        yield OffloadingEvent(keys=to_keys([4, 5, 6]), medium="B", removed=True)
-
-    runner.manager.take_events.side_effect = take_events
-    events = list(runner.scheduler_connector.take_events())
-    assert len(events) == 2
-    event = events[0]
-    assert isinstance(event, BlockStored)
-    assert event.block_hashes == to_hashes([1, 2, 3])
-    assert event.block_size == 0
-    assert event.medium == "A"
-    assert event.token_ids == []
-    assert event.parent_block_hash is None
-    assert event.lora_id is None
-    assert event.lora_name is None
-    event = events[1]
-    assert isinstance(event, BlockRemoved)
-    assert event.block_hashes == to_hashes([4, 5, 6])
-    assert event.medium == "B"
-
 
 @pytest.mark.parametrize("async_scheduling", [True, False])
 def test_request_preemption(request_runner, async_scheduling: bool):

tests/v1/kv_connector/unit/offloading_connector/utils.py

@@ -14,7 +14,12 @@ from tests.v1.kv_connector.unit.utils import (
     create_vllm_config,
 )
 from vllm import SamplingParams
-from vllm.config import KVTransferConfig, VllmConfig, set_current_vllm_config
+from vllm.config import (
+    KVEventsConfig,
+    KVTransferConfig,
+    VllmConfig,
+    set_current_vllm_config,
+)
 from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorRole
 from vllm.distributed.kv_transfer.kv_connector.v1.offloading.common import (
     OffloadingConnectorMetadata,
@@ -198,6 +203,9 @@ class RequestRunner:
             "spec_module_path": "tests.v1.kv_connector.unit.offloading_connector.utils",  # noqa: E501
             # Preserve legacy behavior for tests; new opt-in tests override.
             "offload_prompt_only": False,
+            # Exercise the self-describing KV events path by default;
+            # opt-out tests override this to cover the legacy placeholders.
+            "self_describing_kv_events": True,
         }
         if block_size_factor > 1:
             extra_config["block_size"] = block_size * block_size_factor
@@ -209,6 +217,13 @@ class RequestRunner:
             kv_role="kv_both",
             kv_connector_extra_config=extra_config,
         )
+        vllm_config.kv_events_config = KVEventsConfig(
+            # Enable so the offloading events tracker is active, but use the
+            # null publisher: these tests drain take_events directly and a
+            # real ZMQ publisher would bind a port per test.
+            enable_kv_cache_events=True,
+            publisher="null",
+        )
 
         if kv_cache_groups is None:
             kv_cache_groups = [

tests/v1/kv_connector/unit/test_multi_connector.py

@@ -219,6 +219,7 @@ def test_multi_example_connector_consistency():
         enforce_eager=True,
         gpu_memory_utilization=0.5,
         kv_transfer_config=kv_transfer_config,
+        async_scheduling=False,
     )
     # Run generation - this should trigger saving KV cache
     # Use a single prompt to avoid race conditions depending on the order of scheduling

tests/v1/kv_connector/unit/test_offloading_connector.py

@@ -138,6 +138,10 @@ def _wait_for_prefix_cache_reset(llm: LLM) -> None:
 
 
 def _latency_test(llm: LLM, subscriber: MockSubscriber | None):
+    # TODO: Reintroduce latency test on ROCm once MRV2 supports cross
+    # layer KV Cache. See https://github.com/vllm-project/vllm/pull/45947
+    if current_platform.is_rocm():
+        return
     sampling_params = SamplingParams(max_tokens=1)
 
     num_times_cpu_better_than_cold = 0

tests/v1/kv_offload/cpu/test_manager.py

@@ -294,25 +294,25 @@ def test_cpu_manager():
     # prepare store with no space ([2, 3] is being loaded)
     assert cpu_manager.prepare_store(to_keys([6, 7, 8]), _EMPTY_REQ_CTX) is None
 
-    # complete load [2, 3]
+    # complete load [2, 3]. Load changes the eviction list, making 2, 3 recent.
     cpu_manager.complete_load(to_keys([2, 3]), _EMPTY_REQ_CTX)
 
-    # prepare store [6, 7, 8] -> evicts [2, 3, 4] (oldest)
+    # prepare store [6, 7, 8] -> evicts [4, 5, 2] (oldest)
     prepare_store_output = cpu_manager.prepare_store(to_keys([6, 7, 8]), _EMPTY_REQ_CTX)
     verify_store_output(
         prepare_store_output,
         ExpectedPrepareStoreOutput(
             keys_to_store=[6, 7, 8],
-            store_block_ids=[3, 2, 1],
-            evicted_keys=[2, 3, 4],
+            store_block_ids=[1, 0, 3],
+            evicted_keys=[4, 5, 2],
         ),
     )
 
     # complete store [6, 7, 8]
     cpu_manager.complete_store(to_keys([6, 7, 8]), _EMPTY_REQ_CTX)
 
-    # touch [5, 6, 7] (move to end of LRU order)
-    cpu_manager.touch(to_keys([5, 6, 7]), _EMPTY_REQ_CTX)
+    # touch [3, 6, 7] (move to end of LRU order)
+    cpu_manager.touch(to_keys([3, 6, 7]), _EMPTY_REQ_CTX)
 
     # prepare store [7, 9] -> evicts [8] (oldest following previous touch)
     prepare_store_output = cpu_manager.prepare_store(to_keys([9]), _EMPTY_REQ_CTX)
@@ -320,7 +320,7 @@ def test_cpu_manager():
         prepare_store_output,
         ExpectedPrepareStoreOutput(
             keys_to_store=[9],
-            store_block_ids=[1],
+            store_block_ids=[3],
             evicted_keys=[8],
         ),
     )
@@ -335,7 +335,7 @@ def test_cpu_manager():
     verify_events(
         cpu_manager.take_events(),
         expected_stores=({3, 4, 5}, {6, 7, 8}),
-        expected_evictions=({2, 3, 4}, {8}),
+        expected_evictions=({4, 5, 2}, {8}),
     )
 
 

tests/v1/kv_offload/tiering/test_tiering_offloading.py

@@ -295,6 +295,8 @@ class TestTieringOffloadingManager:
         self.manager.prepare_store(blocks, _CTX)
         self.manager.complete_store(blocks, _CTX, success=True)
         self._simulate_on_schedule_end()
+        # for secondary tiers to drain jobs, so primary tier's blocks are evictable.
+        self._simulate_on_schedule_end()
 
         self.secondary_tier1.touch = MagicMock(wraps=self.secondary_tier1.touch)
         self.secondary_tier2.touch = MagicMock(wraps=self.secondary_tier2.touch)
@@ -303,7 +305,7 @@ class TestTieringOffloadingManager:
         self.manager.touch(blocks, _CTX)
 
         # Verify touch was called on primary tier (check LRU order)
-        primary_keys = list(self.primary_tier._policy.blocks.keys())
+        primary_keys = list(self.primary_tier._policy.evictable_blocks.keys())
         assert primary_keys[-3:] == list(reversed(blocks))
 
