Merge branch 'main' into clean_logs

Author: XkunW

.github/workflows/docker.yml

@@ -45,7 +45,7 @@ jobs:
           images: vectorinstitute/vector-inference
 
       - name: Build and push Docker image
-        uses: docker/build-push-action@14487ce63c7a62a4a324b0bfb37086795e31c6c1
+        uses: docker/build-push-action@1dc73863535b631f98b2378be8619f83b136f4a0
         with:
           context: .
           file: ./Dockerfile

.github/workflows/unit_tests.yml

@@ -72,7 +72,7 @@ jobs:
           uv run pytest tests/test_imports.py
 
       - name: Upload coverage to Codecov
-        uses: codecov/codecov-action@v5.4.2
+        uses: codecov/codecov-action@v5.4.3
         with:
           token: ${{ secrets.CODECOV_TOKEN }}
           file: ./coverage.xml

.pre-commit-config.yaml

@@ -17,7 +17,7 @@ repos:
     - id: check-toml
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: 'v0.11.8'
+    rev: 'v0.11.11'
     hooks:
     - id: ruff
       args: [--fix, --exit-non-zero-on-fix]

README.md

@@ -7,6 +7,7 @@
 [![code checks](https://github.com/VectorInstitute/vector-inference/actions/workflows/code_checks.yml/badge.svg)](https://github.com/VectorInstitute/vector-inference/actions/workflows/code_checks.yml)
 [![docs](https://github.com/VectorInstitute/vector-inference/actions/workflows/docs.yml/badge.svg)](https://github.com/VectorInstitute/vector-inference/actions/workflows/docs.yml)
 [![codecov](https://codecov.io/github/VectorInstitute/vector-inference/branch/main/graph/badge.svg?token=NI88QSIGAC)](https://app.codecov.io/github/VectorInstitute/vector-inference/tree/main)
+[![vLLM](https://img.shields.io/badge/vllm-0.8.5.post1-blue)](https://docs.vllm.ai/en/v0.8.5.post1/index.html)
 ![GitHub License](https://img.shields.io/github/license/VectorInstitute/vector-inference)
 
 This repository provides an easy-to-use solution to run inference servers on [Slurm](https://slurm.schedmd.com/overview.html)-managed computing clusters using [vLLM](https://docs.vllm.ai/en/latest/). **All scripts in this repository runs natively on the Vector Institute cluster environment**. To adapt to other environments, update the environment variables in [`vec_inf/client/slurm_vars.py`](vec_inf/client/slurm_vars.py), and the model config for cached model weights in [`vec_inf/config/models.yaml`](vec_inf/config/models.yaml) accordingly.
@@ -17,7 +18,7 @@ If you are using the Vector cluster environment, and you don't need any customiz
 ```bash
 pip install vec-inf
 ```
-Otherwise, we recommend using the provided [`Dockerfile`](Dockerfile) to set up your own environment with the package
+Otherwise, we recommend using the provided [`Dockerfile`](Dockerfile) to set up your own environment with the package. The latest image has `vLLM` version `0.8.5.post1`.
 
 ## Usage
 
@@ -85,7 +86,7 @@ models:
     vllm_args:
       --max-model-len: 1010000
       --max-num-seqs: 256
-      --compilation-confi: 3
+      --compilation-config: 3
 ```
 
 You would then set the `VEC_INF_CONFIG` path using:
@@ -94,7 +95,11 @@ You would then set the `VEC_INF_CONFIG` path using:
 export VEC_INF_CONFIG=/h/<username>/my-model-config.yaml
 ```
 
-Note that there are other parameters that can also be added to the config but not shown in this example, check the [`ModelConfig`](vec_inf/client/config.py) for details.
+**NOTE**
+* There are other parameters that can also be added to the config but not shown in this example, check the [`ModelConfig`](vec_inf/client/config.py) for details.
+* Check [vLLM Engine Arguments](https://docs.vllm.ai/en/stable/serving/engine_args.html) for the full list of available vLLM engine arguments, the default parallel size for any parallelization is default to 1, so none of the sizes were set specifically in this example
+* For GPU partitions with non-Ampere architectures, e.g. `rtx6000`, `t4v2`, BF16 isn't supported. For models that have BF16 as the default type, when using a non-Ampere GPU, use FP16 instead, i.e. `--dtype: float16`
+* Setting `--compilation-config` to `3` currently breaks multi-node model launches, so we don't set them for models that require multiple nodes of GPUs.
 
 #### Other commands
 
@@ -161,8 +166,9 @@ Once the inference server is ready, you can start sending in inference requests.
     },
     "prompt_logprobs":null
 }
+
 ```
-**NOTE**: For multimodal models, currently only `ChatCompletion` is available, and only one image can be provided for each prompt.
+**NOTE**: Certain models don't adhere to OpenAI's chat template, e.g. Mistral family. For these models, you can either change your prompt to follow the model's default chat template or provide your own chat template via `--chat-template: TEMPLATE_PATH`.
 
 ## SSH tunnel from your local device
 If you want to run inference from your local device, you can open a SSH tunnel to your cluster environment like the following:

docs/index.md

@@ -10,4 +10,4 @@ If you are using the Vector cluster environment, and you don't need any customiz
 pip install vec-inf
 ```
 
-Otherwise, we recommend using the provided [`Dockerfile`](https://github.com/VectorInstitute/vector-inference/blob/main/Dockerfile) to set up your own environment with the package.
+Otherwise, we recommend using the provided [`Dockerfile`](https://github.com/VectorInstitute/vector-inference/blob/main/Dockerfile) to set up your own environment with the package. The latest image has `vLLM` version `0.8.5.post1`.

