[Ascend][RFC] Mainline Capability Adaptation Checklist (Phase 1 + Phase 2)

[CalvinXKY]

Summary

This RFC tracks the Ascend mainline (ascend branch) capability adaptation checklist for Vime.

• Phase 1 - basic capability adaptation: reproducible images, accuracy-gated training modes, and platform capabilities that must be validated before scaling to MoE, multi-node, multimodal, and agentic workloads.
• Phase 2 - expansion and hardening: CI guards, performance, mainline feature parity, broader model coverage, advanced features, and new Ascend hardware.

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

Closed items below are annotated with the primary merged PR(s) that landed the work (where applicable).

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:

• #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.

Relationship to upstream RFCs

• [Ascend][RFC] vime-ascend Build and Roadmap #243 - vime-ascend ecosystem roadmap (A2/A3/A5, TQ, R3, P2P, multimodal/MoE, agentic). This checklist covers a subset: infrastructure + acceptance gates + Phase-2 expansion.
• [RFC] NPU support roadmap. #51 - NPU support matrix (images, weight transfer, parallelism, VLM, R3, agentic). This checklist covers the subset required for mainline basic capabilities and Phase-2 follow-ups.

Global Acceptance Gate

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

• 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

Merge policy (ascend protected branch)

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

---

Phase 1 - Basic Capability Adaptation

Core Capabilities (Mainline Prerequisites)

• Reproducible image build; no runtime manual patching - #266 (follow-ups #271, #287, #305, #373) - @Meihan-chen, @miracle0517
• Qwen3-4B non-colocate (train / rollout split) - #212, #267, #275 - @Meihan-chen, @miracle0517
• Qwen3-4B colocate - #285 (patch follow-up #318) - @CalvinXKY, @Fulin-Gao

Ascend Platform Capabilities

• Save / resume - #366 - @yuxinshan
• R3 (Rollout Routing Replay) - #49, #292 - wukaiyuan, @yuxinshan
• Multi-node 2x8 - #323, #322, #68 - @miracle0517
• Multimodal model support - #314 - @Meihan-chen, @floatlibai
• A5 hardware validation - equivalent smoke / regression on A5 ([Ascend][RFC] vime-ascend Build and Roadmap #243) - @yuxinshan
• CI tests - #282, #316, #349 - zeyu, @Meihan-chen
• CI check adjustments - #349, #361

Model: Qwen3-30B-A3B (MoE)

• Qwen3-30B-A3B end-to-end - #254, #269, #323 - @floatlibai (A3), @yuxinshan (A2)

TQ (Training Quality) - GPU / NPU Alignment

• GPU baseline - same config + data subset; metrics archived - @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)

• ReTool - tool-call RL (Qwen3-4B) - #333
• tau-bench - tool-call agentic eval (Qwen3-4B) - #297
• Search-R1 - retrieval-augmented rollout (Qwen2.5-3B) - #311
• Search Agent - search-integrated agentic rollout on NPU - #311
• Multi-Agent RL (Qwen3-30B-A3B) - #298
• Coding-Agent SWE - subset grading correct; trainable token segments match GPU ([Ascend][RFC] vime-ascend Build and Roadmap #243 Code Agent; examples/coding_agent_rl)
• SWE-style on Modal - agentic SWE rollouts with Modal cloud sandbox ([RFC] NPU support roadmap. #51)
• RLinf - enterprise agentic scenario smoke on NPU ([Ascend][RFC] vime-ascend Build and Roadmap #243)
• Qwen3-VL Geo3K multi-turn - #314, #341
• Fully-async rollout - #368

Phase 1 Suggested Dependencies

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

Phase 1 One-Page Checklist

• Reproducible image build - #266
• Qwen3-4B non-colocate - #212 / #267 / #275
• Qwen3-4B colocate - #285
• Qwen3-30B-A3B - #254 / #269 / #323
• Save / resume - #366
• R3 - #49 / #292
• Multi-node 2x8 - #323 / #322 / #68
• Multimodal - #314
• A5 hardware validation
• CI tests + CI check adjustments - #282 / #316 / #349 / #361
• TQ: GPU baseline + NPU alignment
• ReTool - #333
• tau-bench - #297
• Search-R1 - #311
• Search Agent - #311
• Multi-Agent - #298
• Coding-Agent SWE
• SWE-style on Modal
• RLinf
• Qwen3-VL Geo3K multi-turn - #314 / #341
• Fully-async - #368

---

Phase 2 - Expansion and Hardening

Phase 1 basic mainline adaptation is largely complete. Phase 2 focuses on CI guards, performance, mainline feature parity, broader model coverage, advanced features, and new Ascend SKUs.