         # Verify touch was propagated to all secondary tiers

vllm/distributed/kv_transfer/kv_connector/v1/offloading/events.py

@@ -0,0 +1,286 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Self-describing KV cache events for the offloading connector.
+
+The OffloadingManager identifies an offloaded chunk only by its OffloadKey,
+so its raw events carry no token ids, parent hash, or block size.
+:class:`OffloadingEventsTracker` snapshots each chunk's full ``BlockStored``
+payload while the ``Request`` is alive and publishes stores as block-granular
+payloads: a chunk event may carry multiple constituent per-block hashes, and
+evictions fan out to the same hashes. Chunks overlapping a non-chunk-aligned
+shared prefix re-announce the shared hashes once per chunk; consumers are
+expected to deduplicate (reference-count) repeated store/remove announcements
+of the same hash. Opt-in via
+``kv_connector_extra_config["self_describing_kv_events"]``; inert unless
+KV cache events are enabled. See the PR description for the full design.
+"""
+
+from collections.abc import Iterable
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, NamedTuple
+
+from vllm.distributed.kv_events import BlockRemoved, BlockStored, KVCacheEvent
+from vllm.logger import init_logger
+from vllm.v1.core.kv_cache_utils import BlockHash, maybe_convert_block_hash
+from vllm.v1.kv_cache_interface import (
+    KVCacheGroupSpec,
+    get_kv_cache_spec_kind,
+    get_kv_cache_spec_sliding_window,
+)
+from vllm.v1.kv_offload.base import (
+    OffloadingEvent,
+    OffloadingKVEventsConfig,
+    OffloadKey,
+    get_offload_block_hash,
+    get_offload_group_idx,
+)
+from vllm.v1.request import Request
+
+if TYPE_CHECKING:
+    from vllm.distributed.kv_transfer.kv_connector.v1.offloading.scheduler import (
+        GroupOffloadConfig,
+    )
+
+logger = init_logger(__name__)
+
+
+class OffloadingEventGroupSpec(NamedTuple):
+    kv_cache_spec_kind: str | None
+    kv_cache_spec_sliding_window: int | None
+
+
+def get_offloading_event_group_spec(
+    kv_cache_group: KVCacheGroupSpec,
+) -> OffloadingEventGroupSpec:
+    kv_cache_spec = kv_cache_group.kv_cache_spec
+    return OffloadingEventGroupSpec(
+        kv_cache_spec_kind=get_kv_cache_spec_kind(kv_cache_spec).value,
+        kv_cache_spec_sliding_window=get_kv_cache_spec_sliding_window(kv_cache_spec),
+    )
+
+
+@dataclass(slots=True)
+class _OffloadEventMetadata:
+    """BlockStored payload snapshot for one OffloadKey, captured at store
+    time and kept until the matching eviction event. ``medium`` is forwarded
+    from the OffloadingEvent."""
+
+    # The chunk's constituent block hashes; the last one is the OffloadKey.
+    block_hashes: tuple[BlockHash, ...]
+    parent_block_hash: BlockHash | None
+    token_ids: tuple[int, ...]
+    block_size: int
+    lora_id: int | None
+    lora_name: str | None
+    # Deferred: needs the same incremental curr_mm_idx handling as GPU events.
+    extra_keys: tuple[tuple[Any, ...] | None, ...] | None
+    group_idx: int
+    kv_cache_spec: OffloadingEventGroupSpec
+
+
+class OffloadingEventsTracker:
+    """Tracks offloaded chunks' KV event payloads from store to eviction.
+
+    The scheduler calls :meth:`record_store` from ``_build_store_jobs``
+    while the ``Request`` is available, and routes the manager's raw
+    :class:`OffloadingEvent` stream through :meth:`take_events`. All state
+    is bounded by the CPU pool capacity and cleared by :meth:`reset`.
+    """
+
+    def __init__(self, config: OffloadingKVEventsConfig):
+        self.config = config
+        self.self_describing_enabled = (
+            config.enable_kv_cache_events and config.self_describing_kv_events
+        )
+
+        # OffloadKey -> payload snapshot, kept until the eviction event so
+        # BlockRemoved can fan out. Bounded: one entry per offloaded chunk.
+        self._pending_event_metadata: dict[OffloadKey, _OffloadEventMetadata] = {}
+
+    def record_store(
+        self,
+        req: Request,
+        group_config: "GroupOffloadConfig",
+        offload_block_idx: int,
+        offload_key: OffloadKey,
+    ) -> None:
+        """Snapshot the KV cache event payload for one offloaded chunk.
+
+        No-op when self-describing event capture is disabled or for
+        sliding-window / SSM groups, which keep the legacy placeholder payload.
+        """
+        if not self.self_describing_enabled:
+            return
+        if group_config.sliding_window_size_in_blocks is not None:
+            return
+        meta = self._build_event_metadata(req, group_config, offload_block_idx)
+        self._pending_event_metadata[offload_key] = meta
+
+    def take_events(self, events: Iterable[OffloadingEvent]) -> Iterable[KVCacheEvent]:
+        """Translate raw OffloadingEvents into self-describing KV events.
+
+        Complete metadata is available only for full-attention groups when
+        the tracker is enabled. Other shapes retain the legacy placeholder
+        payload so consumers can ignore them.
+
+        Yields:
+            ``BlockStored`` or ``BlockRemoved`` events corresponding to
+            the underlying :class:`OffloadingEvent` stream.
+        """
+        for event in events:
+            if event.removed:
+                yield from self._take_removed_event(event)
+            else:
+                yield from self._take_stored_event(event)
+
+    def reset(self) -> None:
+        """Drop all tracked state; pending payloads are stale after a
+        manager cache reset."""
+        self._pending_event_metadata.clear()
+
+    def _build_event_metadata(
+        self,
+        req: Request,
+        group_config: "GroupOffloadConfig",
+        offload_block_idx: int,
+    ) -> _OffloadEventMetadata:
+        """Build the payload snapshot for one offloaded chunk: its
+        constituent per-block hashes, the whole chunk's tokens, and the
+        per-block ``block_size``."""
+        hbf = group_config.hash_block_size_factor
+        assert hbf > 0
+        assert offload_block_idx >= 0
+        # per-block token count (= the GPU/hash block size)
+        sub_block_size = group_config.offloaded_block_size // hbf
+        # chunk c covers hash-blocks [c*hbf, (c+1)*hbf); its tail block's hash
+        # is the chunk's OffloadKey.
+        first_hash_idx = offload_block_idx * hbf
+        last_hash_idx = first_hash_idx + hbf
+        assert first_hash_idx >= 0
+        assert last_hash_idx <= len(req.block_hashes)
+        chunk_hashes: list[BlockHash] = []
+        for block_hash in req.block_hashes[first_hash_idx:last_hash_idx]:
+            assert block_hash is not None
+            chunk_hashes.append(block_hash)
+        assert len(chunk_hashes) == hbf
+
+        if group_config.sliding_window_size_in_blocks is not None:
+            # record_store filters these out before calling this helper.
+            raise AssertionError("self-describing events only support full attention")
+
+        parent_block_hash: BlockHash | None
+        if first_hash_idx == 0:
+            parent_block_hash = None
+        else:
+            parent_block_hash = req.block_hashes[first_hash_idx - 1]
+            assert parent_block_hash is not None
+
+        tok_start = offload_block_idx * group_config.offloaded_block_size
+        tok_end = tok_start + group_config.offloaded_block_size
+        assert tok_end <= len(req.all_token_ids)
+        token_ids = tuple(req.all_token_ids[tok_start:tok_end])
+
+        lora_id: int | None = None
+        lora_name: str | None = None
+        if req.lora_request is not None:
+            lora_id = req.lora_request.adapter_id
+            lora_name = req.lora_request.name
+
+        return _OffloadEventMetadata(
+            block_hashes=tuple(chunk_hashes),
+            parent_block_hash=parent_block_hash,
+            token_ids=token_ids,
+            block_size=sub_block_size,
+            lora_id=lora_id,
+            lora_name=lora_name,
+            extra_keys=None,
+            group_idx=group_config.group_idx,
+            kv_cache_spec=group_config.kv_event_group_spec,
+        )
+
+    def _placeholder_stored(self, key: OffloadKey, medium: str) -> BlockStored:
+        return BlockStored(
+            block_hashes=[
+                maybe_convert_block_hash(BlockHash(get_offload_block_hash(key)))
+            ],
+            parent_block_hash=None,
+            token_ids=[],
+            lora_id=None,
+            block_size=0,
+            medium=medium,
+            lora_name=None,
+            group_idx=get_offload_group_idx(key),
+        )
+
+    def _take_stored_event(self, event: OffloadingEvent) -> Iterable[KVCacheEvent]:
+        # Metadata is read, NOT popped: the entry must survive until the
+        # eviction event so BlockRemoved can fan out to the same hashes.
+        # Events are self-contained (own parent), so key order is free.
+        for key in event.keys:
+            meta = self._pending_event_metadata.get(key)
+            if meta is None:
+                if self.self_describing_enabled:
+                    # Expected for unsupported shapes; warn once only.
+                    logger.warning_once(
+                        "OffloadingEventsTracker: no event metadata for "
+                        "offload key during BlockStored emission; emitting a "
+                        "placeholder payload. Expected for non-full-attention "
+                        "groups; otherwise indicates a missing populate path."
+                    )
+                yield self._placeholder_stored(key, event.medium)
+                continue
+
+            yield BlockStored(
+                block_hashes=list(
+                    maybe_convert_block_hash(h) for h in meta.block_hashes
+                ),
+                parent_block_hash=(
+                    maybe_convert_block_hash(meta.parent_block_hash)
+                    if meta.parent_block_hash is not None
+                    else None
+                ),
+                token_ids=list(meta.token_ids),
+                block_size=meta.block_size,
+                lora_id=meta.lora_id,
+                medium=event.medium,
+                lora_name=meta.lora_name,
+                extra_keys=(
+                    list(meta.extra_keys) if meta.extra_keys is not None else None
+                ),
+                group_idx=meta.group_idx,
+                kv_cache_spec_kind=meta.kv_cache_spec.kv_cache_spec_kind,
+                kv_cache_spec_sliding_window=(
+                    meta.kv_cache_spec.kv_cache_spec_sliding_window
+                ),
+            )
+
+    def _take_removed_event(self, event: OffloadingEvent) -> Iterable[KVCacheEvent]:
+        # Keep group_idx unambiguous if a manager batch spans groups.
+        by_group: dict[int, list] = {}
+        for key in event.keys:
+            meta = self._pending_event_metadata.pop(key, None)
+            if meta is not None:
+                group_idx = meta.group_idx
+                by_group.setdefault(group_idx, []).extend(
+                    maybe_convert_block_hash(h) for h in meta.block_hashes
+                )
+            else:
+                if self.self_describing_enabled:
+                    logger.warning_once(
+                        "OffloadingEventsTracker: no event metadata for "
+                        "offload key during BlockRemoved emission; emitting a "
+                        "placeholder removal. Expected if the matching store "
+                        "used the legacy placeholder payload; otherwise "
+                        "indicates missing store metadata."
+                    )
+                group_idx = get_offload_group_idx(key)
+                by_group.setdefault(group_idx, []).append(
+                    maybe_convert_block_hash(BlockHash(get_offload_block_hash(key)))
+                )
+
+        for group_idx, hashes in by_group.items():
+            yield BlockRemoved(
+                block_hashes=hashes,
+                medium=event.medium,
+                group_idx=group_idx,
+            )

vllm/distributed/kv_transfer/kv_connector/v1/offloading/scheduler.py

@@ -5,7 +5,7 @@ from dataclasses import dataclass, field
 from itertools import islice
 from typing import Any, NamedTuple
 
-from vllm.distributed.kv_events import BlockRemoved, BlockStored, KVCacheEvent
+from vllm.distributed.kv_events import KVCacheEvent
 from vllm.distributed.kv_transfer.kv_connector.utils import yield_req_data
 from vllm.distributed.kv_transfer.kv_connector.v1.base import KVConnectorMetadata
 from vllm.distributed.kv_transfer.kv_connector.v1.offloading.common import (
@@ -14,6 +14,11 @@ from vllm.distributed.kv_transfer.kv_connector.v1.offloading.common import (
     ReqId,
     TransferJob,
 )
+from vllm.distributed.kv_transfer.kv_connector.v1.offloading.events import (
+    OffloadingEventGroupSpec,
+    OffloadingEventsTracker,
+    get_offloading_event_group_spec,
+)
 from vllm.distributed.kv_transfer.kv_connector.v1.offloading.metrics import (
     OffloadingConnectorStats,
     _TransferMetricName,
@@ -36,7 +41,6 @@ from vllm.v1.kv_offload.base import (
     OffloadPolicy,
     ReqContext,
     RequestOffloadingContext,
-    get_offload_block_hash,
     make_offload_key,
 )
 from vllm.v1.outputs import KVConnectorOutput
@@ -69,6 +73,9 @@ class GroupOffloadConfig(NamedTuple):
     gpu_block_size: int
     offloaded_block_size: int
     hash_block_size_factor: int
+    # KV cache spec metadata propagated onto emitted BlockStored events so
+    # KV-aware consumers can classify and filter the group.
+    kv_event_group_spec: OffloadingEventGroupSpec
     # None below means full attention
     sliding_window_size_in_blocks: int | None
     # Number of this group's offloaded blocks per full-attention alignment
@@ -200,6 +207,9 @@ class SchedulerOffloadConfig(NamedTuple):
                     alignment_block_count=_alignment_block_count(
                         gpu_block_size * spec.block_size_factor, sw
                     ),
+                    kv_event_group_spec=get_offloading_event_group_spec(
+                        spec.kv_cache_config.kv_cache_groups[idx]
+                    ),
                     is_eagle_group=idx in eagle_groups,
                 )
                 for idx, gpu_block_size in enumerate(spec.gpu_block_size)
@@ -361,6 +371,8 @@ class OffloadingConnectorScheduler:
         # be freed before a request finishes).
         self._block_id_to_pending_jobs: dict[int, set[int]] = {}
 
+        self._events_tracker = OffloadingEventsTracker(spec.kv_events_config)
+
     def _generate_job_id(self) -> int:
         job_id = self._job_counter
         self._job_counter += 1
@@ -934,6 +946,11 @@ class OffloadingConnectorScheduler:
                         continue
 
                     offloaded_block_idx = start_block_idx + idx
+
+                    self._events_tracker.record_store(
+                        req, group_config, offloaded_block_idx, offload_key
+                    )
+
                     gpu_block_idx = offloaded_block_idx * block_size_factor
                     for i in range(block_size_factor):
                         block_id = block_ids[gpu_block_idx + i]
@@ -1184,25 +1201,17 @@ class OffloadingConnectorScheduler:
         return False, None
 
     def take_events(self) -> Iterable[KVCacheEvent]:
-        """Take the KV cache events from the connector.
+        """Drain pending KV cache events.
 
