[docker] update megatron-bridge and add qwen3.6 tests

.github/workflows/pr-test.yml

@@ -201,7 +201,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        info: [{"num_gpus": 8, "test_file": "test_quick_start_glm4_9B.py"}, {"num_gpus": 8, "test_file": "test_glm4.7_30B_A3B_pd_mooncake.py"}, {"num_gpus": 8, "test_file": "test_qwen3_30B_A3B.py", "use_deepep": "1", "use_fp8_rollout": "1"}, {"num_gpus": 8, "test_file": "test_qwen3_30B_A3B_pd_mooncake.py", "use_deepep": "1"}, {"enable_eval": "0", "num_gpus": 8, "test_file": "test_qwen3_30B_A3B_r3.py", "use_deepep": "1", "use_fp8_rollout": "1"}, {"enable_eval": "0", "num_gpus": 8, "test_file": "test_qwen3_30B_A3B_r3.py"}, {"num_gpus": 8, "test_file": "test_qwen3_4B_ppo.py"}, {"num_gpus": 8, "test_file": "test_qwen3_4B_ppo_train_critic_only.py"}, {"num_gpus": 8, "test_file": "test_moonlight_16B_A3B.py"}, {"enable_eval": "0", "num_gpus": 8, "test_file": "test_moonlight_16B_A3B_r3.py"}, {"num_gpus": 8, "test_file": "test_mimo_7B_mtp_only_grad.py"}, {"num_gpus": 8, "test_file": "test_qwen2.5_0.5B_debug_rollout_then_train.py"}, {"num_gpus": 8, "test_file": "test_qwen2.5_0.5B_opd_sglang.py"}]
+        info: [{"num_gpus": 8, "test_file": "test_quick_start_glm4_9B.py"}, {"num_gpus": 8, "test_file": "test_glm4.7_30B_A3B_pd_mooncake.py"}, {"num_gpus": 8, "test_file": "test_qwen3_30B_A3B.py", "use_deepep": "1", "use_fp8_rollout": "1"}, {"num_gpus": 8, "test_file": "test_qwen3.6_35B_A3B_pd_mooncake.py", "use_deepep": "1"}, {"enable_eval": "0", "num_gpus": 8, "test_file": "test_qwen3_30B_A3B_r3.py", "use_deepep": "1", "use_fp8_rollout": "1"}, {"enable_eval": "0", "num_gpus": 8, "test_file": "test_qwen3_30B_A3B_r3.py"}, {"num_gpus": 8, "test_file": "test_qwen3_4B_ppo.py"}, {"num_gpus": 8, "test_file": "test_qwen3_4B_ppo_train_critic_only.py"}, {"num_gpus": 8, "test_file": "test_moonlight_16B_A3B.py"}, {"enable_eval": "0", "num_gpus": 8, "test_file": "test_moonlight_16B_A3B_r3.py"}, {"num_gpus": 8, "test_file": "test_mimo_7B_mtp_only_grad.py"}, {"num_gpus": 8, "test_file": "test_qwen2.5_0.5B_debug_rollout_then_train.py"}, {"num_gpus": 8, "test_file": "test_qwen2.5_0.5B_opd_sglang.py"}]
     defaults:
       run:
         working-directory: ${{ github.workspace }}
@@ -515,7 +515,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        info: [{"num_gpus": 4, "test_file": "test_qwen3.5_0.8B_gsm8k_async_short.py"}, {"num_gpus": 4, "test_file": "test_qwen3.5_0.8B_gsm8k_short.py"}, {"num_gpus": 8, "test_file": "test_quick_start_glm4_9B.py"}, {"num_gpus": 8, "test_file": "test_glm4.7_30B_A3B_pd_mooncake.py"}, {"num_gpus": 8, "test_file": "test_qwen3_30B_A3B.py"}, {"num_gpus": 8, "test_file": "test_qwen3_30B_A3B_pd_mooncake.py", "use_deepep": "1"}, {"num_gpus": 8, "test_file": "test_qwen3_4B_ppo.py"}, {"num_gpus": 8, "test_file": "test_moonlight_16B_A3B.py"}, {"num_gpus": 8, "test_file": "test_mimo_7B_mtp_only_grad.py"}, {"num_gpus": 8, "test_file": "test_qwen3_0.6B_parallel_check.py"}, {"num_gpus": 8, "test_file": "test_qwen3_4B_ckpt.py"}, {"num_gpus": 8, "test_args": "--async-save", "test_file": "test_qwen3_4B_ckpt.py"}, {"num_gpus": 8, "test_file": "test_qwen2.5_0.5B_debug_rollout_then_train.py"}, {"num_gpus": 8, "test_file": "test_qwen2.5_0.5B_opd_sglang.py"}]
+        info: [{"num_gpus": 4, "test_file": "test_qwen3.5_0.8B_gsm8k_async_short.py"}, {"num_gpus": 4, "test_file": "test_qwen3.5_0.8B_gsm8k_short.py"}, {"num_gpus": 8, "test_file": "test_quick_start_glm4_9B.py"}, {"num_gpus": 8, "test_file": "test_glm4.7_30B_A3B_pd_mooncake.py"}, {"num_gpus": 8, "test_file": "test_qwen3_30B_A3B.py"}, {"num_gpus": 8, "test_file": "test_qwen3.6_35B_A3B_pd_mooncake.py", "use_deepep": "1"}, {"num_gpus": 8, "test_file": "test_qwen3_4B_ppo.py"}, {"num_gpus": 8, "test_file": "test_moonlight_16B_A3B.py"}, {"num_gpus": 8, "test_file": "test_mimo_7B_mtp_only_grad.py"}, {"num_gpus": 8, "test_file": "test_qwen3_0.6B_parallel_check.py"}, {"num_gpus": 8, "test_file": "test_qwen3_4B_ckpt.py"}, {"num_gpus": 8, "test_args": "--async-save", "test_file": "test_qwen3_4B_ckpt.py"}, {"num_gpus": 8, "test_file": "test_qwen2.5_0.5B_debug_rollout_then_train.py"}, {"num_gpus": 8, "test_file": "test_qwen2.5_0.5B_opd_sglang.py"}]
     defaults:
       run:
         working-directory: ${{ github.workspace }}

