v0.4.0: A First Step Toward Greater Scale2026-10-09
OpenAI's Navier–Stokes research effort involved a group of roughly 10,000 concurrent agents and used approximately 130 billion output tokens during problem-solving. For us, it offered a concrete glimpse of how large agent workloads can become—and prompted us to think more ambitiously about the RL infrastructure that trains them.
Preparing for that scale means giving rollout orchestration room to grow across nodes, making data flow durable, and preserving useful work when training fails. It also means focusing our development effort so that algorithm exploration and training at scale share a foundation we can keep improving.
slime v0.4.0 is a first step in that direction. We are focusing on synchronous and fully asynchronous training, distributing rollout orchestration across the cluster, introducing persistent queues and tensor storage with straw, and enabling Megatron to recover independently of healthy SGLang serving resources. These changes lay the groundwork for higher concurrency and longer runs, with training correctness remaining central to slime's design.
One training lifecycle, two modes
We are focusing development on two training modes:
- Synchronous training for exploring algorithmic limits and validating correctness.
- Fully asynchronous training for keeping rollout and training resources busy at larger scale.
Both modes now use train.py, with asynchronous scheduling handled by the rollout implementation. We have removed the legacy one-step off-policy driver, train_async.py, giving both workflows a shared training lifecycle to maintain and improve. ([#2391](https://github.com/THUDM/slime/pull/2391))
Alongside this simplification, we are expanding the tools available for stable training. v0.4.0 adds score centering, using a REINFORCE objective with support for top-p replay and the built-in importance-sampling corrections. Enabled with --use-score-centering, it provides another tool for addressing training–inference mismatch as we scale rollout and training. ([#2405](https://github.com/THUDM/slime/pull/2405), [#2406](https://github.com/THUDM/slime/pull/2406))
A distributed, persistent rollout pipeline
Increasing rollout capacity also requires orchestration to grow with the cluster. With a new storing backend straw enabled, fully asynchronous rollout now runs one rollout worker per Ray node with CPU resources, each managing its own generation pool and asyncio event loop. This distributes orchestration across the CPU resources already available in the cluster and reduces the bottleneck of a single process. Training batches can collect completed groups without waiting for every worker. ([#2410](https://github.com/THUDM/slime/pull/2410))
This distributed execution is backed by [straw](https://github.com/zhuzilin/straw), a library providing durable queues and shared tensor storage on a filesystem. Rollout workers publish payloads to shared storage, and training ranks read their assigned data by reference. Bulk data between rollout and training moves off Ray’s object-store transport, while Ray continues to carry control messages and references.
The same data layer persists prompt tasks, partial rollouts, completed samples, and converted training batches. Coordinating this progress with model checkpoints establishes a consistent recovery boundary for asynchronous training and supports checkpoint rollback. Samples and tensors are packed into large append-only files to reduce small-file and shared-filesystem metadata overhead. ([#2410](https://github.com/THUDM/slime/pull/2410), [#2427](https://github.com/THUDM/slime/pull/2427))
Select straw with --rollout-data-transport straw. JuiceFS is the currently supported shared-filesystem target; other network filesystems require deployment-specific validation. See the [straw guide](https://github.com/THUDM/slime/blob/v0.4.0/docs/en/advanced/straw.md) for configuration and recovery details.
Recover training while preserving serving and rollout work
As RL jobs grow, failures and configuration mistakes become more expensive. v0.4.0 separates the lifetime of Megatron training from the SGLang serving cluster, allowing healthy engines and routers to remain running after a training failure.
You can adjust supported trainer parallelism or memory settings and resubmit training to the same live Ray cluster. With Straw or saved rollout debug data, the new training attempt restores the last committed checkpoint and replays retained batches, reusing accepted samples and conversion results.
This reduces the cost of training retries by preserving serving resources and completed rollout work. See the [recovery guide](https://github.com/THUDM/slime/blob/v0.4.0/docs/en/advanced/fault-tolerance.md) for the supported restart workflow. ([#2444](https://github.com/THUDM/slime/pull/2444))
Other Improvements
- PipelineRL: --flush-cache-interval allows unfinished requests and their KV caches to survive selected weight updates, reducing repeated rollout work. Periodic full refreshes remain configurable. This mode requires separate training and rollout GPUs. ([#2437](https://github.com/THUDM/slime/pull/2437))
- Updated training stack: The Docker and Conda environments now use Megatron Core 0.19.2, Transformer Engine 2.18.0, Flash Linear Attention 0.5.2, and NCCL 2.30.7. The Megatron patch is also smaller, with more integration logic maintained in slime. ([#2448](https://github.com/THUDM/slime/pull/2448))
- Streaming and agent rollouts: Added support for cumulative and incremental generation streams, partial-prefix retention, and cancellation of individual requests. Agent timeout cleanup now targets the router’s SGLang workers. ([#2272](https://github.com/THUDM/slime/pull/2272), [#2340](https://github.com/THUDM/slime/pull/2340))
- Accelerator backends: Added Biren SUPA and Ascend NPU integrations to slime’s accelerator abstraction, including runtime initialization, device mapping, and communication backend selection. ([#2393](https://github.com/THUDM/slime/pull/2393), [#2424](https://github.com/THUDM/slime/pull/2424))
- Correctness fixes: Corrected linear-attention gradients under context parallelism, preserved reference log probabilities over the full vocabulary during actor top-p replay, and fixed checkpoint shard selection in disk-delta weight synchronization. ([#2443](https://github.com/THUDM/slime/pull/2443), [#2434](https://github.com/THUDM/slime/pull/2434), [#2442](https://github.com/THUDM/slime/pull/2442))
What's Changed
- feat: support streaming external rollouts by @Aphoh in [#2272](https://github.com/THUDM/slime/pull/2272)
- fix(agent): abort timed-out SGLang requests via router workers by @Arlo-mt in [#2340](https://github.com/THUDM/slime/pull/2340)
- sync from internal by @zhuzilin in [#2390](https
v0.3.22026-08-28
We are excited to announce the release of slime v0.3.2!
This release focuses on GLM-5 train/rollout alignment, community-driven correctness fixes, and continued codebase cleanup.
Some noteworthy improvements include:
Fully aligned GLM-5 training and rollout**
* Added deterministic alignment between Megatron training and SGLang rollout for GLM-5, covering DeepEP, DeepGEMM, DSA sparse attention, and FP8 KV cache.
* The aligned path has been validated on large-scale GLM-5.3 training workloads.
Many community-contributed fixes and optimizations**
* Fixed issues across fully-async rollout, PPO/OPD, tool-use tokenization, dataset processing, evaluation, distributed training, and logging.
Cleaner and easier-to-maintain codebase**
* Continued refactoring and removing legacy/dead code across rollout, actor initialization, CI, and training utilities.
* Added a dedicated slime/observability module for logging, debugging, and tracing utilities.
CUDA 13 support**
* Added an official CUDA 13 / cu130 environment for Blackwell systems.
