Merge pull request #268 from runpod-workers/fix/spec-config-0-to-none

fix: spec config env vars should be none if zero

.runpod/hub.json

@@ -6,20 +6,10 @@
   "iconUrl": "https://registry.npmmirror.com/@lobehub/icons-static-png/latest/files/dark/vllm-color.png",
   "config": {
     "runsOn": "GPU",
-    "containerDiskInGb": 200,
-    "gpuIds": "ADA_80_PRO, AMPERE_80",
+    "containerDiskInGb": 150,
+    "gpuIds": "ADA_80_PRO,AMPERE_80",
     "gpuCount": 1,
-    "allowedCudaVersions": [
-      "12.9",
-      "12.8",
-      "12.7",
-      "12.6",
-      "12.5",
-      "12.4",
-      "12.3",
-      "12.2",
-      "12.1"
-    ],
+    "allowedCudaVersions": ["12.9", "12.8"],
     "presets": [
       {
         "name": "deepseek-ai/deepseek-r1-distill-llama-8b",
@@ -191,41 +181,13 @@
           "advanced": true
         }
       },
-      {
-        "key": "QUANTIZATION_PARAM_PATH",
-        "input": {
-          "name": "Quantization Param Path",
-          "type": "string",
-          "description": "Path to the JSON file containing the KV cache scaling factors.",
-          "advanced": true
-        }
-      },
       {
         "key": "MAX_MODEL_LEN",
         "input": {
           "name": "Max Model Length",
           "type": "number",
           "description": "Model context length.",
-          "advanced": true
-        }
-      },
-      {
-        "key": "GUIDED_DECODING_BACKEND",
-        "input": {
-          "name": "Guided Decoding Backend",
-          "type": "string",
-          "description": "Which engine will be used for guided decoding by default.",
-          "options": [
-            {
-              "label": "outlines",
-              "value": "outlines"
-            },
-            {
-              "label": "lm-format-enforcer",
-              "value": "lm-format-enforcer"
-            }
-          ],
-          "default": "outlines",
+          "default": null,
           "advanced": true
         }
       },
@@ -245,17 +207,8 @@
               "value": "mp"
             }
           ],
-          "advanced": true
-        }
-      },
-      {
-        "key": "WORKER_USE_RAY",
-        "input": {
-          "name": "Worker Use Ray",
-          "type": "boolean",
-          "description": "Deprecated, use --distributed-executor-backend=ray.",
-          "default": false,
-          "advanced": true
+          "advanced": true,
+          "default": "mp"
         }
       },
       {
@@ -317,26 +270,6 @@
           "advanced": true
         }
       },
-      {
-        "key": "USE_V2_BLOCK_MANAGER",
-        "input": {
-          "name": "Use V2 Block Manager",
-          "type": "boolean",
-          "description": "Use BlockSpaceMangerV2.",
-          "default": false,
-          "advanced": true
-        }
-      },
-      {
-        "key": "NUM_LOOKAHEAD_SLOTS",
-        "input": {
-          "name": "Num Lookahead Slots",
-          "type": "number",
-          "description": "Experimental scheduling config necessary for speculative decoding.",
-          "default": 0,
-          "advanced": true
-        }
-      },
       {
         "key": "SEED",
         "input": {
@@ -362,6 +295,7 @@
           "name": "Max Num Batched Tokens",
           "type": "number",
           "description": "Maximum number of batched tokens per iteration.",
+          "default": null,
           "advanced": true
         }
       },
@@ -422,53 +356,6 @@
           "advanced": true
         }
       },
-      {
-        "key": "ROPE_SCALING",
-        "input": {
-          "name": "RoPE Scaling",
-          "type": "string",
-          "description": "RoPE scaling configuration in JSON format.",
-          "advanced": true
-        }
-      },
-      {
-        "key": "ROPE_THETA",
-        "input": {
-          "name": "RoPE Theta",
-          "type": "number",
-          "description": "RoPE theta. Use with rope_scaling.",
-          "advanced": true
-        }
-      },
-      {
-        "key": "TOKENIZER_POOL_SIZE",
-        "input": {
-          "name": "Tokenizer Pool Size",
-          "type": "number",
-          "description": "Size of tokenizer pool to use for asynchronous tokenization.",
-          "default": 0,
-          "advanced": true
-        }
-      },
-      {
-        "key": "TOKENIZER_POOL_TYPE",
-        "input": {
-          "name": "Tokenizer Pool Type",
-          "type": "string",
-          "description": "Type of tokenizer pool to use for asynchronous tokenization.",
-          "default": "ray",
-          "advanced": true
-        }
-      },
-      {
-        "key": "TOKENIZER_POOL_EXTRA_CONFIG",
-        "input": {
-          "name": "Tokenizer Pool Extra Config",
-          "type": "string",
-          "description": "Extra config for tokenizer pool.",
-          "advanced": true
-        }
-      },
       {
         "key": "ENABLE_LORA",
         "input": {
@@ -499,16 +386,6 @@
           "advanced": true
         }
       },
-      {
-        "key": "LORA_EXTRA_VOCAB_SIZE",
-        "input": {
-          "name": "LoRA Extra Vocab Size",
-          "type": "number",
-          "description": "Maximum size of extra vocabulary for LoRA adapters.",
-          "default": 256,
-          "advanced": true
-        }
-      },
       {
         "key": "LORA_DTYPE",
         "input": {
@@ -537,15 +414,6 @@
           "advanced": true
         }
       },
-      {
-        "key": "LONG_LORA_SCALING_FACTORS",
-        "input": {
-          "name": "Long LoRA Scaling Factors",
-          "type": "string",
-          "description": "Specify multiple scaling factors for LoRA adapters.",
-          "advanced": true
-        }
-      },
       {
         "key": "MAX_CPU_LORAS",
         "input": {
@@ -625,6 +493,34 @@
           "advanced": true
         }
       },
+      {
+        "key": "SPECULATIVE_CONFIG",
+        "input": {
+          "name": "Speculative Config (JSON)",
+          "type": "string",
+          "description": "Full speculative decoding configuration as a JSON string. Overrides individual speculative env vars.",
+          "advanced": true
+        }
+      },
+      {
+        "key": "SPECULATIVE_METHOD",
+        "input": {
+          "name": "Speculative Method",
+          "type": "string",
+          "description": "Speculative decoding method to use.",
+          "options": [
+            { "label": "None", "value": "" },
+            { "label": "Draft Model", "value": "draft_model" },