.github/workflows/pr-test.yml.j2

@@ -22,7 +22,7 @@
         {'test_file': 'test_quick_start_glm4_9B.py', 'num_gpus': 8},
         {'test_file': 'test_glm4.7_30B_A3B_pd_mooncake.py', 'num_gpus': 8},
         {'test_file': 'test_qwen3_30B_A3B.py', 'num_gpus': 8, 'use_deepep': '1', 'use_fp8_rollout': '1'},
-        {'test_file': 'test_qwen3_30B_A3B_pd_mooncake.py', 'num_gpus': 8, 'use_deepep': '1'},
+        {'test_file': 'test_qwen3.6_35B_A3B_pd_mooncake.py', 'num_gpus': 8, 'use_deepep': '1'},
         {'test_file': 'test_qwen3_30B_A3B_r3.py', 'num_gpus': 8, 'use_deepep': '1', 'use_fp8_rollout': '1', 'enable_eval': '0'},
         {'test_file': 'test_qwen3_30B_A3B_r3.py', 'num_gpus': 8, 'enable_eval': '0'},
         {'test_file': 'test_qwen3_4B_ppo.py', 'num_gpus': 8},
@@ -69,7 +69,7 @@
         {'test_file': 'test_quick_start_glm4_9B.py', 'num_gpus': 8},
         {'test_file': 'test_glm4.7_30B_A3B_pd_mooncake.py', 'num_gpus': 8},
         {'test_file': 'test_qwen3_30B_A3B.py', 'num_gpus': 8},
-        {'test_file': 'test_qwen3_30B_A3B_pd_mooncake.py', 'num_gpus': 8, 'use_deepep': '1'},
+        {'test_file': 'test_qwen3.6_35B_A3B_pd_mooncake.py', 'num_gpus': 8, 'use_deepep': '1'},
         {'test_file': 'test_qwen3_4B_ppo.py', 'num_gpus': 8},
         {'test_file': 'test_moonlight_16B_A3B.py', 'num_gpus': 8},
         {'test_file': 'test_mimo_7B_mtp_only_grad.py', 'num_gpus': 8},

README.md

@@ -11,7 +11,7 @@
 2.  **Flexible Data Generation**: Enables arbitrary training data generation workflows through custom data generation interfaces and server-based engines.
 
 slime is the RL-framework behind [GLM-5.1](https://z.ai/blog/glm-5.1), [GLM-5](https://z.ai/blog/glm-5), [GLM-4.7](https://z.ai/blog/glm-4.7), [GLM-4.6](https://z.ai/blog/glm-4.6), [GLM-4.5](https://z.ai/blog/glm-4.5) and apart from models from Z.ai, we also supports the following models:
-- Qwen3 series (Qwen3Next, Qwen3MoE, Qwen3), Qwen2.5 series;
+- Qwen series (Qwen3.6, Qwen3.5, Qwen3Next, Qwen3MoE, Qwen3, Qwen2.5);
 - DeepSeek V3 series (DeepSeek V3, V3.1, DeepSeek R1);
 - Llama 3.
 

README_zh.md

@@ -11,7 +11,7 @@
 2. **灵活的数据生成**：通过自定义数据生成接口以及 server based engine，实现任意的数据训练数据生成流程。
 
 slime 是 [GLM-5.1](https://z.ai/blog/glm-5.1)、[GLM-5](https://z.ai/blog/glm-5)、[GLM-4.7](https://z.ai/blog/glm-4.7)、[GLM-4.6](https://z.ai/blog/glm-4.6)、[GLM-4.5](https://z.ai/blog/glm-4.5) 背后的 RL 训练框架，除此之外，slime 还支持:
-- Qwen3 系列 (Qwen3Next, Qwen3MoE, Qwen3), Qwen2.5 系列；
+- Qwen 系列 (Qwen3.6、Qwen3.5、Qwen3Next、Qwen3MoE、Qwen3、Qwen2.5)；
 - DeepSeek V3 系列 (DeepSeek V3, V3.1, DeepSeek R1)；
 - Llama 3。
 

docker/Dockerfile

@@ -54,8 +54,8 @@ RUN git clone https://github.com/NVIDIA/Megatron-LM.git --recursive && \
     cd Megatron-LM && git checkout ${MEGATRON_COMMIT} && \
     pip install -e .
 
-RUN pip install git+https://github.com/fzyzcjy/torch_memory_saver.git@dc6876905830430b5054325fa4211ff302169c6b --no-cache-dir --force-reinstall
-RUN pip install git+https://github.com/fzyzcjy/Megatron-Bridge.git@dev_rl --no-build-isolation
+RUN pip install git+https://github.com/fzyzcjy/torch_memory_saver.git@d64a639 --no-cache-dir --force-reinstall
+RUN pip install git+https://github.com/radixark/Megatron-Bridge.git@bridge --no-deps --no-build-isolation
 RUN pip install nvidia-modelopt[torch]>=0.37.0 --no-build-isolation
 
 # This patch from masahi will be included in later Triton releases

docker/version.txt

@@ -1 +1 @@
-nightly-dev-20260430a
+nightly-dev-20260430b

docs/en/developer_guide/ci.md

@@ -6,7 +6,7 @@ slime uses GitHub Actions for CI. Tests are triggered by **PR labels** — addin
 
 The workflow is defined in `.github/workflows/pr-test.yml` (auto-generated from `pr-test.yml.j2`). Each CI job:
 
-1. Runs on a self-hosted GPU runner inside a Docker container (`slimerl/slime:latest`).
+1. Runs on a self-hosted GPU runner via `docker run`; most tests use `slimerl/slime:latest`, while image validation uses `slimerl/slime-test:latest`.
 2. Installs slime with `pip install -e . --no-deps`.
 3. Acquires the required GPUs via `tests/ci/gpu_lock_exec.py --count <num_gpus>`.
 4. Executes the test file: `python <test_path>.py` or `python tests/<test_file>.py`, depending on whether the test lives under `tests/` or a subdirectory such as `tests/plugin_contracts/`.
@@ -56,7 +56,7 @@ Since this includes every test, it consumes significant GPU time — use it spar
 This is the primary label for validating Megatron-backend changes. It covers:
 