What's Changed
* [docker] support cuda 13 by @zhuzilin in https://github.com/THUDM/slime/pull/2257
* feat(glm5): align Megatron DeepEP training with SGLang rollout by @zhuzilin in https://github.com/THUDM/slime/pull/2262
* fix: preserve consecutive tool responses in Qwen3 SFT tokenization by @HelanHu in https://github.com/THUDM/slime/pull/2264
* fix(rollout): restore partial continuation token budget by @Daucloud in https://github.com/THUDM/slime/pull/2261
fix(tools): clamp block max in blockfp8 to avoid NaN weights from all-zero blocks by @hobostay in https://github.com/THUDM/slime/pull/2256
* docs: correct reverse KL definition in OPD guide by @XiXiphus in https://github.com/THUDM/slime/pull/2254
* fix: forward dual-clip PPO epsilon by @alexqdh in https://github.com/THUDM/slime/pull/2247
fix: cast gpuid to int in sort_key to prevent lexicographic ordering by @qq1243196045 in https://github.com/THUDM/slime/pull/2246
* fix: restore negative dataset slice bounds (path@[-100:]) by @keepkeen in https://github.com/THUDM/slime/pull/2241
fix: keep dataset order in filterlong_prompt for mixed multimodal data by @keepkeen in https://github.com/THUDM/slime/pull/2237
* perf: vectorize REINFORCE++ discounted returns by @morluto in https://github.com/THUDM/slime/pull/2205
* docs(readme): add Dressage to Chinese ecosystem by @zhangdw156 in https://github.com/THUDM/slime/pull/2138
docs(examples): list codingagent_rl in examples/README by @aoshen02 in https://github.com/THUDM/slime/pull/2133
* [Doc] Clarify PPO/Critic docs after #1856 by @coding-famer in https://github.com/THUDM/slime/pull/2189
* fix(mtp): support multi-head MTP loss logging (mtp-num-layers > 1) by @ZiyiTsang in https://github.com/THUDM/slime/pull/2132
Fix placement group crash for external engines under debugrollout_only by @coding-famer in https://github.com/THUDM/slime/pull/2170
fix: restore args.ckptstep after loadothercheckpoint by @keepkeen in https://github.com/THUDM/slime/pull/2243
fix: clear execand_wait's spawn lock between logical invocations by @keepkeen in https://github.com/THUDM/slime/pull/2239
* fix: pair --log-correct-samples rewards with the DP-local samples by @keepkeen in https://github.com/THUDM/slime/pull/2234
* fix: whiten advantages over the DP group that includes context parallel by @keepkeen in https://github.com/THUDM/slime/pull/2235
* fix: don't overwrite an explicitly set --start-rollout-id by @keepkeen in https://github.com/THUDM/slime/pull/2236
* Fix tau-bench token deltas for reasoning templates by @DongzhuoranZhou in https://github.com/THUDM/slime/pull/2213
* fix: stop the fully-async rollout dropping completed groups by @keepkeen in https://github.com/THUDM/slime/pull/2238
fix: honor every eval.defaults key and restore per-dataset stop / minnew_tokens by @keepkeen in https://github.com/THUDM/slime/pull/2242
* fix(npu): bracket IPv6 hosts in distributed init methods by @EazyReal in https://github.com/THUDM/slime/pull/2199
* Refactor --save-debug-train-data by @zhuzilin in https://github.com/THUDM/slime/pull/2266
fix transformue8m0 in fp8 convert by @lilei199908 in https://github.com/THUDM/slime/pull/2271
* [ROCm] Support the INT4 QAT kernel on ROCm by @LZ-QWQ in https://github.com/THUDM/slime/pull/2274
* Fix model convert when use latest megatron by @alexqdh in https://github.com/THUDM/slime/pull/2267
* Add args check for --save-debug-train-data by @zhuzilin in https://github.com/THUDM/slime/pull/2276
* fix: improve compatibility with older SGLang versions by @ForAxel in https://github.com/THUDM/slime/pull/2286
* feat: add backend-aware MUSA support by @ForAxel in https://github.com/THUDM/slime/pull/2216
* cleanup by @zhuzilin in https://github.com/THUDM/slime/pull/2294
* fix(train): skip optimizer and scheduler for eval-only by @EazyReal in https://github.com/THUDM/slime/pull/2296
* fix(ppo): preserve raw KL so rollout/kl logging is correct by @EazyReal in https://github.com/THUDM/slime/pull/2114
* fix(opd): score teacher logprobs at rollout temperature, not 0 by @EazyReal in https://github.com/THUDM/slime/pull/2085
* fix: reject non-positive rollout temperature at parse time by @EazyReal in https://github.com/THUDM/slime/pull/2297
* [NFC] Add observability subfolder by @zhuzilin in https://github.com/THUDM/slime/pull/2298
* docs: remove stale delta NCCL recommendation by @liuhao-labs in https://github.com/THUDM/slime/pull/2312
Remove megatronpatch for memory optimization by @zhuzilin in https://github.com/THUDM/slime/pull/2316
* [ci] Clean up tests by @zhuzilin in https://github.com/THUDM/slime/pull/2317
[ci] move fanouttest_helpers to tests/ by @zhuzilin in https://github.com/THUDM/slime/pull/2318
* [cleanup] remove dead code and merge never visited branches by @zhuzilin in https://github.com/THUDM/slime/pull/2320
[cleanup] extract createweight_updater to make actor's init func cleaner by @zhuzilin in https://github.com/THUDM/slime/pull/2321
[cleanup] Remove rolloutvalidation.py by @zhuzilin in https://github.com/THUDM/slime/pull/2322
* [cleanup] Refactor rollout.py by @zhuzilin in https://github.com/THUDM/slime/pull/2323
* [doc] update doc by @zhuzilin in https://github.com/THUDM/slime/pull/2326
* feat: allow forcing UE8M0 FP8 scales by @zhuzilin in https://github.com/THUDM/slime/pull/2327
* [docs] fix out-dated doc by @zhuzilin in https://github.com/THUDM/slime/pull/2330
* [release] bump to v0.3.2 by @zhuzilin in https://github.com/THUDM/slime/pull/2334
New Contributors
* @HelanHu made their first contribution in https://github.com/THUDM/slime/pull/2264
* @hobostay made their first contribution in https://github.com/THUDM/slime/pull/2256
* @XiXiphus made their first contribution in https://github.com/THUDM/slime/pull/2254
* @alexqdh made their first contribution in https://github.com/THUDM/slime/pull/2247
* @qq1243196045 made their first contribution in https://github.com/THUDM/slime/pull/2246
* @keepkeen made their first contribution in https://github.com/THUDM/slime/pull/2241
* @morluto made their first contribution in https://github.com/THUDM/slime/pull/2205
* @zhangdw156 made their first contribution in https://github.com/THUDM/slime/pull/2138
* @ZiyiTsang made their first contribution in https://github.com/THUDM/slime/pull/2132
* @LZ-QWQ made their first contribution in https://github.com/THUDM/slime/pull/2274
* @ForAxel made their first contribution in https://github.com/THUDM/slime/pull/2286
* @liuhao-labs made their first contribution in https://github.com/THUDM/slime/pull/2312
Full Changelog: https://github.com/THUDM/slime/compare/v0.3.1...v0.3.2
v0.3.12026-08-06
We are excited to announce the release of slime v0.3.1!
Unlike previous releases, v0.3.1 does not focus on major new features. As increasingly capable AI agents further improve our engineering productivity, we are now able to revisit and remove many compromises that were previously difficult to address. This release brings improvements across memory efficiency, training performance, weight updates, rollout alignment, and system robustness.
Some noteworthy improvements include:
More memory-efficient colocated training**
* Reduced Megatron offload overhead by optimizing world-group offloading.
Added support for PYTORCHALLOCCONF, including expandablesegments:True.
* Added --release-train to release and reload training workers between iterations, leaving more GPU memory available for rollout inference and HiCache.
Faster and more flexible weight updates**
Optimized the weight-update path and parallel updatefrom_disk.
* Added faster raw Hugging Face checkpoint saving with node-level writers.
* Fully supported --rollout-external-engine-addrs and zero-GPU rollout-router startup.
Improved disaggregated rollout with disk-level delta weight synchronization and engine-side /pullweights.
Better rollout and training alignment**
* Added top-p masking so training can reproduce the effective sampling distribution used during rollout.
* Improved prefill-decode disaggregation support for R3 routed experts.
Better training performance**
* Fused PPO logprob and entropy computation and optimized the vocabulary-parallel operators to reduce memory usage.
* Added FLOPs-balanced micro-batching to better balance variable-length samples across microbatches.
* Overlapped data loading with SGLang initialization to reduce startup overhead.
Miscellaneous improvements**
* Continued improving coding-agent RL, rollout hooks, CI, documentation, and distributed-training stability.
* Internalized the necessary model-conversion components and removed Megatron Bridge support to simplify the slime codebase. For broader model support, we recommend using radixark/miles.