CI Guards

• Colocate mode: function + accuracy CI (colocate smoke + accuracy gate; build on #285)
• Expand CI model coverage (Qwen3.5 / GLM) @Fulin-Gao
• Image-build CI (patch protection): Dockerfile.npu + docker/npu_patch/ apply cleanly; build fails on missing / broken patches - #361

Performance Optimization

• Qwen3 Dense @miracle0517
• Qwen3 MoE @miracle0517
• Qwen3.5 @Fulin-Gao [NPU] Remove enforce-eager for Qwen3.5-35B-A3B #417
• K2.5/3 @yuxinshan
• GLM @is-not

Mainline Feature Sync

• Colocate performance optimizations - verify on NPU - #340 / #345
• Speculative decoding performance optimizations - verify on NPU - #351
• Commit / PR sync into Ascend mainline (CI green) @Meihan-chen [NPU] Sync Ascend with main and consolidate NPU adaptations #413

Model Support

• Qwen3.5 @Fulin-Gao
  • 27B / 35B function
  • Accuracy issues (GDN module); Loss NaN
• M 2.5
  • 4B (trimmed model) - #143
  • accuracy
• K2 / K3 @yuxinshan
  • e2e test
  • accuracy
• GLM @is-not @fanbingbing
  • 4.7 feat: support GLM-4.7-Flash (MLA + MoE) on Ascend NPU #403
    • e2e test
    • accuracy
  • 5 / 5.2
• DeepSeek-R1

Feature Support

• CP > 1 [NPU] [Spike] Support context parallel #433
• TQ NPU
• Speculative decoding
  • DFSPARK @wuxiang
  • DFLASH @momo609
  • eagle @momo609
  • MTP
• Delta weight transfer @momo609 -> wukaiyuan feat(ascend): add disk checkpoint serving patches #431
• Gateway
• LoRA - #336
• Omni
  • GPU
  • NPU
• Diffusion
• AgentENV

New Hardware Support

• A5 - Qwen3
• A5 - low precision

[floatlibai]

A few questions on the coding_agent_rl example @CalvinXKY :

• I don't see any GPU experimental results in the README — is there actual training/eval data (reward curve, resolve rate, etc.) for this example?
• Where does swe_train.jsonl come from? The README only shows its schema, not a source.
• Is Qwen3.6-35B-A3B required, or can a smaller model be used to validate this example end-to-end?

[CalvinXKY]

A few questions on the coding_agent_rl example @CalvinXKY :

• I don't see any GPU experimental results in the README — is there actual training/eval data (reward curve, resolve rate, etc.) for this example?
• Where does swe_train.jsonl come from? The README only shows its schema, not a source.
• Is Qwen3.6-35B-A3B required, or can a smaller model be used to validate this example end-to-end?

Currently, on the GPU, the coding scenario includes validation/testing use cases, but does not yet include training use cases. The reference test case is: https://github.com/vllm-project/vime/tree/main/examples/coding_agent_rl
Upstream, we are referring to Slime's agentic RL scenario, which ties together all the relevant components in RL. I think our coding scenario does not necessarily need to strictly follow the GPU version; the key is to first integrate it with the agentic RL scenario, and then see how to validate accuracy using a small-model scenario.

[floatlibai]

Currently, on the GPU, the coding scenario includes validation/testing use cases, but does not yet include training use cases. The reference test case is: https://github.com/vllm-project/vime/tree/main/examples/coding_agent_rl Upstream, we are referring to Slime's agentic RL scenario, which ties together all the relevant components in RL. I think our coding scenario does not necessarily need to strictly follow the GPU version; the key is to first integrate it with the agentic RL scenario, and then see how to validate accuracy using a small-model scenario.

