owen/worker-vllm
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge branch 'main' into feat/update-vllm-v0.15.0

Browse Source
This commit is contained in:
velaraptor-runpod
2026-02-13 03:01:05 -06:00
parent c3cf43b228 c45ac42acd
commit 2e8c251447
7 changed files with 128 additions and 134 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.runpod/hub.json
-20
View File
@@ -645,16 +645,6 @@
"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": {
@@ -792,16 +782,6 @@
"advanced": true
}
},
{
"key": "ENABLE_LOG_REQUESTS",
"input": {
"name": "Enable Log Requests",
"type": "boolean",
"description": "Enables vLLM request logging",
"default": false,
"advanced": true
}
},
{
"key": "ENABLE_AUTO_TOOL_CHOICE",
"input": {
Dockerfile
+15 -12
View File
@@ -1,21 +1,17 @@
FROM nvidia/cuda:12.9.1-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-12.9/compat/
RUN ldconfig /usr/local/cuda-12.8/compat/
ENV RAY_METRICS_EXPORT_ENABLED=0 \
RAY_DISABLE_USAGE_STATS=1 \
TOKENIZERS_PARALLELISM=false \
RAYON_NUM_THREADS=4
# 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 vLLM first to avoid PyTorch version conflicts
RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip install --upgrade pip && \
python3 -m pip install "vllm[flashinfer]==0.15.0" --extra-index-url https://download.pytorch.org/whl/cu129
# Install remaining Python dependencies
# 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 -r /requirements.txt
@@ -38,7 +34,14 @@ 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"
builder/requirements.txt
+2 -3
View File
@@ -1,7 +1,7 @@
ray
pandas
pyarrow
runpod>=1.8,<2.0
runpod
huggingface-hub
packaging
typing-extensions>=4.8.0
@@ -11,5 +11,4 @@ hf-transfer
transformers>=4.56.0,<5
bitsandbytes>=0.45.0
kernels
torch>=2.9.1
torch-c-dlpack-ext
torch-c-dlpack-ext
docs/configuration.md
+40 -50
View File
@@ -25,6 +25,7 @@ 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. |
| `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. |
| `MAX_NUM_BATCHED_TOKENS` | None | `int` | Maximum number of batched tokens per iteration. |
@@ -45,7 +46,7 @@ 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:** When using LoRA with serverless deployments, the OpenAI serving engines are initialized on the first request (deferred initialization) to avoid event loop conflicts. LoRA adapter count is logged at startup.
> **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
@@ -86,11 +87,9 @@ If `SPECULATIVE_CONFIG` is set, it takes priority over individual env vars. When
| `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. |
| `SCHEDULER_DELAY_FACTOR` | 0.0 | `float` | Apply a delay before scheduling next prompt. |
| `ENABLE_CHUNKED_PREFILL` | False | `bool` | Enable chunked prefill requests. |
| `ATTENTION_BACKEND` | None | `str` | Attention backend to use (e.g., `FLASH_ATTN`, `XFORMERS`, `FLASHINFER`). |
| `ASYNC_SCHEDULING` | False | `bool` | Enable async scheduling for improved throughput. |
| `STREAM_INTERVAL` | 0 | `float` | Interval in seconds between streaming responses. |
| `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
@@ -112,21 +111,21 @@ The way this works is that the first request will have a batch size of `DEFAULT_
## OpenAI Compatibility Settings
| Variable | Default | Type/Choices | Description |
| --------------------------------------- | ----------- | ---------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `RAW_OPENAI_OUTPUT` | `1` | boolean as `int` | Enables raw OpenAI SSE format string output when streaming. **Required** to be enabled (which it is by default) for OpenAI compatibility. |
| `OPENAI_SERVED_MODEL_NAME_OVERRIDE` | `None` | `str` | Overrides the name of the served model from model repo/path to specified name, which you will then be able to use the value for the `model` parameter when making OpenAI requests |
| `OPENAI_RESPONSE_ROLE` | `assistant` | `str` | Role of the LLM's Response in OpenAI Chat Completions. |
| `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 chat templates from incoming requests to override the server default. |