What's Changed
* [docs] Add step-by-step debug tutorial by @liyc-ai in https://github.com/THUDM/slime/pull/2001
* fix(multi-agent): preserve rollout logprobs by @Jiang020609 in https://github.com/THUDM/slime/pull/1995
Revert "rename rolloutids to group_ids" by @zhuzilin in https://github.com/THUDM/slime/pull/2013
* Fully support --rollout-external-engine-addrs by @zhuzilin in https://github.com/THUDM/slime/pull/2016
* Accelerate raw HF save with node writers by @zhuzilin in https://github.com/THUDM/slime/pull/2020
Support updatefrom_disk by @zhuzilin in https://github.com/THUDM/slime/pull/2021
* Add docs for external servers by @zhuzilin in https://github.com/THUDM/slime/pull/2022
* feat: add --balance-by-flops for FLOPs-balanced micro-batching by @HaoDong0027 in https://github.com/THUDM/slime/pull/2017
* remove abundant function by @zhuzilin in https://github.com/THUDM/slime/pull/2028
use balanceby_flops as balance data across mbs by @zhuzilin in https://github.com/THUDM/slime/pull/2029
[examples]: add qwen3.5-9b model config and fullyasync example by @demouo in https://github.com/THUDM/slime/pull/2030
* Log progress while waiting for placement group by @luccabb in https://github.com/THUDM/slime/pull/2024
Allow only set rolloutid for prompt that return multiple responses by @zhuzilin in https://github.com/THUDM/slime/pull/2031
* Only upload per sample stats to wandb by @zhuzilin in https://github.com/THUDM/slime/pull/2027
* [docs] Add AgentCore RL Toolkit to ecosystem list by @lyzustc in https://github.com/THUDM/slime/pull/2034
* Don't upload 'count' to wandb by @zhuzilin in https://github.com/THUDM/slime/pull/2035
* [docker] always re-register mooncake addr during offloading by @zhuzilin in https://github.com/THUDM/slime/pull/2041
* [docker] update sgl-router by @zhuzilin in https://github.com/THUDM/slime/pull/2042
* Revert "[docker] always re-register mooncake addr during offloading" by @zhuzilin in https://github.com/THUDM/slime/pull/2046
* docs: add vime to the ecosystem section in README by @CalvinXKY in https://github.com/THUDM/slime/pull/2045
* support rich image config for vlm by @demouo in https://github.com/THUDM/slime/pull/2044
* [docs] add Miles to slime ecosystem by @yueming-yuan in https://github.com/THUDM/slime/pull/2047
Set RAYUSE_UVLOOP=0 for Ray actors by @zhuzilin in https://github.com/THUDM/slime/pull/2050
* [ci] clean up ci by @zhuzilin in https://github.com/THUDM/slime/pull/2055
* Use /v1/loads to re-abort server by @zhuzilin in https://github.com/THUDM/slime/pull/2056
* [docs] update miles description and chinese version by @yueming-yuan in https://github.com/THUDM/slime/pull/2058
* Allow zero-GPU rollout router startup by @zhuzilin in https://github.com/THUDM/slime/pull/2057
* [docker] expose sglang load inflight details by @zhuzilin in https://github.com/THUDM/slime/pull/2070
* fix(search-r1): stop generation at </search> and </answer> by @yichuan-w in https://github.com/THUDM/slime/pull/2036
* [docker] upgrade sglang to v0.5.13 by @zhuzilin in https://github.com/THUDM/slime/pull/2072
* cleanup by @zhuzilin in https://github.com/THUDM/slime/pull/2080
* sync from internal and cleanup by @zhuzilin in https://github.com/THUDM/slime/pull/2081
* [algo] Add CISPO advantage estimator (MiniMax-M1) by @EazyReal in https://github.com/THUDM/slime/pull/2067
* Overlapping data loading and sglang initialization by @zhuzilin in https://github.com/THUDM/slime/pull/2082
* Add ci badge to readme by @zhuzilin in https://github.com/THUDM/slime/pull/2086
* Fix CI badge by @zhuzilin in https://github.com/THUDM/slime/pull/2087
Add rolloutdata_transport nixl by @zhuzilin in https://github.com/THUDM/slime/pull/2088
* Support GLM-5.2 by @zhuzilin in https://github.com/THUDM/slime/pull/2093
* [coding-agent-rl] Refactor coding-agent RL: turn-node TrajectoryManager + pluggable harness layer by @jingshenghang in https://github.com/THUDM/slime/pull/2005
* Remove bshd support by @zhuzilin in https://github.com/THUDM/slime/pull/2100
* docs: drop dangling Dr.GRPO custom-reducer example reference by @EazyReal in https://github.com/THUDM/slime/pull/2096
Support topp mask by @zhuzilin in https://github.com/THUDM/slime/pull/2102
fix(examples/tau-bench): use RunConfig.agentstrategy in TAU_CONFIGS by @CalvinXKY in https://github.com/THUDM/slime/pull/2101
* Fix wrongly removed RouterArgs registry by @zhuzilin in https://github.com/THUDM/slime/pull/2107
Extract appendresponse_tokens to Sample by @zhuzilin in https://github.com/THUDM/slime/pull/2108
Extract appendresponse_tokens to Sample by @zhuzilin in https://github.com/THUDM/slime/pull/2110
* [codex] Add Dressage to ecosystem README by @huang3eng in https://github.com/THUDM/slime/pull/2111
* sync from internal by @zhuzilin in https://github.com/THUDM/slime/pull/2118
* Fix bug on non-float reward by @zhuzilin in https://github.com/THUDM/slime/pull/2121
* fix(agent) SWE coding-agent RL stability bugs (abort handling, session cleanup) by @jingshenghang in https://github.com/THUDM/slime/pull/2124
feat(codingagentrl): select claudecode/codex harness+adapter pair via SWE_AGENT by @jingshenghang in https://github.com/THUDM/slime/pull/2125
Fix parallel updatefrom_disk in megatron server by @zhuzilin in https://github.com/THUDM/slime/pull/2143
[docker] fix topp mask speed issue by @zhuzilin in https://github.com/THUDM/slime/pull/2145
* feat(gemma4): add Gemma4 dense and MoE support by @EazyReal in https://github.com/THUDM/slime/pull/2135
* fix: handle empty colocated weight buckets by @EazyReal in https://github.com/THUDM/slime/pull/2134
* perf: fuse PPO logprob entropy computation by @zhuzilin in https://github.com/THUDM/slime/pull/2144
*
v0.3.02026-05-31
We are excited to announce the release of slime v0.3.0!
This release marks a major step toward agent-first reinforcement learning. While slime has supported multi-turn and agentic workloads from early versions, v0.3.0 turns these scattered capabilities into a more coherent foundation: reusable agent components, safer trajectory handling, black-box coding-agent RL examples, fully async training as a first-class path, variable global batch-size training, and lower host-memory usage that enables more flexible rollout-inference configurations.
Major Updates
Agent-first slime/agent module**
* Added a dedicated slime/agent module with reusable components for sandbox-based agents, trajectory merging, rollout grouping, and OpenAI/Anthropic-compatible adapters.
Black-box coding-agent RL example**
Added codingagent_rl, a minimal sandbox-based coding-agent RL example for training agents that interact with real environments through tools, files, commands, retries, and multi-step trajectories.
Variable global batch-size training**
* Added support for variable global batch sizes, reducing the need for padding or trimming and better supporting compacted rollouts, filtered samples, subagents, and overlong trajectory handling.
Fully async training as a first-class path**
* Moved the fully async training workflow into the main codebase, making it easier to support agent workloads with heterogeneous rollout latency and long-horizon trajectories.
Host-memory optimizations**
* Reduced host-memory overhead by removing several large parameter-copy paths and improving memory saver integration, leaving more room for HiCache, offload, rollout buffers, and agent trajectories.
PPO refactor and actor-critic colocation**
* Refactored PPO and critic handling, improved value-head/offload paths, and reduced GPU usage through more efficient actor-critic colocation.
Other Notable Changes
Added delta weight sync** with disk and NCCL transports, moving slime closer to practical disaggregated training/inference deployment.
Added FlashQLA backend** support for Qwen GDN and improved Qwen3.5 / Qwen3.6 compatibility.
* Added --save-hf support for raw mode and improved HF checkpoint conversion.
Environment and CI
* Upgraded SGLang to v0.5.12.post1.
* Upgraded Megatron to 1dcf0dafa.
* Added more CPU-only unit tests for faster correctness checks.
* Added e2e and validation coverage for delta weight update, GPU placement, multi-sample training, server-group GPU indices, and train-rollout log-prob consistency.
What's Changed
* pass critic role through to create RayTrainGroup by @znculee in https://github.com/THUDM/slime/pull/1797
* fix qwen3.5 397B converting error when enable expert parallel by @xutianming in https://github.com/THUDM/slime/pull/1799
* fix(geo3k-vlm-sft): remove --apply-chat-template from SFT launch script by @DongzhuoranZhou in https://github.com/THUDM/slime/pull/1791
Add host memory metrics to availablememory function by @peterjc123 in https://github.com/THUDM/slime/pull/1764
* [WIP] fix loss oom by @lilei199908 in https://github.com/THUDM/slime/pull/1788
* sync from internal by @zhuzilin in https://github.com/THUDM/slime/pull/1805
* sync from internal by @zhuzilin in https://github.com/THUDM/slime/pull/1807
* feat: add npu patch for qwen3-vl-8b grpo & ppo by @cjy0x in https://github.com/THUDM/slime/pull/1750
fix missing positionids in log-prob forward step by @znculee in https://github.com/THUDM/slime/pull/1809
* feat: add support for including missing weights from origin HF checkp… by @peterjc123 in https://github.com/THUDM/slime/pull/1812
[Fix] Initialize gradnorm before found_inf skip path by @kaysonyu in https://github.com/THUDM/slime/pull/1762
[conda] Add install custom sgl-router to buildconda.sh by @zhuzilin in https://github.com/THUDM/slime/pull/1813
Revert nograd for entropy to prevent comm stuck in dsa by @zhuzilin in https://github.com/THUDM/slime/pull/1822
Add fallback for getseqlenbalancedpartitions by @zhuzilin in https://github.com/THUDM/slime/pull/1823
* docs: add Relax to notable projects in README by @Yangruipis in https://github.com/THUDM/slime/pull/1834
Bugfix: use cpu instead of cuda in converttorchdistto_hf.py when --add-missing-from-origin-hf is set by @coding-famer in https://github.com/THUDM/slime/pull/1828
* [fix] eval sample logging when sample is a list by @mathewjhan in https://github.com/THUDM/slime/pull/1836
[Fix] Fix cuda-python pin in buildconda.sh by @kaysonyu in https://github.com/THUDM/slime/pull/1827
* fix entropy bug and update code by @lilei199908 in https://github.com/THUDM/slime/pull/1846
Revert "Add fallback for getseqlenbalancedpartitions" by @zhuzilin in https://github.com/THUDM/slime/pull/1848
* fix by @lilei199908 in https://github.com/THUDM/slime/pull/1849
Add support for NVIDIA DGX Spark (GB10 / sm121a, arm64) by @boots-coder in https://github.com/THUDM/slime/pull/1835
* refactor/ppo by @lilei199908 in https://github.com/THUDM/slime/pull/1856
* [docker] cleanup sglang patch by @zhuzilin in https://github.com/THUDM/slime/pull/1859
* Rename critic config to megatron config by @zhuzilin in https://github.com/THUDM/slime/pull/1866
[Fix] Use Ray ObjectRef await instead of asyncio.tothread in distributed POST by @ryang-max in https://github.com/THUDM/slime/pull/1873
chore: include length context in slicelogprobwith_cp assert by @leofan-lab in https://github.com/THUDM/slime/pull/1862
* [docker] upgrade megatron to 1dcf0dafa by @zhuzilin in https://github.com/THUDM/slime/pull/1867
* fix ppo value head load bugs by @lilei199908 in https://github.com/THUDM/slime/pull/1878
* [docker] upgrade sglang to v0.5.10.post1 by @zhuzilin in https://github.com/THUDM/slime/pull/1874
* [docker] update megatron-bridge and add qwen3.6 tests by @zhuzilin in https://github.com/THUDM/slime/pull/1884
Fix(checkpoint): add resume/pause in savemodel() for offload_train (fixes #1886) by @Procrastinatorrrr in https://github.com/THUDM/slime/pull/1888
* fix ppo value offload bugs by @lilei199908 in https://github.com/THUDM/slime/pull/1882
* fix qwen3.6 hf config validation bug by @zhuzilin in https://github.com/THUDM/slime/pull/1889
* Add missing metrics to log by @zhuzilin in https://github.com/THUDM/slime/pull/1890
fix(qwen3next): use torch.getdefaultdtype() — getcurrentdtype do… by @HeatherLiuzh in https://github.com/THUDM/slime/pull/1883
* Fix location error in install script by @selfanti in https://github.com/THUDM/slime/pull/1877
* Only allow --allgather-cp for DSA model by @zhuzilin in https://github.com/THUDM/slime/pull/1891
* Migrate internal feature by @zhuzilin in https://github.com/THUDM/slime/pull/1897
* [Fix] Fix distributed POST actor concurrency split by @kaysonyu in https://github.com/THUDM/slime/pull/1880
Fix CI: update rolloutdata_postprocess plugin contract for new call site by @jingshenghang in https://github.com/THUDM/slime/pull/1902
* Patch Megatron TP grad coalesce to chunked all-reduce by @jingshenghang in https://github.com/THUDM/slime/pull/1899
* fix: harden retool rollout against multi-turn / retry desync by @leofan-lab in https://github.com/THUDM/slime/pull/1861
* [docker] update torch memory saver by @zhuzilin in https://github.com/THUDM/slime/pull/1916
* add critic wandb config by @lilei199908 in https://github.com/THUDM/slime/pull/1919
Move fullyasync example to main codebase by @zhuzilin in https://github.com/THUDM/slime/pull/1920
* Add example for streaming output by @zhuzilin in https://github.com/THUDM/slime/pull/1921
* Reduce host memory with upgraded tms by @zhuzilin in https://github.com/THUDM/slime/pull/1924
* Move micro-batch scheduling from training side to rollout side by @zhuzilin in https://github.com/THUDM/slime/pull/1926
* save host menmroy by @lilei199908 in https://github.com/THUDM/slime/pul
v0.2.42026-03-29
v0.2.4 is here! Thanks to everyone who contributed to this release.