-        Returns:
-            A list of KV cache events.
+        Complete metadata is available only when self-describing KV events
+        are enabled, and only for full-attention groups. Other shapes retain
+        the previous placeholder payload so consumers can ignore them.
+
+        Yields:
+            ``BlockStored`` or ``BlockRemoved`` events corresponding to
+            the underlying :class:`OffloadingEvent` stream.
         """
-        for event in self.manager.take_events():
-            block_hashes = [get_offload_block_hash(key) for key in event.keys]
-            if event.removed:
-                yield BlockRemoved(block_hashes=block_hashes, medium=event.medium)
-            else:
-                yield BlockStored(
-                    block_hashes=block_hashes,
-                    parent_block_hash=None,
-                    token_ids=[],
-                    lora_id=None,
-                    block_size=0,
-                    medium=event.medium,
-                    lora_name=None,
-                )
+        yield from self._events_tracker.take_events(self.manager.take_events())
 
     def reset_cache(self) -> None:
         """Reset the offloading manager cache, evicting all stored blocks."""
@@ -1238,6 +1247,10 @@ class OffloadingConnectorScheduler:
         self._jobs.clear()
         self._block_id_to_pending_jobs.clear()
 
+        # The manager pool is empty; pending event payloads and announced
+        # reference counts are stale.
+        self._events_tracker.reset()
+
         # Note: _current_batch_jobs_to_flush is intentionally NOT cleared.
         # The load flush IDs collected above must be delivered to workers.
         if self._blocks_being_loaded is not None:

vllm/entrypoints/anthropic/api_router.py

@@ -61,7 +61,7 @@ def translate_error_response(response: ErrorResponse) -> JSONResponse:
 async def create_messages(request: AnthropicMessagesRequest, raw_request: Request):
     handler = messages(raw_request)
     if handler is None:
-        base_server = raw_request.app.state.openai_serving_tokenization
+        base_server = raw_request.app.state.serving_tokenization
         error = base_server.create_error_response(
             NotImplementedError("The model does not support Messages API")
         )
@@ -107,7 +107,7 @@ async def create_messages(request: AnthropicMessagesRequest, raw_request: Reques
 async def count_tokens(request: AnthropicCountTokensRequest, raw_request: Request):
     handler = messages(raw_request)
     if handler is None:
-        base_server = raw_request.app.state.openai_serving_tokenization
+        base_server = raw_request.app.state.serving_tokenization
         error = base_server.create_error_response(
             NotImplementedError("The model does not support Messages API")
         )

vllm/entrypoints/openai/api_server.py

@@ -34,7 +34,7 @@ from vllm.entrypoints.openai.models.serving import OpenAIServingModels
 from vllm.entrypoints.serve.elastic_ep.middleware import ScalingMiddleware
 from vllm.entrypoints.serve.render.serving import OpenAIServingRender
 from vllm.entrypoints.serve.sagemaker.api_router import sagemaker_standards_bootstrap
-from vllm.entrypoints.serve.tokenize.serving import OpenAIServingTokenization
+from vllm.entrypoints.serve.tokenize.serving import ServingTokenization
 from vllm.entrypoints.serve.utils.api_utils import (
     cli_env_setup,
     log_non_default_args,
@@ -376,8 +376,7 @@ async def init_app_state(
         log_error_stack=args.log_error_stack,
     )
 
-    state.openai_serving_tokenization = OpenAIServingTokenization(
-        engine_client,
+    state.serving_tokenization = ServingTokenization(
         state.openai_serving_models,
         state.openai_serving_render,
         request_logger=request_logger,
@@ -461,9 +460,15 @@ async def init_render_app_state(
     )
 
     state.openai_serving_models = model_registry
-
-    # Expose tokenization via the render handler (no engine required).
-    state.openai_serving_tokenization = state.openai_serving_render
+    state.serving_tokenization = ServingTokenization(
+        model_registry,
+        state.openai_serving_render,
+        request_logger=request_logger,
+        chat_template=resolved_chat_template,
+        chat_template_content_format=args.chat_template_content_format,
+        default_chat_template_kwargs=args.default_chat_template_kwargs,
+        trust_request_chat_template=args.trust_request_chat_template,
+    )
 
     state.vllm_config = vllm_config
     # Disable stats logging — there is no engine to poll.

vllm/entrypoints/openai/engine/serving.py

@@ -5,25 +5,19 @@ import time
 from collections.abc import Awaitable, Mapping
 from dataclasses import dataclass, field
 from http import HTTPStatus
-from typing import Any, ClassVar, Generic, Protocol, TypeAlias, TypeVar
+from typing import ClassVar, Generic, TypeVar
 
 from fastapi import Request
 from pydantic import ConfigDict
 from starlette.datastructures import Headers
 
-import vllm.envs as envs
-from vllm.config import ModelConfig
 from vllm.engine.protocol import EngineClient
-from vllm.entrypoints.chat_utils import ChatTemplateContentFormatOption
 from vllm.entrypoints.generate.beam_search.online import BeamSearchOnlineMixin
 from vllm.entrypoints.openai.chat_completion.protocol import (
-    BatchChatCompletionRequest,
     ChatCompletionRequest,
-    ChatCompletionResponse,
 )
 from vllm.entrypoints.openai.completion.protocol import (
     CompletionRequest,
-    CompletionResponse,
 )
 from vllm.entrypoints.openai.engine.protocol import (
     ErrorResponse,
@@ -31,81 +25,22 @@ from vllm.entrypoints.openai.engine.protocol import (
 )
 from vllm.entrypoints.openai.models.serving import OpenAIServingModels
 from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
-from vllm.entrypoints.serve.disagg.protocol import GenerateRequest, GenerateResponse
-from vllm.entrypoints.serve.tokenize.protocol import (
-    DetokenizeRequest,
-    TokenizeChatRequest,
-    TokenizeCompletionRequest,
-    TokenizeResponse,
-)
-from vllm.entrypoints.serve.utils.error_response import create_error_response
+from vllm.entrypoints.serve.engine.serving import BaseServing
+from vllm.entrypoints.serve.engine.typing import AnyRequest
 from vllm.entrypoints.serve.utils.request_logger import RequestLogger
-from vllm.entrypoints.speech_to_text.transcription.protocol import (
-    TranscriptionRequest,
-    TranscriptionResponse,
-)
-from vllm.entrypoints.speech_to_text.translation.protocol import TranslationRequest
-from vllm.inputs import EngineInput, PromptType
+from vllm.inputs import EngineInput
 from vllm.logger import init_logger
 from vllm.logprobs import Logprob, PromptLogprobs
 from vllm.lora.request import LoRARequest
-from vllm.renderers import ChatParams, TokenizeParams
-from vllm.renderers.inputs.preprocess import (
-    extract_prompt_components,
-    extract_prompt_len,
-)
-from vllm.sampling_params import BeamSearchParams, SamplingParams
 from vllm.tokenizers import TokenizerLike
 from vllm.tracing import (
     contains_trace_headers,
     extract_trace_headers,
     log_tracing_disabled_warning,
 )
-from vllm.utils import random_uuid
 
 logger = init_logger(__name__)
 
-
-class RendererRequest(Protocol):
-    def build_tok_params(self, model_config: ModelConfig) -> TokenizeParams:
-        raise NotImplementedError
-
-
-class RendererChatRequest(RendererRequest, Protocol):
-    def build_chat_params(
-        self,
-        default_template: str | None,
-        default_template_content_format: ChatTemplateContentFormatOption,
-    ) -> ChatParams:
-        raise NotImplementedError
-
-
-CompletionLikeRequest: TypeAlias = (
-    CompletionRequest | TokenizeCompletionRequest | DetokenizeRequest
-)
-
-ChatLikeRequest: TypeAlias = (
-    ChatCompletionRequest | BatchChatCompletionRequest | TokenizeChatRequest
-)
-
-SpeechToTextRequest: TypeAlias = TranscriptionRequest | TranslationRequest
-
-AnyRequest: TypeAlias = (
-    CompletionLikeRequest
-    | ChatLikeRequest
-    | SpeechToTextRequest
-    | ResponsesRequest
-    | GenerateRequest
-)
-
-AnyResponse: TypeAlias = (
-    CompletionResponse
-    | ChatCompletionResponse
-    | TranscriptionResponse
-    | TokenizeResponse
-    | GenerateResponse
-)
-
 RequestT = TypeVar("RequestT", bound=AnyRequest)
 _T = TypeVar("_T")
 
@@ -122,7 +57,7 @@ class ServeContext(Generic[RequestT]):
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
 