| `RETURN_TOKENS_AS_TOKEN_IDS` | `false` | `bool` | Return token IDs instead of token strings in responses. |
| `EXCLUDE_TOOLS_WHEN_TOOL_CHOICE_NONE` | `false` | `bool` | When tool_choice is 'none', exclude tool definitions from the prompt. |
| `ENABLE_PROMPT_TOKENS_DETAILS` | `false` | `bool` | Enable detailed prompt token usage in responses. |
| `ENABLE_FORCE_INCLUDE_USAGE` | `false` | `bool` | Force include usage information in all streaming responses. |
| `ENABLE_LOG_OUTPUTS` | `false` | `bool` | Log model outputs for debugging. |
| `LOG_ERROR_STACK` | `false` | `bool` | Log full error stack traces. |
| Variable | Default | Type/Choices | Description |
| ----------------------------------- | ----------- | ---------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `RAW_OPENAI_OUTPUT` | `1` | boolean as `int` | Enables raw OpenAI SSE format string output when streaming. **Required** to be enabled (which it is by default) for OpenAI compatibility. |
| `OPENAI_SERVED_MODEL_NAME_OVERRIDE` | `None` | `str` | Overrides the name of the served model from model repo/path to specified name, which you will then be able to use the value for the `model` parameter when making OpenAI requests |
| `OPENAI_RESPONSE_ROLE` | `assistant` | `str` | Role of the LLM's Response in OpenAI Chat Completions. |
| `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
@@ -134,7 +133,18 @@ 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. |
| `ENABLE_LOG_REQUESTS` | False | `bool` | Enables vLLM request logging. |
| `ENABLE_LOG_REQUESTS` | False | `bool` | Enables vLLM request logging. (Replaces deprecated `DISABLE_LOG_REQUESTS` in vLLM 0.15.0) |
## Advanced Settings
| Variable | Default | Type | Description |
| --------------------------- | ------- | ------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
| `MODEL_LOADER_EXTRA_CONFIG` | None | `dict` | Extra config for model loader. |
| `PREEMPTION_MODE` | None | `str` | If 'recompute', the engine performs preemption-aware recomputation. If 'save', the engine saves activations into the CPU memory as preemption happens. |
| `PREEMPTION_CHECK_PERIOD` | 1.0 | `float` | How frequently the engine checks if a preemption happens. |
| `PREEMPTION_CPU_CAPACITY` | 2 | `float` | The percentage of CPU memory used for the saved activations. |
| `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. |
## Docker Build Arguments
@@ -149,31 +159,11 @@ 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 / Migration | 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 |
| `DISABLE_LOG_REQUESTS` | `ENABLE_LOG_REQUESTS` | Logic inverted: `DISABLE_LOG_REQUESTS=true` → `ENABLE_LOG_REQUESTS=false` |
| `VLLM_ATTENTION_BACKEND` | `ATTENTION_BACKEND` | Use new variable name |
| `WORKER_USE_RAY` | `DISTRIBUTED_EXECUTOR_BACKEND=ray` | Removed in vLLM 0.15.0 |
| `USE_V2_BLOCK_MANAGER` | _(removed)_ | V2 block manager is now the default |
| `NUM_LOOKAHEAD_SLOTS` | _(removed)_ | No longer a separate config |
| `ROPE_SCALING` | _(removed)_ | Use model config directly |
| `ROPE_THETA` | _(removed)_ | Use model config directly |
| `TOKENIZER_POOL_SIZE` | _(removed)_ | Removed in vLLM 0.15.0 |
| `TOKENIZER_POOL_TYPE` | _(removed)_ | Removed in vLLM 0.15.0 |
| `TOKENIZER_POOL_EXTRA_CONFIG` | _(removed)_ | Removed in vLLM 0.15.0 |
| `QUANTIZATION_PARAM_PATH` | _(removed)_ | Removed in vLLM 0.15.0 |
| `LORA_EXTRA_VOCAB_SIZE` | _(removed)_ | Removed in vLLM 0.15.0 |
| `LONG_LORA_SCALING_FACTORS` | _(removed)_ | Removed in vLLM 0.15.0 |
| `GUIDED_DECODING_BACKEND` | _(removed)_ | Removed in vLLM 0.15.0 |
| `SPEC_DECODING_ACCEPTANCE_METHOD` | `SPECULATIVE_CONFIG` (JSON) | Use JSON config |
| `TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_THRESHOLD` | `SPECULATIVE_CONFIG` (JSON) | Use JSON config |
| `TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_ALPHA` | `SPECULATIVE_CONFIG` (JSON) | Use JSON config |
| `SPECULATIVE_DRAFT_TENSOR_PARALLEL_SIZE` | `SPECULATIVE_CONFIG` (JSON) or individual env | Still supported via env var |
| `PREEMPTION_MODE` | _(removed)_ | Removed in vLLM 0.15.0 |
| `PREEMPTION_CHECK_PERIOD` | _(removed)_ | Removed in vLLM 0.15.0 |