Major Updates
In addition to a broad set of bug fixes and stability improvements, v0.2.4 brings several major updates:
- Profiling and observability improvements
Added a rollout trace timeline viewer and W&B reporting for dynamic ITL / TTFT percentile metrics.
- Router stack unified on sgl-router
Consolidated the router stack onto sgl-router and removed slime-router.
- Expanded multimodal and model support
Improved support for GLM-4.6V / GLM4V, Multimodal OPD, and Qwen3.5-related workflows.
Other Notable Changes
- Fixed CUDA IPC cache leaks during weight updates
- Fixed SP/CP gradient inflation in FLA layers
What's Changed
* feat: add GLM-4.6V MoE VL bridge with CP support by @zhuzilin in https://github.com/THUDM/slime/pull/1715
fix: resolve ropetheta from rope_parameters dict in HF config validation by @zhuzilin in https://github.com/THUDM/slime/pull/1720
* [docker] patches for glm4.6v, kimi k2.5 and dsa cp only by @zhuzilin in https://github.com/THUDM/slime/pull/1722
* Fix CUDA IPC cache leaks during weight updates by @zhuzilin in https://github.com/THUDM/slime/pull/1731
* [docker] update megatron by @zhuzilin in https://github.com/THUDM/slime/pull/1729
* [docker] Fix IndexCache with mla model by @zhuzilin in https://github.com/THUDM/slime/pull/1736
* [slime-router] support pd disaggregation and remove radix tree middleware by @zhuzilin in https://github.com/THUDM/slime/pull/1735
* Fix glm4v megatron bridge by @zhuzilin in https://github.com/THUDM/slime/pull/1738
* [docker] update sglang patch by @zhuzilin in https://github.com/THUDM/slime/pull/1743
* feat: GLM4V multimodal support improvements by @zhuzilin in https://github.com/THUDM/slime/pull/1745
* feat: placeholder worker type, metrics router, and GPQA letter range by @zhuzilin in https://github.com/THUDM/slime/pull/1746
always enablemetrics and remove dp context by @zhuzilin in https://github.com/THUDM/slime/pull/1747
* fix: resolve SP/CP gradient inflation in FLA (linear attention) layers by @zhuzilin in https://github.com/THUDM/slime/pull/1748
* Update MTP example configs, rename GLM-4.5 to GLM-4.7, clean scripts by @zhuzilin in https://github.com/THUDM/slime/pull/1749
* Support qwen3.5 loss mask for multi-turn SFT by @huang3eng in https://github.com/THUDM/slime/pull/1742
fix: propagate moetokendispatchertype in bridge model provider by @nanjiangwill in https://github.com/THUDM/slime/pull/1737
fix: resolve ropetheta from rope_parameters in DeepseekV32Bridge by @stevewx in https://github.com/THUDM/slime/pull/1734
* chore: translate remaining Chinese comments to English by @WangHong-yang in https://github.com/THUDM/slime/pull/1726
* feat: add Qwen3.5-4B model support by @shihaohou in https://github.com/THUDM/slime/pull/1721
fix: httputils. disable system proxy for internal SGLang httpx clients by @DongzhuoranZhou in https://github.com/THUDM/slime/pull/1714
* fix: auto-detect GPUs in qwen3-4b script by @ailuntz in https://github.com/THUDM/slime/pull/1700
fix: quote $MOELAYER_FREQ by @lawrence-harmonic in https://github.com/THUDM/slime/pull/1689
disable router healthcheck and allow prompt_data is None by @zhuzilin in https://github.com/THUDM/slime/pull/1751
* small fix on qwen3-235b-a22b launch script by @Zhuohao-Li in https://github.com/THUDM/slime/pull/1719
* sync internal bugfix by @zhuzilin in https://github.com/THUDM/slime/pull/1765
* Fix uploading sglang metrics to wandb by @zhuzilin in https://github.com/THUDM/slime/pull/1768
* use zhuzilin/sgl-router for sglang-router by @zhuzilin in https://github.com/THUDM/slime/pull/1770
* [docker] update sgl-router by @zhuzilin in https://github.com/THUDM/slime/pull/1772
* [Multimodal] Add Multimodal OPD support by @coding-famer in https://github.com/THUDM/slime/pull/1760
* refactor: remove slime router by @zhuzilin in https://github.com/THUDM/slime/pull/1773
* Add rollout trace timeline viewer by @zhuzilin in https://github.com/THUDM/slime/pull/1776
* [Fix] Fix duplicate Megatron LR scheduler resume when optimizer state is not loaded by @kaysonyu in https://github.com/THUDM/slime/pull/1775
* Support FP8 conversion for Qwen3.5 by @peterjc123 in https://github.com/THUDM/slime/pull/1769
* fix typo by @albaNnaksqr in https://github.com/THUDM/slime/pull/1759
* [Fix]Fix some bugs/clean up by @coding-famer in https://github.com/THUDM/slime/pull/1756
(fix):not have encoderonly attr cause run failed by @wangyufak in https://github.com/THUDM/slime/pull/1741
New Contributors
* @stevewx made their first contribution in https://github.com/THUDM/slime/pull/1734
* @WangHong-yang made their first contribution in https://github.com/THUDM/slime/pull/1726
* @shihaohou made their first contribution in https://github.com/THUDM/slime/pull/1721
* @DongzhuoranZhou made their first contribution in https://github.com/THUDM/slime/pull/1714
* @ailuntz made their first contribution in https://github.com/THUDM/slime/pull/1700
* @peterjc123 made their first contribution in https://github.com/THUDM/slime/pull/1769
* @albaNnaksqr made their first contribution in https://github.com/THUDM/slime/pull/1759
* @wangyufak made their first contribution in https://github.com/THUDM/slime/pull/1741
Full Changelog: https://github.com/THUDM/slime/compare/v0.2.3...v0.2.4
v0.2.32026-03-12
v0.2.3 is here! Thanks to everyone who contributed to this release.
Major Updates
In addition to a broad set of bug fixes and stability improvements, v0.2.3 brings several major updates:
YAML-based sglangconfig support for engine group configuration
This makes rollout setup much more flexible: you can now configure different parallelism strategies for PD disaggregation, enable EPD-style deployments, or even serve multiple heterogeneous models within one rollout setup more cleanly.
* Expanded model support, including GLM5, GLM-4.7-Flash and Qwen3.5.
* Dependency and runtime updates, including SGLang v0.5.9 docker support and multiple fixes for PD, NSA, HiCache, CP+PP, etc.
