[Ascend][RFC] Mainline Basic Capability Adaptation Checklist

[CalvinXKY]

Summary

This RFC defines the basic capability adaptation checklist for the Ascend mainline (ascend branch) in Vime. It tracks near-term, verifiable milestones—reproducible images, accuracy-gated training modes, and platform capabilities—that must be validated before scaling to MoE, multi-node, multimodal, and agentic workloads.

This checklist does not replace the broader Ascend roadmaps. It focuses on what must be hardened on top of the current protected mainline.

Related RFCs

• [Ascend][RFC] vime-ascend Build and Roadmap #243
• [RFC] NPU support roadmap #51

Background

The ascend branch is now the protected Ascend mainline. Two foundational merges have landed:

PR	Description	Status
#266	Dockerfile.npu + docker/npu_patch/ — reproducible NPU image build and colocate patches	Merged
#267	Rebuild ascend on NPU implementation — training/rollout backend, weight transfer, NPU attention, scripts, unit tests	Merged

Implication: further changes land on a protected branch under review policy (see below). This issue tracks what must be validated and hardened on top of #266 / #267 before basic capabilities can be considered complete.

Relationship to upstream RFCs

Upstream RFC	Scope	This checklist
#243	vime-ascend ecosystem roadmap (A2/A3/A5, TQ, R3, P2P, multimodal/MoE, agentic)	Subset: infrastructure + basic acceptance gates
#51	NPU support matrix (images, weight transfer, parallelism, VLM, R3, agentic)	Subset: items required for mainline basic capabilities

Global Acceptance Gate

Every item below must run successfully and meet accuracy / alignment requirements:

Gate	Criteria
Logprob alignment	logprob diff / KL within agreed thresholds; no spikes
Reproducibility	fixed seed + fixed data subset → reward / loss reproducible
Checkpointing	save → load → resume → metrics continuous
Weight sync	after IPC (colocate) or HCCL (non-colocate) sync, rollout logprob matches training forward

TODO List

Core Capabilities (Mainline Prerequisites)

Status	Item	Acceptance Criteria	Notes	Owner
[ ]	Reproducible image build; no runtime manual patching	Fresh container imports full stack; colocate smoke (1 step) passes; diff not worse than baseline	Built from #266; A2 / A3	@Meihan-chen, @miracle0517
[ ]	Qwen3-4B non-colocate (train / rollout split)	1 full rollout + 1 training step; HCCL weight sync aligned; logprob diff acceptable	#51	@Meihan-chen, @miracle0517
[ ]	Qwen3-4B colocate	≥10 steps; weight sync stable (~seconds); logprob diff acceptable	#243 A2 colocate	@CalvinXKY, @Fulin-Gao

Ascend Platform Capabilities

Status	Item	Acceptance Criteria	Notes	Owner
[ ]	Save / resume	save → load → continue training; next-step loss continuous		@yuxinshan
[ ]	R3 (Rollout Routing Replay)	MoE or designated model smoke; no numerical anomalies; reward curve + reference case	#51; #243	wukaiyuan, @yuxinshan
[ ]	Multi-node 2×8	≥1 step; loss consistent across ranks; TP / DP / EP / PP supported	#243	@miracle0517
[ ]	Multimodal model support	VLM train/infer path working; reward or accuracy within tolerance vs GPU	#51	@Meihan-chen, @floatlibai
[ ]	A5 hardware validation	Equivalent smoke / regression on A5	#243	@yuxinshan
[ ]	CI tests	NPU smoke / regression in CI; failures block merge	TBD	yangzeyu
[ ]	CI check adjustments	Aligned with NPU stack, patches, image tags

Merge policy (ascend protected branch)

• Large PRs: merge via review meeting
• Small PRs: ≥2 reviewers required

Model: Qwen3-30B-A3B (MoE)

Status	Item	Acceptance Criteria	Notes	Owner
[ ]	Qwen3-30B-A3B end-to-end	1 rollout + 1 training step; MoE healthy; no NaN	A3 first; A2 follow-up	@floatlibai (A3), @yuxinshan (A2)

TQ (Training Quality) — GPU / NPU Alignment

Status	Item	Acceptance Criteria	Notes	Owner
[ ]	GPU baseline	Same config + data subset; metrics archived	Reference for NPU	@miracle0517
[ ]	NPU alignment	Same seed vs GPU baseline; diff / reward within tolerance		@miracle0517

Scale-up (after multi-node + multimodal are ready):

• Cluster size: >2 nodes
• Dataset: production-scale volume
• Deliverable: full training guide / reference recipe

Agentic Workloads (@Fulin-Gao, @momo609)

Status	Item	Acceptance Criteria	Notes	Owner
[ ]	ReTool — tool-call RL (Qwen3-4B)	Executable tool sandbox rollout; reward curve acceptable	#51; examples/retool
[ ]	τ-bench — tool-call agentic eval (Qwen3-4B)	Agentic eval smoke passes; scoring matches GPU reference	#51, #243; examples/tau-bench, #142
[ ]	Search-R1 — retrieval-augmented rollout (Qwen2.5-3B)	Search server integrated; rollout completes; reward reasonable	#51; examples/search-r1
[ ]	Search Agent	Search-integrated agentic rollout on NPU	#243
[ ]	Multi-Agent RL (Qwen3-30B-A3B)	Reward weighting matches GPU reference	#51; examples/multi_agent
[ ]	Coding-Agent SWE	Subset grading correct; trainable token segments match GPU	#243 Code Agent; examples/coding_agent_rl
[ ]	SWE-style on Modal	Agentic SWE rollouts with Modal cloud sandbox	#51
[ ]	RLinf	Enterprise agentic scenario smoke on NPU	#243
[ ]	Qwen3-VL Geo3K multi-turn	Reward curve acceptable; tool parsing correct; multi-turn loss_mask aligned	#51; examples/geo3k_vlm_multi_turn
[ ]	Fully-async rollout	Same seed as sync; reward / loss within tolerance

Suggested Dependencies

#266 / #267 merged (protected ascend mainline)
  → Reproducible image validation (no runtime patching)
  → Qwen3-4B non-colocate / colocate
  → Accuracy gates (save-resume, weight sync)
  → Qwen3-30B-A3B MoE → Multi-node / R3 → Multimodal → Agentic → TQ at scale

One-Page Checklist (Appendix)

• Reproducible image build; no runtime manual patching
• Qwen3-4B non-colocate: 1 cycle, HCCL weight sync, logprob diff acceptable
• Qwen3-4B colocate: ≥10 steps, logprob diff acceptable
• Qwen3-30B-A3B: 1 rollout + 1 training step, MoE healthy
• Save / resume: loss continuous
• R3: MoE or designated model smoke, no numerical anomalies
• Multi-node 2×8: 1 step, consistent loss across ranks, TP/DP/EP/PP supported
• Multimodal model support
• A5 hardware validation
• CI tests + CI check adjustments
• TQ: GPU baseline + NPU alignment
• ReTool (Qwen3-4B): tool-call RL smoke
• τ-bench (Qwen3-4B): agentic eval smoke
• Search-R1 (Qwen2.5-3B): retrieval-augmented rollout
• Search Agent
• Multi-Agent (Qwen3-30B-A3B): reward weighting matches GPU
• Coding-Agent SWE: subset grading correct
• SWE-style on Modal
• RLinf
• Qwen3-VL Geo3K multi-turn: reward curve acceptable
• Fully-async: same seed as sync, within tolerance