 - Dense models: GLM4-9B, Qwen3-4B (PPO)
-- MoE models: Qwen3-30B-A3B (with/without DeepEP + FP8), Moonlight-16B-A3B
+- MoE models: Qwen3-30B-A3B (with DeepEP + FP8), Qwen3.6-35B-A3B PD + Mooncake, Moonlight-16B-A3B
 - Specialized: MiMo-7B MTP, Qwen2.5-0.5B debug rollout-then-train, OPD with sglang teacher
 
 All tests use 8 GPUs. If you are modifying Megatron training logic, loss computation, or checkpoint conversion, this is the label to use.

docs/en/get_started/quick_start.md

@@ -71,7 +71,7 @@ hf download --repo-type dataset zhuzilin/aime-2024 \
 
 When using Megatron as the training backend, you need to first convert Hugging Face format model weights to Megatron `torch_dist` format.
 
-First, load the configuration file of the target model. The `slime/scripts/models` directory contains configuration files for supported models. You need to `source` the corresponding model script to load the configuration parameters into the current environment. Here we use GLM4-9B model as an example, and it's similar for Qwen3-4B, GLM-4.7-Flash, Qwen3-30B-A3B, etc.
+First, load the configuration file of the target model. The `slime/scripts/models` directory contains configuration files for supported models. You need to `source` the corresponding model script to load the configuration parameters into the current environment. Here we use GLM4-9B model as an example, and it's similar for Qwen3-4B, Qwen3.5, Qwen3.6, GLM-4.7-Flash, Qwen3-30B-A3B, etc.
 
 ```bash
 cd /root/slime

docs/zh/developer_guide/ci.md

@@ -6,7 +6,7 @@ slime 使用 GitHub Actions 进行 CI。测试通过 **PR label** 触发——
 
 工作流定义在 `.github/workflows/pr-test.yml`（由 `pr-test.yml.j2` 自动生成）。每个 CI 任务会：
 
-1. 在自托管 GPU runner 上以 Docker 容器（`slimerl/slime:latest`）运行。
+1. 在自托管 GPU runner 上通过 `docker run` 运行；大多数测试使用 `slimerl/slime:latest`，镜像验证使用 `slimerl/slime-test:latest`。
 2. 通过 `pip install -e . --no-deps` 安装 slime。
 3. 通过 `tests/ci/gpu_lock_exec.py --count <num_gpus>` 获取所需数量的 GPU。
 4. 执行测试文件：`python <test_path>.py` 或 `python tests/<test_file>.py`。如果测试位于 `tests/plugin_contracts/` 这样的子目录，CI 也会自动处理。
@@ -56,7 +56,7 @@ slime 使用 GitHub Actions 进行 CI。测试通过 **PR label** 触发——
 这是验证 Megatron 后端改动的主要 label，覆盖：
 
 - Dense 模型：GLM4-9B、Qwen3-4B（PPO）
-- MoE 模型：Qwen3-30B-A3B（有/无 DeepEP + FP8）、Moonlight-16B-A3B
+- MoE 模型：Qwen3-30B-A3B（DeepEP + FP8）、Qwen3.6-35B-A3B PD + Mooncake、Moonlight-16B-A3B
 - 特殊场景：MiMo-7B MTP、Qwen2.5-0.5B debug rollout-then-train、OPD（sglang teacher 模式）
 
 所有测试使用 8 张 GPU。如果你正在修改 Megatron 训练逻辑、loss 计算或 checkpoint 转换，应该使用这个 label。

docs/zh/get_started/quick_start.md

@@ -70,7 +70,7 @@ hf download --repo-type dataset zhuzilin/aime-2024 \
 
 当使用 Megatron 作为训练后端时，需要先将 Hugging Face 格式的模型权重转换为 Megatron `torch_dist` 格式。
 
-首先，加载目标模型的配置文件。`slime/scripts/models` 目录下包含了支持模型的配置文件。需要 `source` 对应模型的脚本，将配置参数加载到当前环境中。此处我们以 GLM4-9B 模型为例子，对于 Qwen3-4B，GLM-4.7-Flash，Qwen3-30B-A3B，是类似的。
+首先，加载目标模型的配置文件。`slime/scripts/models` 目录下包含了支持模型的配置文件。需要 `source` 对应模型的脚本，将配置参数加载到当前环境中。此处我们以 GLM4-9B 模型为例子，对于 Qwen3-4B、Qwen3.5、Qwen3.6、GLM-4.7-Flash、Qwen3-30B-A3B，是类似的。
 
