Update README.md for 0.3.2

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "vllm-base-image/vllm"]
+	path = vllm-base-image/vllm
+	url = https://github.com/runpod/vllm-fork-for-sls-worker.git

Dockerfile

@@ -1,5 +1,5 @@
 ARG WORKER_CUDA_VERSION=11.8.0
-FROM runpod/worker-vllm:base-0.3.1-cuda${WORKER_CUDA_VERSION} AS vllm-base
+FROM runpod/worker-vllm:base-0.3.2-cuda${WORKER_CUDA_VERSION} AS vllm-base
 
 RUN apt-get update -y \
     && apt-get install -y python3-pip

README.md

@@ -4,7 +4,7 @@
 
 Deploy Blazing-fast LLMs powered by [vLLM](https://github.com/vllm-project/vllm) on RunPod Serverless in a few clicks.
 
-<p>Worker Version: 0.3.1 | vLLM Version: 0.3.2</p>
+<p>Worker Version: 0.3.2 | vLLM Version: 0.3.3</p>
 
 [![CD | Docker-Build-Release](https://github.com/runpod-workers/worker-vllm/actions/workflows/docker-build-release.yml/badge.svg)](https://github.com/runpod-workers/worker-vllm/actions/workflows/docker-build-release.yml)
 
@@ -58,7 +58,9 @@ Deploy Blazing-fast LLMs powered by [vLLM](https://github.com/vllm-project/vllm)
 
 ### Option 1: Deploy Any Model Using Pre-Built Docker Image [Recommended]
 > [!TIP]
-> This is the recommended way to deploy your model, as it does not require you to build a Docker image, upload heavy models to DockerHub and wait for workers to download them. Instead, use this option to deploy your model in a few clicks. For even more convenience, attach a network storage volume to your Endpoint, which will download the model once and share it across all workers.
+> This is the quickest and easiest way to tes your model, as it does not require you to build a Docker image, upload heavy models to DockerHub and wait for workers to download them. You can use this option to deploy your model in a few clicks. For even more convenience, attach a network storage volume to your Endpoint, which will download the model once and share it across all workers.
+> 
+> However, for actual deployment, it is recommended that you build an image with the model baked in, which is described in Option 2 - this will ensure the fastest load speeds.
 
 We now offer a pre-built Docker Image for the vLLM Worker that you can configure entirely with Environment Variables when creating the RunPod Serverless Endpoint:
 
@@ -70,8 +72,8 @@ Below is a summary of the available RunPod Worker images, categorized by image s
 
 | CUDA Version | Stable Image Tag                  | Development Image Tag             | Note                                                        |
 |--------------|-----------------------------------|-----------------------------------|----------------------------------------------------------------------|
-| 11.8.0       | `runpod/worker-vllm:0.3.1-cuda11.8.0`        | `runpod/worker-vllm:dev-cuda11.8.0`   | Available on all RunPod Workers without additional selection needed. |
-| 12.1.0       | `runpod/worker-vllm:0.3.1-cuda12.1.0` | `runpod/worker-vllm:dev-cuda12.1.0` | When creating an Endpoint, select CUDA Version 12.2 and 12.1 in the filter. |
+| 11.8.0       | `runpod/worker-vllm:0.3.2-cuda11.8.0`        | `runpod/worker-vllm:dev-cuda11.8.0`   | Available on all RunPod Workers without additional selection needed. |
+| 12.1.0       | `runpod/worker-vllm:0.3.2-cuda12.1.0` | `runpod/worker-vllm:dev-cuda12.1.0` | When creating an Endpoint, select CUDA Version 12.2 and 12.1 in the filter. |
 
 This table provides a quick reference to the image tags you should use based on the desired CUDA version and image stability (Stable or Development). Ensure to follow the selection note for CUDA 12.1.0 compatibility.
 
