Migrate cuda to runtime

The CUDA variant is absorbed into the base image. The CUDA SDK was
already installed at runtime by devinstall_software.sh; the only
build-time additions were Nsight GUI dependencies and the CUDA_VERSION
env var. Nsight GUI deps now install conditionally with INSTALL_TOOLS.

- Delete Dockerfile.cuda and cuda-image.yml CI workflow
- Move Nsight GUI deps (libglvnd-opengl, etc.) to devinstall_software.sh
  under INSTALL_TOOLS + CUDA_VERSION
- Remove IMAGE_TAG version embedding for cuda in triton-dev-containers.sh
- Resolve -d cuda to the base image
- Remove cuda-image build target from Makefile

stack-info: PR: #2, branch: hinriksnaer/stack/17

Author: hinriksnaer

.github/workflows/cuda-image.yml

@@ -48,14 +48,12 @@ IMAGE_REPO = quay.io/triton-dev-containers
 
 # Image name definitions (clean and extensible)
 BASE_IMAGE_NAME = base
-CUDA_IMAGE_NAME = cuda
 ROCM_IMAGE_NAME = rocm
 
 # Image tags
 IMAGE_TAG      := centos$(CENTOS_VERSION)
 
 BASE_IMAGE_TAG := $(IMAGE_TAG)
-CUDA_IMAGE_TAG := $(CUDA_VERSION)-$(IMAGE_TAG)
 ROCM_IMAGE_TAG := $(ROCM_VERSION)-$(IMAGE_TAG)
 
 # ------------------------------------------------------------------------------
@@ -134,7 +132,7 @@ define build-image
 endef
 
 .PHONY: build-images
-build-images: cuda-image rocm-image ## Build all container images
+build-images: base-image rocm-image ## Build all container images
 
 define base_image_build_args
 --build-arg CENTOS_VERSION=$(CENTOS_VERSION) \
@@ -152,11 +150,6 @@ define image_build_args
 --build-arg BASE_IMAGE_TAG=$(BASE_IMAGE_TAG)
 endef
 
-.PHONY: cuda-image
-cuda-image: dockerfiles/Dockerfile.cuda | base-image ## Build a CUDA container image
-	$(call build-image,$(CUDA_IMAGE_NAME),$(CUDA_IMAGE_TAG),$(image_build_args) \
-		--build-arg BUILD_CUDA_VERSION=$(CUDA_VERSION),$<)
-
 .PHONY: rocm-image
 rocm-image: dockerfiles/Dockerfile.rocm | base-image ## Build a ROCm container image
 	$(call build-image,$(ROCM_IMAGE_NAME),$(ROCM_IMAGE_TAG),$(image_build_args) \
@@ -273,8 +266,8 @@ base-run: ## Run the base container
 	@./triton-dev-containers.sh $(RUNTIME_ARGS) -d base
 
 .PHONY: cuda-run
-cuda-run: ## Run the CUDA container image
-	@./triton-dev-containers.sh $(CUDA_RUNTIME_ARGS) -d $(CUDA_IMAGE_NAME)
+cuda-run: ## Run a CUDA container
+	@./triton-dev-containers.sh $(CUDA_RUNTIME_ARGS) -d cuda
 
 .PHONY: cpu-run
 cpu-run: ## Run a CPU container
@@ -285,8 +278,8 @@ rocm-run: ## Run the ROCm container image
 	@./triton-dev-containers.sh $(ROCM_RUNTIME_ARGS) -d $(ROCM_IMAGE_NAME)
 
 .PHONY: helion-cuda-run
-helion-cuda-run: ## Run the Helion CUDA container image
-	@./triton-dev-containers.sh $(CUDA_RUNTIME_ARGS) -d $(CUDA_IMAGE_NAME) -k helion
+helion-cuda-run: ## Run the Helion CUDA container
+	@./triton-dev-containers.sh $(CUDA_RUNTIME_ARGS) -d cuda -k helion
 
 .PHONY: helion-cpu-run
 helion-cpu-run: ## Run the Helion CPU container
@@ -297,8 +290,8 @@ helion-rocm-run: ## Run the Helion ROCm container image
 	@./triton-dev-containers.sh $(ROCM_RUNTIME_ARGS) -d $(ROCM_IMAGE_NAME) -k helion
 
 .PHONY: triton-cuda-run
-triton-cuda-run: ## Run the Triton CUDA container image
-	@./triton-dev-containers.sh $(CUDA_RUNTIME_ARGS) -d $(CUDA_IMAGE_NAME) -k triton
+triton-cuda-run: ## Run the Triton CUDA container
+	@./triton-dev-containers.sh $(CUDA_RUNTIME_ARGS) -d cuda -k triton
 