| `PREEMPTION_CPU_CAPACITY` | _(removed)_ | Removed in vLLM 0.15.0 |
| `MAX_LOG_LEN` | _(removed)_ | Removed in vLLM 0.15.0 |
| `DISABLE_LOGGING_REQUEST` | `ENABLE_LOG_REQUESTS` | Removed in vLLM 0.15.0 |
| `MAX_SEQ_LEN_TO_CAPTURE` | _(removed)_ | Removed in vLLM 0.15.0 |
| 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
+31 -13
View File
@@ -6,7 +6,6 @@ import time
from typing import AsyncGenerator, Optional
from dotenv import load_dotenv
from vllm import AsyncLLMEngine
from vllm.entrypoints.logger import RequestLogger
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
@@ -17,10 +16,10 @@ 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):
@@ -179,11 +178,21 @@ class OpenAIvLLMEngine(vLLMEngine):
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()
# 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"Deferring OpenAI engine initialization with {len(self.lora_adapters)} LoRA adapter(s) until first request")
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("Deferring OpenAI engine initialization until first request")
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'):
@@ -208,12 +217,21 @@ class OpenAIvLLMEngine(vLLMEngine):
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):
logging.info("Initializing OpenAI serving engines...")
self.model_config = self.llm.model_config
self.base_model_paths = [
BaseModelPath(name=self.served_model_name, model_path=self.engine_args.model)
]
@@ -237,13 +255,13 @@ class OpenAIvLLMEngine(vLLMEngine):
request_logger=None,
chat_template=chat_template,
chat_template_content_format="auto",
reasoning_parser=os.getenv('REASONING_PARSER', "") or None,
enable_auto_tools=os.getenv('ENABLE_AUTO_TOOL_CHOICE', 'false').lower() == 'true',
tool_parser=os.getenv('TOOL_CALL_PARSER', "") or None,
enable_prompt_tokens_details=os.getenv('ENABLE_PROMPT_TOKENS_DETAILS', 'false').lower() == 'true',
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=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',
@@ -261,10 +279,10 @@ class OpenAIvLLMEngine(vLLMEngine):
if hasattr(self.chat_engine, 'warmup'):
await self.chat_engine.warmup()
logging.info("OpenAI serving engines initialized successfully")
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
+2
View File
@@ -15,6 +15,7 @@ RENAME_ARGS_MAP = {
DEFAULT_ARGS = {
"disable_log_stats": os.getenv('DISABLE_LOG_STATS', '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)),
@@ -39,6 +40,7 @@ DEFAULT_ARGS = {
"disable_sliding_window": os.getenv('DISABLE_SLIDING_WINDOW', 'False').lower() == 'true',
"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
src/handler.py
+38 -36
View File
@@ -1,53 +1,55 @@
import multiprocessing
import os
import sys
import multiprocessing
import traceback
import runpod
from runpod import RunPodLogger
from utils import JobInput
from engine import vLLMEngine, OpenAIvLLMEngine
log = RunPodLogger()
# Prevent re-initialization in vLLM worker subprocesses
if multiprocessing.current_process().name == "MainProcess":
vllm_engine = vLLMEngine()
openai_engine = OpenAIvLLMEngine(vllm_engine)
else:
vllm_engine = None
openai_engine = None
vllm_engine = None
openai_engine = None
async def handler(job):
job_input = JobInput(job["input"])
engine = openai_engine if job_input.openai_route else vllm_engine
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:
err_str = str(e)
is_cuda_error = "CUDA" in err_str or "cuda" in err_str
is_oom = "out of memory" in err_str.lower()
error_str = str(e)
full_traceback = traceback.format_exc()
if is_cuda_error and not is_oom:
log.error(f"CUDA error (non-OOM), exiting for worker recycle: {e}")
traceback.print_exc()
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)
elif is_oom:
log.error(f"CUDA OOM error: {e}")
traceback.print_exc()
yield {"error": str(e)}
else:
log.error(f"Handler error: {e}")
traceback.print_exc()
yield {"error": str(e)}
runpod.serverless.start(
{
"handler": handler,
"concurrency_modifier": lambda x: vllm_engine.max_concurrency if vllm_engine else 1,
"return_aggregate_stream": True,
}
)
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,
}
)