Thank you, but I still wonder where to get the swe_train.jsonl dataset specifically — is it derived from an existing dataset (e.g. SWE-bench Pro), or generated via a script, or anything else?

[CalvinXKY]

Currently, on the GPU, the coding scenario includes validation/testing use cases, but does not yet include training use cases. The reference test case is: https://github.com/vllm-project/vime/tree/main/examples/coding_agent_rl Upstream, we are referring to Slime's agentic RL scenario, which ties together all the relevant components in RL. I think our coding scenario does not necessarily need to strictly follow the GPU version; the key is to first integrate it with the agentic RL scenario, and then see how to validate accuracy using a small-model scenario.

Thank you, but I still wonder where to get the swe_train.jsonl dataset specifically — is it derived from an existing dataset (e.g. SWE-bench Pro), or generated via a script, or anything else?

Using the SWE-bench Pro dataset is fine.

[CalvinXKY]

Framework Comparison

vime ascend code using #417 (commit:40e53d454f3f71297621f31183bb0c9a0a39bc08)

Case1: Qwen3.6 (2026.9.15)

Result

Summary

Environment, Versions & Parameters

Item	VIME	VERL Graph	VERL Megatron Graph+	AReaL
Completed steps	90 (0–89)	102 (1–102)	26 (1–26)	33 (1–33)
Device	Ascend NPU ×16 / 1 node	Ascend NPU ×16 / 1 node	Ascend NPU ×16 / 1 node	Ascend NPU ×16 / 1 node
Train–rollout mode	colocate	colocate	colocate	colocation (rollout→actor)
Scheduler	Ray	Ray	Ray	Ray
Python	3.12.13	3.11 (PyTorch-2.10.0)	3.11 (PyTorch-2.10.0)	3.11.15
torch_npu	transfer_to_npu (runtime)	PyTorch-2.10.0 env	same	2.10.0.post2
RL framework path	/root/vime	verl (main_ppo)	same	AReaL (930-baseline/AReaL)
Train backend	Megatron	Megatron (ppo_megatron_trainer)	Megatron	Megatron (d1p4t4)
Rollout engine	vLLM Ascend	vLLM	vLLM	vLLM (vllm:d4p1t4)
vLLM version (log)	0.22.1+empty	not printed in log head	not printed in log head	not printed in log head
Weight sync	vLLM native + sleep mode	checkpoint engine (bucket 6144 MB)	same	AWEX (ipc_backend=npu)
Model	Qwen3.6-27B	Qwen3.6-27B	Qwen3.6-27B	Qwen3.6-27B
Dataset	dapo-math-17k	dapo-math-17k	dapo-math-17k	dapo-math-17k
Algorithm	GRPO	GRPO	GRPO	GRPO
dtype	bf16	bfloat16	bfloat16	bfloat16
LR	1e-6	1e-6	1e-6	Adam (script/YAML override)
Max response / new tokens	14336	14336	14336	14336
Max model / context len	16384	prompt 2048 + max_model_len 16384	same	train max_length 2048
n samples / prompt	8	8	8	8
Prompt batch size	rollout_batch_size 16	train_batch_size 16	16	batch_size 16
Trajectories / step (est.)	16×8=128	16×8=128	128	16×8=128
Actor TP / PP / CP	4 / 4 / 1	4 / 4 / 1	4 / 4 / 1	4 / 4 (d1p4t4)
Recompute / grad ckpt	full / uniform / 1 layer	full / uniform / 1 layer	same	gradient_checkpointing=true
Optimizer offload	Adam + CPU offload	param/optim/grad offload=True	same	MindSpeed swap_optimizer=true
Train micro / mini batch	global_batch_size 128	ppo_mini_batch_size 1, micro=1	same	ppo_n_minibatches 1, max_tokens_per_mb=16384
vLLM layout (script)	TP4 × DP4, 16 NPUs/engine	GEN_TP=4	GEN_TP=4	vllm:d4p1t4
vLLM mem util	0.8 (log dump)	0.7	0.7	0.7
CUDA Graph mode	ACL graph disabled (ENABLE_ACL_GRAPH=0)	FULL_AND_PIECEWISE	FULL_DECODE_ONLY	FULL_DECODE_ONLY (YAML)
KL / entropy	kl=0.0, entropy=0.0	use_kl_loss=False, entropy=0	same	kl_ctl=0.0
Clip	eps 0.2 / high 0.28	(trainer defaults + script)	same	eps_clip 0.2
Planned scale	num_epoch=2	total_epochs=2	total_epochs=2	total_train_steps=200

