Merge pull request #272 from runpod-workers/feat/vllm-0.16.0

feat: Update to 0.16.0

.runpod/README.md

@@ -28,6 +28,8 @@ All behaviour is controlled through environment variables:
 | `OPENAI_SERVED_MODEL_NAME_OVERRIDE` | Override served model name in API                 |                     | String                                                             |
 | `MAX_CONCURRENCY`                   | Maximum concurrent requests                       | 300                 | Integer                                                            |
 
+**Pass any vLLM engine arg** not listed above by setting an env var with the **UPPERCASED** field name (e.g. `MAX_MODEL_LEN=4096`, `ENABLE_CHUNKED_PREFILL=true`). The worker auto-discovers all `AsyncEngineArgs` fields from env. See the [vLLM engine args docs](https://docs.vllm.ai/en/latest/configuration/engine_args) for all available options.
+
 For complete configuration options, see the [full configuration documentation](https://github.com/runpod-workers/worker-vllm/blob/main/docs/configuration.md).
 
 ## API Usage

.runpod/hub.json

@@ -280,15 +280,6 @@
           "advanced": true
         }
       },
-      {
-        "key": "NUM_GPU_BLOCKS_OVERRIDE",
-        "input": {
-          "name": "Num GPU Blocks Override",
-          "type": "number",
-          "description": "If specified, ignore GPU profiling result and use this number of GPU blocks.",
-          "advanced": true
-        }
-      },
       {
         "key": "MAX_NUM_BATCHED_TOKENS",
         "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)
+# Install vLLM with FlashInfer - use CUDA 12.8 PyTorch wheels (compatible with vLLM 0.15.1)
 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]==0.16.0" --extra-index-url https://download.pytorch.org/whl/cu129
 
 
 

README.md

@@ -59,6 +59,16 @@ Configure worker-vllm using environment variables:
 | `OPENAI_SERVED_MODEL_NAME_OVERRIDE` | Override served model name in API                 |                     | String                                                             |
 | `MAX_CONCURRENCY`                   | Maximum concurrent requests                       | 30                  | Integer                                                            |
 
+**Pass any vLLM engine arg** not listed above by setting an environment variable with the **UPPERCASED** field name (same names vLLM uses). The worker auto-discovers all `AsyncEngineArgs` fields from env. For example:
+
+| Environment Variable      | vLLM Engine Arg          | Example Value |
+| ------------------------- | ------------------------ | ------------- |
+| `MAX_MODEL_LEN`           | `max_model_len`          | `4096`        |
+| `ENFORCE_EAGER`           | `enforce_eager`          | `true`        |
+| `ENABLE_CHUNKED_PREFILL`  | `enable_chunked_prefill` | `true`        |
+
+Any env var whose name matches a valid `AsyncEngineArgs` field (uppercased) is applied automatically. Backward-compat aliases: `MODEL_NAME`, `TOKENIZER_NAME`, `MAX_CONTEXT_LEN_TO_CAPTURE`. This lets you configure any vLLM option without waiting for explicit worker support.
+
 For the complete list of all available environment variables, examples, and detailed descriptions: **[Configuration](docs/configuration.md)**
 
 ## Option 2: Build Docker Image with Model Inside
@@ -80,6 +90,7 @@ 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 +100,20 @@ 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

@@ -156,6 +156,29 @@ The way this works is that the first request will have a batch size of `DEFAULT_
 | `DISABLE_LOGGING_REQUEST`   | False   | `bool`  | Disable logging requests.                                                                                                                              |
 | `MAX_LOG_LEN`               | None    | `int`   | Max number of prompt characters or prompt ID numbers being printed in log.                                                                             |
 