@@ -88,7 +90,7 @@ This table provides a quick reference to the image tags you should use based on
 **LLM Settings**
 | `MODEL_NAME`**\***                        | -                    | `str`                                         | Hugging Face Model Repository (e.g., `openchat/openchat-3.5-1210`). |
 | `MODEL_REVISION`                    | `None`               | `str`                                         |Model revision(branch) to load. |
-| `MAX_MODEL_LENGTH`                  | Model's maximum      | `int`                                         |Maximum number of tokens for the engine to handle per request. |
+| `MAX_MODEL_LEN`                  | Model's maximum      | `int`                                         |Maximum number of tokens for the engine to handle per request. |
 | `BASE_PATH`                         | `/runpod-volume`     | `str`                                         |Storage directory for Huggingface cache and model. Utilizes network storage if attached when pointed at `/runpod-volume`, which will have only one worker download the model once, which all workers will be able to load. If no network volume is present, creates a local directory within each worker. |
 | `LOAD_FORMAT`                       | `auto`               | `str`                                         |Format to load model in. |
 | `HF_TOKEN`                          | -                    | `str`                                         |Hugging Face token for private and gated models. |

builder/download_model.py

@@ -45,7 +45,6 @@ if __name__ == "__main__":
     with open("/local_model_path.txt", "w") as f:
         f.write(model_folder)
 
-    if tokenizer != model:
-        tokenizer_folder = download_extras_or_tokenizer(tokenizer, download_dir, revisions["tokenizer"])
-        with open("/local_tokenizer_path.txt", "w") as f:
-            f.write(tokenizer_folder)
+    tokenizer_folder = download_extras_or_tokenizer(tokenizer, download_dir, revisions["tokenizer"])
+    with open("/local_tokenizer_path.txt", "w") as f:
+        f.write(tokenizer_folder)

builder/requirements.txt

@@ -6,4 +6,5 @@ runpod==1.6.2
 huggingface-hub
 packaging
 typing-extensions==4.7.1
-pydantic
+pydantic
+pydantic-settings

src/config.py

@@ -39,7 +39,7 @@ class EngineConfig:
             "trust_remote_code": bool(int(os.getenv("TRUST_REMOTE_CODE", 0))),
             "gpu_memory_utilization": float(os.getenv("GPU_MEMORY_UTILIZATION", 0.95)),
             "max_parallel_loading_workers": None if device_count() > 1 or not os.getenv("MAX_PARALLEL_LOADING_WORKERS") else int(os.getenv("MAX_PARALLEL_LOADING_WORKERS")),
-            "max_model_len": int(os.getenv("MAX_MODEL_LENGTH")) if os.getenv("MAX_MODEL_LENGTH") else None,
+            "max_model_len": int(os.getenv("MAX_MODEL_LEN")) if os.getenv("MAX_MODEL_LEN") else None,
             "tensor_parallel_size": device_count(),
             "seed": int(os.getenv("SEED")) if os.getenv("SEED") else None,
             "kv_cache_dtype": os.getenv("KV_CACHE_DTYPE"),

src/constants.py

@@ -1,30 +1,4 @@
-from typing import Union
-
 DEFAULT_BATCH_SIZE = 50
 DEFAULT_MAX_CONCURRENCY = 300
 DEFAULT_BATCH_SIZE_GROWTH_FACTOR = 3
-DEFAULT_MIN_BATCH_SIZE = 1
-
-SAMPLING_PARAM_TYPES = {
-    "n": int,
-    "best_of": int,
-    "presence_penalty": float,
-    "frequency_penalty": float,
-    "repetition_penalty": float,
-    "temperature": Union[float, int],
-    "top_p": float,
-    "top_k": int,
-    "min_p": float,
-    "use_beam_search": bool,
-    "length_penalty": float,
-    "early_stopping": Union[bool, str],
-    "stop": Union[str, list],
-    "stop_token_ids": list,
-    "ignore_eos": bool,
-    "max_tokens": int,
-    "logprobs": int,
-    "prompt_logprobs": int,
-    "skip_special_tokens": bool,
-    "spaces_between_special_tokens": bool,
-    "include_stop_str_in_output": bool
-}
+DEFAULT_MIN_BATCH_SIZE = 1

src/engine.py

@@ -6,7 +6,7 @@ from dotenv import load_dotenv
 from torch.cuda import device_count
 from typing import AsyncGenerator
 
-from vllm import AsyncLLMEngine, AsyncEngineArgs, SamplingParams
+from vllm import AsyncLLMEngine, AsyncEngineArgs
 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
@@ -16,7 +16,6 @@ from constants import DEFAULT_MAX_CONCURRENCY, DEFAULT_BATCH_SIZE, DEFAULT_BATCH
 from tokenizer import TokenizerWrapper
 from config import EngineConfig
 