Other Notable Changes
* Added consistent hashing routing for multi-turn rollout
* Removed FSDP support so we can focus maintenance effort on the training and rollout paths we actively invest in
What's Changed
* fix: fix sglang regression by @nanjiangwill in https://github.com/THUDM/slime/pull/1363
* [docker] upgrade fla to 0.4.1 by @zhuzilin in https://github.com/THUDM/slime/pull/1452
Allow passing ppsize by @zhuzilin in https://github.com/THUDM/slime/pull/1454
* Revert "fix: fix sglang regression" by @zhuzilin in https://github.com/THUDM/slime/pull/1457
* update #1457 by @nanjiangwill in https://github.com/THUDM/slime/pull/1458
* [model] Add support for GLM4.7 Flash by @zhuzilin in https://github.com/THUDM/slime/pull/1460
* [script] Add example script for GLM4.7 Flash by @zhuzilin in https://github.com/THUDM/slime/pull/1467
* add lb default by @lilei199908 in https://github.com/THUDM/slime/pull/1465
* support non-symmetric int4 qat by @zhuzilin in https://github.com/THUDM/slime/pull/1472
* add convert hf to int4 without calibration dataset by @xieck13 in https://github.com/THUDM/slime/pull/1489
* [docker] fix nsa + hicache by @zhuzilin in https://github.com/THUDM/slime/pull/1494
* fix compute logprobs menmory leak bug by @lilei199908 in https://github.com/THUDM/slime/pull/1506
[Fix] support converting torchdist to hf for qwen3vl dense model by @p1k0pan in https://github.com/THUDM/slime/pull/1491
renamed qwen3-vl.py to qwen3vl.py to fix typo by @gxlvera in https://github.com/THUDM/slime/pull/1512
* sync internal features by @zhuzilin in https://github.com/THUDM/slime/pull/1513
* [docker] allow retract to 0 req during PD by @zhuzilin in https://github.com/THUDM/slime/pull/1515
* Pass through correct Megatron model provider PP args by @hari-hm in https://github.com/THUDM/slime/pull/1486
* fix: support mtp for qwen3-next by @huang3eng in https://github.com/THUDM/slime/pull/1503
* [Doc] Add doc for slime router by @Hecate0821 in https://github.com/THUDM/slime/pull/1499
* Add support for vlm checkpoints conversion by @cklxx in https://github.com/THUDM/slime/pull/1475
* fix: use aread() to fully consume HTTP response body by @ann-qin-lu in https://github.com/THUDM/slime/pull/1488
* Add response.aclose() and fix lint by @zhuzilin in https://github.com/THUDM/slime/pull/1520
* [docker] alleviate pd memory leakage by @zhuzilin in https://github.com/THUDM/slime/pull/1525
* Update convert to INT4 script path by @gxlvera in https://github.com/THUDM/slime/pull/1528
[docker] add allocextendtorchfallback for long context by @zhuzilin in https://github.com/THUDM/slime/pull/1530
[docker] remove seqlen in getkandstritonkernel by @zhuzilin in https://github.com/THUDM/slime/pull/1531
* [Feature] Add megatron version for on policy distillation by @yitianlian in https://github.com/THUDM/slime/pull/1538
Make DataSource implement len__ to standardize the data source contract by @TSunny007 in https://github.com/THUDM/slime/pull/1518
* [Fix] Add fsdp assert for OPD by @yitianlian in https://github.com/THUDM/slime/pull/1545
* [bug] fix: gracefully handle datapoints with no multimodal input in a multimodal dataset by @hleehlee-amazon in https://github.com/THUDM/slime/pull/1535
* fix sglang hicache nsa bugs by @lilei199908 in https://github.com/THUDM/slime/pull/1549
* [docker] fix nsa retract by @zhuzilin in https://github.com/THUDM/slime/pull/1566
* [Multimodal] make multimodal processing robust by @coding-famer in https://github.com/THUDM/slime/pull/1516
[Fix] Fix multimodaltrain_inputs handling for mixed text-multimodal datasets by @coding-famer in https://github.com/THUDM/slime/pull/1559
[Feature] add converttorchdisttohfbridge.py by @coding-famer in https://github.com/THUDM/slime/pull/1573
* [Fix] Fix resuming from Megatron checkpoint when using bridge by @coding-famer in https://github.com/THUDM/slime/pull/1577
* [Feaure] Add Profile Config by @yitianlian in https://github.com/THUDM/slime/pull/1561
* [Feature] Add consistent hashing routing policy for rollout by @yitianlian in https://github.com/THUDM/slime/pull/1588
* [Fix] Minor fix for support distribute mode by @yitianlian in https://github.com/THUDM/slime/pull/1589
* [bug] fix: fix type hint in Sample class by @hleehlee-amazon in https://github.com/THUDM/slime/pull/1551
* Add support for GLM5 by @zhuzilin in https://github.com/THUDM/slime/pull/1599
fix(examples): update strandssglang example to strands-sglang v0.2.x API by @Lawhy in https://github.com/THUDM/slime/pull/1593
* refactor: separate sglang argparse from megatron with two-phase parsing by @zhuzilin in https://github.com/THUDM/slime/pull/1600
* Add issue templates by @zhuzilin in https://github.com/THUDM/slime/pull/1602
* [cleanup] remove metric checker and long tests by @zhuzilin in https://github.com/THUDM/slime/pull/1603
fix: use getattr for sglang params when rolloutonly is disabled by @zhuzilin in https://github.com/THUDM/slime/pull/1604
refactor: extract RolloutServerGroup and startrollout_server by @zhuzilin in https://github.com/THUDM/slime/pull/1605
* refactor: delegate engine state and operations to RolloutServerGroup by @zhuzilin in https://github.com/THUDM/slime/pull/1606
* fix: sync docs with current implementation by @zhuzilin in https://github.com/THUDM/slime/pull/1608
* fix: restrict SymPy parser scope to prevent arbitrary code execution by @Hardik-369 in https://github.com/THUDM/slime/pull/1587
* refactor: split PD disaggregation into separate EngineGroups by @zhuzilin in https://github.com/THUDM/slime/pull/1609
* feat: move OPD to slime/rollout, add CI test and docs by @zhuzilin in https://github.com/THUDM/slime/pull/1610
* docs: add CI section to developer guide (EN + ZH) by @zhuzilin in https://github.com/THUDM/slime/pull/1612
* refactor: make EngineGroup ops non-blocking and batch ray.get at RolloutServer level by @zhuzilin in https://github.com/THUDM/slime/pull/1613
Fix: disable allowabbrev in preparse_mode to prevent --load collision by @coding-famer in https://github.com/THUDM/slime/pull/1616
* fix: skip memory check for non-communication query functions in ReloadableProcessGroup by @zhuzilin in https://github.com/THUDM/slime/pull/1625
* Fix: Handle quantization formats during weight synchronization in Megatron bridge. by @GeLee-Q in https://github.com/THUDM/slime/pull/1624
* Add retries, which are not built into aiohttp by @joyliu-q in https://github.com/THUDM/slime/pull/1617
* [docker] upgrade to v0.5.9 by @zhuzilin in https://github.com/THUDM/slime/pull/1626
* Upgrade GitHub Actions for Node 24 compatibility by @salmanmkc in https://github.com/THUDM/slime/pull/1590
* add no colocate update critic only by @lilei199908 in https://github.com/THUDM/slime/pull/1567
remove redundant logrollout_data by @zhuzilin in https://github.com/THUDM/slime/pull/1629
* feat: add --sglang-config YAML for engine group configuration by @zhuzilin in https://github.com/THUDM/slime/pull/1614
Fix #1595: pass rolloutid explicitly to offload_train by @yitianlian in https://github.com/THUDM/slime/pull/1631
Fix #1615: tolerate abortrequest connection failures by @yitianlian in https://github.com/THUDM/slime/pull/1632
* [docker] suppo
v0.2.22026-01-18
v0.2.2 is here! Thanks to everyone who contributed to this release.