+## UPPERCASED env vars: Pass any engine arg
+
+Any vLLM `AsyncEngineArgs` field can be set via an environment variable using the **UPPERCASED** field name (the same names vLLM uses). The worker auto-discovers all fields from env — no prefix.
+
+**Format:** `<FIELD_NAME_UPPERCASED>=<value>` (e.g. `MAX_MODEL_LEN=4096`)
+
+**Examples:**
+
+| Environment Variable     | vLLM Engine Arg          | Value Example |
+| ------------------------ | ------------------------ | ------------- |
+| `MAX_MODEL_LEN`          | `max_model_len`          | `4096`        |
+| `ENFORCE_EAGER`          | `enforce_eager`          | `true`        |
+| `ENABLE_CHUNKED_PREFILL` | `enable_chunked_prefill` | `true`        |
+| `NUM_SCHEDULER_STEPS`    | `num_scheduler_steps`    | `8`           |
+| `TOKENIZER_POOL_SIZE`    | `tokenizer_pool_size`   | `4`           |
+
+**Backward-compat aliases:** `MODEL_NAME` → `model`, `TOKENIZER_NAME` → `tokenizer`, `MAX_CONTEXT_LEN_TO_CAPTURE` → `max_seq_len_to_capture`, `MODEL_REVISION` → `revision`.
+
+**Notes:**
+- Only valid `AsyncEngineArgs` fields are applied. Unknown keys are silently ignored.
+- Values are automatically cast to the correct type (`int`, `float`, `bool`, `str`, or JSON for `dict`/`list`/`tuple`).
+- For a full list of available engine args, see the [vLLM AsyncEngineArgs documentation](https://docs.vllm.ai/en/latest/configuration/engine_args/).
+
 ## Docker Build Arguments
 
 These variables are used when building custom Docker images with models baked in:

src/engine_args.py

@@ -1,106 +1,169 @@
 import os
 import json
 import logging
+from typing import get_origin, get_args
 from torch.cuda import device_count
 from vllm import AsyncEngineArgs
 from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
 from src.utils import convert_limit_mm_per_prompt
 
-RENAME_ARGS_MAP = {
+# Backward-compat: env var names users already know → engine arg name
+ENV_ALIASES = {
     "MODEL_NAME": "model",
     "MODEL_REVISION": "revision",
     "TOKENIZER_NAME": "tokenizer",
-    "MAX_CONTEXT_LEN_TO_CAPTURE": "max_seq_len_to_capture"
 }
 