-
 class vLLMEngine:
     def __init__(self, engine = None):
         load_dotenv() # For local development
@@ -35,7 +34,7 @@ class vLLMEngine:
         try:
             async for batch in self._generate_vllm(
                 llm_input=job_input.llm_input,
-                validated_sampling_params=job_input.validated_sampling_params,
+                validated_sampling_params=job_input.sampling_params,
                 batch_size=job_input.max_batch_size,
                 stream=job_input.stream,
                 apply_chat_template=job_input.apply_chat_template,
@@ -45,12 +44,11 @@ class vLLMEngine:
             ):
                 yield batch
         except Exception as e:
-            yield create_error_response(str(e)).model_dump()
+            yield {"error": create_error_response(str(e)).model_dump()}
 
     async def _generate_vllm(self, llm_input, validated_sampling_params, batch_size, stream, apply_chat_template, request_id, batch_size_growth_factor, min_batch_size: str) -> AsyncGenerator[dict, None]:
         if apply_chat_template or isinstance(llm_input, list):
             llm_input = self.tokenizer.apply_chat_template(llm_input)
-        validated_sampling_params = SamplingParams(**validated_sampling_params)
         results_generator = self.llm.generate(llm_input, validated_sampling_params, request_id)
         n_responses, n_input_tokens, is_first_output = validated_sampling_params.n, 0, True
         last_output_texts, token_counters = ["" for _ in range(n_responses)], {"batch": 0, "total": 0}

src/utils.py

@@ -1,10 +1,9 @@
 import logging
 from http import HTTPStatus
 from typing import Any, Dict
-from constants import SAMPLING_PARAM_TYPES
 from vllm.utils import random_uuid
 from vllm.entrypoints.openai.protocol import ErrorResponse
-
+from vllm import SamplingParams
 
 logging.basicConfig(level=logging.INFO)
 
@@ -25,20 +24,6 @@ def count_physical_cores():
 
     return len(cores)
 
-def validate_sampling_params(params: Dict[str, Any]) -> Dict[str, Any]:
-    validated_params = {}
-    invalid_params = []
-    for key, value in params.items():
-        expected_type = SAMPLING_PARAM_TYPES.get(key)
-        if expected_type and isinstance(value, expected_type):
-            validated_params[key] = value
-        else:
-            invalid_params.append(key)
-        
-    if len(invalid_params) > 0:
-        logging.warning("Ignoring invalid sampling params: %s", invalid_params)
-        
-    return validated_params
 
 class JobInput:
     def __init__(self, job):
@@ -47,7 +32,7 @@ class JobInput:
         self.max_batch_size = job.get("max_batch_size")
         self.apply_chat_template = job.get("apply_chat_template", False)
         self.use_openai_format = job.get("use_openai_format", False)
-        self.validated_sampling_params = validate_sampling_params(job.get("sampling_params", {}))
+        self.sampling_params = SamplingParams(**job.get("sampling_params", {}))
         self.request_id = random_uuid()
         batch_size_growth_factor = job.get("batch_size_growth_factor")
         self.batch_size_growth_factor = float(batch_size_growth_factor) if batch_size_growth_factor else None 
@@ -78,4 +63,6 @@ class BatchSize:
 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)
+                            code=status_code.value)
+    
+    

vllm-base-image/Dockerfile

@@ -17,25 +17,16 @@ ARG WORKER_CUDA_VERSION
 RUN apt-get update -y \
     && apt-get install -y python3-pip git
 
-RUN if [ "${WORKER_CUDA_VERSION}" = "12.1.0" ]; then \
-        ldconfig /usr/local/cuda-12.1/compat/; \
-    fi
-
 # Set working directory
 WORKDIR /vllm-installation
 
 # Install build and runtime dependencies
-COPY vllm-${WORKER_CUDA_VERSION}/requirements.txt requirements.txt
+COPY vllm/requirements-${WORKER_CUDA_VERSION}.txt  requirements.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install -r requirements.txt
 