Major Updates
In addition to multiple memory and performance improvements, v0.2.2 adds support for:
- Int4-QAT training
- Full R3 (Rollout Routing Replay) support with DeepEP and MTP
- Dependency upgrades: SGLang v0.5.7 and the Megatron dev branch
What's Changed
* add ckpt load save ci by @lilei199908 in https://github.com/THUDM/slime/pull/1104
* Add --rollout-all-samples-process-path for RLVE by @zhuzilin in https://github.com/THUDM/slime/pull/1107
* feat: support Qwen3 Moe BackEnd Kernel by @attack204 in https://github.com/THUDM/slime/pull/1071
* fix max response/context/prompt len by @lilei199908 in https://github.com/THUDM/slime/pull/1110
* fix max len by @lilei199908 in https://github.com/THUDM/slime/pull/1112
* [docker] remove amem and support deepep + r3 by @zhuzilin in https://github.com/THUDM/slime/pull/1115
* [Fix] Fix early return in init rollout engine by @yitianlian in https://github.com/THUDM/slime/pull/1118
* [Fix] Add sglang patch for weight version update by @yitianlian in https://github.com/THUDM/slime/pull/1119
* fix: improve tokenization by @nanjiangwill in https://github.com/THUDM/slime/pull/1113
* [Feature] Add CI test for weight version update by @yitianlian in https://github.com/THUDM/slime/pull/1120
* [docker] optimize r3 with base64 encode by @zhuzilin in https://github.com/THUDM/slime/pull/1124
* [docker] fix r3 gather buffer by @zhuzilin in https://github.com/THUDM/slime/pull/1129
* [docker] support mtp for r3 by @zhuzilin in https://github.com/THUDM/slime/pull/1131
* [Fix] Fix some bugs in retool example by @yitianlian in https://github.com/THUDM/slime/pull/1130
Add finalizemodelgradswithemptycache by @zhuzilin in https://github.com/THUDM/slime/pull/1133
* Feat: add usage docs for fsdp by @lin0303-siyuan in https://github.com/THUDM/slime/pull/1092
* Reserve more ports for new sglang dp attn impl by @zhuzilin in https://github.com/THUDM/slime/pull/1142
* Blog: fix the path of the Blog's architecture image by @ShanningZhuang in https://github.com/THUDM/slime/pull/1125
* Support async save and add extra save at the end of the training by @zhuzilin in https://github.com/THUDM/slime/pull/1143
* fix: fix GemmeRMSNorm.forward() bug by @nanjiangwill in https://github.com/THUDM/slime/pull/1121
* [WIP][FSDP] Support FSDP for Qwen3Next by @rucnyz in https://github.com/THUDM/slime/pull/1116
* Megatron VLM Support (1/N) by @Zhuohao-Li in https://github.com/THUDM/slime/pull/1123
* Update deprecated huggingface-cli and fix broken links by @Lyken17 in https://github.com/THUDM/slime/pull/1147
Added FSDP checkpoint handling to converttorchdistto_hf.py by @cklxx in https://github.com/THUDM/slime/pull/1101
* minor fix for megatron compatibility by @zhuzilin in https://github.com/THUDM/slime/pull/1149
Remove configmapping to use megatron-bridge by @zhuzilin in https://github.com/THUDM/slime/pull/1166
* Avoids repeated work. by @qqwqqw689 in https://github.com/THUDM/slime/pull/1163
Make tools/converttorchdistto_hf.py not rely on megatron by @zhuzilin in https://github.com/THUDM/slime/pull/1167
* support converting dpsk mtp layer by @zhuzilin in https://github.com/THUDM/slime/pull/1169
* [FSDP] Add Masked importance sampling by @zijiexia in https://github.com/THUDM/slime/pull/1122
* [TIS/MIS] fix and add better metric by @ChangyiYang in https://github.com/THUDM/slime/pull/1174
* Fix optimizer schedule resume by @lr-tsinghua11 in https://github.com/THUDM/slime/pull/1152
* [docker] upgrade to megatron dev branch by @zhuzilin in https://github.com/THUDM/slime/pull/1153
* Minor fix by @lancerts in https://github.com/THUDM/slime/pull/1165
* Fix forward of Qwen3VLTextRotaryEmbedding in Megatron-Bridge by @zhuzilin in https://github.com/THUDM/slime/pull/1179
* Reuse the text llm config for qwen3 vl models by @zhuzilin in https://github.com/THUDM/slime/pull/1180
* Don't save AutoBridge in args by @zhuzilin in https://github.com/THUDM/slime/pull/1181
* [Fix] Fix port error in PD disaggregation setting by @yitianlian in https://github.com/THUDM/slime/pull/1175
Fix prompt type bug in generatewith_search within examples/search-r1 by @jiahe7ay in https://github.com/THUDM/slime/pull/1182
* feat: support Qwen3 VL MoE by @nanjiangwill in https://github.com/THUDM/slime/pull/1171
* [Fix] Minor fix by @yitianlian in https://github.com/THUDM/slime/pull/1183
* Set parallel config for megatron bridge by @zhuzilin in https://github.com/THUDM/slime/pull/1184
Fix tools/converthftotorch_dist.py by @zhuzilin in https://github.com/THUDM/slime/pull/1186
* Don't calculate entropy grad when coef is 0 by @zhuzilin in https://github.com/THUDM/slime/pull/1185
* Disable routing replay for critic by @zhuzilin in https://github.com/THUDM/slime/pull/1187
* Revert "Don't calculate entropy grad when coef is 0" by @zhuzilin in https://github.com/THUDM/slime/pull/1189
* Fix qwen3next for megatron dev branch by @zhuzilin in https://github.com/THUDM/slime/pull/1190
* fix: fix logging for rollout by @nanjiangwill in https://github.com/THUDM/slime/pull/1188
* sync internal features by @zhuzilin in https://github.com/THUDM/slime/pull/1192
Fix checkweights api by @zhuzilin in https://github.com/THUDM/slime/pull/1194
* Add --custom-rollout-log-function-path and --custom-eval-rollout-log-function-path by @zhuzilin in https://github.com/THUDM/slime/pull/1196
* [Feature] Add more logging for health monitor by @yitianlian in https://github.com/THUDM/slime/pull/1195
* fix: SFT tools support by @maoquan-ms in https://github.com/THUDM/slime/pull/1198
* [Featuren] Change default value of rollout health check by @yitianlian in https://github.com/THUDM/slime/pull/1197
* Megatron VLM Support w/ SFT (2/N) by @Zhuohao-Li in https://github.com/THUDM/slime/pull/1150
* tiny fix for sft script after tokenizer improvement by @Zhuohao-Li in https://github.com/THUDM/slime/pull/1201
* tests: add test for multi turn loss mask by @maoquan-ms in https://github.com/THUDM/slime/pull/1204
* Always pass loss masks to model by @zhuzilin in https://github.com/THUDM/slime/pull/1205
* [on-policy distillation] update reward function to fix potential token mismatches by @ahxt in https://github.com/THUDM/slime/pull/1128
* Add ci for mtp by @zhuzilin in https://github.com/THUDM/slime/pull/1207
* Fix mla tflops by @lilei199908 in https://github.com/THUDM/slime/pull/1209
* update docs by @zhuzilin in https://github.com/THUDM/slime/pull/1211
* update docs by @zhuzilin in https://github.com/THUDM/slime/pull/1214
* [Feature] Support 0.3.0 sglang router for fault tolerance by @yitianlian in https://github.com/THUDM/slime/pull/1215
* sync internal features by @zhuzilin in https://github.com/THUDM/slime/pull/1216
* feat: add custom logic for processing list[list[Sample]] to training data by @nanjiangwill in https://github.com/THUDM/slime/pull/1218
add int4quant cuda kernel by @Hyaloid in https://github.com/THUDM/slime/pull/1220
* update doc by @zhuzilin in https://github.com/THUDM/slime/pull/1224
* Improve AMD tutorial with complete model/data setup workflow by @Vivicai1005 in https://github.com/THUDM/slime/pull/1212
* update megatron patch by @zhuzilin in https://github.com/THUDM/slime/pull/1228
* sync from internal by @zhuzilin in https://github.com/THUDM/slime/pull/1229
* fix model saving bug in megatron by @zhuzilin in https://github.com/THUDM/slime/pull/1230
* add new status by @nanjiangwill in https://github.com/THUDM/slime/pull/1219
* update customization docs by @nanjiangwill in https://github.com/THUDM/slime/pull/1233
* Revert data processing of VLM by @zhuzilin in https://github.com/THUDM/slime/pull/1232
* [VLM] optimize VLM processing by @nanjiangwill in https://github.com/THUDM/slime/pull/1234
feat: add custom pgloss reducer by @ChangyiYang in https://github.com/THUDM/slime/pull/1235
* fix: typo "sgalng" → "sglang" in ROCm D
v0.2.12025-12-12
Thanks to the incredible support and contributions from our community — v0.2.1 is here!
Major Updates
- VLM + FSDP: true on-policy training on Qwen3-VL (dense).