+# Literal defaults from original worker (used when env/local do not set a value)
 DEFAULT_ARGS = {
-    "disable_log_stats": os.getenv('DISABLE_LOG_STATS', 'False').lower() == 'true',
+    "disable_log_stats": False,
-    # disable_log_requests is deprecated, use enable_log_requests instead
+    "enable_log_requests": False,
-    "enable_log_requests": os.getenv('ENABLE_LOG_REQUESTS', 'False').lower() == 'true',
+    "gpu_memory_utilization": 0.95,
-    "gpu_memory_utilization": float(os.getenv('GPU_MEMORY_UTILIZATION', 0.95)),
+    "pipeline_parallel_size": 1,
-    "pipeline_parallel_size": int(os.getenv('PIPELINE_PARALLEL_SIZE', 1)),
+    "tensor_parallel_size": 1,
-    "tensor_parallel_size": int(os.getenv('TENSOR_PARALLEL_SIZE', 1)),
+    "skip_tokenizer_init": False,
-    "served_model_name": os.getenv('SERVED_MODEL_NAME', None),
+    "tokenizer_mode": "auto",
-    "tokenizer": os.getenv('TOKENIZER', None),
+    "trust_remote_code": False,
-    "skip_tokenizer_init": os.getenv('SKIP_TOKENIZER_INIT', 'False').lower() == 'true',
+    "load_format": "auto",
-    "tokenizer_mode": os.getenv('TOKENIZER_MODE', 'auto'),
+    "dtype": "auto",
-    "trust_remote_code": os.getenv('TRUST_REMOTE_CODE', 'False').lower() == 'true',
+    "kv_cache_dtype": "auto",
-    "download_dir": os.getenv('DOWNLOAD_DIR', None),
+    "seed": 0,
-    "load_format": os.getenv('LOAD_FORMAT', 'auto'),
+    "worker_use_ray": False,
-    "config_format": os.getenv('CONFIG_FORMAT', 'auto'),
+    "block_size": 16,
-    "dtype": os.getenv('DTYPE', 'auto'),
+    "enable_prefix_caching": False,
-    "kv_cache_dtype": os.getenv('KV_CACHE_DTYPE', 'auto'),
+    "disable_sliding_window": False,
-    "quantization_param_path": os.getenv('QUANTIZATION_PARAM_PATH', None),
+    "swap_space": 4,
-    "seed": int(os.getenv('SEED', 0)),
+    "cpu_offload_gb": 0,
-    "max_model_len": int(os.getenv('MAX_MODEL_LEN', 0)) or None,
+    "max_num_seqs": 256,
-    "worker_use_ray": os.getenv('WORKER_USE_RAY', 'False').lower() == 'true',
+    "max_logprobs": 20,
-    "distributed_executor_backend": os.getenv('DISTRIBUTED_EXECUTOR_BACKEND', None),
+    "enforce_eager": False,
-    "max_parallel_loading_workers": int(os.getenv('MAX_PARALLEL_LOADING_WORKERS', 0)) or None,
+    "max_seq_len_to_capture": 8192,
-    "block_size": int(os.getenv('BLOCK_SIZE', 16)),
+    "disable_custom_all_reduce": False,
-    "enable_prefix_caching": os.getenv('ENABLE_PREFIX_CACHING', 'False').lower() == 'true',
+    "tokenizer_pool_size": 0,
-    "disable_sliding_window": os.getenv('DISABLE_SLIDING_WINDOW', 'False').lower() == 'true',
+    "tokenizer_pool_type": "ray",
-    # attention_backend replaces deprecated VLLM_ATTENTION_BACKEND env var
+    "enable_lora": False,
-    "attention_backend": os.getenv('ATTENTION_BACKEND', None),
+    "max_loras": 1,
-    # Enabled by default for improved throughput. Set to False to disable if experiencing issues
+    "max_lora_rank": 16,
-    "async_scheduling": None if os.getenv('ASYNC_SCHEDULING') is None else os.getenv('ASYNC_SCHEDULING', 'True').lower() == 'true',
+    "enable_prompt_adapter": False,
-    # Controls how often to yield streaming results
+    "max_prompt_adapters": 1,
-    "stream_interval": int(os.getenv('STREAM_INTERVAL', 1)),
+    "max_prompt_adapter_token": 0,
-    "swap_space": int(os.getenv('SWAP_SPACE', 4)),  # GiB
+    "fully_sharded_loras": False,
-    "cpu_offload_gb": int(os.getenv('CPU_OFFLOAD_GB', 0)),  # GiB
+    "lora_extra_vocab_size": 256,
-    # vLLM defaults None to 2048; keep 0 as None to let vLLM auto-calculate
+    "lora_dtype": "auto",