-class OpenAIServing(BeamSearchOnlineMixin):
+class OpenAIServing(BaseServing, BeamSearchOnlineMixin):
     request_id_prefix: ClassVar[str] = """
     A short string prepended to every request’s ID.
     """
@@ -135,15 +70,14 @@ class OpenAIServing(BeamSearchOnlineMixin):
         request_logger: RequestLogger | None,
         return_tokens_as_token_ids: bool = False,
     ):
-        super().__init__()
+        super().__init__(
+            models=models,
+            model_config=engine_client.model_config,
+            request_logger=request_logger,
+        )
 
         self.engine_client = engine_client
-        self.models = models
-
-        self.request_logger = request_logger
         self.return_tokens_as_token_ids = return_tokens_as_token_ids
-
-        self.model_config = engine_client.model_config
         self.renderer = engine_client.renderer
         self.input_processor = engine_client.input_processor
         vllm_config = getattr(engine_client, "vllm_config", None)
@@ -163,15 +97,6 @@ class OpenAIServing(BeamSearchOnlineMixin):
             # Never fail server startup over the fingerprint.
             self.system_fingerprint = None
 
-    @staticmethod
-    def create_error_response(
-        message: str | Exception,
-        err_type: str = "BadRequestError",
-        status_code: HTTPStatus = HTTPStatus.BAD_REQUEST,
-        param: str | None = None,
-    ) -> ErrorResponse:
-        return create_error_response(message, err_type, status_code, param)
-
     def create_streaming_error_response(
         self,
         message: str | Exception,
@@ -208,167 +133,6 @@ class OpenAIServing(BeamSearchOnlineMixin):
             status_code=e.status_code,
         )
 
-    async def _check_model(
-        self,
-        request: AnyRequest,
-    ) -> ErrorResponse | None:
-        error_response = None
-
-        if self._is_model_supported(request.model):
-            return None
-        if request.model in self.models.lora_requests:
-            return None
-        if (
-            envs.VLLM_ALLOW_RUNTIME_LORA_UPDATING
-            and request.model
-            and (load_result := await self.models.resolve_lora(request.model))
-        ):
-            if isinstance(load_result, LoRARequest):
-                return None
-            if (
-                isinstance(load_result, ErrorResponse)
-                and load_result.error.code == HTTPStatus.BAD_REQUEST.value
-            ):
-                error_response = load_result
-
-        return error_response or self.create_error_response(
-            message=f"The model `{request.model}` does not exist.",
-            err_type="NotFoundError",
-            status_code=HTTPStatus.NOT_FOUND,
-            param="model",
-        )
-
-    def _get_active_default_mm_loras(self, request: AnyRequest) -> LoRARequest | None:
-        """Determine if there are any active default multimodal loras."""
-        # TODO: Currently this is only enabled for chat completions
-        # to be better aligned with only being enabled for .generate
-        # when run offline. It would be nice to support additional
-        # tasks types in the future.
-        message_types = self._get_message_types(request)
-        default_mm_loras = set()
-
-        for lora in self.models.lora_requests.values():
-            # Best effort match for default multimodal lora adapters;
-            # There is probably a better way to do this, but currently
-            # this matches against the set of 'types' in any content lists
-            # up until '_', e.g., to match audio_url -> audio
-            if lora.lora_name in message_types:
-                default_mm_loras.add(lora)
-
-        # Currently only support default modality specific loras if
-        # we have exactly one lora matched on the request.
-        if len(default_mm_loras) == 1:
-            return default_mm_loras.pop()
-        return None
-
-    def _maybe_get_adapters(
-        self,
-        request: AnyRequest,
-        supports_default_mm_loras: bool = False,
-    ) -> LoRARequest | None:
-        if request.model in self.models.lora_requests:
-            return self.models.lora_requests[request.model]
-
-        # Currently only support default modality specific loras
-        # if we have exactly one lora matched on the request.
-        if supports_default_mm_loras:
-            default_mm_lora = self._get_active_default_mm_loras(request)
-            if default_mm_lora is not None:
-                return default_mm_lora
-
-        if self._is_model_supported(request.model):
-            return None
-
-        # if _check_model has been called earlier, this will be unreachable
-        raise ValueError(f"The model `{request.model}` does not exist.")
-
-    def _get_message_types(self, request: AnyRequest) -> set[str]:
-        """Retrieve the set of types from message content dicts up
-        until `_`; we use this to match potential multimodal data
-        with default per modality loras.
-        """
-        message_types: set[str] = set()
-
-        if not hasattr(request, "messages"):
-            return message_types
-
-        messages = request.messages
-        if messages is None or isinstance(messages, (str, bytes)):
-            return message_types
-
-        for message in messages:
-            if (
-                isinstance(message, dict)
-                and "content" in message
-                and isinstance(message["content"], list)
-            ):
-                for content_dict in message["content"]:
-                    if "type" in content_dict:
-                        message_types.add(content_dict["type"].split("_")[0])
-        return message_types
-
-    def _validate_chat_template(
-        self,
-        request_chat_template: str | None,
-        chat_template_kwargs: dict[str, Any] | None,
-        trust_request_chat_template: bool,
-    ) -> ErrorResponse | None:
-        if not trust_request_chat_template and (
-            request_chat_template is not None
-            or (
-                chat_template_kwargs
-                and chat_template_kwargs.get("chat_template") is not None
-            )
-        ):
-            return self.create_error_response(
-                "Chat template is passed with request, but "
-                "--trust-request-chat-template is not set. "
-                "Refused request with untrusted chat template."
-            )
-        return None
-
-    @staticmethod
-    def _prepare_extra_chat_template_kwargs(
-        request_chat_template_kwargs: dict[str, Any] | None = None,
-        default_chat_template_kwargs: dict[str, Any] | None = None,
-    ) -> dict[str, Any]:
-        """Helper to merge server-default and request-specific chat template kwargs."""
-        request_chat_template_kwargs = request_chat_template_kwargs or {}
-        if default_chat_template_kwargs is None:
-            return request_chat_template_kwargs
-        # Apply server defaults first, then request kwargs override.
-        return default_chat_template_kwargs | request_chat_template_kwargs
-
-    def _extract_prompt_components(self, prompt: PromptType | EngineInput):
-        return extract_prompt_components(self.model_config, prompt)
-
-    def _extract_prompt_text(self, prompt: PromptType | EngineInput):
-        return self._extract_prompt_components(prompt).text
-
-    def _extract_prompt_len(self, prompt: EngineInput):
-        return extract_prompt_len(self.model_config, prompt)
-
-    def _log_inputs(
-        self,
-        request_id: str,
-        inputs: PromptType | EngineInput,
-        params: SamplingParams | BeamSearchParams | None,
-        lora_request: LoRARequest | None,
-    ) -> None:
-        if self.request_logger is None:
-            return
-
-        components = self._extract_prompt_components(inputs)
-
-        self.request_logger.log_inputs(
-            request_id,
-            components.text,
-            components.token_ids,
-            components.embeds,
-            params=params,
-            lora_request=lora_request,
-        )
-
     async def _get_trace_headers(
         self,
         headers: Headers,
@@ -383,18 +147,6 @@ class OpenAIServing(BeamSearchOnlineMixin):
 
         return None
 
-    @staticmethod
-    def _base_request_id(
-        raw_request: Request | None, default: str | None = None
-    ) -> str | None:
-        """Pulls the request id to use from a header, if provided"""
-        if raw_request is not None and (
-            (req_id := raw_request.headers.get("X-Request-Id")) is not None
-        ):
-            return req_id
-
-        return random_uuid() if default is None else default
-
     @staticmethod
     def _get_data_parallel_rank(raw_request: Request | None) -> int | None:
         """Pulls the data parallel rank from a header, if provided"""
@@ -464,13 +216,6 @@ class OpenAIServing(BeamSearchOnlineMixin):
 
         return tokenizer.decode([token_id])
 
-    def _is_model_supported(self, model_name: str | None) -> bool:
-        if not model_name:
-            return True
-        if envs.VLLM_SKIP_MODEL_NAME_VALIDATION:
-            return True
-        return self.models.is_base_model(model_name)
-
 
 def format_token_id_placeholder(token_id: int) -> str:
     return f"token_id:{token_id}"

vllm/entrypoints/openai/models/serving.py

@@ -42,6 +42,10 @@ class OpenAIModelRegistry:
     ) -> None:
         self.model_config = model_config
         self.base_model_paths = base_model_paths
+        self.lora_requests: dict[str, LoRARequest] = {}
+
+    def model_name(self, lora_request: LoRARequest | None = None) -> str:
+        return self.base_model_paths[0].name
 
     def is_base_model(self, model_name: str) -> bool:
         return any(model.name == model_name for model in self.base_model_paths)
@@ -72,6 +76,9 @@ class OpenAIModelRegistry:
             ]
         )
 
+    async def resolve_lora(self, lora_name: str):
+        raise RuntimeError("The OpenAIModelRegistry has no LoRA support.")
+
 
 class OpenAIServingModels:
     """Shared instance to hold data about the loaded base model(s) and adapters.

vllm/entrypoints/pooling/base/io_processor.py

@@ -12,7 +12,7 @@ from vllm.entrypoints.chat_utils import (
     ChatTemplateContentFormatOption,
     ConversationMessage,
 )
-from vllm.entrypoints.openai.engine.serving import RendererChatRequest, RendererRequest
+from vllm.entrypoints.serve.engine.typing import RendererChatRequest, RendererRequest
 from vllm.inputs import EngineInput, SingletonPrompt
 from vllm.renderers import BaseRenderer, TokenizeParams, merge_kwargs
 from vllm.renderers.inputs.preprocess import parse_model_prompt, prompt_to_seq

vllm/entrypoints/pooling/base/serving.py

@@ -52,7 +52,7 @@ class PoolingServingBase(ABC):
         self.engine_client = engine_client
         self.models = models
         self.model_config = models.model_config
-        self.renderer = models.renderer
+        self.renderer = engine_client.renderer
         self.vllm_config = engine_client.vllm_config
         self.max_model_len = self.model_config.max_model_len
         self.request_logger = request_logger
@@ -61,7 +61,7 @@ class PoolingServingBase(ABC):
         self.chat_template_config = chat_template_config
 
         # Shared thread pool executor for preprocessing and postprocessing.
-        self._executor: Executor = models.renderer._executor
+        self._executor: Executor = self.renderer._executor
         self._preprocessing_async = make_async(
             self._preprocessing, executor=self._executor
         )

vllm/entrypoints/serve/disagg/api_router.py

@@ -20,7 +20,7 @@ from vllm.entrypoints.serve.disagg.protocol import (
 from vllm.entrypoints.serve.disagg.serving import (
     ServingTokens,
 )
-from vllm.entrypoints.serve.tokenize.serving import OpenAIServingTokenization
+from vllm.entrypoints.serve.tokenize.serving import ServingTokenization
 from vllm.entrypoints.serve.utils.api_utils import (
     load_aware_call,
     validate_json_request,
@@ -31,8 +31,8 @@ from vllm.logger import init_logger
 logger = init_logger(__name__)
 
