From f78df6de2d8faf647d7a8c7f6a32e4f0394bdc71 Mon Sep 17 00:00:00 2001 From: ooooo <3164076421@qq.com> Date: Thu, 10 Sep 2026 23:36:25 +0800 Subject: [PATCH] docs(links): fix Relax repository URLs MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit # 📝 Documentation ## Fix repository links - Point all six GenRM documentation links to redai-studio/Relax. - Update the English and Chinese pages together while preserving the target paths. --- docs/en/examples/generative-reward-model.md | 6 +++--- docs/zh/examples/generative-reward-model.md | 6 +++--- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/docs/en/examples/generative-reward-model.md b/docs/en/examples/generative-reward-model.md index ad65b4930..7ac5d17d9 100644 --- a/docs/en/examples/generative-reward-model.md +++ b/docs/en/examples/generative-reward-model.md @@ -98,7 +98,7 @@ Under `--colocate`, how Rollout and GenRM cohabit the shared bundles has three f └─────────────────────────────────────────────────┘ ``` -All three colocate sub-modes reclaim every GPU for the Actor during training. Rollout produces candidate responses and (for Split and Shared / Co-resident) sends each one over HTTP to GenRM inline. In Shared / Defer-swap the HTTP call is batched once per rollout step from a userland `custom_reward_post_process` function; see [`examples/generate_reward_model/README.md`](https://github.com/xhs-tech/Relax/blob/main/examples/generate_reward_model/README.md) for the split-vs-defer trade-off matrix. +All three colocate sub-modes reclaim every GPU for the Actor during training. Rollout produces candidate responses and (for Split and Shared / Co-resident) sends each one over HTTP to GenRM inline. In Shared / Defer-swap the HTTP call is batched once per rollout step from a userland `custom_reward_post_process` function; see [`examples/generate_reward_model/README.md`](https://github.com/redai-studio/Relax/blob/main/examples/generate_reward_model/README.md) for the split-vs-defer trade-off matrix. ## Scripts @@ -106,7 +106,7 @@ All three colocate sub-modes reclaim every GPU for the Actor during training. Ro | :---------------------------------------------- | :------------------------- | :----------------------------------------------------------------------------- | | `run-qwen3-4B-8xgpu-colocated.sh` | Split (small GenRM) | Qwen3-4B policy + small GenRM on 8 GPU; disjoint bundles, inline reward | | `run-qwen35-35B-A3B-16xgpu-genrm-397B-split.sh` | Split (large GenRM) | 35B-A3B policy + 397B FP8 GenRM on 16 GPU; 8+8 disjoint shards, inline reward | -| `run-qwen35-35B-A3B-16xgpu-genrm-397B-defer.sh` | Shared / Defer-swap | 35B-A3B policy + 397B FP8 GenRM on 16 GPU; shared bundles, two-phase sleep-wake swap, batched reward via [`post_process_genrm_swap.py`](https://github.com/xhs-tech/Relax/blob/main/examples/generate_reward_model/post_process_genrm_swap.py) | +| `run-qwen35-35B-A3B-16xgpu-genrm-397B-defer.sh` | Shared / Defer-swap | 35B-A3B policy + 397B FP8 GenRM on 16 GPU; shared bundles, two-phase sleep-wake swap, batched reward via [`post_process_genrm_swap.py`](https://github.com/redai-studio/Relax/blob/main/examples/generate_reward_model/post_process_genrm_swap.py) | | `run-qwen3-4B-8xgpu-async.sh` | (Fully Async) | Independent GPU pools per role; rollout & training fully overlapped | ### Resource Layout @@ -268,7 +268,7 @@ On `--colocate` with GenRM, the GPU layout picks Split vs Shared automatically: | `rollout_num_gpus == genrm_num_gpus == actor_total` | **Shared** (same bundles) | | Anything else | Rejected at startup with a clear error | -Within Shared, the default is **Co-resident** (both engines held via `mem_fraction_static` split). Adding `--rm-type dummy` + `--defer-reward-to-post-process` + `--custom-reward-post-process-path` switches it to **Defer-swap** — sequenced sleep-wake, one engine holds full memory at a time. See the [example README](https://github.com/xhs-tech/Relax/blob/main/examples/generate_reward_model/README.md) for when to prefer defer-swap. +Within Shared, the default is **Co-resident** (both engines held via `mem_fraction_static` split). Adding `--rm-type dummy` + `--defer-reward-to-post-process` + `--custom-reward-post-process-path` switches it to **Defer-swap** — sequenced sleep-wake, one engine holds full memory at a time. See the [example README](https://github.com/redai-studio/Relax/blob/main/examples/generate_reward_model/README.md) for when to prefer defer-swap. ::: ::: warning Set `mem_fraction_static` in Shared / Co-resident diff --git a/docs/zh/examples/generative-reward-model.md b/docs/zh/examples/generative-reward-model.md index cb07887a2..875f45c57 100644 --- a/docs/zh/examples/generative-reward-model.md +++ b/docs/zh/examples/generative-reward-model.md @@ -98,7 +98,7 @@ Relax 中 GenRM 有两种顶层部署模式： └─────────────────────────────────────────────────┘ ``` -三种 colocate 子模式在训练阶段都把全部 GPU 归还给 Actor。Split 与 Shared / Co-resident 走 inline reward：Rollout 每生成一个候选就通过 HTTP 单发给 GenRM。Shared / Defer-swap 把 HTTP 调用改成每轮 rollout 后由 userland `custom_reward_post_process` 一次性批量发出；split 与 defer-swap 的完整取舍见 [`examples/generate_reward_model/README.md`](https://github.com/xhs-tech/Relax/blob/main/examples/generate_reward_model/README.md)。 +三种 colocate 子模式在训练阶段都把全部 GPU 归还给 Actor。Split 与 Shared / Co-resident 走 inline reward：Rollout 每生成一个候选就通过 HTTP 单发给 GenRM。Shared / Defer-swap 把 HTTP 调用改成每轮 rollout 后由 userland `custom_reward_post_process` 一次性批量发出；split 与 defer-swap 的完整取舍见 [`examples/generate_reward_model/README.md`](https://github.com/redai-studio/Relax/blob/main/examples/generate_reward_model/README.md)。 ## 脚本 @@ -106,7 +106,7 @@ Relax 中 GenRM 有两种顶层部署模式： | :---------------------------------------------- | :---------------------- | :-------------------------------------------------------------------------------- | | `run-qwen3-4B-8xgpu-colocated.sh` | Split（小 GenRM） | Qwen3-4B policy + 小 GenRM 共 8 GPU；不相交 bundle，inline reward | | `run-qwen35-35B-A3B-16xgpu-genrm-397B-split.sh` | Split（大 GenRM） | 35B-A3B policy + 397B FP8 GenRM 共 16 GPU；8+8 不相交分片，inline reward | -| `run-qwen35-35B-A3B-16xgpu-genrm-397B-defer.sh` | Shared / Defer-swap | 35B-A3B policy + 397B FP8 GenRM 共 16 GPU；共享 bundle，两阶段 sleep-wake 切换，批量 reward，实现见 [`post_process_genrm_swap.py`](https://github.com/xhs-tech/Relax/blob/main/examples/generate_reward_model/post_process_genrm_swap.py) | +| `run-qwen35-35B-A3B-16xgpu-genrm-397B-defer.sh` | Shared / Defer-swap | 35B-A3B policy + 397B FP8 GenRM 共 16 GPU；共享 bundle，两阶段 sleep-wake 切换，批量 reward，实现见 [`post_process_genrm_swap.py`](https://github.com/redai-studio/Relax/blob/main/examples/generate_reward_model/post_process_genrm_swap.py) | | `run-qwen3-4B-8xgpu-async.sh` | （Fully Async） | 每个角色独占 GPU 池；rollout 与训练完全并行 | ### 资源分配 @@ -267,7 +267,7 @@ python3 relax/entrypoints/train.py \ | `rollout_num_gpus == genrm_num_gpus == actor_total` | **Shared**（同一批 bundle） | | 其他 | 启动时报错拒绝 | -Shared 内部默认是 **Co-resident**（两个引擎按 `mem_fraction_static` 同时驻留）。再加上 `--rm-type dummy` + `--defer-reward-to-post-process` + `--custom-reward-post-process-path` 就切成 **Defer-swap**——sleep-wake 串行，每次只有一个引擎占显存。何时优先 defer-swap 见 [示例 README](https://github.com/xhs-tech/Relax/blob/main/examples/generate_reward_model/README.md)。 +Shared 内部默认是 **Co-resident**（两个引擎按 `mem_fraction_static` 同时驻留）。再加上 `--rm-type dummy` + `--defer-reward-to-post-process` + `--custom-reward-post-process-path` 就切成 **Defer-swap**——sleep-wake 串行，每次只有一个引擎占显存。何时优先 defer-swap 见 [示例 README](https://github.com/redai-studio/Relax/blob/main/examples/generate_reward_model/README.md)。 ::: ::: warning Shared / Co-resident 必须设置 `mem_fraction_static`