+            { "label": "N-gram", "value": "ngram" },
+            { "label": "EAGLE", "value": "eagle" },
+            { "label": "EAGLE3", "value": "eagle3" },
+            { "label": "Medusa", "value": "medusa" },
+            { "label": "MLP Speculator", "value": "mlp_speculator" }
+          ],
+          "default": "",
+          "advanced": true
+        }
+      },
       {
         "key": "SPECULATIVE_MODEL",
         "input": {
@@ -643,33 +539,6 @@
           "advanced": true
         }
       },
-      {
-        "key": "SPECULATIVE_DRAFT_TENSOR_PARALLEL_SIZE",
-        "input": {
-          "name": "Speculative Draft Tensor Parallel Size",
-          "type": "number",
-          "description": "Number of tensor parallel replicas for the draft model.",
-          "advanced": true
-        }
-      },
-      {
-        "key": "SPECULATIVE_MAX_MODEL_LEN",
-        "input": {
-          "name": "Speculative Max Model Length",
-          "type": "number",
-          "description": "The maximum sequence length supported by the draft model.",
-          "advanced": true
-        }
-      },
-      {
-        "key": "SPECULATIVE_DISABLE_BY_BATCH_SIZE",
-        "input": {
-          "name": "Speculative Disable by Batch Size",
-          "type": "number",
-          "description": "Disable speculative decoding if the number of enqueue requests is larger than this value.",
-          "advanced": true
-        }
-      },
       {
         "key": "NGRAM_PROMPT_LOOKUP_MAX",
         "input": {
@@ -679,53 +548,6 @@
           "advanced": true
         }
       },
-      {
-        "key": "NGRAM_PROMPT_LOOKUP_MIN",
-        "input": {
-          "name": "Ngram Prompt Lookup Min",
-          "type": "number",
-          "description": "Min size of window for ngram prompt lookup in speculative decoding.",
-          "advanced": true
-        }
-      },
-      {
-        "key": "SPEC_DECODING_ACCEPTANCE_METHOD",
-        "input": {
-          "name": "Speculative Decoding Acceptance Method",
-          "type": "string",
-          "description": "Specify the acceptance method for draft token verification in speculative decoding.",
-          "options": [
-            {
-              "label": "rejection_sampler",
-              "value": "rejection_sampler"
-            },
-            {
-              "label": "typical_acceptance_sampler",
-              "value": "typical_acceptance_sampler"
-            }
-          ],
-          "default": "rejection_sampler",
-          "advanced": true
-        }
-      },
-      {
-        "key": "TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_THRESHOLD",
-        "input": {
-          "name": "Typical Acceptance Sampler Posterior Threshold",
-          "type": "number",
-          "description": "Set the lower bound threshold for the posterior probability of a token to be accepted.",
-          "advanced": true
-        }
-      },
-      {
-        "key": "TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_ALPHA",
-        "input": {
-          "name": "Typical Acceptance Sampler Posterior Alpha",
-          "type": "number",
-          "description": "A scaling factor for the entropy-based threshold for token acceptance.",
-          "advanced": true
-        }
-      },
       {
         "key": "MODEL_LOADER_EXTRA_CONFIG",
         "input": {
@@ -736,49 +558,11 @@
         }
       },
       {
-        "key": "PREEMPTION_MODE",
+        "key": "ENABLE_LOG_REQUESTS",
         "input": {
-          "name": "Preemption Mode",
-          "type": "string",
-          "description": "If 'recompute', the engine performs preemption-aware recomputation. If 'save', the engine saves activations into the CPU memory as preemption happens.",
-          "advanced": true
-        }
-      },
-      {
-        "key": "PREEMPTION_CHECK_PERIOD",
-        "input": {
-          "name": "Preemption Check Period",
-          "type": "number",
-          "description": "How frequently the engine checks if a preemption happens.",
-          "default": 1,
-          "advanced": true
-        }
-      },
-      {
-        "key": "PREEMPTION_CPU_CAPACITY",
-        "input": {
-          "name": "Preemption CPU Capacity",
-          "type": "number",
-          "description": "The percentage of CPU memory used for the saved activations.",
-          "default": 2,
-          "advanced": true
-        }
-      },
-      {
-        "key": "MAX_LOG_LEN",
-        "input": {
-          "name": "Max Log Length",
-          "type": "number",
-          "description": "Max number of characters or ID numbers being printed in log.",
-          "advanced": true
-        }
-      },
-      {
-        "key": "DISABLE_LOGGING_REQUEST",
-        "input": {
-          "name": "Disable Logging Request",
+          "name": "Enable Log Requests",
           "type": "boolean",
-          "description": "Disable logging requests.",
+          "description": "Enable vLLM request logging.",
           "default": false,
           "advanced": true
         }
@@ -850,16 +634,6 @@
           "advanced": true
         }
       },
-      {
-        "key": "MAX_SEQ_LEN_TO_CAPTURE",
-        "input": {
-          "name": "CUDA Graph Max Content Length",
-          "type": "number",
-          "description": "Maximum context length covered by CUDA graphs. If a sequence has context length larger than this, we fall back to eager mode",
-          "default": 8192,
-          "advanced": true
-        }
-      },
       {
         "key": "DISABLE_CUSTOM_ALL_REDUCE",
         "input": {
@@ -939,6 +713,16 @@
           "advanced": true
         }
       },
+      {
+        "key": "ENABLE_EXPERT_PARALLEL",
+        "input": {
+          "name": "Enable Expert Parallel",
+          "type": "boolean",
+          "description": "Enable Expert Parallel for MoE models",
+          "default": false,
+          "advanced": true
+        }
+      },
       {
         "key": "MODEL_REVISION",
         "input": {
@@ -958,16 +742,6 @@
           "advanced": true
         }
       },
-      {
-        "key": "DISABLE_LOG_REQUESTS",
-        "input": {
-          "name": "Disable Log Requests",
-          "type": "boolean",
-          "description": "Enables or disables vLLM request logging",
-          "default": true,
-          "advanced": true
-        }
-      },
       {
         "key": "ENABLE_AUTO_TOOL_CHOICE",
         "input": {