-    "max_num_batched_tokens": int(os.getenv('MAX_NUM_BATCHED_TOKENS', 0)) or None,
+    "device": "auto",
-    "max_num_seqs": int(os.getenv('MAX_NUM_SEQS', 256)),
+    "ray_workers_use_nsight": False,
-    "max_logprobs": int(os.getenv('MAX_LOGPROBS', 20)),  # Default value for OpenAI Chat Completions API
+    "num_lookahead_slots": 0,
-    "revision": os.getenv('REVISION', None),
+    "scheduler_delay_factor": 0.0,
-    "code_revision": os.getenv('CODE_REVISION', None),
+    "guided_decoding_backend": "outlines",
-    "rope_scaling": os.getenv('ROPE_SCALING', None),
+    "spec_decoding_acceptance_method": "rejection_sampler",
-    "rope_theta": float(os.getenv('ROPE_THETA', 0)) or None,
+    "stream_interval": 1,
-    "tokenizer_revision": os.getenv('TOKENIZER_REVISION', None),
+
-    "quantization": os.getenv('QUANTIZATION', None),
-    "enforce_eager": os.getenv('ENFORCE_EAGER', 'False').lower() == 'true',
-    "max_context_len_to_capture": int(os.getenv('MAX_CONTEXT_LEN_TO_CAPTURE', 0)) or None,
-    "max_seq_len_to_capture": int(os.getenv('MAX_SEQ_LEN_TO_CAPTURE', 8192)),
-    "disable_custom_all_reduce": os.getenv('DISABLE_CUSTOM_ALL_REDUCE', 'False').lower() == 'true',
-    "tokenizer_pool_size": int(os.getenv('TOKENIZER_POOL_SIZE', 0)),
-    "tokenizer_pool_type": os.getenv('TOKENIZER_POOL_TYPE', 'ray'),
-    "tokenizer_pool_extra_config": os.getenv('TOKENIZER_POOL_EXTRA_CONFIG', None),
-    "enable_lora": os.getenv('ENABLE_LORA', 'False').lower() == 'true',
-    "max_loras": int(os.getenv('MAX_LORAS', 1)),
-    "max_lora_rank": int(os.getenv('MAX_LORA_RANK', 16)),
-    "enable_prompt_adapter": os.getenv('ENABLE_PROMPT_ADAPTER', 'False').lower() == 'true',
-    "max_prompt_adapters": int(os.getenv('MAX_PROMPT_ADAPTERS', 1)),
-    "max_prompt_adapter_token": int(os.getenv('MAX_PROMPT_ADAPTER_TOKEN', 0)),
-    "fully_sharded_loras": os.getenv('FULLY_SHARDED_LORAS', 'False').lower() == 'true',
-    "lora_extra_vocab_size": int(os.getenv('LORA_EXTRA_VOCAB_SIZE', 256)),
-    "long_lora_scaling_factors": tuple(map(float, os.getenv('LONG_LORA_SCALING_FACTORS', '').split(','))) if os.getenv('LONG_LORA_SCALING_FACTORS') else None,
-    "lora_dtype": os.getenv('LORA_DTYPE', 'auto'),
-    "max_cpu_loras": int(os.getenv('MAX_CPU_LORAS', 0)) or None,
-    "device": os.getenv('DEVICE', 'auto'),
-    "ray_workers_use_nsight": os.getenv('RAY_WORKERS_USE_NSIGHT', 'False').lower() == 'true',
-    "num_gpu_blocks_override": int(os.getenv('NUM_GPU_BLOCKS_OVERRIDE', 0)) or None,
-    "num_lookahead_slots": int(os.getenv('NUM_LOOKAHEAD_SLOTS', 0)),
-    "model_loader_extra_config": os.getenv('MODEL_LOADER_EXTRA_CONFIG', None),
-    "ignore_patterns": os.getenv('IGNORE_PATTERNS', None),
-    "preemption_mode": os.getenv('PREEMPTION_MODE', None),
-    "scheduler_delay_factor": float(os.getenv('SCHEDULER_DELAY_FACTOR', 0.0)),
-    "enable_chunked_prefill": os.getenv('ENABLE_CHUNKED_PREFILL', None),
-    "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,
-    "ngram_prompt_lookup_max": int(os.getenv('NGRAM_PROMPT_LOOKUP_MAX', 0)) or None,
-    "ngram_prompt_lookup_min": int(os.getenv('NGRAM_PROMPT_LOOKUP_MIN', 0)) or None,
-    "spec_decoding_acceptance_method": os.getenv('SPEC_DECODING_ACCEPTANCE_METHOD', 'rejection_sampler'),
-    "typical_acceptance_sampler_posterior_threshold": float(os.getenv('TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_THRESHOLD', 0)) or None,
-    "typical_acceptance_sampler_posterior_alpha": float(os.getenv('TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_ALPHA', 0)) or None,
-    "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),
 }
 