 .PHONY: triton-cpu-run
 triton-cpu-run: ## Run the Triton CPU container
@@ -309,8 +302,8 @@ triton-rocm-run: ## Run the Triton ROCm container image
 	@./triton-dev-containers.sh $(ROCM_RUNTIME_ARGS) -d $(ROCM_IMAGE_NAME) -k triton
 
 .PHONY: torch-cuda-run
-torch-cuda-run: ## Run the PyTorch CUDA container image
-	@./triton-dev-containers.sh $(CUDA_RUNTIME_ARGS) -d $(CUDA_IMAGE_NAME) -k torch
+torch-cuda-run: ## Run the PyTorch CUDA container
+	@./triton-dev-containers.sh $(CUDA_RUNTIME_ARGS) -d cuda -k torch
 
 .PHONY: torch-cpu-run
 torch-cpu-run: ## Run the PyTorch CPU container
@@ -321,8 +314,8 @@ torch-rocm-run: ## Run the PyTorch ROCm container image
 	@./triton-dev-containers.sh $(ROCM_RUNTIME_ARGS) -d $(ROCM_IMAGE_NAME) -k torch
 
 .PHONY: vllm-cuda-run
-vllm-cuda-run: ## Run the vLLM CUDA container image
-	@./triton-dev-containers.sh $(CUDA_RUNTIME_ARGS) -d $(CUDA_IMAGE_NAME) -k vllm
+vllm-cuda-run: ## Run the vLLM CUDA container
+	@./triton-dev-containers.sh $(CUDA_RUNTIME_ARGS) -d cuda -k vllm
 
 .PHONY: vllm-cpu-run
 vllm-cpu-run: ## Run the vLLM CPU container

Makefile

@@ -71,8 +71,6 @@ Arguments that can be added to the build commands below, i.e. `OPTION=VALUE`.
 
 - `CENTOS_VERSION`
   - CentOS Stream image version to use, i.e. `9`
-- `CUDA_VERSION`
-  - Use the CUDA RPM package version, i.e. `12-9`
 - `GOSU_VERSION`
   - Check the gosu GitHub for the desired release
   - <https://github.com/tianon/gosu>
@@ -95,12 +93,6 @@ make build-images [OPTIONS]
 make base-image [OPTIONS]
 ```
 
-#### NVIDIA CUDA container build
-
-```sh
-make cuda-image [OPTIONS]
-```
-
 #### AMD ROCm container build
 
 ```sh
@@ -117,7 +109,7 @@ Arguments that can be added to the run commands below, i.e. `OPTION=VALUE`.
 
 - `CUDA_VERSION`
   - Use the CUDA RPM package version, i.e. `12-9`
-  - Specifies the CUDA image to use
+  - Selects the CUDA SDK version to install at runtime
 - `ROCM_VERSION`
   - Specifies the ROCm image to use
   - Use the ROCm RPM package version, i.e. `6.3.4`

README.md

@@ -125,6 +125,10 @@ EOF
 	fi
 
 	if [ "${INSTALL_TOOLS:-}" = "true" ]; then
+		echo "Installing NVIDIA Nsight GUI dependencies ..."
+		$SUDO dnf -y install libglvnd-opengl libxkbfile \
+			qt5-qtwayland xcb-util-cursor
+
 		echo "Installing NVIDIA Nsight ..."
 		$SUDO dnf -y install cublasmp "cuda-gdb-$CUDA_VERSION" \
 			"cuda-nsight-$CUDA_VERSION" "cuda-nsight-compute-$CUDA_VERSION" \

dockerfiles/Dockerfile.cuda

@@ -283,7 +283,6 @@ set_device_opts() {
 
 		set_env_opt CUDA_VERSION "${DEFAULT_ENV_OPTS["CUDA_VERSION"]}"
 		set_env_opt CUDA_VISIBLE_DEVICES "${DEFAULT_ENV_OPTS["CUDA_VISIBLE_DEVICES"]}"
-		IMAGE_TAG=${ENV_OPTS["CUDA_VERSION"]}-${DEFAULT_IMAGE_TAG}
 		;;
 	cpu)
 		set_env_opt TRITON_CPU_BACKEND 1
@@ -439,8 +438,8 @@ fi
 
 # Resolve image name from target device (cpu uses the base image)
 case ${TARGET_DEVICE,,} in
-	cpu|base) IMAGE_NAME=base ;;
-	*)        IMAGE_NAME=${TARGET_DEVICE} ;;
+	cpu|cuda|base) IMAGE_NAME=base ;;
+	*)             IMAGE_NAME=${TARGET_DEVICE} ;;
 esac
 
 IMAGE=${IMAGE:-${DEFAULT_IMAGE_REPO}/${IMAGE_NAME}:${IMAGE_TAG:-${DEFAULT_IMAGE_TAG}}}