[CalvinXKY]

Case2: Qwen3 (2026.9.18)

Environment, Versions & Parameters

Item	VIME 4B	Slime 4B	VIME 30B-A3B	Slime 30B-A3B	VIME 32B	Slime 32B
Completed steps (step_time)	10 (0–9)	9 (0–8)	10 (0–9)	9 (0–8)	10 (0–9)	9 (0–8)
Model	Qwen3-4B	Qwen3-4B	Qwen3-30B-A3B (MoE)	Qwen3-30B-A3B (MoE)	Qwen3-32B	Qwen3-32B
Device layout	actor 4 NPU	actor 4 NPU	actor 8 NPU	actor 8 NPU	actor 8 NPU	actor 8 NPU
Rollout NPUs	4	4	8	8	8	8
NPUs / engine	4	4	4	4	4	4
Train–rollout mode	colocate=True	colocate=False	colocate=False	colocate=False	colocate=False	colocate=False
RL framework	VIME	Slime	VIME	Slime	VIME	Slime
Train backend	Megatron	Megatron	Megatron	Megatron	Megatron	Megatron
Rollout engine	vLLM	SGLang	vLLM	SGLang	vLLM	SGLang
Rollout parallel	vLLM TP=4 DP=1	SGLang TP=4 DP=1	vLLM TP=4 DP=1	SGLang TP=4 DP=1	vLLM TP=4 DP=1	SGLang TP=4 DP=1
Mem utilization	vLLM 0.6	SGLang mem_frac 0.7	vLLM 0.7	SGLang 0.7	vLLM 0.7	SGLang 0.7
SGLang attention	—	ascend	—	ascend	—	ascend
Algorithm	GRPO	GRPO	GRPO	GRPO	GRPO	GRPO
Reward type	math	math	deepscaler	math	deepscaler	math
LR	1e-6	1e-6	1e-6	1e-6	1e-6	1e-6
Optimizer	Adam + CPU offload	Adam + CPU offload	Adam + CPU offload	Adam + CPU offload	Adam + CPU offload	Adam + CPU offload
Max response len	2048	2048	4096	4096	4096	4096
Temperature	1.0	1.0	1.0	0.8	1.0	0.8
n samples / prompt	8	8	8	8	8	8
Rollout batch size	32	32	32	32	32	32
Global batch size	256	256	256	256	256	256
Trajectories / step	32×8=256	256	256	256	256	256
Actor TP / PP / CP	4 / 1 / 1	4 / 1 / 1	4 / 1 / 1	4 / 1 / 1	8 / 1 / 1	8 / 1 / 1
Expert parallel (EP)	1	1	8	8	1	1
Num experts	—	—	128	128	—	—
Sequence parallel	False	True	True	True	True	True
Recompute	full	full	full	full	full	full
KL / entropy	0 / 0	0 / 0	0 / 0	0 / 0	0 / 0	0 / 0
Clip	0.2 / 0.28	0.2 / 0.28	0.2 / 0.28	0.2 / 0.28	0.2 / 0.28	0.2 / 0.28
Planned rollouts	10	10	10	10	10	10

Note:

1. Across all three model sizes, VIME (vLLM) is faster end-to-end than Slime (SGLang) on mean step time and rollout throughput.
2. 4B is not fully apples-to-apples: VIME uses colocate=True, Slime uses colocate=False. Use colocate mode would raise error on slime-ascend
3. 30B/32B temperature differs: VIME=1.0 vs Slime=0.8; reward type also differs (deepscaler vs math).