-RUN --mount=type=cache,target=/root/.cache/pip \
-    if [ "${WORKER_CUDA_VERSION}" = "11.8.0" ]; then \
-        pip install -U --force-reinstall torch==2.1.2 xformers==0.0.23.post1 --index-url https://download.pytorch.org/whl/cu118; \
-    fi
-
 # Install development dependencies
-COPY vllm-${WORKER_CUDA_VERSION}/requirements-dev.txt requirements-dev.txt
+COPY vllm/requirements-dev.txt requirements-dev.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install -r requirements-dev.txt
 
@@ -45,25 +36,15 @@ FROM dev AS build
 ARG WORKER_CUDA_VERSION
 
 # Install build dependencies
-COPY vllm-${WORKER_CUDA_VERSION}/requirements-build.txt requirements-build.txt
+COPY vllm/requirements-build.txt requirements-build.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install -r requirements-build.txt
 
 # Copy necessary files
-COPY vllm-${WORKER_CUDA_VERSION}/csrc csrc
-COPY vllm-${WORKER_CUDA_VERSION}/setup.py setup.py
-COPY vllm-12.1.0/pyproject.toml pyproject.toml
-COPY vllm-${WORKER_CUDA_VERSION}/vllm/__init__.py vllm/__init__.py
-
-# Conditional installation based on CUDA version
-RUN --mount=type=cache,target=/root/.cache/pip \
-    if [ "${WORKER_CUDA_VERSION}" = "11.8.0" ]; then \
-        pip install -U --force-reinstall torch==2.1.2 xformers==0.0.23.post1 --index-url https://download.pytorch.org/whl/cu118; \
-        rm pyproject.toml; \
-    elif [ "${WORKER_CUDA_VERSION}" != "12.1.0" ]; then \
-        echo "WORKER_CUDA_VERSION not supported"; \
-        exit 1; \
-    fi
+COPY vllm/csrc csrc
+COPY vllm/setup.py setup.py
+COPY vllm/pyproject.toml pyproject.toml
+COPY vllm/vllm/__init__.py vllm/__init__.py
 
 # Set environment variables for building extensions
 ARG torch_cuda_arch_list='7.0 7.5 8.0 8.6 8.9 9.0+PTX'
@@ -72,8 +53,10 @@ ARG max_jobs=48
 ENV MAX_JOBS=${max_jobs}
 ARG nvcc_threads=1024
 ENV NVCC_THREADS=${nvcc_threads}
-
+ENV WORKER_CUDA_VERSION=${WORKER_CUDA_VERSION}
+ENV VLLM_INSTALL_PUNICA_KERNELS=0
 # Build extensions
+RUN ldconfig /usr/local/cuda-$(echo "$WORKER_CUDA_VERSION" | sed 's/\.0$//')/compat/
 RUN python3 setup.py build_ext --inplace
 
 FROM nvidia/cuda:${WORKER_CUDA_VERSION}-runtime-ubuntu22.04 AS vllm-base
@@ -88,19 +71,15 @@ RUN apt-get update -y \
 # Set working directory
 WORKDIR /vllm-installation
 
+
 # Install runtime dependencies
-COPY vllm-${WORKER_CUDA_VERSION}/requirements.txt requirements.txt
+COPY vllm/requirements-${WORKER_CUDA_VERSION}.txt  requirements.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install -r requirements.txt
 
-RUN --mount=type=cache,target=/root/.cache/pip \
-    if [ "${WORKER_CUDA_VERSION}" = "11.8.0" ]; then \
-        pip install -U --force-reinstall torch==2.1.2 xformers==0.0.23.post1 --index-url https://download.pytorch.org/whl/cu118; \
-    fi
-
 # Copy built files from the build stage
 COPY --from=build /vllm-installation/vllm/*.so /vllm-installation/vllm/
-COPY vllm-${WORKER_CUDA_VERSION}/vllm vllm
+COPY vllm/vllm vllm
 
 # Set PYTHONPATH environment variable
 ENV PYTHONPATH="/"

vllm-base/download_required_files.sh

@@ -1,12 +0,0 @@
-#!/bin/bash
-
-git clone https://github.com/runpod/vllm-fork-for-sls-worker.git
-
-cp -r vllm-fork-for-sls-worker vllm-12.1.0
-cp -r vllm-fork-for-sls-worker vllm-11.8.0
-rm -rf vllm-fork-for-sls-worker
-
-cd vllm-11.8.0
-git checkout cuda-11.8
-
-echo "vLLM Base Image Builder Setup Complete."