Merge pull request #185 from runpod-workers/up-0.9.0

Version upgrade

Dockerfile

@@ -12,7 +12,7 @@ RUN --mount=type=cache,target=/root/.cache/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==0.8.4 && \
+RUN python3 -m pip install vllm==0.9.0.1 && \
     python3 -m pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3
 
 # Setup for Option 2: Building the Image with the Model included

README.md

@@ -18,9 +18,9 @@ Deploy OpenAI-Compatible Blazing-Fast LLM Endpoints powered by the [vLLM](https:
 ### 1. UI for Deploying vLLM Worker on RunPod console:
 ![Demo of Deploying vLLM Worker on RunPod console with new UI](media/ui_demo.gif)
 
-### 2. Worker vLLM `v2.5.0` with vLLM `0.8.5` now available under `stable` tags 
+### 2. Worker vLLM `v2.6.0` with vLLM `0.9.0` now available under `stable` tags 
 
-Update v2.5.0 is now available, use the image tag `runpod/worker-v1-vllm:v2.5.0stable-cuda12.1.0`.
+Update v2.6.0 is now available, use the image tag `runpod/worker-v1-vllm:v2.6.0stable-cuda12.1.0`.
 
 ### 3. OpenAI-Compatible [Embedding Worker](https://github.com/runpod-workers/worker-infinity-embedding) Released
 Deploy your own OpenAI-compatible Serverless Endpoint on RunPod with multiple embedding models and fast inference for RAG and more! 
@@ -81,7 +81,7 @@ Below is a summary of the available RunPod Worker images, categorized by image s
 
 | CUDA Version | Stable Image Tag                  | Development Image Tag             | Note                                                        |
 |--------------|-----------------------------------|-----------------------------------|----------------------------------------------------------------------|
-| 12.1.0       | `runpod/worker-v1-vllm:v2.5.0stable-cuda12.1.0` | `runpod/worker-v1-vllm:v2.5.0dev-cuda12.1.0` | When creating an Endpoint, select CUDA Version 12.3, 12.2 and 12.1 in the filter. |
+| 12.1.0       | `runpod/worker-v1-vllm:v2.6.0stable-cuda12.1.0` | `runpod/worker-v1-vllm:v2.6.0dev-cuda12.1.0` | When creating an Endpoint, select CUDA Version 12.3, 12.2 and 12.1 in the filter. |
 
 
 
@@ -139,6 +139,7 @@ Below is a summary of the available RunPod Worker images, categorized by image s
 | `LONG_LORA_SCALING_FACTORS`               | None                  | `tuple`                                    | Specify multiple scaling factors for LoRA adapters. |
 | `MAX_CPU_LORAS`                           | None                  | `int`                                      | Maximum number of LoRAs to store in CPU memory. |
 | `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"}`|
 | `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. |

docker-bake.hcl

@@ -7,7 +7,7 @@ variable "REPOSITORY" {
 }
 
 variable "BASE_IMAGE_VERSION" {
-  default = "v2.4.0stable"
+  default = "v2.6.0stable"
 }
 
 group "all" {

src/engine.py

@@ -122,31 +122,40 @@ class OpenAIvLLMEngine(vLLMEngine):
         super().__init__(vllm_engine)
         self.served_model_name = os.getenv("OPENAI_SERVED_MODEL_NAME_OVERRIDE") 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())
         self.raw_openai_output = bool(int(os.getenv("RAW_OPENAI_OUTPUT", 1)))
 
+    def _load_lora_adapters(self):
+        adapters = []
+        try:
+            adapters = json.loads(os.getenv("LORA_MODULES", '[]'))
+        except Exception as e:
+            logging.info(f"---Initialized adapter json load error: {e}")
+
+        for i, adapter in enumerate(adapters):
+            try:
+                adapters[i] = LoRAModulePath(**adapter)
+                logging.info(f"---Initialized adapter: {adapter}")
+            except Exception as e:
+                logging.info(f"---Initialized adapter not worked: {e}")
+                continue
+        return adapters
+
     async def _initialize_engines(self):
         self.model_config = await self.llm.get_model_config()
         self.base_model_paths = [
             BaseModelPath(name=self.engine_args.model, model_path=self.engine_args.model)
         ]
 
-        lora_modules = os.getenv('LORA_MODULES', None)
-        if lora_modules is not None:
-            try:
-                lora_modules = json.loads(lora_modules)
-                lora_modules = [LoRAModulePath(**lora_modules)]
-            except:
-                lora_modules = None
-
         self.serving_models = OpenAIServingModels(
             engine_client=self.llm,
             model_config=self.model_config,
             base_model_paths=self.base_model_paths,
-            lora_modules=None,
+            lora_modules=self.lora_adapters,
             prompt_adapters=None,
         )
-
+        await self.serving_models.init_static_loras()
         self.chat_engine = OpenAIServingChat(
             engine_client=self.llm, 
             model_config=self.model_config,
@@ -156,7 +165,7 @@ class OpenAIvLLMEngine(vLLMEngine):
             chat_template=self.tokenizer.tokenizer.chat_template,
             chat_template_content_format="auto",
             # enable_reasoning=os.getenv('ENABLE_REASONING', 'false').lower() == 'true',
-            # reasoning_parser=None,
+            reasoning_parser= os.getenv('REASONING_PARSER', "") or None,
             # return_token_as_token_ids=False,
             enable_auto_tools=os.getenv('ENABLE_AUTO_TOOL_CHOICE', 'false').lower() == 'true',
             tool_parser=os.getenv('TOOL_CALL_PARSER', "") or None,

src/engine_args.py

@@ -122,7 +122,7 @@ def get_local_args():
         local_args = json.load(f)
 
     if local_args.get("MODEL_NAME") is None:
-        raise ValueError("Model name not found in /local_model_args.json. There was a problem when baking the model in.")
+        logging.warning("Model name not found in /local_model_args.json. There maybe was a problem when baking the model in.")
 