.runpod/tests_json

@@ -38,16 +38,6 @@
         "value": "HuggingFaceTB/SmolLM2-135M-Instruct"
       }
     ],
-    "allowedCudaVersions": [
-      "12.9",
-      "12.8",
-      "12.7",
-      "12.6",
-      "12.5",
-      "12.4",
-      "12.3",
-      "12.2",
-      "12.1"
-    ]
+    "allowedCudaVersions": ["12.9", "12.8", "12.7", "12.6", "12.5"]
   }
 }

Dockerfile

@@ -1,20 +1,21 @@
-FROM nvidia/cuda:12.1.0-base-ubuntu22.04 
+FROM nvidia/cuda:12.8.0-base-ubuntu22.04 
 
 RUN apt-get update -y \
     && apt-get install -y python3-pip
 
-RUN ldconfig /usr/local/cuda-12.1/compat/
+RUN ldconfig /usr/local/cuda-12.8/compat/
 
-# Install Python dependencies
+# Install vLLM with FlashInfer - use CUDA 12.8 PyTorch wheels (compatible with vLLM 0.15.0)
+RUN python3 -m pip install --upgrade pip && \
+    python3 -m pip install "vllm[flashinfer]==0.15.0" --extra-index-url https://download.pytorch.org/whl/cu128
+
+
+
+# Install additional Python dependencies (after vLLM to avoid PyTorch version conflicts)
 COPY builder/requirements.txt /requirements.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install --upgrade pip && \
     python3 -m pip install --upgrade -r /requirements.txt
 
-# Install vLLM (switching back to pip installs since issues that required building fork are fixed and space optimization is not as important since caching) and FlashInfer 
-RUN python3 -m pip install vllm==0.11.0 && \
-    python3 -m pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3
-
 # Setup for Option 2: Building the Image with the Model included
 ARG MODEL_NAME=""
 ARG TOKENIZER_NAME=""
@@ -22,6 +23,7 @@ ARG BASE_PATH="/runpod-volume"
 ARG QUANTIZATION=""
 ARG MODEL_REVISION=""
 ARG TOKENIZER_REVISION=""
+ARG VLLM_NIGHTLY="false"
 
 ENV MODEL_NAME=$MODEL_NAME \
     MODEL_REVISION=$MODEL_REVISION \
@@ -32,10 +34,22 @@ ENV MODEL_NAME=$MODEL_NAME \
     HF_DATASETS_CACHE="${BASE_PATH}/huggingface-cache/datasets" \
     HUGGINGFACE_HUB_CACHE="${BASE_PATH}/huggingface-cache/hub" \
     HF_HOME="${BASE_PATH}/huggingface-cache/hub" \
-    HF_HUB_ENABLE_HF_TRANSFER=0 
+    HF_HUB_ENABLE_HF_TRANSFER=0 \
+    # Suppress Ray metrics agent warnings (not needed in containerized environments)
+    RAY_METRICS_EXPORT_ENABLED=0 \
+    RAY_DISABLE_USAGE_STATS=1 \
+    # Prevent rayon thread pool panic in containers where ulimit -u < nproc
+    # (tokenizers uses Rust's rayon which tries to spawn threads = CPU cores)
+    TOKENIZERS_PARALLELISM=false \
+    RAYON_NUM_THREADS=4
 
 ENV PYTHONPATH="/:/vllm-workspace"
 
+RUN if [ "${VLLM_NIGHTLY}" = "true" ]; then \
+    pip install -U vllm --pre --index-url https://pypi.org/simple --extra-index-url https://wheels.vllm.ai/nightly && \
+    apt-get update && apt-get install -y git && rm -rf /var/lib/apt/lists/* && \
+    pip install git+https://github.com/huggingface/transformers.git; \
+fi
 
 COPY src /src
 RUN --mount=type=secret,id=HF_TOKEN,required=false \

builder/requirements.txt

@@ -1,14 +1,14 @@
 ray
 pandas
 pyarrow
-runpod~=1.7.7
+runpod
 huggingface-hub
 packaging
 typing-extensions>=4.8.0
 pydantic
 pydantic-settings
 hf-transfer
-transformers>=4.55.0
+transformers>=4.57.0
 bitsandbytes>=0.45.0
 kernels
-torch==2.6.0
+torch-c-dlpack-ext

docs/configuration.md

@@ -28,7 +28,6 @@ Complete guide to all environment variables and configuration options for worker
 | `RAY_WORKERS_USE_NSIGHT`       | False               | `bool`                                                      | If specified, use nsight to profile Ray workers.                                |
 | `ENABLE_PREFIX_CACHING`        | False               | `bool`                                                      | Enables automatic prefix caching.                                               |
 | `DISABLE_SLIDING_WINDOW`       | False               | `bool`                                                      | Disables sliding window, capping to sliding window size.                        |
-| `USE_V2_BLOCK_MANAGER`         | False               | `bool`                                                      | Use BlockSpaceMangerV2.                                                         |
 | `NUM_LOOKAHEAD_SLOTS`          | 0                   | `int`                                                       | Experimental scheduling config necessary for speculative decoding.              |
 | `SEED`                         | 0                   | `int`                                                       | Random seed for operations.                                                     |
 | `NUM_GPU_BLOCKS_OVERRIDE`      | None                | `int`                                                       | If specified, ignore GPU profiling result and use this number of GPU blocks.    |
@@ -57,24 +56,36 @@ Complete guide to all environment variables and configuration options for worker
 | `FULLY_SHARDED_LORAS`       | False   | `bool`                                     | Enable fully sharded LoRA layers.                                                                         |
 | `LORA_MODULES`              | `[]`    | `list[dict]`                               | Add lora adapters from Hugging Face `[{"name": "xx", "path": "xxx/xxxx", "base_model_name": "xxx/xxxx"}]` |
 
+> **Note (Serverless)**: When LoRA adapters are configured via `LORA_MODULES`, initialization is deferred to the first request to ensure compatibility with RunPod Serverless. This means the first request will include LoRA loading time. Subsequent requests are unaffected. Check logs for "LoRA mode: X adapter(s) will load on first request" at startup.
+
 ## Speculative Decoding Settings
 
