chore(sync): publish SFT, diffusion and agentic updates

.gitleaks.toml

@@ -70,13 +70,15 @@ tags = ["internal", "token"]
 [[rules]]
 id = "private-ip-10-range"
 description = "Detects hardcoded private IP addresses in 10.0.0.0/8 range (RFC 1918)"
-regex = '''\b10\.\d{1,3}\.\d{1,3}\.\d{1,3}\b'''
+regex = '''\b10\.\d{1,3}\.\d{1,3}\.\d{1,3}\b(?:/\d{1,2}\b)?'''
 tags = ["internal", "ip", "private"]
   [rules.allowlist]
   regexTarget = "match"
   regexes = [
     # 10.0.0.1 is a conventional placeholder IP in examples and tests.
     '''^10\.0\.0\.1$''',
+    # The full RFC 1918 network is public, not a deployment-specific endpoint.
+    '''^10\.0\.0\.0/8$''',
   ]
   paths = ['''scripts/ci/benchmark\.sh$''']
 

.pre-commit-config.yaml

@@ -50,7 +50,7 @@ repos:
     hooks:
       - id: docformatter
         name: docformatter
-        entry: docformatter --in-place --wrap-descriptions 79
+        entry: python .pre-commit-hooks/docformatter_compat.py --in-place --wrap-descriptions 79
         language: python
         types: [python]
         additional_dependencies: ["docformatter==1.3.1"]

.pre-commit-hooks/docformatter_compat.py

@@ -0,0 +1,38 @@
+#!/usr/bin/env python3
+# Copyright (c) 2026 Relax Authors. All Rights Reserved.
+
+"""Run docformatter with correct multiline string token positions."""
+
+from __future__ import annotations
+
+import tokenize
+from collections.abc import Callable, Iterator
+
+
+_original_generate_tokens = tokenize.generate_tokens
+
+
+def _generate_tokens(readline: Callable[[], str]) -> Iterator[tokenize.TokenInfo]:
+    for token in _original_generate_tokens(readline):
+        if token.type == tokenize.STRING and token.start[0] != token.end[0]:
+            # Python 3.12.0 can undercount this column after non-ASCII text.
+            # untokenize otherwise copies the apparent gap after the string,
+            # duplicating source characters, including its closing quotes.
+            end_column = len(token.string.rsplit("\n", 1)[-1])
+            token = token._replace(end=(token.end[0], end_column))
+        yield token
+
+
+def main() -> int:
+    import docformatter
+
+    original = tokenize.generate_tokens
+    tokenize.generate_tokens = _generate_tokens
+    try:
+        return docformatter.main()
+    finally:
+        tokenize.generate_tokens = original
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

.pre-commit-hooks/gitleaks_tracked.py

@@ -90,6 +90,8 @@ def main(argv: Sequence[str] | None = None) -> int:
             stdout=subprocess.PIPE,
             stderr=subprocess.PIPE,
             text=True,
+            encoding="utf-8",
+            errors="replace",
         )
 
         _relay(result.stdout, sys.stdout, snapshot_root)