 
-def tokenization(request: Request) -> OpenAIServingTokenization:
-    return request.app.state.openai_serving_tokenization
+def tokenization(request: Request) -> ServingTokenization:
+    return request.app.state.serving_tokenization
 
 
 def generate_tokens(request: Request) -> ServingTokens | None:

vllm/entrypoints/serve/engine/serving.py

@@ -0,0 +1,193 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from http import HTTPStatus
+
+from fastapi import Request
+
+from vllm import PromptType, SamplingParams, envs
+from vllm.config import ModelConfig
+from vllm.entrypoints.openai.engine.protocol import ErrorResponse
+from vllm.entrypoints.openai.models.serving import (
+    OpenAIModelRegistry,
+    OpenAIServingModels,
+)
+from vllm.entrypoints.serve.engine.typing import AnyRequest
+from vllm.entrypoints.serve.utils.error_response import create_error_response
+from vllm.entrypoints.serve.utils.request_logger import RequestLogger
+from vllm.inputs import EngineInput
+from vllm.lora.request import LoRARequest
+from vllm.renderers.inputs.preprocess import (
+    extract_prompt_components,
+    extract_prompt_len,
+)
+from vllm.sampling_params import BeamSearchParams
+from vllm.utils import random_uuid
+
+
+class BaseServing:
+    def __init__(
+        self,
+        models: OpenAIServingModels | OpenAIModelRegistry,
+        model_config: ModelConfig,
+        request_logger: RequestLogger | None = None,
+    ):
+        self.models = models
+        self.model_config = model_config
+        self.request_logger = request_logger
+
+    async def _check_model(
+        self,
+        request: AnyRequest,
+    ) -> ErrorResponse | None:
+        error_response = None
+
+        if self._is_model_supported(request.model):
+            return None
+        if request.model in self.models.lora_requests:
+            return None
+        if (
+            envs.VLLM_ALLOW_RUNTIME_LORA_UPDATING
+            and request.model
+            and (load_result := await self.models.resolve_lora(request.model))
+        ):
+            if isinstance(load_result, LoRARequest):
+                return None
+            if (
+                isinstance(load_result, ErrorResponse)
+                and load_result.error.code == HTTPStatus.BAD_REQUEST.value
+            ):
+                error_response = load_result
+
+        return error_response or self.create_error_response(
+            message=f"The model `{request.model}` does not exist.",
+            err_type="NotFoundError",
+            status_code=HTTPStatus.NOT_FOUND,
+            param="model",
+        )
+
+    def _is_model_supported(self, model_name: str | None) -> bool:
+        if not model_name:
+            return True
+        if envs.VLLM_SKIP_MODEL_NAME_VALIDATION:
+            return True
+        return self.models.is_base_model(model_name)
+
+    @staticmethod
+    def create_error_response(
+        message: str | Exception,
+        err_type: str = "BadRequestError",
+        status_code: HTTPStatus = HTTPStatus.BAD_REQUEST,
+        param: str | None = None,
+    ) -> ErrorResponse:
+        return create_error_response(message, err_type, status_code, param)
+
+    def _extract_prompt_components(self, prompt: PromptType | EngineInput):
+        return extract_prompt_components(self.model_config, prompt)
+
+    def _extract_prompt_text(self, prompt: PromptType | EngineInput):
+        return self._extract_prompt_components(prompt).text
+
+    def _extract_prompt_len(self, prompt: EngineInput):
+        return extract_prompt_len(self.model_config, prompt)
+
+    def _log_inputs(
+        self,
+        request_id: str,
+        inputs: PromptType | EngineInput,
+        params: SamplingParams | BeamSearchParams | None,
+        lora_request: LoRARequest | None,
+    ) -> None:
+        if self.request_logger is None:
+            return
+
+        components = self._extract_prompt_components(inputs)
+
+        self.request_logger.log_inputs(
+            request_id,
+            components.text,
+            components.token_ids,
+            components.embeds,
+            params=params,
+            lora_request=lora_request,
+        )
+
+    @staticmethod
+    def _base_request_id(
+        raw_request: Request | None, default: str | None = None
+    ) -> str | None:
+        """Pulls the request id to use from a header, if provided"""
+        if raw_request is not None and (
+            (req_id := raw_request.headers.get("X-Request-Id")) is not None
+        ):
+            return req_id
+
+        return random_uuid() if default is None else default
+
+    def _get_message_types(self, request: AnyRequest) -> set[str]:
+        """Retrieve the set of types from message content dicts up
+        until `_`; we use this to match potential multimodal data
+        with default per modality loras.
+        """
+        message_types: set[str] = set()
+
+        if not hasattr(request, "messages"):
+            return message_types
+
+        messages = request.messages
+        if messages is None or isinstance(messages, (str, bytes)):
+            return message_types
+
+        for message in messages:
+            if (
+                isinstance(message, dict)
+                and "content" in message
+                and isinstance(message["content"], list)
+            ):
+                for content_dict in message["content"]:
+                    if "type" in content_dict:
+                        message_types.add(content_dict["type"].split("_")[0])
+        return message_types
+
+    def _get_active_default_mm_loras(self, request: AnyRequest) -> LoRARequest | None:
+        """Determine if there are any active default multimodal loras."""
+        # TODO: Currently this is only enabled for chat completions
+        # to be better aligned with only being enabled for .generate
+        # when run offline. It would be nice to support additional
+        # tasks types in the future.
+        message_types = self._get_message_types(request)
+        default_mm_loras = set()
+
+        for lora in self.models.lora_requests.values():
+            # Best effort match for default multimodal lora adapters;
+            # There is probably a better way to do this, but currently
+            # this matches against the set of 'types' in any content lists
+            # up until '_', e.g., to match audio_url -> audio
+            if lora.lora_name in message_types:
+                default_mm_loras.add(lora)
+
+        # Currently only support default modality specific loras if
+        # we have exactly one lora matched on the request.
+        if len(default_mm_loras) == 1:
+            return default_mm_loras.pop()
+        return None
+
+    def _maybe_get_adapters(
+        self,
+        request: AnyRequest,
+        supports_default_mm_loras: bool = False,
+    ) -> LoRARequest | None:
+        if request.model in self.models.lora_requests:
+            return self.models.lora_requests[request.model]
+
+        # Currently only support default modality specific loras
+        # if we have exactly one lora matched on the request.
+        if supports_default_mm_loras:
+            default_mm_lora = self._get_active_default_mm_loras(request)
+            if default_mm_lora is not None:
+                return default_mm_lora
+
+        if self._is_model_supported(request.model):
+            return None
+
+        # if _check_model has been called earlier, this will be unreachable
+        raise ValueError(f"The model `{request.model}` does not exist.")

vllm/entrypoints/serve/engine/typing.py

@@ -0,0 +1,71 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Protocol, TypeAlias
+
+from vllm.config import ModelConfig
+from vllm.entrypoints.chat_utils import ChatTemplateContentFormatOption
+from vllm.entrypoints.openai.chat_completion.protocol import (
+    BatchChatCompletionRequest,
+    ChatCompletionRequest,
+    ChatCompletionResponse,
+)
+from vllm.entrypoints.openai.completion.protocol import (
+    CompletionRequest,
+    CompletionResponse,
+)
+from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
+from vllm.entrypoints.serve.disagg.protocol import GenerateRequest, GenerateResponse
+from vllm.entrypoints.serve.tokenize.protocol import (
+    DetokenizeRequest,
+    TokenizeChatRequest,
+    TokenizeCompletionRequest,
+    TokenizeResponse,
+)
+from vllm.entrypoints.speech_to_text.transcription.protocol import (
+    TranscriptionRequest,
+    TranscriptionResponse,
+)
+from vllm.entrypoints.speech_to_text.translation.protocol import TranslationRequest
+from vllm.renderers import ChatParams, TokenizeParams
+
+
+class RendererRequest(Protocol):
+    def build_tok_params(self, model_config: ModelConfig) -> TokenizeParams:
+        raise NotImplementedError
+
+
+class RendererChatRequest(RendererRequest, Protocol):
+    def build_chat_params(
+        self,
+        default_template: str | None,
+        default_template_content_format: ChatTemplateContentFormatOption,
+    ) -> ChatParams:
+        raise NotImplementedError
+
+
+CompletionLikeRequest: TypeAlias = (
+    CompletionRequest | TokenizeCompletionRequest | DetokenizeRequest
+)
+
+ChatLikeRequest: TypeAlias = (
+    ChatCompletionRequest | BatchChatCompletionRequest | TokenizeChatRequest
+)
+
+SpeechToTextRequest: TypeAlias = TranscriptionRequest | TranslationRequest
+
+AnyRequest: TypeAlias = (
+    CompletionLikeRequest
+    | ChatLikeRequest
+    | SpeechToTextRequest
+    | ResponsesRequest
+    | GenerateRequest
+)
+
+AnyResponse: TypeAlias = (
+    CompletionResponse
+    | ChatCompletionResponse
+    | TranscriptionResponse
+    | TokenizeResponse
+    | GenerateResponse
+)

vllm/entrypoints/serve/instrumentator/basic.py

@@ -5,8 +5,7 @@ from fastapi import APIRouter, Request
 from fastapi.responses import JSONResponse
 
 from vllm.engine.protocol import EngineClient
-from vllm.entrypoints.openai.engine.serving import OpenAIServing
-from vllm.entrypoints.serve.tokenize.serving import OpenAIServingTokenization
+from vllm.entrypoints.serve.tokenize.serving import ServingTokenization
 from vllm.logger import init_logger
 from vllm.version import __version__ as VLLM_VERSION
 
@@ -15,13 +14,13 @@ router = APIRouter()
 logger = init_logger(__name__)
 
 
-def base(request: Request) -> OpenAIServing:
+def base(request: Request) -> ServingTokenization:
     # Reuse the existing instance
     return tokenization(request)
 
