feat: update dockerfile to 12.9.1 (#267)

* feat: update dockerfile to 12.9.1

* update readme on VLLM_NIGHTLY build arg

.runpod/hub.json

@@ -187,6 +187,7 @@
           "name": "Max Model Length",
           "type": "number",
           "description": "Model context length.",
+          "default": null,
           "advanced": true
         }
       },
@@ -206,7 +207,8 @@
               "value": "mp"
             }
           ],
-          "advanced": true
+          "advanced": true,
+          "default": "mp"
         }
       },
       {
@@ -293,6 +295,7 @@
           "name": "Max Num Batched Tokens",
           "type": "number",
           "description": "Maximum number of batched tokens per iteration.",
+          "default": null,
           "advanced": true
         }
       },
@@ -490,6 +493,61 @@
           "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": {
+          "name": "Speculative Model",
+          "type": "string",
+          "description": "The name of the draft model to be used in speculative decoding.",
+          "advanced": true
+        }
+      },
+      {
+        "key": "NUM_SPECULATIVE_TOKENS",
+        "input": {
+          "name": "Num Speculative Tokens",
+          "type": "number",
+          "description": "The number of speculative tokens to sample from the draft model.",
+          "advanced": true
+        }
+      },
+      {
+        "key": "NGRAM_PROMPT_LOOKUP_MAX",
+        "input": {
+          "name": "Ngram Prompt Lookup Max",
+          "type": "number",
+          "description": "Max size of window for ngram prompt lookup in speculative decoding.",
+          "advanced": true
+        }
+      },
       {
         "key": "MODEL_LOADER_EXTRA_CONFIG",
         "input": {

Dockerfile

@@ -1,13 +1,13 @@
-FROM nvidia/cuda:12.8.0-base-ubuntu22.04 
+FROM nvidia/cuda:12.9.1-base-ubuntu22.04 
 
 RUN apt-get update -y \
     && apt-get install -y python3-pip
 
-RUN ldconfig /usr/local/cuda-12.8/compat/
+RUN ldconfig /usr/local/cuda-12.9/compat/
 
 # 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
+    python3 -m pip install "vllm[flashinfer]==0.15.0" --extra-index-url https://download.pytorch.org/whl/cu129
 
 
 
@@ -23,6 +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 \
@@ -44,6 +45,11 @@ ENV MODEL_NAME=$MODEL_NAME \
 
 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 \

README.md

@@ -80,6 +80,7 @@ To build an image with the model baked in, you must specify the following docker
   - `WORKER_CUDA_VERSION`: `12.1.0` (`12.1.0` is recommended for optimal performance).
   - `TOKENIZER_NAME`: Tokenizer repository if you would like to use a different tokenizer than the one that comes with the model. (default: `None`, which uses the model's tokenizer)
   - `TOKENIZER_REVISION`: Tokenizer revision to load (default: `main`).
+  - `VLLM_NIGHTLY`: Set to `true` to replace the pinned vLLM release with the latest nightly build and the latest `transformers` from source. Useful for testing unreleased vLLM features. (default: `false`)
 
 For the remaining settings, you may apply them as environment variables when running the container. Supported environment variables are listed in the [Environment Variables](#environment-variables) section.
 
@@ -89,6 +90,20 @@ For the remaining settings, you may apply them as environment variables when run
 docker build -t username/image:tag --build-arg MODEL_NAME="openchat/openchat_3.5" --build-arg BASE_PATH="/models" .
 ```
 
+### Example: Building with vLLM Nightly
+
+To use the latest unreleased vLLM build (installs from the nightly wheel index and `transformers` from source):
+
+```bash
+docker build -t username/image:tag --build-arg VLLM_NIGHTLY=true .
+```
+
+You can combine it with other arguments:
+
+```bash
+docker build -t username/image:tag --build-arg VLLM_NIGHTLY=true --build-arg MODEL_NAME="meta-llama/Llama-3.1-8B-Instruct" --build-arg BASE_PATH="/models" .
+```
+
 ### (Optional) Including Huggingface Token
 
 If the model you would like to deploy is private or gated, you will need to include it during build time as a Docker secret, which will protect it from being exposed in the image and on DockerHub.

docs/configuration.md

@@ -60,22 +60,32 @@ Complete guide to all environment variables and configuration options for worker
 
 ## Speculative Decoding Settings
 
-| Variable                                         | Default             | Type/Choices                                        | Description                                                                               |
+Speculative decoding can be configured in two ways:
-| ------------------------------------------------ | ------------------- | --------------------------------------------------- | ----------------------------------------------------------------------------------------- |
-| `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.                    |
 
-## 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                                                                                                                         |
 | ------------------------------ | ------- | --------------- | ----------------------------------------------------------------------------------------------------------------------------------- |

src/engine.py

@@ -175,7 +175,7 @@ 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()
 
@@ -233,7 +233,7 @@ class OpenAIvLLMEngine(vLLMEngine):
     async def _initialize_engines(self):
         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(

src/engine_args.py

@@ -100,6 +100,115 @@ DEFAULT_ARGS = {
     "disable_logprobs_during_spec_decoding": os.getenv('DISABLE_LOGPROBS_DURING_SPEC_DECODING', None),
     "otlp_traces_endpoint": os.getenv('OTLP_TRACES_ENDPOINT', None),
 }
+
+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)
@@ -182,11 +291,26 @@ def get_engine_args():
     #     os.environ["VLLM_ATTENTION_BACKEND"] = "FLASHINFER"
     #     logging.info("Using FLASHINFER for gemma-2 model.")
     
-    # When max_num_batched_tokens is None (env var was 0), set to max_model_len
+    # Set max_num_batched_tokens to max_model_len for unlimited batching.
-    # to preserve "unlimited" behavior. vLLM defaults None to 2048.
+    # vLLM defaults max_num_batched_tokens to 2048 when None, which is too low.
-    if args.get("max_num_batched_tokens") is None and args.get("max_model_len") is not None:
+
-        args["max_num_batched_tokens"] = args["max_model_len"]
+    if args.get("max_model_len") == 0:
-        logging.info(f"Setting max_num_batched_tokens to max_model_len ({args['max_model_len']}) for unlimited batching.")
+        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'):
@@ -207,4 +331,9 @@ def get_engine_args():
         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/utils.py

@@ -6,7 +6,7 @@ from time import time
 
 try:
     from vllm.utils import random_uuid
-    from vllm.entrypoints.openai.engine.protocol import ErrorResponse, RequestResponseMetadata
+    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")
@@ -87,9 +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,
+    return ErrorResponse(error=ErrorInfo(message=message,
-                            type=err_type,
+                                         type=err_type,
-                            code=status_code.value)
+                                         code=status_code.value))
     
 def get_int_bool_env(env_var: str, default: bool) -> bool:
     return int(os.getenv(env_var, int(default))) == 1