+
+def _resolve_field_type(field_type: type) -> type:
+    """Resolve Optional/Union to the concrete type for conversion."""
+    origin = get_origin(field_type)
+    args = get_args(field_type) if hasattr(field_type, "__args__") else ()
+    if origin is not None:
+        # Optional[X] is Union[X, None]; X | None is UnionType
+        non_none = [a for a in args if a is not type(None)]
+        if non_none:
+            return non_none[0]
+    return field_type
+
+
+def _convert_env_value_to_field_type(value: str, field_name: str, field_type: type):
+    """Convert env var string to the type expected by AsyncEngineArgs for this field."""
+    val = value.strip() if isinstance(value, str) else value
+    if val in ("", "None", "none"):
+        args = get_args(field_type) if hasattr(field_type, "__args__") else ()
+        if type(None) in (args or ()):
+            return None
+        raise ValueError("empty value not allowed for non-optional field")
+    effective_type = _resolve_field_type(field_type)
+    # bool
+    if effective_type is bool:
+        return str(val).lower() in ("true", "1", "yes", "on")
+    # int
+    if effective_type is int:
+        return int(val)
+    # float
+    if effective_type is float:
+        return float(val)
+    # str
+    if effective_type is str:
+        return str(val)
+    # dict, list, or complex (try JSON)
+    origin = get_origin(effective_type)
+    if effective_type in (dict, list) or origin in (dict, list):
+        try:
+            return json.loads(val)
+        except json.JSONDecodeError:
+            return val
+    # tuple (e.g. long_lora_scaling_factors) — comma-separated or JSON array
+    if effective_type is tuple or origin is tuple:
+        args = get_args(field_type) if hasattr(field_type, "__args__") else ()
+        elem_types = [a for a in args if a is not Ellipsis]
+        elem_type = elem_types[0] if elem_types else str
+        try:
+            parsed = json.loads(val)
+            if isinstance(parsed, list):
+                return tuple(elem_type(x) for x in parsed)
+        except (json.JSONDecodeError, TypeError):
+            pass
+        return tuple(elem_type(x.strip()) for x in str(val).split(",") if x.strip())
+    # Fallback: try int, float, then str
+    try:
+        return int(val)
+    except ValueError:
+        pass
+    try:
+        return float(val)
+    except ValueError:
+        pass
+    return str(val)
+
+
+def _get_args_from_env_auto_discover() -> dict:
+    """Auto-discover engine args from env vars using UPPERCASED field names.
+
+    For every field in AsyncEngineArgs, check os.getenv(FIELD_NAME).
+    E.g. MAX_MODEL_LEN=4096 -> max_model_len=4096.
+    Uses same type conversion as before; supports all vLLM engine args without manual listing.
+    """
+    args = {}
+    valid_fields = AsyncEngineArgs.__dataclass_fields__
+    for field_name, field in valid_fields.items():
+        env_key = field_name.upper()
+        value = os.environ.get(env_key)
+        if value is None:
+            continue
+        try:
+            args[field_name] = _convert_env_value_to_field_type(
+                value, field_name, field.type
+            )
+        except (ValueError, TypeError, json.JSONDecodeError) as e:
+            logging.warning(
+                "Skip env %s=%r: %s", env_key, value, e
+            )
+    return args
+
+
+def _apply_env_aliases(args: dict) -> None:
+    """Apply ENV_ALIASES: if MODEL_NAME etc. are set, set the target engine arg."""
+    valid_fields = AsyncEngineArgs.__dataclass_fields__
+    for alias, target in ENV_ALIASES.items():
+        value = os.environ.get(alias)
+        if value is None or target not in valid_fields:
+            continue
+        try:
+            args[target] = _convert_env_value_to_field_type(
+                value, target, valid_fields[target].type
+            )
+        except (ValueError, TypeError, json.JSONDecodeError) as e:
+            logging.warning("Skip env alias %s=%r: %s", alias, value, e)
+
 def get_speculative_config():
     """Build speculative decoding configuration from environment variables.
 