 
-def tokenization(request: Request) -> OpenAIServingTokenization:
-    return request.app.state.openai_serving_tokenization
+def tokenization(request: Request) -> ServingTokenization:
+    return request.app.state.serving_tokenization
 
 
 def engine_client(request: Request) -> EngineClient:

vllm/entrypoints/serve/tokenize/api_router.py

@@ -9,16 +9,14 @@ from fastapi.exceptions import RequestValidationError
 from fastapi.responses import JSONResponse
 from typing_extensions import assert_never
 
-from vllm.entrypoints.openai.engine.protocol import (
-    ErrorResponse,
-)
+from vllm.entrypoints.openai.engine.protocol import ErrorResponse
 from vllm.entrypoints.serve.tokenize.protocol import (
     DetokenizeRequest,
     DetokenizeResponse,
     TokenizeRequest,
     TokenizeResponse,
 )
-from vllm.entrypoints.serve.tokenize.serving import OpenAIServingTokenization
+from vllm.entrypoints.serve.tokenize.serving import ServingTokenization
 from vllm.entrypoints.serve.utils.api_utils import (
     validate_json_request,
     with_cancellation,
@@ -28,8 +26,8 @@ from vllm.logger import init_logger
 logger = init_logger(__name__)
 
 
-def tokenization(request: Request) -> OpenAIServingTokenization:
-    return request.app.state.openai_serving_tokenization
+def tokenization(request: Request) -> ServingTokenization:
+    return request.app.state.serving_tokenization
 
 
 router = APIRouter()

vllm/entrypoints/serve/tokenize/serving.py

@@ -5,11 +5,13 @@ from typing import Any, Final
 
 from fastapi import Request
 
-from vllm.engine.protocol import EngineClient
 from vllm.entrypoints.chat_utils import ChatTemplateContentFormatOption
 from vllm.entrypoints.openai.engine.protocol import ErrorResponse
-from vllm.entrypoints.openai.engine.serving import OpenAIServing
-from vllm.entrypoints.openai.models.serving import OpenAIServingModels
+from vllm.entrypoints.openai.models.serving import (
+    OpenAIModelRegistry,
+    OpenAIServingModels,
+)
+from vllm.entrypoints.serve.engine.serving import BaseServing
 from vllm.entrypoints.serve.render.serving import OpenAIServingRender
 from vllm.entrypoints.serve.tokenize.protocol import (
     DetokenizeRequest,
@@ -27,11 +29,10 @@ from vllm.tokenizers import TokenizerLike
 logger = init_logger(__name__)
 
 
-class OpenAIServingTokenization(OpenAIServing):
+class ServingTokenization(BaseServing):
     def __init__(
         self,
-        engine_client: EngineClient,
-        models: OpenAIServingModels,
+        models: OpenAIServingModels | OpenAIModelRegistry,
         openai_serving_render: OpenAIServingRender,
         *,
         request_logger: RequestLogger | None,
@@ -41,11 +42,12 @@ class OpenAIServingTokenization(OpenAIServing):
         trust_request_chat_template: bool = False,
     ) -> None:
         super().__init__(
-            engine_client=engine_client,
             models=models,
+            model_config=openai_serving_render.model_config,
             request_logger=request_logger,
         )
 
+        self.renderer = openai_serving_render.renderer
         self.openai_serving_render = openai_serving_render
         self.chat_template = chat_template
         self.chat_template_content_format: Final = chat_template_content_format

vllm/entrypoints/speech_to_text/base/serving.py

@@ -22,8 +22,9 @@ from vllm.entrypoints.openai.engine.protocol import (
     RequestResponseMetadata,
     UsageInfo,
 )
-from vllm.entrypoints.openai.engine.serving import OpenAIServing, SpeechToTextRequest
+from vllm.entrypoints.openai.engine.serving import OpenAIServing
 from vllm.entrypoints.openai.models.serving import OpenAIServingModels
+from vllm.entrypoints.serve.engine.typing import SpeechToTextRequest
 from vllm.entrypoints.serve.utils.api_utils import get_max_tokens
 from vllm.entrypoints.serve.utils.request_logger import RequestLogger
 from vllm.exceptions import VLLMValidationError

vllm/model_executor/layers/attention/attention.py

@@ -581,7 +581,14 @@ class Attention(nn.Module, AttentionLayerBase):
     def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec | None:
         # Block size may get updated after model loading, refresh it
         block_size = vllm_config.cache_config.block_size
-        # Should not be called for enc-dec or encoder-only attention.
+        # Encoder-only attention is prefill-only and keeps no autoregressive KV
+        # cache. In hybrid models (e.g. Qwen3.5 / ColQwen3.5: GatedDeltaNet
+        # linear_attention interleaved with full_attention) the runner iterates
+        # every attention module to build the KV-cache spec, so an ENCODER_ONLY
+        # full_attention layer reaches here; it contributes no KV cache group.
+        if self.attn_type in (AttentionType.ENCODER_ONLY, AttentionType.ENCODER):
+            return None
+        # Should not be called for enc-dec attention.
         assert self.attn_type == AttentionType.DECODER
         quant_mode = get_kv_quant_mode(self.kv_cache_dtype)
         if self.sliding_window is not None:

vllm/model_executor/layers/fused_moe/experts/xpu_moe.py

@@ -67,6 +67,7 @@ class XPUExperts(mk.FusedMoEExpertsModular):
         self.is_mxfp4 = False
         self.is_block_fp8 = False
         self.is_mxfp8 = False
+        self.gemm1_clamp_limit = quant_config.gemm1_clamp_limit
         self.fused_moe_impl: XpuFusedMoe | None = None
 
     @property
@@ -176,6 +177,7 @@ class XPUExperts(mk.FusedMoEExpertsModular):
                 is_mxfp4=self.is_mxfp4,
                 is_mxfp8=self.is_mxfp8,
                 is_block_fp8=self.is_block_fp8,
+                gemm1_clamp_limit=self.gemm1_clamp_limit,
             )
         assert self.fused_moe_impl is not None
         self.fused_moe_impl.apply(

vllm/model_executor/layers/mhc.py

@@ -6,9 +6,25 @@ import torch
 # import vllm.model_executor.kernels.mhc  # noqa: F401
 import vllm.model_executor.kernels.mhc as mhc_kernels
 from vllm.model_executor.custom_op import CustomOp
+from vllm.platforms import current_platform
 from vllm.utils.import_utils import has_tilelang
 
-HAS_TILELANG = has_tilelang()
+
+def _has_tilelang_mhc() -> bool:
+    if not has_tilelang():
+        return False
+    if current_platform.is_cuda():
+        return True
+    if current_platform.is_rocm():
+        from vllm.platforms.rocm import on_gfx942
+
+        # TileLang MHC currently produces incorrect results on gfx942. Keep
+        # gfx942 on the existing torch/triton fallbacks until that path is fixed.
+        return not on_gfx942()
+    return False
+
+
+HAS_TILELANG_MHC = _has_tilelang_mhc()
 
 
 # --8<-- [start:mhc_pre]
@@ -89,7 +105,7 @@ class MHCPreOp(CustomOp):
         #         sinkhorn_repeat,
         #     )
         # else:
-        if HAS_TILELANG:
+        if HAS_TILELANG_MHC:
             return torch.ops.vllm.mhc_pre_tilelang(
                 residual,
                 fn,
@@ -224,7 +240,7 @@ class MHCPostOp(CustomOp):
         #         comb_res_mix,
         #     )
         # else:
-        if HAS_TILELANG:
+        if HAS_TILELANG_MHC:
             return torch.ops.vllm.mhc_post_tilelang(
                 x, residual, post_layer_mix, comb_res_mix
             )
@@ -310,7 +326,7 @@ class HCHeadOp(CustomOp):
         outer_shape = hidden_states.shape[:-2]
         hs_flat = hidden_states.view(-1, hc_mult, hidden_size)
 
-        if HAS_TILELANG:
+        if HAS_TILELANG_MHC:
             out = torch.ops.vllm.hc_head_fused_kernel_tilelang(
                 hs_flat,
                 hc_fn,
@@ -447,7 +463,26 @@ class MHCFusedPostPreOp(CustomOp):
         norm_weight: torch.Tensor | None = None,
         norm_eps: float = 0.0,
     ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
-        return torch.ops.vllm.mhc_fused_post_pre_tilelang(
+        if HAS_TILELANG_MHC:
+            return torch.ops.vllm.mhc_fused_post_pre_tilelang(
+                x,
+                residual,
+                post_layer_mix,
+                comb_res_mix,
+                fn,
+                hc_scale,
+                hc_base,
+                rms_eps,
+                hc_pre_eps,
+                hc_sinkhorn_eps,
+                hc_post_mult_value,
+                sinkhorn_repeat,
+                n_splits,
+                tile_n,
+                norm_weight,
+                norm_eps,
+            )
+        return self.forward_native(
             x,
             residual,
             post_layer_mix,

vllm/model_executor/layers/quantization/moe_wna16.py

@@ -426,6 +426,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
 
             device = get_tp_group().device
             tp_rank = get_tensor_model_parallel_rank()
+            tp_size = layer.moe_config.moe_parallel_config.tp_size
             loaded_weight = loaded_weight.to(device)
             shard_size = layer.intermediate_size_per_partition
 
@@ -464,9 +465,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
                 )
 
             if "w13_qzeros" in weight_name:
-                tensor = loaded_weight.view(layer.tp_size, -1, loaded_weight.size(1))[
-                    tp_rank
-                ]
+                tensor = loaded_weight.view(tp_size, -1, loaded_weight.size(1))[tp_rank]
                 if shard_id == "w1":
                     param.data[expert_id, : shard_size // 2] = tensor
                 else:
@@ -474,7 +473,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
                 return True if return_success else None
             elif "w2_qzeros" in weight_name:
                 param.data[expert_id] = loaded_weight.view(
-                    loaded_weight.size(0), layer.tp_size, -1
+                    loaded_weight.size(0), tp_size, -1
                 )[:, tp_rank]
                 return True if return_success else None
             else:

vllm/model_executor/models/colqwen3_5.py

@@ -15,6 +15,7 @@ Based on: Qwen3.5 backbone with custom text projection
 
 Target models:
 - athrael-soju/colqwen3.5-4.5B-v3
+- vultr/VultronRetrieverPrime-Qwen3.5-8B
 """
 
 from collections.abc import Iterable, Mapping
@@ -166,12 +167,19 @@ class ColQwen3_5Model(
             or 128  # default from reference implementation
         )
 
