[RFC] VIME Roadmap

[gcanlin]

[Roadmap]

Consolidate pre-existing code

Port remaining content from slime (fork point 8ef1fb47 → snapshot 44d29ee) that hasn't landed on vime main.

• Delta weight sync — #278
  Bandwidth-optimized non-colocate weight transfer from slime #1806/#1946/#1991. Bytewise-diff encoding (indices / gap-deltas / zstd), sparse NCCL broadcast or disk-safetensors transport. Includes:

  • update_weight_from_distributed_delta.py — core delta encoder/decoder + NCCL publisher/subscriber
  • delta_io.py — standalone DeltaEncoding/DeltaParam/DeltaSpec dataclasses
  • actor.py / data.py / train_metric_utils.py — delta mode branch + extra_metrics plumbing to wandb
  • arguments.py — 7 CLI args (--update-weight-mode, --update-weight-transport, --update-weight-encoding, etc.)
  • types.py — rollout_id field on Sample
  • docs (en/zh), example script, unit test

• Run scripts — #260
  Complete slime-exact translation of all 29 training scripts following SGLANG_TO_VLLM_TRANSLATION.md. 23 new scripts + 6 existing updated. Covers dense, MoE, FP8, SFT, multi-node, and speculative decoding configurations.

• Other Missing examples
  Slime examples not yet on vime main (need new PR):

  • eval_multi_task/ — multi-task evaluation pipeline
  • on_policy_distillation/ — OPD example scripts
  • retool/ — ReTool tool-call RL example
  • search-r1/ — retrieval-augmented rollout example

Rollout Performance & Workload Enablement

Disaggregation (PD / EPD / OPD)

• EPD (encoder prefill disaggregation) P2 — VLM encoder server split from decoder prefill; composes with PD. (Net-new in vime; no slime counterpart.)
  Acceptance: new vLLM-only test gated on multimodal landing.

• Residual rollout overhead profiling (py-spy RolloutManager) — the remaining ~3-4% gap after router + logprobs fixes is in framework overhead, not engine/router

• LoRA support — #228

Runtime State Switching

• NCCL m2n weight sync backend — #159
• Shard-level P2P weight transfer — #161

RL Scaling

• TransferQueue for rollout-to-training — #265

Multi hardware

NPU/Ascend

• Qwen3-30B EP — #269
• NPU Dockerfile — #256
• NPU Ascend rebuild — #261
• NPU examples — #164

AMD ROCm

• Initial ROCm support (ROCm 7.0.2, MI350/MI355X) — #273
  Dockerfile.rocm + megatron nofork patch + Qwen3-8B colocate/async run scripts

---