 ```bash
 cd /root/slime

slime/ray/rollout.py

@@ -123,7 +123,6 @@ def start_engines(self, port_cursors: dict[int, int] | None = None) -> tuple[lis
                     "SLIME_ENABLE_PROFILING": "true",
                 }.items()
             }
-
             rollout_engine = RolloutRayActor.options(
                 num_cpus=num_cpus,
                 num_gpus=num_gpus,

tests/test_glm4.7_30B_A3B_pd_mooncake.py

@@ -122,6 +122,10 @@ def execute():
         "--sglang-cuda-graph-max-bs 8 "
         "--sglang-max-running-requests 16 "
         "--sglang-disaggregation-transfer-backend mooncake "
+        "--sglang-speculative-algorithm EAGLE "
+        "--sglang-speculative-num-steps 3 "
+        "--sglang-speculative-eagle-topk 1 "
+        "--sglang-speculative-num-draft-tokens 4 "
         "--sglang-watchdog-timeout 1200 "
         "--sglang-router-request-timeout-secs 1200 "
         "--sglang-enable-metrics "

tests/test_qwen3_30B_A3B_pd_mooncake.py → tests/test_qwen3.6_35B_A3B_pd_mooncake.py

@@ -4,9 +4,10 @@
 import slime.utils.external_utils.command_utils as U
 
 
-MODEL_NAME = "Qwen3-30B-A3B"
-MODEL_TYPE = "qwen3-30B-A3B"
+MODEL_NAME = "Qwen3.6-35B-A3B"
+MODEL_TYPE = "qwen3.5-35B-A3B"
 NUM_GPUS = 8
+TORCH_DIST_CKPT = f"/root/models/{MODEL_NAME}_torch_dist"
 
 
 def prepare():
@@ -18,20 +19,21 @@ def prepare():
         model_name=MODEL_NAME,
         megatron_model_type=MODEL_TYPE,
         num_gpus_per_node=NUM_GPUS,
+        dir_dst="/root/models",
     )
 
 
 def execute():
     debug_data_path = os.environ.get("DEBUG_ROLLOUT_DATA") or tempfile.mktemp(
-        prefix="qwen3_30b_a3b_pd_rollout_", suffix=".pt"
+        prefix="qwen3_6_35b_a3b_pd_rollout_", suffix=".pt"
     )
     try:
         os.remove(debug_data_path)
     except FileNotFoundError:
         pass
     print(f"Saving debug rollout data to {debug_data_path}")
 
-    ckpt_args = f"--hf-checkpoint /root/models/{MODEL_NAME} " f"--ref-load /root/{MODEL_NAME}_torch_dist "
+    ckpt_args = f"--hf-checkpoint /root/models/{MODEL_NAME} " f"--ref-load {TORCH_DIST_CKPT} "
 
     rollout_args = (
         "--prompt-data /root/datasets/dapo-math-17k/dapo-math-17k.jsonl "
@@ -57,7 +59,7 @@ def execute():
     )
 
     perf_args = (
-        "--tensor-model-parallel-size 4 "
+        "--tensor-model-parallel-size 2 "
         "--sequence-parallel "
         "--pipeline-model-parallel-size 1 "
         "--context-parallel-size 2 "
@@ -100,6 +102,12 @@ def execute():
         "--sglang-cuda-graph-bs 1 2 4 8 16 24 32 "
         "--sglang-max-running-requests 512 "
         "--prefill-num-servers 1 "
+        "--sglang-disaggregation-transfer-backend mooncake "
+        "--sglang-speculative-algorithm EAGLE "
+        "--sglang-speculative-num-steps 3 "
+        "--sglang-speculative-eagle-topk 1 "
+        "--sglang-speculative-num-draft-tokens 4 "
+        "--sglang-mamba-scheduler-strategy extra_buffer "
         "--sglang-enable-metrics "
     )