+        # ColPali defines `custom_text_proj = nn.Linear(hidden, dim)`, i.e.
+        # bias=True by default, and the trained ColQwen3.5 checkpoints ship a
+        # `custom_text_proj.bias`. Construct with a bias and zero-initialize it:
+        # a (legacy) bias-less checkpoint then behaves identically to bias=False,
+        # while load_weights() below picks up a trained bias instead of silently
+        # dropping it (which shifts every per-token vector and the MaxSim ranking).
         self.custom_text_proj = nn.Linear(
             hidden_size,
             self.embed_dim,
-            bias=False,
+            bias=True,
             dtype=head_dtype,
         )
+        nn.init.zeros_(self.custom_text_proj.bias)
 
         pooler_config = vllm_config.model_config.pooler_config
         assert pooler_config is not None

vllm/model_executor/models/config.py

@@ -627,6 +627,20 @@ class Qwen3_5ForConditionalGenerationConfig(VerifyAndUpdateConfig):
             )
 
 
+class ColQwen3_5Config(Qwen3_5ForConditionalGenerationConfig):
+    """ColQwen3.5 (late-interaction retrieval) inherits Qwen3.5's mamba cache
+    handling and additionally serves BIDIRECTIONAL attention: ColPali-style
+    document/query encoding attends over the whole sequence, not causally. Set
+    is_causal=False so Qwen3NextAttention builds its full_attention layers with
+    AttentionType.ENCODER_ONLY (the linear_attention GatedDeltaNet layers are
+    unaffected). Generation arches keep the parent (causal) and are untouched.
+    """
+
+    @staticmethod
+    def verify_and_update_model_config(model_config: "ModelConfig") -> None:
+        model_config.hf_config.is_causal = False
+
+
 class SnowflakeGteNewModelConfig(VerifyAndUpdateConfig):
     @staticmethod
     def verify_and_update_model_config(model_config: "ModelConfig") -> None:
@@ -656,7 +670,7 @@ class VoyageQwen3BidirectionalEmbedModelConfig(VerifyAndUpdateConfig):
 
 MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
     "ColBERTJinaRobertaModel": JinaRobertaModelConfig,
-    "ColQwen3_5": Qwen3_5ForConditionalGenerationConfig,
+    "ColQwen3_5": ColQwen3_5Config,
     "DeepseekV4ForCausalLM": DeepseekV4ForCausalLMConfig,
     "DeepseekV32ForCausalLM": DeepseekV32ForCausalLM,
     "DiffusionGemmaForBlockDiffusion": DiffusionGemmaModelForBlockDiffusionConfig,  # noqa: E501

vllm/model_executor/models/minimax_vl_01.py

@@ -1,385 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from collections.abc import Iterable, Mapping
-from typing import Annotated, Literal, TypeAlias
-
-import torch
-import torch.nn as nn
-from transformers import BatchFeature, PretrainedConfig
-from transformers.models.llava_next.modeling_llava_next import (
-    get_anyres_image_grid_shape,
-    unpad_image,
-)
-
-from vllm.config import VllmConfig
-from vllm.model_executor.layers.activation import get_act_fn
-from vllm.model_executor.layers.linear import ColumnParallelLinear, RowParallelLinear
-from vllm.model_executor.layers.quantization import QuantizationConfig
-from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig
-from vllm.sequence import IntermediateTensors
-from vllm.utils.tensor_schema import TensorSchema, TensorShape
-
-from .clip import CLIPVisionModel
-from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
-from .llava import (
-    BaseLlavaMultiModalProcessor,
-    LlavaDummyInputsBuilder,
-    init_vision_tower_for_llava,
-)
-from .llava_next import LlavaNextProcessingInfo
-from .pixtral import PixtralHFVisionModel
-from .siglip import SiglipVisionModel
-from .utils import (
-    AutoWeightsLoader,
-    init_vllm_registered_model,
-    maybe_prefix,
-)
-
-
-class MiniMaxVL01ImagePixelInputs(TensorSchema):
-    """
-    Dimensions:
-        - bn: Batch size * number of images
-        - np: Number of patches + 1
-        - c: Number of channels (3)
-        - h: Height
-        - w: Width
-
-    Note that `num_patches` may be different per batch and image,
-    in which case the data is passed as a list instead of a batched tensor.
-    """
-
-    type: Literal["pixel_values"] = "pixel_values"
-    pixel_values: Annotated[
-        torch.Tensor | list[torch.Tensor],
-        TensorShape("bn", "np", 3, "h", "w", dynamic_dims={"np", "h", "w"}),
-    ]
-
-    image_sizes: Annotated[torch.Tensor | None, TensorShape("bn", 2)]
-    # This should be in `(height, width)` format.
-
-
-class MiniMaxVL01ImageEmbeddingInputs(TensorSchema):
-    """
-    Dimensions:
-        - bn: Batch size * number of images
-        - ifs: Image feature size
-        - hs: Hidden size (must match language model backbone)
-    """
-
-    type: Literal["image_embeds"] = "image_embeds"
-    data: Annotated[torch.Tensor, TensorShape("bn", "ifs", "hs")]
-
-
-MiniMaxVL01ImageInputs: TypeAlias = (
-    MiniMaxVL01ImagePixelInputs | MiniMaxVL01ImageEmbeddingInputs
-)
-
-
-class MiniMaxVL01MultiModalProjector(nn.Module):
-    def __init__(
-        self,
-        vision_hidden_size: int,
-        text_hidden_size: int,
-        projector_hidden_act: str,
-        multimodal_projector_bias: bool,
-        quant_config: QuantizationConfig | None = None,
-        prefix: str = "",
-    ):
-        super().__init__()
-
-        self.linear_1 = ColumnParallelLinear(
-            vision_hidden_size,
-            text_hidden_size,
-            bias=multimodal_projector_bias,
-            quant_config=quant_config,
-            prefix=f"{prefix}.linear_1",
-        )
-        self.act = get_act_fn(projector_hidden_act)
-        self.linear_2 = RowParallelLinear(
-            text_hidden_size,
-            text_hidden_size,
-            bias=multimodal_projector_bias,
-            quant_config=quant_config,
-            prefix=f"{prefix}.linear_2",
-        )
-
-    def forward(self, image_features: torch.Tensor) -> torch.Tensor:
-        hidden_states, _ = self.linear_1(image_features)
-        hidden_states = self.act(hidden_states)
-        hidden_states, _ = self.linear_2(hidden_states)
-        return hidden_states
-
-
-class MiniMaxVL01DummyInputsBuilder(LlavaDummyInputsBuilder):
-    pass
-
-
-class MiniMaxVL01ProcessingInfo(LlavaNextProcessingInfo):
-    def get_hf_config(self):  # Need to override the config type
-        return self.ctx.get_hf_config(PretrainedConfig)
-
-    def get_hf_processor(self, **kwargs: object):
-        hf_processor = self.ctx.get_hf_processor(**kwargs)
-        image_processor = hf_processor.image_processor
-        image_processor.anyres_preprocess = image_processor.anyres_for_vllm_preprocess
-
-        return hf_processor
-
-    def get_supported_mm_limits(self) -> Mapping[str, int | None]:
-        return {"image": None}
-
-
-class MiniMaxVL01MultiModalProcessor(
-    BaseLlavaMultiModalProcessor[MiniMaxVL01ProcessingInfo]
-):
-    def _call_hf_processor(
-        self,
-        prompt: str,
-        mm_data: Mapping[str, object],
-        mm_kwargs: Mapping[str, object],
-        tok_kwargs: Mapping[str, object],
-    ) -> BatchFeature:
-        processed_outputs = super()._call_hf_processor(
-            prompt=prompt,
-            mm_data=mm_data,
-            mm_kwargs=mm_kwargs,
-            tok_kwargs=tok_kwargs,
-        )
-
-        pixel_values = processed_outputs.get("pixel_values")
-        if pixel_values is not None:
-            # Avoid padding since we need the output for each image to be
-            # independent of other images for the cache to work correctly
-            image_sizes = processed_outputs["image_sizes"]
-            assert len(pixel_values) == len(image_sizes)
-
-            processed_outputs["pixel_values"] = [
-                p[:, :h, :w] for p, (h, w) in zip(pixel_values, image_sizes)
-            ]
-
-        return processed_outputs
-
-    def _get_mm_fields_config(
-        self,
-        hf_inputs: BatchFeature,
-        hf_processor_mm_kwargs: Mapping[str, object],
-    ) -> Mapping[str, MultiModalFieldConfig]:
-        return {
-            "pixel_values": MultiModalFieldConfig.batched("image"),
-            "image_sizes": MultiModalFieldConfig.batched("image"),
-            "image_embeds": MultiModalFieldConfig.batched("image"),
-        }
-
-
-@MULTIMODAL_REGISTRY.register_processor(
-    MiniMaxVL01MultiModalProcessor,
-    info=MiniMaxVL01ProcessingInfo,
-    dummy_inputs=MiniMaxVL01DummyInputsBuilder,
-)
-class MiniMaxVL01ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    packed_modules_mapping = {
-        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
-        "gate_up_proj": ["gate_proj", "up_proj"],
-    }
-
-    @classmethod
-    def get_placeholder_str(cls, modality: str, i: int) -> str | None:
-        if modality.startswith("image"):
-            return "<image>"
-
-        raise ValueError("Only image modality is supported")
-
-    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
-        super().__init__()
-
-        config = vllm_config.model_config.hf_config
-        quant_config = vllm_config.quant_config
-        multimodal_config = vllm_config.model_config.multimodal_config
-
-        self.config = config
-        self.multimodal_config = multimodal_config
-
-        with self._mark_tower_model(vllm_config, "image"):
-            self.vision_tower = init_vision_tower_for_llava(
-                config,
-                quant_config=quant_config,
-                require_post_norm=False,
-                prefix=maybe_prefix(prefix, "vision_tower"),
-            )
-            self.multi_modal_projector = MiniMaxVL01MultiModalProjector(
-                vision_hidden_size=config.vision_config.hidden_size,
-                text_hidden_size=config.text_config.hidden_size,
-                projector_hidden_act=config.projector_hidden_act,
-                multimodal_projector_bias=True,
-                quant_config=quant_config,
-                prefix=maybe_prefix(prefix, "multi_modal_projector"),
-            )
-            self.image_newline = nn.Parameter(
-                torch.empty(config.text_config.hidden_size)
-            )
-
-        with self._mark_language_model(vllm_config):
-            self.language_model = init_vllm_registered_model(
-                vllm_config=vllm_config,
-                hf_config=config.text_config,
-                prefix=maybe_prefix(prefix, "language_model"),
-            )
-
-        self.vision_feature_layer = config.vision_feature_layer
-        self.vocab_size = config.text_config.vocab_size
-        self.pad_token_id = -1
-        if self.config.text_config.pad_token_id is not None:
-            self.pad_token_id = self.config.text_config.pad_token_id
-
-        self.make_empty_intermediate_tensors = (
-            self.language_model.make_empty_intermediate_tensors
-        )
-
-    def _image_pixels_to_features(
-        self,
-        vision_tower: CLIPVisionModel | SiglipVisionModel | PixtralHFVisionModel,
-        pixel_values: torch.Tensor | list[torch.Tensor],
-    ) -> torch.Tensor | tuple[torch.Tensor, ...]:
-        # NOTE: we skip the step to select the vision feature layer since
-        # this is already done inside the vision tower
-        feature_select_strategy = self.config.vision_feature_select_strategy
-        return tuple(
-            vision_tower(p, feature_select_strategy=feature_select_strategy)
-            for p in pixel_values
-        )
-
-    # adapted from https://huggingface.co/MiniMaxAI/MiniMax-VL-01/blob/main/modeling_minimax_vl_01.py#L616-L631
-    def pack_image_features(
-        self, image_features: list[torch.Tensor], image_sizes: torch.Tensor
-    ):
-        new_image_features = []
-        for image_idx, image_feature in enumerate(image_features):
-            if image_feature.shape[0] > 1:
-                base_image_feature = image_feature[0]
-                image_feature = image_feature[1:]
-                height = width = (
-                    self.config.vision_config.image_size
-                    // self.config.vision_config.patch_size
-                )
-                if height * width != base_image_feature.shape[0]:
-                    raise ValueError(
-                        "The number of patches is not consistent with the image size."
-                    )
-                num_patch_height, num_patch_width = get_anyres_image_grid_shape(
-                    image_sizes[image_idx],
-                    self.config.image_grid_pinpoints,
-                    self.config.vision_config.image_size,
-                )
-
-                image_feature = image_feature.view(
-                    num_patch_height, num_patch_width, height, width, -1
-                )
-                image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
-                image_feature = image_feature.flatten(1, 2).flatten(2, 3)
-                image_feature = unpad_image(image_feature, image_sizes[image_idx])
-
-                image_feature = torch.cat(
-                    (
-                        image_feature,
-                        self.image_newline[:, None, None]
-                        .expand(*image_feature.shape[:-1], 1)
-                        .to(image_feature.dtype),
-                    ),
-                    dim=-1,
-                )
-                image_feature = image_feature.flatten(1, 2).transpose(0, 1)
-                image_feature = torch.cat((base_image_feature, image_feature), dim=0)
-            else:
-                image_feature = image_feature[0]
-                image_feature = torch.cat(
-                    (image_feature, self.image_newline[None].to(image_feature)), dim=0
-                )
-            new_image_features.append(image_feature)
-        return new_image_features
-
-    def _process_image_pixels(
-        self,
-        inputs: MiniMaxVL01ImagePixelInputs,
-    ) -> torch.Tensor | tuple[torch.Tensor, ...]:
-        pixel_values = inputs["pixel_values"]
-        return self._image_pixels_to_features(self.vision_tower, pixel_values)
-
-    def _process_image_input(
-        self,
-        image_input: MiniMaxVL01ImageInputs,
-    ) -> torch.Tensor | tuple[torch.Tensor, ...]:
-        if image_input["type"] == "image_embeds":
-            return image_input["data"]
-
-        image_features = self._process_image_pixels(image_input)
-
-        if isinstance(image_features, torch.Tensor):
-            return self.multi_modal_projector(image_features)
-
-        feature_sizes = [image_feature.shape[0] for image_feature in image_features]
-
-        image_embeds = self.multi_modal_projector(torch.cat(image_features))
-        image_embeds = torch.split(image_embeds, feature_sizes)
-        image_sizes = image_input.get("image_sizes")
-        return self.pack_image_features(image_embeds, image_sizes)
-
-    def _parse_and_validate_image_input(
-        self, **kwargs: object
-    ) -> MiniMaxVL01ImageInputs | None:
-        pixel_values = kwargs.pop("pixel_values", None)
-        image_sizes = kwargs.pop("image_sizes", None)
-        image_embeds = kwargs.pop("image_embeds", None)
-
-        if pixel_values is None and image_embeds is None:
-            return None
-
-        if pixel_values is not None and image_sizes is not None:
-            return MiniMaxVL01ImagePixelInputs(
-                type="pixel_values",
-                pixel_values=pixel_values,
-                image_sizes=image_sizes,
-            )
-
-        if image_embeds is not None:
-            return MiniMaxVL01ImageEmbeddingInputs(
-                type="image_embeds",
-                data=image_embeds,
-            )
-
-        raise AssertionError("This line should be unreachable.")
-
-    def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
-        image_input = self._parse_and_validate_image_input(**kwargs)
-        if image_input is None:
-            return []
-
-        return self._process_image_input(image_input)
-
-    def forward(
-        self,
-        input_ids: torch.Tensor | None,
-        positions: torch.Tensor,
-        intermediate_tensors: IntermediateTensors | None = None,
-        inputs_embeds: torch.Tensor | None = None,
-        **kwargs: object,
-    ) -> torch.Tensor | IntermediateTensors:
-        if intermediate_tensors is not None:
-            inputs_embeds = None
-
-        hidden_states = self.language_model.model(
-            input_ids, positions, intermediate_tensors, inputs_embeds=inputs_embeds
-        )
-
-        return hidden_states
-
-    def compute_logits(
-        self,
-        hidden_states: torch.Tensor,
-    ) -> torch.Tensor | None:
-        return self.language_model.compute_logits(hidden_states)
-
-    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
-        loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)

