[Ceng23333]

Summary

Enable MiniCPM5 on QY with FlashAttention, longctx MoE/RoPE, CUDA-graph capture safety, FULL P+D graph campaign, VocabParallelLMHead + InfiniCCL NCCL TP path, and MoE bf16/warpnk decode surpass.

Targets refactor/adopt-modern-infini-stack. Content matches ITW InfiniLM tip / tag v2026.11.12-qy-moe-bf16-warpnk.

Motivation / related

• Depends on Ops #1038, RT Paged Attention合入pr，待验证 #54, CCL QY机器测试7B模型 #90, Core #1583.
• Supersedes working tip of feat(qy): MiniCPM5 FA, longctx MoE/RoPE, CUDA-graph capture safety #597 (left open).

Type of change

• feat
• style (scripts/format.py apply vs refactor base)

Platforms affected

• QY

Validation

• scripts/format.py --ref origin/refactor/adopt-modern-infini-stack --check clean (CI Check Format green).
• Component stack smokes on Denglin QY (DENGLIN_DEVICES=3): Ops add/gemm 1633 passed; RT ctest 13/13; CCL WITH_QY+NCCL link OK; Core gitlinks retargeted.
• Greedy / EvalScope: not re-run on this host in this pass (full InfiniLM xmake rebuild against new pins not completed here; prior ITW pin v2026.11.12-qy-moe-bf16-warpnk remains the integration reference). See smoke comments.

Checklist

• Check Format green
• Conventional Commits title/branch
• Full greedy + EvalScope P1/N10 on PR-head pins (deferred; stated above)

Stack PR index (merge order)

1. InfiniOps: feat(qy): QY ops backend with FA and longctx MoE/RoPE kernels (2026.11.12) InfiniOps#1038 — 71e85162ab96
2. InfiniRT: feat(qy): enable QY CUDA runtime device backend (2026.11.12) InfiniRT#54 — 047340f45efa
3. InfiniCCL: feat(qy): enable InfiniCCL NCCL backend for Denglin QY (2026.11.12) InfiniCCL#90 — 691c3092c4c9
4. InfiniCore: chore(qy): pin InfiniRT/Ops/CCL for QY MoE bf16-warpnk (2026.11.12) InfiniCore#1583 — 0b69ee9002bd
5. InfiniLM: feat(qy): MiniCPM5 FA/longctx/cudagraph + MoE bf16/warpnk surpass (2026.11.12) #618 — 36beae5376f4 → base refactor/adopt-modern-infini-stack

ITW pin: https://github.com/Ceng23333/InfiniTensorWorktree/tree/v2026.11.12-qy-moe-bf16-warpnk

[Ceng23333]

QY required smoke (PR head)

• Head SHA: 36beae5376f4f61c79d6d01277f6bed211e69dd6
• Host: Denglin QY, DENGLIN_DEVICES=3
• Command: component stack smokes (Ops/RT/CCL) + format check; greedy/EvalScope deferred
• Result: PARTIAL PASS
  • Check Format: PASS
  • Ops add/gemm / RT ctest / CCL QY+NCCL link: PASS (sibling heads)
  • Greedy + EvalScope mixed-4096: not re-run (full InfiniLM xmake rebuild vs new pins not completed on this host; ITW v2026.11.12-qy-moe-bf16-warpnk remains integration reference)
• Log: nov_milestone/logs/prhead-verify-20261113/ (ops/rt/ccl/core logs)

Sibling PR heads: RT 047340f45efa, Ops 71e85162ab96, CCL 691c3092c4c9; Core 0b69ee9002bd.

[Ceng23333]

QY required smoke (PR head) — detailed

• Head SHA: 36beae5376f4f61c79d6d01277f6bed211e69dd6
• Host: Denglin QY, DENGLIN_DEVICES=3
• Sibling heads: RT 047340f45efa, Ops 71e85162ab96, CCL 691c3092c4c9, Core 0b69ee9002bd
• Overall: PARTIAL PASS (format + component smokes; greedy/EvalScope deferred)
• Logs: nov_milestone/logs/prhead-verify-20261113/ops-pytest-add-gemm.log, rt-ctest.log, ccl-build.log, core-retarget.txt

Build / stack

Step	Detail	Result
scripts/format.py --check vs refactor base	CI Check Format	PASS
InfiniRT ctest (QY)	13/13	PASS
InfiniOps m0 add/gemm	1633 passed / 6 skipped	PASS
InfiniCCL WITH_QY+NCCL	libinfiniccl.so	PASS
Core gitlinks	retargeted to heads above	PASS
InfiniLM xmake + greedy hello	full rebuild vs new pins	SKIPPED (time/scope; ITW tag remains reference)
EvalScope mixed-4096 P1/N10	—	SKIPPED (depends on greedy path)

Why greedy/EvalScope skipped

Minimum bar called for stack --test + greedy; component --test equivalents are green. Completing InfiniLM xmake against the post-format pins on this host was deferred after component verification.