-| Variable                                         | Default             | Type/Choices                                        | Description                                                                               |
-| ------------------------------------------------ | ------------------- | --------------------------------------------------- | ----------------------------------------------------------------------------------------- |
-| `SCHEDULER_DELAY_FACTOR`                         | 0.0                 | `float`                                             | Apply a delay before scheduling next prompt.                                              |
-| `ENABLE_CHUNKED_PREFILL`                         | False               | `bool`                                              | Enable chunked prefill requests.                                                          |
-| `SPECULATIVE_MODEL`                              | None                | `str`                                               | The name of the draft model to be used in speculative decoding.                           |
-| `NUM_SPECULATIVE_TOKENS`                         | None                | `int`                                               | The number of speculative tokens to sample from the draft model.                          |
-| `SPECULATIVE_DRAFT_TENSOR_PARALLEL_SIZE`         | None                | `int`                                               | Number of tensor parallel replicas for the draft model.                                   |
-| `SPECULATIVE_MAX_MODEL_LEN`                      | None                | `int`                                               | The maximum sequence length supported by the draft model.                                 |
-| `SPECULATIVE_DISABLE_BY_BATCH_SIZE`              | None                | `int`                                               | Disable speculative decoding if the number of enqueue requests is larger than this value. |
-| `NGRAM_PROMPT_LOOKUP_MAX`                        | None                | `int`                                               | Max size of window for ngram prompt lookup in speculative decoding.                       |
-| `NGRAM_PROMPT_LOOKUP_MIN`                        | None                | `int`                                               | Min size of window for ngram prompt lookup in speculative decoding.                       |
-| `SPEC_DECODING_ACCEPTANCE_METHOD`                | 'rejection_sampler' | ['rejection_sampler', 'typical_acceptance_sampler'] | Specify the acceptance method for draft token verification in speculative decoding.       |
-| `TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_THRESHOLD` | None                | `float`                                             | Set the lower bound threshold for the posterior probability of a token to be accepted.    |
-| `TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_ALPHA`     | None                | `float`                                             | A scaling factor for the entropy-based threshold for token acceptance.                    |
+Speculative decoding can be configured in two ways:
 
-## System Performance Settings
+### Option 1: JSON Configuration
+
+Set `SPECULATIVE_CONFIG` to a JSON string with your full speculative decoding configuration:
+
+```bash
+SPECULATIVE_CONFIG='{"method": "ngram", "num_speculative_tokens": 5, "prompt_lookup_max": 4}'
+```
+
+### Option 2: Individual Environment Variables
+
+| Variable                                 | Default | Type/Choices                                                       | Description                                                                               |
+| ---------------------------------------- | ------- | ------------------------------------------------------------------ | ----------------------------------------------------------------------------------------- |
+| `SPECULATIVE_METHOD`                     | None    | ['draft_model', 'ngram', 'eagle', 'eagle3', 'medusa', 'mlp_speculator'] | Speculative decoding method to use.                                                       |
+| `SPECULATIVE_MODEL`                      | None    | `str`                                                              | The name of the draft model to be used in speculative decoding.                           |
+| `NUM_SPECULATIVE_TOKENS`                 | None    | `int`                                                              | The number of speculative tokens to sample from the draft model.                          |
+| `SPECULATIVE_DRAFT_TENSOR_PARALLEL_SIZE` | None    | `int`                                                              | Number of tensor parallel replicas for the draft model.                                   |
+| `SPECULATIVE_MAX_MODEL_LEN`              | None    | `int`                                                              | The maximum sequence length supported by the draft model.                                 |
+| `SPECULATIVE_DISABLE_BY_BATCH_SIZE`      | None    | `int`                                                              | Disable speculative decoding if the number of enqueue requests is larger than this value. |
+| `NGRAM_PROMPT_LOOKUP_MAX`                | None    | `int`                                                              | Max size of window for ngram prompt lookup in speculative decoding.                       |
+| `NGRAM_PROMPT_LOOKUP_MIN`                | None    | `int`                                                              | Min size of window for ngram prompt lookup in speculative decoding.                       |
+
+If `SPECULATIVE_CONFIG` is set, it takes priority over individual env vars. When using individual env vars without `SPECULATIVE_METHOD`, the method is auto-detected from the model name or configuration.
+
+## Scheduling & Performance Settings
 
 | Variable                       | Default | Type/Choices    | Description                                                                                                                         |
 | ------------------------------ | ------- | --------------- | ----------------------------------------------------------------------------------------------------------------------------------- |
@@ -85,6 +96,10 @@ Complete guide to all environment variables and configuration options for worker
 | `ENFORCE_EAGER`                | False   | `bool`          | Always use eager-mode PyTorch. If False(`0`), will use eager mode and CUDA graph in hybrid for maximal performance and flexibility. |
 | `MAX_SEQ_LEN_TO_CAPTURE`       | `8192`  | `int`           | Maximum context length covered by CUDA graphs. When a sequence has context length larger than this, we fall back to eager mode.     |
 | `DISABLE_CUSTOM_ALL_REDUCE`    | `0`     | `int`           | Enables or disables custom all reduce.                                                                                              |
+| `ENABLE_EXPERT_PARALLEL`       | `False` | `bool`          | Enable Expert Parallel for MoE models.                                                                                              |
+| `ATTENTION_BACKEND`            | `None`  | `str`           | Attention backend to use (e.g., `FLASH_ATTN`, `FLASHINFER`, `TRITON_FLASH_ATTN`). Replaces deprecated `VLLM_ATTENTION_BACKEND`.     |
+| `ASYNC_SCHEDULING`             | `None`  | `bool`          | Enable async scheduling (overlaps engine scheduling with GPU execution). Default: enabled in vLLM 0.14.0+. Set to `false` to disable. |
+| `STREAM_INTERVAL`              | `1`     | `int`           | Controls how often to yield streaming results. Lower = more frequent updates.                                                        |
 
 ## Tokenizer Settings
 
@@ -114,6 +129,13 @@ The way this works is that the first request will have a batch size of `DEFAULT_
 | `ENABLE_AUTO_TOOL_CHOICE`           | `false`     | `bool`           | Enables automatic tool selection for supported models. Set to `true` to activate.                                                                                                                                 |
 | `TOOL_CALL_PARSER`                  | `None`      | `str`            | Specifies the parser for tool calls. Options: `mistral`, `hermes`, `llama3_json`, `llama4_json`, `llama4_pythonic`, `granite`, `granite-20b-fc`, `deepseek_v3`, `internlm`, `jamba`, `phi4_mini_json`, `pythonic` |
 | `REASONING_PARSER`                  | `None`      | `str`            | Parser for reasoning-capable models (enables reasoning mode). Examples: `deepseek_r1`, `qwen3`, `granite`, `hunyuan_a13b`. Leave unset to disable.                                                                |
+| `TRUST_REQUEST_CHAT_TEMPLATE`       | `false`     | `bool`           | Allow clients to send custom chat templates in API requests. **Security consideration:** Only enable if you trust your API clients.                                                                               |
+| `RETURN_TOKENS_AS_TOKEN_IDS`        | `false`     | `bool`           | Return token IDs instead of decoded text strings in responses.                                                                                                                                                     |
+| `EXCLUDE_TOOLS_WHEN_TOOL_CHOICE_NONE` | `false`   | `bool`           | Exclude tool definitions from the prompt when `tool_choice` is set to `none`.                                                                                                                                      |
+| `ENABLE_PROMPT_TOKENS_DETAILS`      | `false`     | `bool`           | Include detailed prompt token information in API responses.                                                                                                                                                        |
+| `ENABLE_FORCE_INCLUDE_USAGE`        | `false`     | `bool`           | Always include usage statistics in API responses, even when not requested.                                                                                                                                         |
+| `ENABLE_LOG_OUTPUTS`                | `false`     | `bool`           | Log model outputs for debugging purposes.                                                                                                                                                                          |
+| `LOG_ERROR_STACK`                   | `false`     | `bool`           | Include full stack traces in error responses for debugging.                                                                                                                                                        |
 