- PD-disaggregation support during rollout
- DP-attention support in rollout routing replay (R3)
- Upgraded to SGLang v0.5.6
What's Changed
* extract mla update weight logic out by @zhuzilin in https://github.com/THUDM/slime/pull/960
* support do all evals together by @zhuzilin in https://github.com/THUDM/slime/pull/959
* Add --rollout-sample-filter-path by @zhuzilin in https://github.com/THUDM/slime/pull/961
* [FSDP] Optimize FSDP2 Model Loading with Rank-0 Broadcast by @Hecate0821 in https://github.com/THUDM/slime/pull/915
Add sample.removesample by @zhuzilin in https://github.com/THUDM/slime/pull/977
* add --eval-max-prompt-len by @zhuzilin in https://github.com/THUDM/slime/pull/978
Add args check for maxcontext_len by @zhuzilin in https://github.com/THUDM/slime/pull/979
Remove hard coded balanceabs_threshold by @zhuzilin in https://github.com/THUDM/slime/pull/981
Tiny fix fp8cast_bf16 not copying chat template by @fzyzcjy in https://github.com/THUDM/slime/pull/964
* Super tiny install dnsutils in dockerfile by @fzyzcjy in https://github.com/THUDM/slime/pull/965
* Super tiny sanity check checkpoint dir by @fzyzcjy in https://github.com/THUDM/slime/pull/966
Fix converthftotorch_dist OOM by @fzyzcjy in https://github.com/THUDM/slime/pull/967
* Tiny support using environment variables in addition to arguments for all scripts by @fzyzcjy in https://github.com/THUDM/slime/pull/968
* Super tiny increase default timeout sec by @fzyzcjy in https://github.com/THUDM/slime/pull/969
* Fix random port in use error even though already have free port detection by @fzyzcjy in https://github.com/THUDM/slime/pull/970
* Super tiny enable draft-weights-cpu-backup to avoid MTP acc len issue by @fzyzcjy in https://github.com/THUDM/slime/pull/971
* Add generation function for benchmarking purpose by @fzyzcjy in https://github.com/THUDM/slime/pull/972
* Support zero host or device memory waste for weight update by @fzyzcjy in https://github.com/THUDM/slime/pull/973
* Add fp8 kv cache and tis in qwen3 30b a3b script by @fzyzcjy in https://github.com/THUDM/slime/pull/974
* Add GB200, MTP, benchmark, fp8 rollout mode to glm script by @fzyzcjy in https://github.com/THUDM/slime/pull/975
* [FSDP] Add private func indicator for better usage by @PopSoda2002 in https://github.com/THUDM/slime/pull/982
* [Bugfix] Rename save model by @PopSoda2002 in https://github.com/THUDM/slime/pull/983
Fix: resolve variable shadowing bug in setupmodelandoptimizer by @fangzhensheng in https://github.com/THUDM/slime/pull/963
* remove unnecessary optimizer init by @zhuzilin in https://github.com/THUDM/slime/pull/984
* [release] bump to v0.2.0.post1 by @zhuzilin in https://github.com/THUDM/slime/pull/986
* fix scaling of per token loss by @zhuzilin in https://github.com/THUDM/slime/pull/987
* Add strands-agents example by @Lawhy in https://github.com/THUDM/slime/pull/976
* Add nemo skills evaluation by @guapisolo in https://github.com/THUDM/slime/pull/989
* [1/N] Tiny execute Ruff auto lint by @fzyzcjy in https://github.com/THUDM/slime/pull/991
* [2/N] Tiny manually fix for Ruff default ruleset and add to pre-commit by @fzyzcjy in https://github.com/THUDM/slime/pull/992
* [3/N] Enable B ruleset in Ruff by @fzyzcjy in https://github.com/THUDM/slime/pull/993
* [4/N] Tiny enable UP ruleset in Ruff by @fzyzcjy in https://github.com/THUDM/slime/pull/994
* Super tiny further fix lint error by @fzyzcjy in https://github.com/THUDM/slime/pull/995
* Add DataSource and --data-source-path by @zhuzilin in https://github.com/THUDM/slime/pull/912
* Fix per token loss scale and add e2e ci by @zhuzilin in https://github.com/THUDM/slime/pull/990
* [FSDP] Add script for FSDP Qwen3-4B by @Hecate0821 in https://github.com/THUDM/slime/pull/988
Fixed bug in checking maxlength for SFT #997 by @Surya-Gunukula in https://github.com/THUDM/slime/pull/998
* [ci] Add CI to make sure all dense parallel gives the same grad norm by @zhuzilin in https://github.com/THUDM/slime/pull/1000
* [Feature] Add off-policy sequence masking algorithm proposed in DeepSeek v3.2 by @yitianlian in https://github.com/THUDM/slime/pull/999
[FSDP][3/N] support trueon_policy training for FSDP2 by @zhuzilin in https://github.com/THUDM/slime/pull/1001
* fix lint by @zhuzilin in https://github.com/THUDM/slime/pull/1002
Fix bare except clause and remove redundant computation in ppoutils by @lancerts in https://github.com/THUDM/slime/pull/1007
* fix: FSDP runnable for Qwen3-30b-a3b by @yueming-yuan in https://github.com/THUDM/slime/pull/1010
* move tis function outside by @zhuzilin in https://github.com/THUDM/slime/pull/1014
* Add backward impl for SiluAndMulFunction and MoeSumReduceFunction by @zhuzilin in https://github.com/THUDM/slime/pull/1015
refactor: expose computemetricsfromsamples as public by @lancerts in https://github.com/THUDM/slime/pull/1012
* Fix evaluation parameter parsing by @guapisolo in https://github.com/THUDM/slime/pull/1005
* pre-commit run --all-files by @lancerts in https://github.com/THUDM/slime/pull/1021
* fix: update deprecated import path in mcore2hf script by @Chen-GX in https://github.com/THUDM/slime/pull/1003
* [FSDP] Add gpt oss 20b script by @PopSoda2002 in https://github.com/THUDM/slime/pull/996
* Fix mimo speculative decoding oom by @guapisolo in https://github.com/THUDM/slime/pull/1024
* [FSDP, VLM] feat: add vlm training for FSDP by @nanjiangwill in https://github.com/THUDM/slime/pull/501
* [rollout] support disable trim samples when converting rollout samples to train datas by @GGGGGGXY in https://github.com/THUDM/slime/pull/1016
* Backward compatible for older megatron version by @zhuzilin in https://github.com/THUDM/slime/pull/1028
* extract all sglang deps in megatron actor to one file by @zhuzilin in https://github.com/THUDM/slime/pull/1029
* feat: Add Unbiased KL Estimation from DeepSeek-V3.2 by @kekmodel in https://github.com/THUDM/slime/pull/1004
* refactor: extract duplicated checkpoint interval logic into reusable helper by @lancerts in https://github.com/THUDM/slime/pull/1027
Fix typo in sglangrollout.py comment by @ChenmienTan in https://github.com/THUDM/slime/pull/980
* fix ci for nodes with proxy by @zhuzilin in https://github.com/THUDM/slime/pull/1035
[FSPP] fix args error in applyfsdp2 function by @ChangyiYang in https://github.com/THUDM/slime/pull/1041
* [FSDP] Support lr scheduler by @ChangyiYang in https://github.com/THUDM/slime/pull/1040
* [Fix] Fix some bugs when on/offload model by @yitianlian in https://github.com/THUDM/slime/pull/1038
Improve debug output formatting in replayreward_fn.py by @lancerts in https://github.com/THUDM/slime/pull/1033
* Support pd disaggregation with p and d of same config by @zhuzilin in https://github.com/THUDM/slime/pull/1046
* [rollout] Truncate last token for rollout routing replay by @Hecate0821 in https://github.com/THUDM/slime/pull/1045
* fix: modernize type hint and add distributed init checks in utils by @lancerts in https://github.com/THUDM/slime/pull/1049
* Fix the padding of rollout routing replay experts by @zhuzilin in https://github.com/THUDM/slime/pull/1052
* update sglang to 0.5.6 by @lilei199908 in https://github.com/THUDM/slime/pull/1051
* [docker] fix cudnn version by @zhuzilin in https://github.com/THUDM/slime/pull/1066
* [docker] fix megatron cpu adam load issue by @zhuzilin in https://github.com/THUDM/slime/pull/1070
* fix(examples): correct quotes and comment out ray cleanup commands in Qwen3-30B-A3B FP8 script by @pandengyao in https://github.com/THUDM/slime/pull/1069
* Fix typos and improve clarity in documentation and code comments by @lancerts in https://github.com/THUDM/slime/pull/1067
* fix: remove redundant gc.collect()
v0.2.0.post12025-12-01
Fix critical bug mentioned in #958.
What's Changed
* extract mla update weight logic out by @zhuzilin in https://github.com/THUDM/slime/pull/960
* support do all evals together by @zhuzilin in https://github.com/THUDM/slime/pull/959
* Add --rollout-sample-filter-path by @zhuzilin in https://github.com/THUDM/slime/pull/961
* [FSDP] Optimize FSDP2 Model Loading with Rank-0 Broadcast by @Hecate0821 in https://github.com/THUDM/slime/pull/915
Add sample.removesample by @zhuzilin in https://github.com/THUDM/slime/pull/977
* add --eval-max-prompt-len by @zhuzilin in https://github.com/THUDM/slime/pull/978
Add args check for maxcontext_len by @zhuzilin in https://github.com/THUDM/slime/pull/979
Remove hard coded balanceabs_threshold by @zhuzilin in https://github.com/THUDM/slime/pull/981
Tiny fix fp8cast_bf16 not copying chat template by @fzyzcjy in https://github.com/THUDM/slime/pull/964
* Super tiny install dnsutils in dockerfile by @fzyzcjy in https://github.com/THUDM/slime/pull/965
* Super tiny sanity check checkpoint dir by @fzyzcjy in https://github.com/THUDM/slime/pull/966
Fix converthftotorch_dist OOM by @fzyzcjy in https://github.com/THUDM/slime/pull/967
* Tiny support using environment variables in addition to arguments for all scripts by @fzyzcjy in https://github.com/THUDM/slime/pull/968
* Super tiny increase default timeout sec by @fzyzcjy in https://github.com/THUDM/slime/pull/969
* Fix random port in use error even though already have free port detection by @fzyzcjy in https://github.com/THUDM/slime/pull/970
* Super tiny enable draft-weights-cpu-backup to avoid MTP acc len issue by @fzyzcjy in https://github.com/THUDM/slime/pull/971
* Add generation function for benchmarking purpose by @fzyzcjy in https://github.com/THUDM/slime/pull/972
* Support zero host or device memory waste for weight update by @fzyzcjy in https://github.com/THUDM/slime/pull/973
* Add fp8 kv cache and tis in qwen3 30b a3b script by @fzyzcjy in https://github.com/THUDM/slime/pull/974
* Add GB200, MTP, benchmark, fp8 rollout mode to glm script by @fzyzcjy in https://github.com/THUDM/slime/pull/975
* [FSDP] Add private func indicator for better usage by @PopSoda2002 in https://github.com/THUDM/slime/pull/982
* [Bugfix] Rename save model by @PopSoda2002 in https://github.com/THUDM/slime/pull/983
Fix: resolve variable shadowing bug in setupmodelandoptimizer by @fangzhensheng in https://github.com/THUDM/slime/pull/963
New Contributors
* @fangzhensheng made their first contribution in https://github.com/THUDM/slime/pull/963
Full Changelog: https://github.com/THUDM/slime/compare/v0.2.0...v0.2.0.post1
v0.2.02025-11-28
We are thrilled to announce the release of slime v0.2.0! Thanks to the incredible support and contributions from our community, slime has gained significant features and substantial performance enhancements in this version.
