[Ascend][RFC] vime-ascend Build and Roadmap

[Fulin-Gao]

vime-ascend Ecosystem Roadmap

The Huawei Modelarts Team is delighted to announce that we will actively join the development of vime-ascend, with all contributions built upon the existing ascend branch.

Building on our existing test cases, we have extended testing to cover the A2, A3 and A5 hardware platforms. For A2 and A3, mainstream models including the Qwen series have been verified under both separated-card and shared-card modes. We have also ramped up development and testing for multi-node and distributed capabilities.

Our core roadmap and strategic initiatives are outlined below:

1. Hardware Stability & Infrastructure

Focuses on hardware-specific optimization, functional validation, and maximizing compute utilization across generations.

• A2 & A3 Hardening: Standardize and harden container images, basic execution paths, and functional stability.

  • A2 — Qwen3-4B: non-colocate, reward converges [NPU][Spike] Steps and Test Results for Running Qwen3-4B on NPU（A2） #157 feat(npu): Enhance Dockerfile.npu with verified vllm commit and run script #287
  • A3 — Qwen3-32B: non-colocate, reward converges [NPU][Spike] Steps and Test Results for Running Qwen3-32B on NPU（A3） #165 [NPU][A3] Add patch workflow and Docker support for Atlas 800I A3 #163
  • A2 — Qwen3-4B: colocate, reward converges [NPU][Spike] Colocate Mode Test (Qwen3-4B GRPO) on Ascend 910B1 (A2) #255 fix(npu): adapt colocate patch for vllm-ascend and vllm version upgrade #318
  • A3 — Qwen3-32B: colocate, reward converges refactor(vllm): use native IPC engine for colocated weight sync #345

• Next-Gen Hardware (A5) Enablement: Kickstart core capability development and early-access optimization on Huawei’s next-generation hardware to unleash its exceptional raw performance.

  • A5 — Qwen3-8B

2. High-Performance Distributed & Serving Scaling

Focuses on minimizing communication overheads, optimizing distributed request routing, and maximizing throughput for large-scale serving clusters.

• TransferQueue: Introduce an optional direct data path to streamline rollout-to-training transfers and bypass Ray ObjectRef overhead.

  • TransferQueue: on Qwen3-4B [RFC] Integrating TransferQueue into Vime: Decoupling Rollout and Training with a Streaming Data Bus #227 [Feature] TransferQueue Integration for Rollout-to-Training #242 [Spike] [TQ] Transfer Queue test on H800 (Qwen3-4B GRPO) #244

• Routing Replay (R3): Introduce robust routing and replay mechanisms to enhance traffic dispatching efficiency and system resilience.

  • Routing Replay (R3) [RFC] Routing Replay (R3) for vLLM Rollout Engine #32

• Weight Transfer: Introduce a peer-to-peer (P2P) weight transfer mechanism to optimize large-scale model synchronization and eliminate communication bottlenecks.

  • Peer-to-Peer (P2P): which achieved ideal results on AReaL (Awex) for a 1T-parameter model on A3 supernodes—directly into the vime-ascend framework. [Feature] shard-level P2P weight transfer (Phase 1, #160) #161

3. Multimodal/MoE & Advanced Scenarios

Focuses on extending vime ascend to complex workloads like multimodal, MoE, and agentic workflows.

• Multimodal/MoE RL: Migrate existing multimodal/MoE reinforcement learning capabilities seamlessly to the vime ascend ecosystem.
  • Qwen3-30B-A3B [Spike] Phased Qwen3-30B-A3B #14 [NPU] replace 30b-npu script with ray-based launch method #323
  • Qwen3-VL-30B-A3B
  • Qwen3.5-27B
  • Qwen3.5-35B-A3B [NPU] Add Qwen3.5-35B-A3B support #389 [NPU] Add Qwen3.5-35B-A3B support #409
  • MiniMax M2.5 [NPU] Add MiniMax M2.5 support #391
  • GLM-4.7-Flash feat: support GLM-4.7-Flash (MLA + MoE) on Ascend NPU #403
• Knowledge Distillation: Implement online teacher-student training paradigms to boost learning efficiency and model alignment.
  • On-Policy Distillation (OPD): Student trains against a teacher inference server; currently the test uses an sglang teacher
• Customer-Driven Ecosystems: Drive R&D and exploratory implementation for advanced enterprise scenarios:
  • RLinf
  • Code Agent: Slime ref examples/retool and examples/coding_agent_rl.
  • Search Agent: Slime ref examples/search-r1. [NPU] Add Search-R1 example #311
  • Retool: Slime ref examples/retool. [NPU] Add Retool example #333
  • tau_bench: Slime ref examples/tau-bench and slime-agentic ref agentic/ToolOrchestra. [NPU] add tau bench example for npu #297
  • MemAgent: slime-agentic ref agentic/memagent. [Example] Add MemAgent long-context RL example (mem_agent) #291
  • AgentFlow: slime-agentic ref agentic/agentflow.