vllm/model_executor/models/qwen3_next.py

@@ -62,6 +62,7 @@ from vllm.model_executor.models.utils import sequence_parallel_chunk
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs.qwen3_next import Qwen3NextConfig
+from vllm.v1.attention.backend import AttentionType
 
 from .interfaces import (
     EagleModelMixin,
@@ -267,6 +268,15 @@ class Qwen3NextAttention(nn.Module):
             dual_chunk_attention_config=self.dual_chunk_attention_config,
         )
 
+        # Late-interaction retrieval models (e.g. ColQwen3.5) run BIDIRECTIONAL
+        # attention on the full_attention layers; they set config.is_causal=False
+        # via a VerifyAndUpdateConfig handler. Generation models leave is_causal
+        # unset (-> causal/DECODER), so this is a no-op for them. Mirrors qwen3.py.
+        attn_type = (
+            AttentionType.DECODER
+            if getattr(config, "is_causal", True)
+            else AttentionType.ENCODER_ONLY
+        )
         self.attn = Attention(
             self.num_heads,
             self.head_dim,
@@ -275,6 +285,7 @@ class Qwen3NextAttention(nn.Module):
             cache_config=cache_config,
             quant_config=quant_config,
             prefix=f"{prefix}.attn",
+            attn_type=attn_type,
             **{
                 "layer_idx": extract_layer_index(prefix),
                 "dual_chunk_attention_config": self.dual_chunk_attention_config,

vllm/model_executor/models/registry.py

@@ -159,9 +159,6 @@ _TEXT_GENERATION_MODELS = {
     "MellumForCausalLM": ("mellum", "MellumForCausalLM"),
     "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
     "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
-    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
-    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
-    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
     "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
     "MiniMaxM3SparseForCausalLM": (
         "vllm.models.minimax_m3",
@@ -490,10 +487,6 @@ _MULTIMODAL_MODELS = {
         "vllm.models.minimax_m3",
         "MiniMaxM3SparseForConditionalGeneration",
     ),
-    "MiniMaxVL01ForConditionalGeneration": (
-        "minimax_vl_01",
-        "MiniMaxVL01ForConditionalGeneration",
-    ),
     "MiniCPMO": ("minicpmo", "MiniCPMO"),
     "MiniCPMV": ("minicpmv", "MiniCPMV"),
     "MiniCPMV4_6ForConditionalGeneration": (
@@ -735,6 +728,10 @@ _PREVIOUSLY_SUPPORTED_MODELS = {
     "XverseForCausalLM": "0.23.0",
     "Dots1ForCausalLM": "0.23.0",
     "BambaForCausalLM": "0.23.0",
+    "MiniMaxForCausalLM": "0.23.0",
+    "MiniMaxText01ForCausalLM": "0.23.0",
+    "MiniMaxM1ForCausalLM": "0.23.0",
+    "MiniMaxVL01ForConditionalGeneration": "0.23.0",
 }
 
 _OOT_SUPPORTED_MODELS = {

vllm/models/deepseek_v4/amd/model.py

@@ -27,6 +27,7 @@ from vllm.model_executor.layers.linear import (
 )
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.mhc import (
+    HAS_TILELANG_MHC,
     HCHeadOp,
     MHCFusedPostPreOp,
     MHCPostOp,
@@ -51,7 +52,6 @@ from vllm.model_executor.models.utils import (
 from vllm.models.deepseek_v4.amd.rocm import DeepseekV4ROCMAiterMLAAttention
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
-from vllm.utils.import_utils import has_tilelang
 
 
 class DeepseekV4MLP(nn.Module):
@@ -303,7 +303,7 @@ class DeepseekV4DecoderLayer(nn.Module):
         self.mhc_pre = MHCPreOp()
         self.mhc_post = MHCPostOp()
         self.mhc_fused_post_pre = MHCFusedPostPreOp()
-        self.has_tilelang = has_tilelang()
+        self.has_tilelang = HAS_TILELANG_MHC
 
     def hc_pre(
         self,
@@ -513,7 +513,7 @@ class DeepseekV4Model(nn.Module):
             requires_grad=False,
         )
         self.hc_head_op = HCHeadOp()
-        self.has_tilelang = has_tilelang()
+        self.has_tilelang = HAS_TILELANG_MHC
         # Pre-hc_head residual stream buffer for the MTP draft. Stable
         # address (outside the cudagraph pool) so the copy_ in forward()
         # refreshes it correctly across captured shapes.