@@ -122,9 +185,14 @@ def get_speculative_config():
     # 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')
+    _num_spec_tokens = os.getenv('NUM_SPECULATIVE_TOKENS')
-    ngram_max = os.getenv('NGRAM_PROMPT_LOOKUP_MAX')
+    _ngram_max = os.getenv('NGRAM_PROMPT_LOOKUP_MAX')
-    ngram_min = os.getenv('NGRAM_PROMPT_LOOKUP_MIN')
+    _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
@@ -150,11 +218,11 @@ def get_speculative_config():
     if spec_model:
         config['model'] = spec_model
     if num_spec_tokens:
-        config['num_speculative_tokens'] = int(num_spec_tokens)
+        config['num_speculative_tokens'] = num_spec_tokens
     if ngram_max:
-        config['prompt_lookup_max'] = int(ngram_max)
+        config['prompt_lookup_max'] = ngram_max
     if ngram_min:
-        config['prompt_lookup_min'] = int(ngram_min)
+        config['prompt_lookup_min'] = ngram_min
 
     draft_tp = os.getenv('SPECULATIVE_DRAFT_TENSOR_PARALLEL_SIZE')
     if draft_tp:
@@ -186,6 +254,7 @@ def get_speculative_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:
@@ -204,25 +273,19 @@ def _resolve_max_model_len(model, trust_remote_code=False, revision=None):
         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)
 
-def match_vllm_args(args):
+def _local_args_to_engine_args(local: dict) -> dict:
-    """Rename args to match vllm by:
+    """Map local args (e.g. from /local_model_args.json) to engine arg names and filter."""
-    1. Renaming keys to lower case
+    valid = AsyncEngineArgs.__dataclass_fields__
-    2. Renaming keys to match vllm
+    out = {}
-    3. Filtering args to match vllm's AsyncEngineArgs
+    for k, v in local.items():
+        target = ENV_ALIASES.get(k, k.lower().replace("-", "_"))
+        if target not in valid or v in (None, "", "None"):
+            continue
+        out[target] = v
+    return out
 
-    Args:
-        args (dict): Dictionary of args
 
-    Returns:
-        dict: Dictionary of args with renamed keys
-    """
-    renamed_args = {RENAME_ARGS_MAP.get(k, k): v for k, v in args.items()}
-    matched_args = {k: v for k, v in renamed_args.items() if k in AsyncEngineArgs.__dataclass_fields__}
-    return {k: v for k, v in matched_args.items() if v not in [None, "", "None"]}
 def get_local_args():
     """
     Retrieve local arguments from a JSON file.
@@ -245,24 +308,37 @@ def get_local_args():
 
     return local_args
 def get_engine_args():
-    # Start with default args
+    # Start with worker custom defaults (only where we differ from vLLM)
-    args = DEFAULT_ARGS
+    args = dict(DEFAULT_ARGS)
 
-    # Get env args that match keys in AsyncEngineArgs
+    # Auto-discover: every AsyncEngineArgs field from env UPPERCASED (e.g. MAX_MODEL_LEN)
-    args.update(os.environ)
+    args.update(_get_args_from_env_auto_discover())
 
-    # Get local args if model is baked in and overwrite env args
+    # Backward-compat aliases (MODEL_NAME → model, etc.)
-    args.update(get_local_args())
+    _apply_env_aliases(args)
+
+    # Local baked-in model overrides
+    local = get_local_args()
+    if local:
+        args.update(_local_args_to_engine_args(local))
+
+    # Filter to valid engine args and drop sentinel empty values
+    valid_fields = AsyncEngineArgs.__dataclass_fields__
+    args = {
+        k: v for k, v in args.items()
+        if k in valid_fields and v not in (None, "", "None")
+    }
+
+    # Special conversion for limit_mm_per_prompt (e.g. "image=1,video=0")
+    limit_mm_env = os.getenv("LIMIT_MM_PER_PROMPT")
+    if limit_mm_env is not None:
+        args["limit_mm_per_prompt"] = convert_limit_mm_per_prompt(limit_mm_env)
 
     # if args.get("TENSORIZER_URI"): TODO: add back once tensorizer is ready
     #     args["load_format"] = "tensorizer"
     #     args["model_loader_extra_config"] = TensorizerConfig(tensorizer_uri=args["TENSORIZER_URI"], num_readers=None)
     #     logging.info(f"Using tensorized model from {args['TENSORIZER_URI']}")
 
-    
-    # Rename and match to vllm args
-    args = match_vllm_args(args)
-
     if args.get("load_format") == "bitsandbytes":
         args["quantization"] = args["load_format"]