AGENTS.md

@@ -8,20 +8,36 @@ Relax 是一个基于 Ray Serve 的大模型强化学习训练框架，支持 Me
 
 ```
 relax/                   核心框架
+├── agentic/             Agentic 层 — Session、pipeline、agent process 与训练导出
 ├── core/                编排层 — 训练循环、服务基类、全局注册表
 ├── components/          组件层 — RL 服务组件（Ray Serve Deployment）
 ├── engine/              引擎层 — Rollout 数据生成、奖励计算、请求路由
 ├── backends/            后端层 — Megatron 训练后端、SGLang 推理引擎
 ├── distributed/         分布式层 — Ray 集群管理、分布式 Checkpoint
 ├── entrypoints/         入口层 — 训练入口脚本
+├── models/              模型层 — 模型专属实现与注册
 └── utils/               基础设施 — 工具函数、指标监控、多模态处理
-tests/                   测试（镜像 relax/ 层级）
+tests/                   模块测试与集成回归测试
 ├── backends/megatron/   Megatron 后端测试（权重转换等）
+├── backends/sglang/     SGLang 后端测试
+├── components/          RL 服务组件测试
+├── core/                Controller、Service 与注册表测试
+├── data/                数据处理与 SFT 数据测试
+├── distributed/checkpoint_service/  分布式 Checkpoint 测试
 ├── distributed/ray/     分布式 / Ray 测试（弹性伸缩等）
 ├── engine/rewards/      奖励函数测试
 ├── engine/rollout/      Rollout 引擎测试（预取、数据源等）
-└── utils/               工具函数测试（HTTP、指标、流式数据集等）
-transfer_queue/          分布式数据传输队列
+├── engine/sft/          SFT 引擎测试
+├── entrypoints/         入口行为测试
+├── examples/            示例级回归测试
+├── integration/         跨模块集成测试
+├── models/              模型专属测试
+├── tools/               工具与辅助脚本测试
+├── utils/               工具函数测试（HTTP、指标、流式数据集等）
+└── test_agentic_rollout.py  Agentic runtime 与 Session 测试
+docs/                   中英文用户文档
+skills/                 仓库开发与运维工作流
+docker/                 训练镜像与依赖 patch
 examples/                用户级示例（deepeyes、OPD 等）
 scripts/                 训练启动脚本 & 模型配置
 configs/env.yaml         运行时环境配置

docker/Dockerfile

@@ -1,7 +1,7 @@
 ARG HTTP_PROXY
 ARG HTTPS_PROXY
 ARG NO_PROXY
-ARG BASE_IMAGE=mirror.ccs.tencentyun.com/lmsysorg/sglang:v0.5.15.post1-cu129
+ARG BASE_IMAGE=mirror.ccs.tencentyun.com/lmsysorg/sglang:v0.5.17-cu129
 ARG TRAIN_IMAGE=train
 FROM ${BASE_IMAGE} as base
 
@@ -38,9 +38,24 @@ RUN pip install nvidia-cudnn-cu12==9.16.0.29
 FROM base as train
 
 RUN MAX_JOBS=64 pip -v install flash-attn==2.7.4.post1 --no-build-isolation --no-cache-dir && \
-    pip install --no-cache-dir flash-linear-attention==0.4.1 && \
+    pip install --no-cache-dir flash-linear-attention==0.4.2 && \
     pip install --no-cache-dir tilelang -f https://tile-ai.github.io/whl/nightly/cu128/
 
+# DeepSeek-V4 DSA deps, both listed under mcore's `no_pypi_wheels` so built from pinned source:
+#   - fast-hadamard-transform: required by the Lightning Indexer (dsa.rotate_activation asserts on it)
+#   - FlashMLA: fused DSA sparse attention; without it the CSA backward is quadratic in seq len
+#     (mcore only allows the fusion on SM90 when dsa_indexer_loss_coeff == 0)
+ARG FLASH_MLA_COMMIT=b7643bd54521f563b839b98289b5cd048c062ba2
+RUN MAX_JOBS=64 pip install --no-cache-dir --no-build-isolation --no-deps \
+        git+https://github.com/Dao-AILab/fast-hadamard-transform.git@f134af63deb2df17e1171a9ec1ea4a7d8604d5ca && \
+    git clone --recurse-submodules https://github.com/deepseek-ai/FlashMLA.git /opt/FlashMLA && \
+    cd /opt/FlashMLA && git checkout ${FLASH_MLA_COMMIT} && \
+    git submodule update --init --recursive && \
+    MAX_JOBS=64 pip install --no-cache-dir --no-build-isolation --no-deps . && \
+    cd / && python -c "import flash_mla; assert hasattr(flash_mla, 'flash_mla_sparse_fwd'), \
+        'flash_mla built but flash_mla_sparse_fwd missing'" && \
+    rm -rf /opt/FlashMLA
+
 # FA3 (Hopper flash-attention), built from source. This commit's _flash_attn_forward carries
 # window_size_left/right to match TE 2.14.1 (docs/draft/sglang-0.5.12-upgrade-plan.md §8.4).
 # The `cp` exposes flash_attn_3.flash_attn_interface, which is what TE imports. BUT this commit's
@@ -70,6 +85,7 @@ RUN MAX_JOBS=64 \
 RUN pip -v install --no-cache-dir --no-build-isolation "transformer_engine[pytorch]==2.14.1" && \
     TMS_CUDA_MAJOR=$(python -c 'import torch; print(torch.version.cuda.split(".")[0])') pip install git+https://github.com/redai-studio/torch_memory_saver.git@afc13785c50119048e2dd8ac497cc9e29ec75bd4 --no-cache-dir --force-reinstall && \
     pip install nvidia-modelopt[torch]>=0.37.0 --no-build-isolation --no-cache-dir && \
+    pip install megatron-energon fsspec==2024.3.1 --no-cache-dir&& \
     pip install "numpy<2" nvidia-cudnn-cu12==9.16.0.29 --no-cache-dir && \
     NVCC_APPEND_FLAGS="--threads 32" \
     pip -v install --disable-pip-version-check --no-cache-dir \
@@ -93,7 +109,7 @@ WORKDIR /root
 ARG PATCH_VERSION=latest
 ARG ENABLE_SGLANG_PATCH=1
 
-ARG MEGATRON_BRIDGE_COMMIT=2faedbf6fe3c422835a44b2b360cadcb2a116a54
+ARG MEGATRON_BRIDGE_COMMIT=af17edf52c514c58c79cd291b04a9b42bc5c57f3
 ENV MEGATRON_BRIDGE_COMMIT=${MEGATRON_BRIDGE_COMMIT} \
     PYTHONPATH=/root/Megatron-LM/
 

docker/patch/latest/megatron.patch

@@ -1 +1 @@
-../megatron/20260506-85bced0ae.patch
+../megatron/20260728-0e6ac576f.patch

docker/patch/latest/sglang.patch

@@ -1 +1 @@
-../sglang/v0.5.15.post1.patch
+../sglang/v0.5.17.patch