ci/build/feat: bump vLLM libs to v0.4.2 and other deps in Dockerfile.ubi (#23)

Changes:
- vLLM v0.4.2 was published today, update our build to use pre-built
libs from their wheel
- bump other dependencies in the image build (base UBI image, miniforge,
flash attention, grpcio-tools, accelerate)
- little cleanup to remove `PYTORCH_` args that are no longer used

---------

Signed-off-by: Travis Johnson <[email protected]>

Author: tjohnson31415

Dockerfile.ubi

@@ -2,11 +2,8 @@
 # docs/source/dev/dockerfile-ubi/dockerfile-ubi.rst
 
 ## Global Args #################################################################
-ARG BASE_UBI_IMAGE_TAG=9.3-1612
+ARG BASE_UBI_IMAGE_TAG=9.4-949.1714662671
 ARG PYTHON_VERSION=3.11
-ARG PYTORCH_INDEX="https://download.pytorch.org/whl"
-# ARG PYTORCH_INDEX="https://download.pytorch.org/whl/nightly"
-ARG PYTORCH_VERSION=2.1.2
 
 # NOTE: This setting only has an effect when not using prebuilt-wheel kernels
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0+PTX"
@@ -149,8 +146,8 @@ RUN microdnf install -y \
     && microdnf clean all
 
 ARG PYTHON_VERSION
-# 0.4.1 is built for CUDA 12.1 and PyTorch 2.1.2
-ARG VLLM_WHEEL_VERSION=0.4.1
+# 0.4.2 is built for CUDA 12.1 and PyTorch 2.3.0
+ARG VLLM_WHEEL_VERSION=0.4.2
 
 RUN curl -Lo vllm.whl https://github.com/vllm-project/vllm/releases/download/v${VLLM_WHEEL_VERSION}/vllm-${VLLM_WHEEL_VERSION}-cp${PYTHON_VERSION//.}-cp${PYTHON_VERSION//.}-manylinux1_x86_64.whl \
     && unzip vllm.whl \
@@ -223,7 +220,7 @@ RUN microdnf install -y git \
 ARG max_jobs=2
 ENV MAX_JOBS=${max_jobs}
 # flash attention version
-ARG flash_attn_version=v2.5.6
+ARG flash_attn_version=v2.5.8
 ENV FLASH_ATTN_VERSION=${flash_attn_version}
 
 WORKDIR /usr/src/flash-attention-v2
@@ -260,9 +257,9 @@ RUN --mount=type=bind,from=flash-attn-builder,src=/usr/src/flash-attention-v2,ta
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install \
         # additional dependencies for the TGIS gRPC server
-        grpcio==1.62.1 \
+        grpcio-tools==1.63.0 \
         # additional dependencies for openai api_server
-        accelerate==0.28.0 \
+        accelerate==0.30.0 \
         # hf_transfer for faster HF hub downloads
         hf_transfer==0.1.6