     logging.info(f"Using baked in model with args: {local_args}")
     os.environ["TRANSFORMERS_OFFLINE"] = "1"

worker-config.json

@@ -1,5 +1,66 @@
 {
   "versions": {
+    "0.9.0": {
+      "imageName": "runpod/worker-v1-vllm:v2.6.0stable-cuda12.1.0",
+      "minimumCudaVersion": "12.1",
+      "categories": [
+        {
+          "title": "LLM Settings",
+          "settings": [
+            "TOKENIZER", "TOKENIZER_MODE", "SKIP_TOKENIZER_INIT", "TRUST_REMOTE_CODE",
+            "DOWNLOAD_DIR", "LOAD_FORMAT", "DTYPE", "KV_CACHE_DTYPE", "QUANTIZATION_PARAM_PATH",
+            "MAX_MODEL_LEN", "GUIDED_DECODING_BACKEND", "DISTRIBUTED_EXECUTOR_BACKEND",
+            "WORKER_USE_RAY", "RAY_WORKERS_USE_NSIGHT", "PIPELINE_PARALLEL_SIZE",
+            "TENSOR_PARALLEL_SIZE", "MAX_PARALLEL_LOADING_WORKERS", "ENABLE_PREFIX_CACHING",
+            "DISABLE_SLIDING_WINDOW", "NUM_LOOKAHEAD_SLOTS",
+            "SEED", "NUM_GPU_BLOCKS_OVERRIDE", "MAX_NUM_BATCHED_TOKENS", "MAX_NUM_SEQS",
+            "MAX_LOGPROBS", "DISABLE_LOG_STATS", "QUANTIZATION", "ROPE_SCALING", "ROPE_THETA",
+            "TOKENIZER_POOL_SIZE", "TOKENIZER_POOL_TYPE", "TOKENIZER_POOL_EXTRA_CONFIG",
+            "ENABLE_LORA", "MAX_LORAS", "MAX_LORA_RANK", "LORA_EXTRA_VOCAB_SIZE",
+            "LORA_DTYPE", "LONG_LORA_SCALING_FACTORS", "MAX_CPU_LORAS", "FULLY_SHARDED_LORAS",
+            "DEVICE", "SCHEDULER_DELAY_FACTOR", "ENABLE_CHUNKED_PREFILL", "SPECULATIVE_MODEL",
+            "NUM_SPECULATIVE_TOKENS", "SPECULATIVE_DRAFT_TENSOR_PARALLEL_SIZE",
+            "SPECULATIVE_MAX_MODEL_LEN", "SPECULATIVE_DISABLE_BY_BATCH_SIZE",
+            "NGRAM_PROMPT_LOOKUP_MAX", "NGRAM_PROMPT_LOOKUP_MIN", "SPEC_DECODING_ACCEPTANCE_METHOD",
+            "TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_THRESHOLD", "TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_ALPHA",
+            "MODEL_LOADER_EXTRA_CONFIG", "PREEMPTION_MODE", "PREEMPTION_CHECK_PERIOD",
+            "PREEMPTION_CPU_CAPACITY", "MAX_LOG_LEN", "DISABLE_LOGGING_REQUEST",
+            "ENABLE_AUTO_TOOL_CHOICE", "TOOL_CALL_PARSER"
+          ]
+        },
+        {
+          "title": "Tokenizer Settings",
+          "settings": [
+            "TOKENIZER_NAME", "TOKENIZER_REVISION", "CUSTOM_CHAT_TEMPLATE"
+          ]
+        },
+        {
+          "title": "System Settings",
+          "settings": [
+            "GPU_MEMORY_UTILIZATION", "MAX_PARALLEL_LOADING_WORKERS", "BLOCK_SIZE",
+            "SWAP_SPACE", "ENFORCE_EAGER", "MAX_SEQ_LEN_TO_CAPTURE", "DISABLE_CUSTOM_ALL_REDUCE"
+          ]
+        },
+        {
+          "title": "Streaming Settings",
+          "settings": [
+            "DEFAULT_BATCH_SIZE", "DEFAULT_MIN_BATCH_SIZE", "DEFAULT_BATCH_SIZE_GROWTH_FACTOR"
+          ]
+        },
+        {
+          "title": "OpenAI Settings",
+          "settings": [
+            "RAW_OPENAI_OUTPUT", "OPENAI_RESPONSE_ROLE", "OPENAI_SERVED_MODEL_NAME_OVERRIDE"
+          ]
+        },
+        {
+          "title": "Serverless Settings",
+          "settings": [
+            "MAX_CONCURRENCY", "DISABLE_LOG_STATS", "DISABLE_LOG_REQUESTS"
+          ]
+        }
+      ]
+    },
     "0.8.5": {
       "imageName": "runpod/worker-v1-vllm:v2.5.0stable-cuda12.1.0",
       "minimumCudaVersion": "12.1",