Skip to content

feat(qy): MiniCPM5 FA/longctx/cudagraph + MoE bf16/warpnk surpass (2026.11.12) - #618

Open
Ceng23333 wants to merge 1 commit into
InfiniTensor:refactor/adopt-modern-infini-stackfrom
Ceng23333:feat/qy-moe-bf16-warpnk-20261112
Open

Ceng23333 wants to merge 1 commit into
InfiniTensor:refactor/adopt-modern-infini-stackfrom
Ceng23333:feat/qy-moe-bf16-warpnk-20261112

Conversation

@Ceng23333

@Ceng23333 Ceng23333 commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

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 (WALL_DECODE_RATIO≈1.009 on 16a3 TP2).

Targets refactor/adopt-modern-infini-stack (base f94d83d). Content matches ITW InfiniLM tip c9d03ec / tag v2026.11.12-qy-moe-bf16-warpnk (applied as one commit onto the current base; orphan pin history cannot be cherry-picked cleanly).

Dependencies (merge order)

  1. InfiniOps feat(qy): QY ops backend with FA and longctx MoE/RoPE kernels (2026.11.12) InfiniOps#1038
  2. InfiniRT feat(qy): enable QY CUDA runtime device backend (2026.11.12) InfiniRT#54
  3. InfiniCCL feat(qy): enable InfiniCCL NCCL backend for Denglin QY (2026.11.12) InfiniCCL#90
  4. InfiniCore chore(qy): pin InfiniRT/Ops/CCL for QY MoE bf16-warpnk (2026.11.12) InfiniCore#1583
  5. This PR

Validation

  • Full QY native build/serve: not run on PR host (no QY SDK).
  • Integration-validated via ITW v2026.11.12-qy-moe-bf16-warpnk (16a3 TP2 wall surpass).

Stack PR index (merge order)

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

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

Supersedes working tip of #597 (left open).

Bring ITW v2026.11.12-qy-moe-bf16-warpnk InfiniLM tip onto
refactor/adopt-modern-infini-stack: QY FA/longctx, FULL P+D graphs,
VocabParallelLMHead + InfiniCCL NCCL, decode FULL MoE tiles/align,
and use_moe_cu size-1 metadata for bf16/warpnk decode.

Co-authored-by: Cursor <cursoragent@cursor.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant