fix(vllm_inference): upgrade vLLM to 0.19.1 for transformers 5.x compat

Signed-off-by: svc-bionemo <267129667+svc-bionemo@users.noreply.github.com>

Author: svc-bionemo

.github/workflows/unit-tests-recipes.yml

@@ -56,7 +56,6 @@ jobs:
             bionemo-recipes/models/**
             bionemo-recipes/recipes/**
             !bionemo-recipes/recipes/*megatron/**
-            !bionemo-recipes/recipes/vllm_inference/**
 
       - id: set-dirs
         name: Determine which directories to run

bionemo-recipes/recipes/vllm_inference/esm2/install_vllm.sh

@@ -9,14 +9,13 @@ echo "Building vLLM for CUDA arch: $ARCH (MAX_JOBS=$MAX_JOBS)"
 
 cd /workspace
 if [ ! -d vllm ]; then
-    git clone --branch v0.15.1 --depth 1 https://github.com/vllm-project/vllm.git
+    git clone --branch v0.19.1 --depth 1 https://github.com/vllm-project/vllm.git
 fi
 cd vllm
 python use_existing_torch.py
 TORCH_CUDA_ARCH_LIST="$ARCH" MAX_JOBS="$MAX_JOBS" \
     uv pip install -r requirements/build.txt --system
 TORCH_CUDA_ARCH_LIST="$ARCH" MAX_JOBS="$MAX_JOBS" \
     uv pip install --no-build-isolation -e . --system
-pip install --upgrade "transformers[torch]"
 
 echo "vLLM installed for arch $ARCH"

bionemo-recipes/recipes/vllm_inference/esm2/tests/test_vllm.py

@@ -22,8 +22,6 @@
 so the HuggingFace paths replicate that post-processing for comparison.
 """
 
-from pathlib import Path
-
 import numpy as np
 import pytest
 import torch
@@ -44,7 +42,6 @@
 
 EXPORT_TAG = "esm2_t6_8M_UR50D"
 REFERENCE_MODEL_ID = "nvidia/esm2_t6_8M_UR50D"
-ESM2_MODEL_DIR = Path(__file__).resolve().parent.parent
 
 SEQUENCES = [
     "LKGHAMCLGCLHMLMCGLLAGAMCGLMKLLKCCGKCLMHLMKAMLGLKCACHHHHLLLHACAAKKLCLGAKLAMGLKLLGAHGKGLKMACGHHMLHLHMH",