 ## Serverless & Concurrency Settings
 
@@ -121,7 +143,7 @@ The way this works is that the first request will have a batch size of `DEFAULT_
 | ---------------------- | ------- | ------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `MAX_CONCURRENCY`      | `30`    | `int`        | Max concurrent requests per worker. vLLM has an internal queue, so you don't have to worry about limiting by VRAM, this is for improving scaling/load balancing efficiency |
 | `DISABLE_LOG_STATS`    | False   | `bool`       | Enables or disables vLLM stats logging.                                                                                                                                    |
-| `DISABLE_LOG_REQUESTS` | False   | `bool`       | Enables or disables vLLM request logging.                                                                                                                                  |
+| `ENABLE_LOG_REQUESTS`  | False   | `bool`       | Enables vLLM request logging. (Replaces deprecated `DISABLE_LOG_REQUESTS` in vLLM 0.15.0)                                                                                  |
 
 ## Advanced Settings
 
@@ -147,7 +169,11 @@ These variables are used when building custom Docker images with models baked in
 
 ⚠️ **The following variables are deprecated and will be removed in future versions:**
 
-| Old Variable                 | New Variable             | Note                  |
-| ---------------------------- | ------------------------ | --------------------- |
-| `MAX_CONTEXT_LEN_TO_CAPTURE` | `MAX_SEQ_LEN_TO_CAPTURE` | Use new variable name |
-| `kv_cache_dtype=fp8_e5m2`    | `kv_cache_dtype=fp8`     | Simplified fp8 format |
+| Old Variable                 | New Variable             | Note                                                                 |
+| ---------------------------- | ------------------------ | -------------------------------------------------------------------- |
+| `MAX_CONTEXT_LEN_TO_CAPTURE` | `MAX_SEQ_LEN_TO_CAPTURE` | Use new variable name                                                |
+| `kv_cache_dtype=fp8_e5m2`    | `kv_cache_dtype=fp8`     | Simplified fp8 format                                                |
+| `USE_V2_BLOCK_MANAGER`       | *(removed)*              | V2 block manager is now the default in vLLM 0.13.0, setting ignored  |
+| `VLLM_ATTENTION_BACKEND`     | `ATTENTION_BACKEND`      | Use new env var name (old still works with deprecation warning)      |
+| `DISABLE_LOG_REQUESTS`       | `ENABLE_LOG_REQUESTS`    | Inverted logic in vLLM 0.15.0 (old still works with deprecation warning) |
+

src/engine.py

@@ -1,24 +1,25 @@
-import os
-import logging
-import json
 import asyncio
+import json
+import logging
+import os
+import time
+from typing import AsyncGenerator, Optional
 
 from dotenv import load_dotenv
-from typing import AsyncGenerator, Optional
-import time
-
 from vllm import AsyncLLMEngine
 from vllm.entrypoints.logger import RequestLogger
-from vllm.entrypoints.openai.serving_chat import OpenAIServingChat
-from vllm.entrypoints.openai.serving_completion import OpenAIServingCompletion
-from vllm.entrypoints.openai.protocol import ChatCompletionRequest, CompletionRequest, ErrorResponse
-from vllm.entrypoints.openai.serving_models import BaseModelPath, LoRAModulePath, OpenAIServingModels
+from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
+from vllm.entrypoints.openai.chat_completion.serving import OpenAIServingChat
+from vllm.entrypoints.openai.completion.protocol import CompletionRequest
+from vllm.entrypoints.openai.completion.serving import OpenAIServingCompletion
+from vllm.entrypoints.openai.engine.protocol import ErrorResponse
+from vllm.entrypoints.openai.models.protocol import BaseModelPath, LoRAModulePath
+from vllm.entrypoints.openai.models.serving import OpenAIServingModels
 
-
-from utils import DummyRequest, JobInput, BatchSize, create_error_response
-from constants import DEFAULT_MAX_CONCURRENCY, DEFAULT_BATCH_SIZE, DEFAULT_BATCH_SIZE_GROWTH_FACTOR, DEFAULT_MIN_BATCH_SIZE
-from tokenizer import TokenizerWrapper
+from constants import DEFAULT_BATCH_SIZE, DEFAULT_BATCH_SIZE_GROWTH_FACTOR, DEFAULT_MAX_CONCURRENCY, DEFAULT_MIN_BATCH_SIZE
 from engine_args import get_engine_args
+from tokenizer import TokenizerWrapper
+from utils import BatchSize, DummyRequest, JobInput, create_error_response
 
 class vLLMEngine:
     def __init__(self, engine = None):
@@ -174,10 +175,24 @@ class vLLMEngine:
 class OpenAIvLLMEngine(vLLMEngine):
     def __init__(self, vllm_engine):
         super().__init__(vllm_engine)
-        self.served_model_name = os.getenv("OPENAI_SERVED_MODEL_NAME_OVERRIDE") or self.engine_args.model
+        self.served_model_name = os.getenv("OPENAI_SERVED_MODEL_NAME_OVERRIDE") or self.engine_args.served_model_name or self.engine_args.model
         self.response_role = os.getenv("OPENAI_RESPONSE_ROLE") or "assistant"
         self.lora_adapters = self._load_lora_adapters()
-        asyncio.run(self._initialize_engines())
+
+        # Always defer OpenAI engine initialization to the first request.
+        # asyncio.run() creates a temporary event loop that gets closed, but async
+        # components (tokenizer pool, serving engines) bind futures to that loop.
+        # When Runpod's serverless handler runs in its own event loop, those futures
+        # are "attached to a different loop" causing RuntimeError.
+        # This affects all configurations, not just LoRA.
+        self._engines_initialized = False
+        if self.lora_adapters:
+            logging.info(f"LoRA mode: {len(self.lora_adapters)} adapter(s) will load on first request")
+            for adapter in self.lora_adapters:
+                logging.info(f"  - {adapter.name}: {adapter.path}")
+        else:
+            logging.info("OpenAI engines will initialize on first request")
+
         # Handle both integer and boolean string values for RAW_OPENAI_OUTPUT
         raw_output_env = os.getenv("RAW_OPENAI_OUTPUT", "1")
         if raw_output_env.lower() in ('true', 'false'):
@@ -201,15 +216,28 @@ class OpenAIvLLMEngine(vLLMEngine):
                 continue
         return adapters
 