Major Updates
- FSDP Backend: Introduced a fully Fully Sharded Data Parallel (FSDP) based training backend for improved scalability.
- PPO Support: Added native support for Proximal Policy Optimization (PPO).
- MTP Training: Enabled training of the MTP (Multi-Token Prediction) during Reinforcement Learning.
- FP8 Full Stack: Support for both FP8 training and FP8 inference.
- Train-Inference Mismatch: Alleviate or even eliminate train-inference mismatch
- Importance Sampling: Custom interface for train-infer importance sampling (e.g., MIS).
- Routing Replay: Added Rollout Routing Replay (R3) and Routing Replay (R2).
- True On-Policy Training: Enabled strictly on-policy training with dense models on the FSDP backend.
- Performance Improvements
- Memory Optimization: CUDA Graphs offload, asystem-amem integration.
- Faster Weight Updates: Significantly accelerated FP8 weight updates.
- Python-based Router: A new slime router implemented in pure Python for accessibility.
- Fault Tolerance: Added robustness with fault tolerance for the rollout engines.
- Custom Configs: Support for passing customized configurations via --config.
- [Experimental] Checkpoint Loading: Added support for Megatron-bridge based checkpoint loading.
- New Examples
- Fully Async Training
- Multi-Agent Scenarios
- On-Policy Distillation
- Retool
What's Changed
[Doc typo] Update amdtutorial.md by @yushengsu-thu in https://github.com/THUDM/slime/pull/246
[bugfix] use fp32 for rolloutlog_probs by @zhuzilin in https://github.com/THUDM/slime/pull/245
* Complete the RayTrainGroup args string docs. by @MrAta in https://github.com/THUDM/slime/pull/248
* Update speculative decoding doc and sglang patch by @guapisolo in https://github.com/THUDM/slime/pull/250
* fix debug-rollout-only by @zyzshishui in https://github.com/THUDM/slime/pull/249
* retool in one commit by @maocheng23 in https://github.com/THUDM/slime/pull/237
* fix: modify the rotary-base of qwen-3b to 1000000 for consistency by @YuchenFan48 in https://github.com/THUDM/slime/pull/252
* update logging and fix typo by @maocheng23 in https://github.com/THUDM/slime/pull/254
* [bugfix] fix read data containing "tools" field by @Maybewuss in https://github.com/THUDM/slime/pull/255
* Revert "[bugfix] fix read data containing "tools" field" by @zhuzilin in https://github.com/THUDM/slime/pull/256
* add shell script for qwen3-32B task by @Gao016 in https://github.com/THUDM/slime/pull/253
* docs: Fix custom interface documentation errors by @GeLee-Q in https://github.com/THUDM/slime/pull/251
* [example] Add fully async example by @zhuzilin in https://github.com/THUDM/slime/pull/258
* added sphinx-based documentation by @FrankLeeeee in https://github.com/THUDM/slime/pull/262
* fixed build error for documentation by @FrankLeeeee in https://github.com/THUDM/slime/pull/263
* [bugfix] Fix bugs on multi samples from one prompt (multi-agent) by @zhuzilin in https://github.com/THUDM/slime/pull/260
* fixed sphinx configuration by @FrankLeeeee in https://github.com/THUDM/slime/pull/264
* [bugfix] fix read data containing "tools" field by @Maybewuss in https://github.com/THUDM/slime/pull/259
* add DeepWiki badge by @richardodliu in https://github.com/THUDM/slime/pull/265
* [doc] add example doc to the website by @zhuzilin in https://github.com/THUDM/slime/pull/267
* [doc] add blogs by @zhuzilin in https://github.com/THUDM/slime/pull/268
Update actorgroup.py by @zlH518 in https://github.com/THUDM/slime/pull/266
* [doc] prettify language convertion toggle by @zhuzilin in https://github.com/THUDM/slime/pull/270
* [example] add an example for multi-agent rl by @yinpeisu in https://github.com/THUDM/slime/pull/269
* [refactor] Add isort back and move global gloo to global util by @zhuzilin in https://github.com/THUDM/slime/pull/273
[refactor] remove oversampling_filter and extract some functions by @zhuzilin in https://github.com/THUDM/slime/pull/278
* [feat] init support for FSDP by @zhuzilin in https://github.com/THUDM/slime/pull/282
* Chatbot entry for Sphinx style docs by @jhinpan in https://github.com/THUDM/slime/pull/284
* Revert "Chatbot entry for Sphinx style docs" by @zhuzilin in https://github.com/THUDM/slime/pull/286
Remove getrolloutdata from actorgroup by @MrAta in https://github.com/THUDM/slime/pull/285
* Add docs logo by @jhinpan in https://github.com/THUDM/slime/pull/283
* [Hardware] AMD Dockerfile update - support up to d4a7741 (Sep 6, 2025) by @yushengsu-thu in https://github.com/THUDM/slime/pull/307
* [feat] init xtuner backend by @zhuzilin in https://github.com/THUDM/slime/pull/310
* [docker] update to sglang 0.5.2rc2 by @zhuzilin in https://github.com/THUDM/slime/pull/313
* Add model version attribute in each sample by @yitianlian in https://github.com/THUDM/slime/pull/271
[nfc] cleanup for weightversion by @zhuzilin in https://github.com/THUDM/slime/pull/314
* Add raw reward metric in fdsp backend by @yitianlian in https://github.com/THUDM/slime/pull/315
fix: check args.save when saveinterval is set. by @SanftMonster in https://github.com/THUDM/slime/pull/308
* Fix comment for --load parameter in checkpoint configuration (Quick Start Doc) by @Arist12 in https://github.com/THUDM/slime/pull/306
[refactor] bind numa and rename numgpuspernode by @zhuzilin in https://github.com/THUDM/slime/pull/316
* [xtuner] unroll TrainingWorker and TrainEngine by @zhuzilin in https://github.com/THUDM/slime/pull/322
* [xtuner] add wandb by @zhuzilin in https://github.com/THUDM/slime/pull/324
[bugfix] fix no weightversion for aborted samples by @zhuzilin in https://github.com/THUDM/slime/pull/327
* [FSDP] Verify FSDP backend availability via uv install / pip install by @Zhuohao-Li in https://github.com/THUDM/slime/pull/325
* Add FSDP extras dependency and import test (#302) by @souhil25 in https://github.com/THUDM/slime/pull/303
fix: small bug fix in the rolloutbuffer_example.sh by @rbao2018 in https://github.com/THUDM/slime/pull/328
[refactor] remove slime/backend/utils and extract slimevalidate_args by @zhuzilin in https://github.com/THUDM/slime/pull/329
* feat: auto configure megatron from hf config. by @SanftMonster in https://github.com/THUDM/slime/pull/312
* Do not read ip if env is provided by @oraluben in https://github.com/THUDM/slime/pull/337
[rmhub] fix groundtruth type error in gradeanswer_verl by @GGGGGGXY in https://github.com/THUDM/slime/pull/336
* [feat] use one global httpx.AsyncClient and remove --use-http2 by @zhuzilin in https://github.com/THUDM/slime/pull/338
* [Refactor] Merge rollout controller into rollout manager by @PopSoda2002 in https://github.com/THUDM/slime/pull/304
* add dockerfile and patch for b200 by @maocheng23 in https://github.com/THUDM/slime/pull/340
* [feat] init support for PPO by @zhuzilin in https://github.com/THUDM/slime/pull/342
* wrong expressions and typo by @ArtificialZeng in https://github.com/THUDM/slime/pull/343
* Add basic VLM data pipeline by @ppraneth in https://github.com/THUDM/slime/pull/335
* [FSDP] Add reference model support for correct KL loss computation #296 by @UbeCc in https://github.com/THUDM/slime/pull/344
* fix incorrect sft loss mask for qwen3 thinking series models. by @luppx in https://github.com/THUDM/slime/pull/330
* feature: ppo by @lilei199908 in https://github.com/THUDM/slime/pull/347
* [FIX] NVLINK detection method in scripts by @JustinTong0323 in https://github.com/THUDM/slime/pull/356
* fix lint by @JustinTong0323 in https://github.com/THUDM/slime/pull/358
* [feat] add --critic-lr and --num-critic-only-steps by @zhuzilin