[Ceng23333]

Summary

Add CPU GetMemorySpace specialization so the post-refactor Infini stack (QY MiniCPM5 pin) can build against InfiniCCL.

Changes

• src/devices/cpu/device_.h: provide GetMemorySpace returning MemorySpace::kHost for CPU device.

Validation

• Cluster / example programs: N/A for this 1-file header specialization (no NCCL path change). Compile-time fix only.
• Cherry-picked from Ceng tip 8dda36c onto current master; clang-format applied.
• Integration-validated via InfiniTensorWorktree release/2026.10.13-qy-minicpm5-16a3-fa-longctx-cudagraph.

Related

• Companion InfiniOps #1002 / InfiniRT feat: support NCCL on Iluvatar #49
• Downstream InfiniCore submodule pin + InfiniLM FA/longctx PR

Made with Cursor

Stack PR index (merge order)

1. InfiniOps: feat(qy): QY ops backend with FA and longctx MoE/RoPE kernels InfiniOps#1002
2. InfiniRT: feat(qy): enable QY CUDA runtime device backend InfiniRT#49
3. InfiniCCL: feat(qy): QY-compatible CPU device header fix #77
4. InfiniCore: chore(qy): pin InfiniRT/Ops/CCL for MiniCPM5 FA/longctx InfiniCore#1575
5. InfiniLM: feat(qy): MiniCPM5 FA, longctx MoE/RoPE, CUDA-graph capture safety InfiniLM#597 → base refactor/adopt-modern-infini-stack

ITW pin: https://github.com/Ceng23333/InfiniTensorWorktree/tree/release/2026.10.13-qy-minicpm5-16a3-fa-longctx-cudagraph

[Ceng23333]

QY required smoke (PR head)

• Head SHA: 13e5f613693c35de09e1266c9dd1a50a60a67e82
• Host: Denglin QY build container
• Command: InfiniLM build_infini_stack.py --backend qy (CCL configure/build/install, TP1 CPU/OMPI)
• Result: PASS (compile/link OK)
• Log: …/logs/prhead-verify/stack-m0-build-test.log

Notes: no QY NCCL / TP2 smoke (out of scope for this PR).

[Ziminli]

This PR has been closed as duplicate work. The same fix has been applied in #79. Your contribution is appreciated and has been acknowledged.

Additionally, please note that InfiniCCL, as a project under the InfiniTensor organization, does not currently accept AI/agent as a co-author.

[Ceng23333]

QY required smoke (PR head) — detailed

• Head SHA: 13e5f613693c35de09e1266c9dd1a50a60a67e82
• Host: Denglin QY build container
• Command: InfiniLM build_infini_stack.py --backend qy (CCL configure / build / install)
• Overall: PASS
• Log: …/logs/prhead-verify/stack-m0-build-test.log, stack-m2-build-test.log

Step	Config	Result
cmake configure	WITH_NCCL=OFF, devices [cpu], backends [ompi]	PASS
cmake --build …/infiniccl	libinfiniccl.so	PASS
cmake --install	prefix lib/libinfiniccl.so + headers + icclrun	PASS
Stack consume (m0)	linked into QY stack with RT 95d30fe / Ops ed0b40d	PASS
Stack consume (m2 FA)	same, full 23-op QY config	PASS

No QY NCCL / TP2 AllReduce smoke (out of scope for this 1-file CPU header PR).

[Ceng23333]

Fresh QY NCCL enablement on current master for ITW v2026.11.12-qy-moe-bf16-warpnk: https://github.com/InfiniTensor/InfiniCCL/pull/null. CPU GetMemorySpace from this PR already landed on master; leaving #77 closed.

[Ceng23333]

Correction: fresh QY NCCL PR is #90 (head Ceng23333/InfiniCCL-infinittensor:feat/qy-nccl-20261112) for ITW v2026.11.12-qy-moe-bf16-warpnk.