+    async def _ensure_engines_initialized(self):
+        """Initialize engines on first request to avoid event loop mismatch.
+
+        In Runpod Serverless, the startup code runs outside the handler's event
+        loop. Deferring initialization to the first request ensures all async
+        components (tokenizer pool, serving engines, LoRA state) are created in
+        the correct event loop context.
+        """
+        if not self._engines_initialized:
+            logging.info("Initializing OpenAI serving engines...")
+            await self._initialize_engines()
+            self._engines_initialized = True
+            logging.info("OpenAI serving engines initialized successfully")
+
     async def _initialize_engines(self):
-        self.model_config = await self.llm.get_model_config()
+        self.model_config = self.llm.model_config
         self.base_model_paths = [
-            BaseModelPath(name=self.engine_args.model, model_path=self.engine_args.model)
+            BaseModelPath(name=self.served_model_name, model_path=self.engine_args.model)
         ]
 
         self.serving_models = OpenAIServingModels(
             engine_client=self.llm,
-            model_config=self.model_config,
             base_model_paths=self.base_model_paths,
             lora_modules=self.lora_adapters,
         )
@@ -222,28 +250,39 @@ class OpenAIvLLMEngine(vLLMEngine):
         
         self.chat_engine = OpenAIServingChat(
             engine_client=self.llm, 
-            model_config=self.model_config,
             models=self.serving_models,
             response_role=self.response_role,
             request_logger=None,
             chat_template=chat_template,
             chat_template_content_format="auto",
-            # enable_reasoning=os.getenv('ENABLE_REASONING', 'false').lower() == 'true',
-            reasoning_parser= os.getenv('REASONING_PARSER', "") or None,
-            # return_token_as_token_ids=False,
+            trust_request_chat_template=os.getenv('TRUST_REQUEST_CHAT_TEMPLATE', 'false').lower() == 'true',
+            return_tokens_as_token_ids=os.getenv('RETURN_TOKENS_AS_TOKEN_IDS', 'false').lower() == 'true',
+            reasoning_parser=os.getenv('REASONING_PARSER', "") or "",
             enable_auto_tools=os.getenv('ENABLE_AUTO_TOOL_CHOICE', 'false').lower() == 'true',
+            exclude_tools_when_tool_choice_none=os.getenv('EXCLUDE_TOOLS_WHEN_TOOL_CHOICE_NONE', 'false').lower() == 'true',
             tool_parser=os.getenv('TOOL_CALL_PARSER', "") or None,
-            enable_prompt_tokens_details=False
+            enable_prompt_tokens_details=os.getenv('ENABLE_PROMPT_TOKENS_DETAILS', 'false').lower() == 'true',
+            enable_force_include_usage=os.getenv('ENABLE_FORCE_INCLUDE_USAGE', 'false').lower() == 'true',
+            enable_log_outputs=os.getenv('ENABLE_LOG_OUTPUTS', 'false').lower() == 'true',
+            log_error_stack=os.getenv('LOG_ERROR_STACK', 'false').lower() == 'true',
         )
         self.completion_engine = OpenAIServingCompletion(
-            engine_client=self.llm, 
-            model_config=self.model_config,
+            engine_client=self.llm,
             models=self.serving_models,
             request_logger=None,
-            # return_token_as_token_ids=False,
+            return_tokens_as_token_ids=os.getenv('RETURN_TOKENS_AS_TOKEN_IDS', 'false').lower() == 'true',
+            enable_prompt_tokens_details=os.getenv('ENABLE_PROMPT_TOKENS_DETAILS', 'false').lower() == 'true',
+            enable_force_include_usage=os.getenv('ENABLE_FORCE_INCLUDE_USAGE', 'false').lower() == 'true',
+            log_error_stack=os.getenv('LOG_ERROR_STACK', 'false').lower() == 'true',
         )
-    
+
+        if hasattr(self.chat_engine, 'warmup'):
+            await self.chat_engine.warmup()
+
     async def generate(self, openai_request: JobInput):
+        # Ensure engines are ready (no-op if already initialized at startup)
+        await self._ensure_engines_initialized()
+
         if openai_request.openai_route == "/v1/models":
             yield await self._handle_model_request()
         elif openai_request.openai_route in ["/v1/chat/completions", "/v1/completions"]:

src/engine_args.py

@@ -15,7 +15,8 @@ RENAME_ARGS_MAP = {
 
 DEFAULT_ARGS = {
     "disable_log_stats": os.getenv('DISABLE_LOG_STATS', 'False').lower() == 'true',
-    "disable_log_requests": os.getenv('DISABLE_LOG_REQUESTS', 'False').lower() == 'true',
+    # disable_log_requests is deprecated, use enable_log_requests instead
+    "enable_log_requests": os.getenv('ENABLE_LOG_REQUESTS', 'False').lower() == 'true',
     "gpu_memory_utilization": float(os.getenv('GPU_MEMORY_UTILIZATION', 0.95)),
     "pipeline_parallel_size": int(os.getenv('PIPELINE_PARALLEL_SIZE', 1)),
     "tensor_parallel_size": int(os.getenv('TENSOR_PARALLEL_SIZE', 1)),
@@ -38,9 +39,15 @@ DEFAULT_ARGS = {
     "block_size": int(os.getenv('BLOCK_SIZE', 16)),
     "enable_prefix_caching": os.getenv('ENABLE_PREFIX_CACHING', 'False').lower() == 'true',
     "disable_sliding_window": os.getenv('DISABLE_SLIDING_WINDOW', 'False').lower() == 'true',
-    "use_v2_block_manager": os.getenv('USE_V2_BLOCK_MANAGER', 'False').lower() == 'true',
+    # attention_backend replaces deprecated VLLM_ATTENTION_BACKEND env var
+    "attention_backend": os.getenv('ATTENTION_BACKEND', None),
+    # Enabled by default for improved throughput. Set to False to disable if experiencing issues
+    "async_scheduling": None if os.getenv('ASYNC_SCHEDULING') is None else os.getenv('ASYNC_SCHEDULING', 'True').lower() == 'true',
+    # Controls how often to yield streaming results
+    "stream_interval": int(os.getenv('STREAM_INTERVAL', 1)),
     "swap_space": int(os.getenv('SWAP_SPACE', 4)),  # GiB
     "cpu_offload_gb": int(os.getenv('CPU_OFFLOAD_GB', 0)),  # GiB
+    # vLLM defaults None to 2048; keep 0 as None to let vLLM auto-calculate
     "max_num_batched_tokens": int(os.getenv('MAX_NUM_BATCHED_TOKENS', 0)) or None,
     "max_num_seqs": int(os.getenv('MAX_NUM_SEQS', 256)),
     "max_logprobs": int(os.getenv('MAX_LOGPROBS', 20)),  # Default value for OpenAI Chat Completions API
@@ -80,6 +87,7 @@ DEFAULT_ARGS = {
     "guided_decoding_backend": os.getenv('GUIDED_DECODING_BACKEND', 'outlines'),
     "speculative_model": os.getenv('SPECULATIVE_MODEL', None),
     "speculative_draft_tensor_parallel_size": int(os.getenv('SPECULATIVE_DRAFT_TENSOR_PARALLEL_SIZE', 0)) or None,
+    "enable_expert_parallel": bool(os.getenv('ENABLE_EXPERT_PARALLEL', 'False').lower() == 'true'),
     "num_speculative_tokens": int(os.getenv('NUM_SPECULATIVE_TOKENS', 0)) or None,
     "speculative_max_model_len": int(os.getenv('SPECULATIVE_MAX_MODEL_LEN', 0)) or None,
     "speculative_disable_by_batch_size": int(os.getenv('SPECULATIVE_DISABLE_BY_BATCH_SIZE', 0)) or None,
@@ -91,8 +99,116 @@ DEFAULT_ARGS = {
     "qlora_adapter_name_or_path": os.getenv('QLORA_ADAPTER_NAME_OR_PATH', None),
     "disable_logprobs_during_spec_decoding": os.getenv('DISABLE_LOGPROBS_DURING_SPEC_DECODING', None),
     "otlp_traces_endpoint": os.getenv('OTLP_TRACES_ENDPOINT', None),
-    "use_v2_block_manager": os.getenv('USE_V2_BLOCK_MANAGER', 'true'),
 }
+
+def get_speculative_config():
+    """Build speculative decoding configuration from environment variables.
+
+    Supports two modes:
+    1. Full JSON config via SPECULATIVE_CONFIG env var
+    2. Individual env vars for common settings
+    """
+    # Option 1: Full JSON configuration
+    spec_config_json = os.getenv('SPECULATIVE_CONFIG')
+    if spec_config_json:
+        try:
+            config = json.loads(spec_config_json)
+            logging.info(f"Using speculative config from SPECULATIVE_CONFIG: {config}")
+            return config
+        except json.JSONDecodeError as e:
+            logging.error(f"Failed to parse SPECULATIVE_CONFIG JSON: {e}")
+            return None
+
+    # Option 2: Build config from individual environment variables
+    spec_method = os.getenv('SPECULATIVE_METHOD')
+    spec_model = os.getenv('SPECULATIVE_MODEL')
+    _num_spec_tokens = os.getenv('NUM_SPECULATIVE_TOKENS')
+    _ngram_max = os.getenv('NGRAM_PROMPT_LOOKUP_MAX')
+    _ngram_min = os.getenv('NGRAM_PROMPT_LOOKUP_MIN')
+
+    # Convert numeric vars to int so '0' (hub.json default) is treated as unset
+    num_spec_tokens = (int(_num_spec_tokens) or None) if _num_spec_tokens else None
+    ngram_max = (int(_ngram_max) or None) if _ngram_max else None
+    ngram_min = (int(_ngram_min) or None) if _ngram_min else None
+
+    if not any([spec_method, spec_model, ngram_max]):
+        return None
+
+    config = {}
+
+    # Determine method
+    if spec_method:
+        config['method'] = spec_method
+    elif ngram_max and not spec_model:
+        config['method'] = 'ngram'
+    elif spec_model:
+        model_lower = spec_model.lower()
+        if 'eagle3' in model_lower:
+            config['method'] = 'eagle3'
+        elif 'eagle' in model_lower:
+            config['method'] = 'eagle'
+        elif 'medusa' in model_lower:
+            config['method'] = 'medusa'
+        else:
+            config['method'] = 'draft_model'
+
+    if spec_model:
+        config['model'] = spec_model
+    if num_spec_tokens:
+        config['num_speculative_tokens'] = num_spec_tokens
+    if ngram_max:
+        config['prompt_lookup_max'] = ngram_max
+    if ngram_min:
+        config['prompt_lookup_min'] = ngram_min
+
+    draft_tp = os.getenv('SPECULATIVE_DRAFT_TENSOR_PARALLEL_SIZE')
+    if draft_tp:
+        config['draft_tensor_parallel_size'] = int(draft_tp)
+
+    spec_max_len = os.getenv('SPECULATIVE_MAX_MODEL_LEN')
+    if spec_max_len:
+        config['max_model_len'] = int(spec_max_len)
+
+    disable_batch = os.getenv('SPECULATIVE_DISABLE_BY_BATCH_SIZE')
+    if disable_batch:
+        config['disable_by_batch_size'] = int(disable_batch)
+
+    spec_quant = os.getenv('SPECULATIVE_QUANTIZATION')
+    if spec_quant:
+        config['quantization'] = spec_quant
+
+    spec_revision = os.getenv('SPECULATIVE_MODEL_REVISION')
+    if spec_revision:
+        config['revision'] = spec_revision
+
+    spec_eager = os.getenv('SPECULATIVE_ENFORCE_EAGER')
+    if spec_eager:
+        config['enforce_eager'] = spec_eager.lower() == 'true'
+
+    if config:
+        logging.info(f"Built speculative config from env vars: {config}")
+        return config
+
+    return None
+
+def _resolve_max_model_len(model, trust_remote_code=False, revision=None):
+    """Resolve max_model_len from the model's HuggingFace config."""
+    try:
+        from transformers import AutoConfig
+        config = AutoConfig.from_pretrained(
+            model,
+            trust_remote_code=trust_remote_code,
+            revision=revision,
+        )
+        for attr in ('max_position_embeddings', 'n_positions', 'max_seq_len', 'seq_length'):
+            val = getattr(config, attr, None)
+            if val is not None:
+                logging.info(f"Resolved max_model_len={val} from model config ({attr})")
+                return val
+    except Exception as e:
+        logging.warning(f"Could not resolve max_model_len from model config: {e}")
+    return None
+
 limit_mm_env = os.getenv('LIMIT_MM_PER_PROMPT')
 if limit_mm_env is not None:
     DEFAULT_ARGS["limit_mm_per_prompt"] = convert_limit_mm_per_prompt(limit_mm_env)
@@ -174,5 +290,50 @@ def get_engine_args():
     # if "gemma-2" in args.get("model", "").lower():
     #     os.environ["VLLM_ATTENTION_BACKEND"] = "FLASHINFER"
     #     logging.info("Using FLASHINFER for gemma-2 model.")
-        
+    
+    # Set max_num_batched_tokens to max_model_len for unlimited batching.
+    # vLLM defaults max_num_batched_tokens to 2048 when None, which is too low.
+
+    if args.get("max_model_len") == 0:
+        args["max_model_len"] = None
+
+    if args.get("max_num_batched_tokens") == 0:
+        args["max_num_batched_tokens"] = None
+
+    if args.get("max_num_batched_tokens") is None:
+        max_model_len = args.get("max_model_len")
+        if max_model_len is None:
+            max_model_len = _resolve_max_model_len(
+                args.get("model"),
+                trust_remote_code=args.get("trust_remote_code", False),
+                revision=args.get("revision"),
+            )
+        if max_model_len is not None:
+            args["max_num_batched_tokens"] = max_model_len
+            logging.info(f"Setting max_num_batched_tokens to {max_model_len}")
+    
+    # VLLM_ATTENTION_BACKEND is deprecated, migrate to attention_backend
+    if os.getenv('VLLM_ATTENTION_BACKEND'):
+        logging.warning(
+            "VLLM_ATTENTION_BACKEND env var is deprecated. "
+            "Use ATTENTION_BACKEND instead (maps to --attention-backend CLI arg)."
+        )
+        if not args.get('attention_backend'):
+            args['attention_backend'] = os.getenv('VLLM_ATTENTION_BACKEND')
+
+    # DISABLE_LOG_REQUESTS is deprecated, use ENABLE_LOG_REQUESTS instead
+    if os.getenv('DISABLE_LOG_REQUESTS'):
+        logging.warning(
+            "DISABLE_LOG_REQUESTS env var is deprecated. "
+            "Use ENABLE_LOG_REQUESTS instead (default: False)."
+        )
+        # Honor old behavior: if DISABLE_LOG_REQUESTS=true, don't enable logging
+        if os.getenv('DISABLE_LOG_REQUESTS', 'False').lower() == 'true':
+            args['enable_log_requests'] = False
+
+    # Add speculative decoding configuration if present
+    speculative_config = get_speculative_config()
+    if speculative_config:
+        args["speculative_config"] = speculative_config
+
     return AsyncEngineArgs(**args)

