feat(BA-3071): add support for DGX spark devices

[kyujin-cho]

resolves #6798 (BA-3071).

This pull request refactors how device allocation and unified device management are handled in the agent resource specification and kernel lifecycle. The main improvements are the introduction of the DeviceView dataclass and the device_list property, which provide a unified and clearer way to access all devices (including unified devices) attached to a kernel. This change simplifies device iteration logic throughout the codebase and ensures unified devices are consistently included in resource specifications and related operations.

Device allocation and unified device management:

• Added the DeviceView dataclass and the device_list property to KernelResourceSpec, enabling a unified view of all devices (including unified devices) attached to a kernel. Unified devices are now tracked via a new unified_devices attribute, and the resource spec serialization/deserialization logic is updated to include this information. (src/ai/backend/agent/resources.py) [1] [2] [3] [4]
• Introduced the SlotTypes.UNIFIED enum value to distinguish unified device slots, supporting the new unified device logic. (src/ai/backend/common/types.py)

API and method refactoring:

• Added the generate_resource_spec async method to the agent, which builds a complete resource spec including unified devices, replacing direct calls to prepare_resource_spec throughout the codebase. (src/ai/backend/agent/agent.py) [1] [2]
• Updated device iteration logic in methods responsible for mounting static binaries, preparing hook mounts, and container hook selection to use the new device_list property instead of iterating over allocations and manually checking for nonzero allocations. This improves clarity and ensures unified devices are handled correctly. (src/ai/backend/agent/agent.py, src/ai/backend/agent/stage/kernel_lifecycle/docker/environ.py, src/ai/backend/agent/stage/kernel_lifecycle/docker/mount/krunner.py) [1] [2] [3]

Documentation and typing improvements:

• Added docstrings and type hints to new and updated methods, clarifying the purpose and usage of resource spec generation and device aggregation. (src/ai/backend/agent/agent.py, src/ai/backend/agent/resources.py) [1] [2]

These changes collectively improve device management consistency, reduce code duplication, and make the agent's resource handling logic easier to maintain and extend.

[Copilot]

Pull Request Overview

This PR adds support for DGX Spark devices by introducing a new "unified" slot type that allows implicit attachment of certain accelerator devices to all created kernels.

Key changes:

• Added UNIFIED slot type enum value to distinguish unified accelerators
• Modified kernel creation logic to automatically attach unified-type accelerators to every kernel
• Removed x86_64 platform-specific wheel build configuration from build script

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
src/ai/backend/common/types.py	Added UNIFIED enum value to SlotTypes for unified accelerator devices
src/ai/backend/agent/agent.py	Added logic in create_kernel to automatically attach unified accelerators to all kernels
scripts/build-wheels.sh	Removed x86_64 platform-specific wheel build configuration

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull Request Overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[kyujin-cho]

@copilot Update PR description to abstract changes introduced in this PR.

[Copilot]

@kyujin-cho I've opened a new pull request, #6850, to work on those changes. Once the pull request is ready, I'll request review from you.