docs/user_guide.md

@@ -91,7 +91,11 @@ You would then set the `VEC_INF_CONFIG` path using:
 export VEC_INF_CONFIG=/h/<username>/my-model-config.yaml
 ```
 
-Note that there are other parameters that can also be added to the config but not shown in this example, check the [`ModelConfig`](https://github.com/VectorInstitute/vector-inference/blob/main/vec_inf/client/config.py) for details.
+**NOTE**
+* There are other parameters that can also be added to the config but not shown in this example, check the [`ModelConfig`](https://github.com/VectorInstitute/vector-inference/blob/main/vec_inf/client/config.py) for details.
+* Check [vLLM Engine Arguments](https://docs.vllm.ai/en/stable/serving/engine_args.html) for the full list of available vLLM engine arguments. The default parallel size for any parallelization defaults to 1, so none of the sizes were set specifically in this example.
+* For GPU partitions with non-Ampere architectures, e.g. `rtx6000`, `t4v2`, BF16 isn't supported. For models that have BF16 as the default type, when using a non-Ampere GPU, use FP16 instead, i.e. `--dtype: float16`.
+* Setting `--compilation-config` to `3` currently breaks multi-node model launches, so we don't set them for models that require multiple nodes of GPUs.
 
 ### `status` command
 
@@ -254,8 +258,7 @@ Once the inference server is ready, you can start sending in inference requests.
 }
 ```
 
-
-**NOTE**: For multimodal models, currently only `ChatCompletion` is available, and only one image can be provided for each prompt.
+**NOTE**: Certain models don't adhere to OpenAI's chat template, e.g. Mistral family. For these models, you can either change your prompt to follow the model's default chat template or provide your own chat template via `--chat-template: TEMPLATE_PATH`.
 
 ## SSH tunnel from your local device
 

examples/README.md

@@ -9,3 +9,4 @@
   - [`logits.py`](logits/logits.py): Python example of getting logits from hosted model.
 - [`api`](api): Examples for using the Python API
   - [`basic_usage.py`](api/basic_usage.py): Basic Python example demonstrating the Vector Inference API
+- [`slurm_dependency`](slurm_dependency): Example of launching a model with `vec-inf` and running a downstream SLURM job that waits for the server to be ready before sending a request.

examples/slurm_dependency/README.md

@@ -0,0 +1,33 @@
+# SLURM Dependency Workflow Example
+
+This example demonstrates how to launch a model server using `vec-inf`, and run a downstream SLURM job that waits for the server to become ready before querying it.
+
+## Files
+
+This directory contains the following:
+
+1. [run_workflow.sh](run_workflow.sh)
+   Launches the model server and submits the downstream job with a dependency, so it starts only after the server job begins running.
+
+2. [downstream_job.sbatch](downstream_job.sbatch)
+   A SLURM job script that runs the downstream logic (e.g., prompting the model).
+
+3. [run_downstream.py](run_downstream.py)
+   A Python script that waits until the inference server is ready, then sends a request using the OpenAI-compatible API.
+
+## What to update
+
+Before running this example, update the following in [downstream_job.sbatch](downstream_job.sbatch):
+
+- `--job-name`, `--output`, and `--error` paths
+- Virtual environment path in the `source` line
+- SLURM resource configuration (e.g., partition, memory, GPU)
+
+Also update the model name in [run_downstream.py](run_downstream.py) to match what you're launching.
+
+## Running the example
+
+First, activate a virtual environment where `vec-inf` is installed. Then, from this directory, run:
+
+```bash
+bash run_workflow.sh

examples/slurm_dependency/downstream_job.sbatch

@@ -0,0 +1,18 @@
+#!/bin/bash
+#SBATCH --job-name=Meta-Llama-3.1-8B-Instruct-downstream
+#SBATCH --partition=a40
+#SBATCH --qos=m2
+#SBATCH --time=08:00:00
+#SBATCH --nodes=1
+#SBATCH --gpus-per-node=1
+#SBATCH --cpus-per-task=4
+#SBATCH --mem=8G
+#SBATCH --output=$HOME/.vec-inf-logs/Meta-Llama-3.1-8B-Instruct-downstream.%j.out
+#SBATCH --error=$HOME/.vec-inf-logs/Meta-Llama-3.1-8B-Instruct-downstream.%j.err
+
+# Activate your environment
+# TODO: update this path to match your venv location
+source $HOME/vector-inference/.venv/bin/activate
+
+# Wait for the server to be ready using the job ID passed as CLI arg
+python run_downstream.py "$SERVER_JOB_ID"

examples/slurm_dependency/run_downstream.py

@@ -0,0 +1,26 @@
+"""Example script to query a launched model via the OpenAI-compatible API."""
+
+import sys
+
+from openai import OpenAI
+
+from vec_inf.client import VecInfClient
+
+
+if len(sys.argv) < 2:
+    raise ValueError("Expected server job ID as the first argument.")
+job_id = int(sys.argv[1])
+
+vi_client = VecInfClient()
+print(f"Waiting for SLURM job {job_id} to be ready...")
+status = vi_client.wait_until_ready(slurm_job_id=job_id)
+print(f"Server is ready at {status.base_url}")
+
+api_client = OpenAI(base_url=status.base_url, api_key="EMPTY")
+resp = api_client.completions.create(
+    model="Meta-Llama-3.1-8B-Instruct",
+    prompt="Where is the capital of Canada?",
+    max_tokens=20,
+)
+
+print(resp)