src/handler.py

@@ -1,22 +1,55 @@
-import os
+import sys
+import multiprocessing
+import traceback
 import runpod
-from utils import JobInput
-from engine import vLLMEngine, OpenAIvLLMEngine
+from runpod import RunPodLogger
+
+log = RunPodLogger()
+
+vllm_engine = None
+openai_engine = None
 
-vllm_engine = vLLMEngine()
-OpenAIvLLMEngine = OpenAIvLLMEngine(vllm_engine)
 
 async def handler(job):
-    job_input = JobInput(job["input"])
-    engine = OpenAIvLLMEngine if job_input.openai_route else vllm_engine
-    results_generator = engine.generate(job_input)
-    async for batch in results_generator:
-        yield batch
+    try:
+        from utils import JobInput
+        job_input = JobInput(job["input"])
+        engine = openai_engine if job_input.openai_route else vllm_engine
+        results_generator = engine.generate(job_input)
+        async for batch in results_generator:
+            yield batch
+    except Exception as e:
+        error_str = str(e)
+        full_traceback = traceback.format_exc()
 
-runpod.serverless.start(
-    {
-        "handler": handler,
-        "concurrency_modifier": lambda x: vllm_engine.max_concurrency,
-        "return_aggregate_stream": True,
-    }
-)
+        log.error(f"Error during inference: {error_str}")
+        log.error(f"Full traceback:\n{full_traceback}")
+
+        # CUDA errors = worker is broken, exit to let RunPod spin up a healthy one
+        if "CUDA" in error_str or "cuda" in error_str:
+            log.error("Terminating worker due to CUDA/GPU error")
+            sys.exit(1)
+
+        yield {"error": error_str}
+
+
+# Only run in main process to prevent re-initialization when vLLM spawns worker subprocesses
+if __name__ == "__main__" or multiprocessing.current_process().name == "MainProcess":
+
+    try:
+        from engine import vLLMEngine, OpenAIvLLMEngine
+
+        vllm_engine = vLLMEngine()
+        openai_engine = OpenAIvLLMEngine(vllm_engine)
+        log.info("vLLM engines initialized successfully")
+    except Exception as e:
+        log.error(f"Worker startup failed: {e}\n{traceback.format_exc()}")
+        sys.exit(1)
+
+    runpod.serverless.start(
+        {
+            "handler": handler,
+            "concurrency_modifier": lambda x: vllm_engine.max_concurrency if vllm_engine else 1,
+            "return_aggregate_stream": True,
+        }
+    )

src/utils.py

@@ -3,11 +3,10 @@ import logging
 from http import HTTPStatus
 from functools import wraps
 from time import time
-from vllm.entrypoints.openai.protocol import RequestResponseMetadata
 
 try:
     from vllm.utils import random_uuid
-    from vllm.entrypoints.openai.protocol import ErrorResponse
+    from vllm.entrypoints.openai.engine.protocol import ErrorResponse, ErrorInfo, RequestResponseMetadata
     from vllm import SamplingParams
 except ImportError:
     logging.warning("Error importing vllm, skipping related imports. This is ONLY expected when baking model into docker image from a machine without GPUs")
@@ -88,9 +87,9 @@ class BatchSize:
             self.current_batch_size = min(self.current_batch_size*self.batch_size_growth_factor, self.max_batch_size)
         
 def create_error_response(message: str, err_type: str = "BadRequestError", status_code: HTTPStatus = HTTPStatus.BAD_REQUEST) -> ErrorResponse:
-    return ErrorResponse(message=message,
-                            type=err_type,
-                            code=status_code.value)
+    return ErrorResponse(error=ErrorInfo(message=message,
+                                         type=err_type,
+                                         code=status_code.value))
     
 def get_int_bool_env(env_var: str, default: bool) -> bool:
     return int(os.getenv(env_var, int(default))) == 1