[Store] Refactor accelerator device registry and staging copies

[Aionw]

Description

Closes #2582.

This PR introduces a Store-side AcceleratorDevice abstraction for local
accelerator memory operations. Vendor-specific pointer query, context switch,
copy, and pinned-host allocation are moved behind per-vendor device
implementations, while Store call sites use the registry's available
accelerator list and pointer-based dispatch.

The latest revision also simplifies the accelerator registry around static
device registration:

• removes the public RegisterAcceleratorDevice entrypoint and keeps
  registration available only through static AcceleratorDeviceRegistrar
  instances
• stores registered devices in a simple list instead of maintaining a separate
  vendor-indexed table
• caches available devices with an atomic immutable snapshot so the common
  RuntimeAccelerators(false) path avoids taking a mutex after initialization
• keeps ensure=true as the explicit refresh path

It also tightens runtime copy behavior:

• skips null pointer queries in FindDeviceForPointer
• removes the unused IsDevicePointer wrapper
• reuses FindDeviceForPointer results at D2H staging sites to avoid querying
  the same pointer twice
• passes explicit copy directions for H2D, D2H, and D2D copies instead of
  relying on kAuto
• fixes Ascend current-device reporting to return the logical device id
• keeps pinned-host allocation fallback in PinnedBufferPool

Module

• Transfer Engine (mooncake-transfer-engine)
• Mooncake Store (mooncake-store)
• Mooncake EP (mooncake-ep)
• Mooncake PG (mooncake-pg)
• Integration (mooncake-integration)
• P2P Store (mooncake-p2p-store)
• Python Wheel (mooncake-wheel)
• Common (mooncake-common)
• Mooncake RL (mooncake-rl)
• CI/CD
• Docs
• Other

Type of Change

• Bug fix
• New feature
• Refactor
• Breaking change
• Documentation update
• Performance improvement
• Other

How Has This Been Tested?

Test commands:

Test by UT

Test results:

• Unit tests pass
• Integration tests pass (if applicable)
• Manual testing done

runtime_accelerator_test passes with 9/9 tests.

Note: targeted pre-commit was run, but the project mooncake-code-format
hook cannot complete in this environment because clang-format-20 is not
installed. The other relevant hooks for the changed files passed.

Checklist

• I have performed a self-review of my own code
• I have formatted my code using ./scripts/code_format.sh
• I have run pre-commit run --all-files and all hooks pass
• I have updated the documentation (if applicable)
• I have added tests to prove my changes are effective
• For changes >500 LOC: I have filed an RFC issue

AI Assistance Disclosure

• No AI tools were used
• AI tools were used (specify below)

This PR was implemented with AI assistance.

AI generated the implementation. Human review covered 100% of the diff.

[gemini-code-assist]

Code Review

This pull request refactors platform-specific accelerator memory management and copy operations into an object-oriented AcceleratorDevice abstraction and registry pattern, replacing direct preprocessor macro checks. The review feedback identifies several critical issues: a restriction in PinnedBufferPool that limits pinned host memory allocation to single-accelerator environments and causes performance degradation in multi-accelerator setups; a compilation error in ascend_accelerator_device.cpp when USE_ASCEND_DIRECT is defined due to a pointer type mismatch; and performance overhead in the CUDA, HIP, and Ascend implementations from repeatedly querying device counts in the hot path, which should instead be cached statically.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[Aionw]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces a unified device abstraction layer (AcceleratorDevice, AcceleratorRegistry, and RuntimeAccelerator) to replace platform-specific conditional compilation across the codebase, updating PinnedBufferPool, Client, FileStorage, and MemcpyWorkerPool to use this new interface. The review feedback highlights several critical improvements: resolving an inconsistency in AscendAcceleratorDevice where CurrentDeviceId() returns a physical ID instead of a logical ID, adding a nullptr check in FindDeviceForPointer to prevent driver errors, optimizing CopyMaybeAccelerator by passing explicit copy directions, refactoring move constructors in PinnedHostBuffer and Buffer to use member initializers, avoiding self-move-assignment in PinnedBufferPool, and handling duplicate device registrations in AcceleratorRegistry consistently.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 51.41243% with 172 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...-store/src/device/cuda_like_accelerator_device.cpp	0.00%	41 Missing ⚠️
mooncake-store/include/pinned_buffer_pool.h	0.00%	40 Missing ⚠️
mooncake-store/include/pinned_host_buffer.h	0.00%	24 Missing ⚠️
mooncake-store/src/device/accelerator_registry.cpp	48.64%	19 Missing ⚠️
mooncake-store/src/file_storage.cpp	0.00%	10 Missing ⚠️
mooncake-store/src/device/accelerator_device.cpp	0.00%	8 Missing ⚠️
mooncake-store/src/transfer_task.cpp	57.89%	8 Missing ⚠️
mooncake-store/src/device/runtime_accelerator.cpp	86.27%	7 Missing ⚠️
mooncake-store/src/real_client.cpp	36.36%	7 Missing ⚠️
mooncake-store/tests/runtime_accelerator_test.cpp	94.17%	6 Missing ⚠️
... and 1 more

📢 Thoughts on this report? Let us know!