Fix GRPO weight-sync hangs and HCCL resource exhaustion on NPU

docs/source_en/Components/Checkpoint Engine/HCCLCheckpointEngine.md

@@ -25,4 +25,11 @@ HCCLCheckpointEngine is specifically designed for Ascend NPU environments:
 - Synchronizing model weights between NPUs
 - Large-scale NPU cluster deployment
 
+## Environment Variables
+
+- `TWINKLE_CKPT_HCCL_META_TIMEOUT_S`:
+  Controls the timeout (in seconds) for the HCCL CheckpointEngine
+  metadata handshake channel (ZMQ REQ/REP).
+  Default is `300`. This value should be an integer greater than `0`.
+
 > In Ascend NPU environments, HCCLCheckpointEngine provides performance comparable to NCCL.

docs/source_en/Components/Sampler/vLLMSampler.md

@@ -69,4 +69,11 @@ sampler = vLLMSampler(
 response = sampler.sample(trajectories, sampling_params=params)
 ```
 
+## Environment Variables
+
+- `TWINKLE_VLLM_IPC_TIMEOUT_S`:
+  Controls the timeout (in seconds) for the IPC channel (ZMQ REQ/REP)
+  between `vLLMSampler` and the vLLM worker extension.
+  Default is `300`. This value must be greater than `0`.
+
 > In RLHF training, vLLMSampler is typically separated from the Actor model, using different hardware resources to avoid interference between inference and training.

docs/source_zh/组件/检查点引擎/HCCLCheckpointEngine.md

@@ -25,4 +25,10 @@ HCCLCheckpointEngine 专门用于昇腾 NPU 环境:
 - 需要在 NPU 间同步模型权重
 - 大规模 NPU 集群部署
 
+## 环境变量
+
+- `TWINKLE_CKPT_HCCL_META_TIMEOUT_S`:
+  控制 HCCL CheckpointEngine 元数据握手通道（ZMQ REQ/REP）的超时时间（秒）。
+  默认值为 `300`。该值应设置为大于 `0` 的整数。
+
 > 在昇腾 NPU 环境中,HCCLCheckpointEngine 提供了与 NCCL 相当的性能。

docs/source_zh/组件/采样器/vLLMSampler.md

@@ -69,4 +69,10 @@ sampler = vLLMSampler(
 response = sampler.sample(trajectories, sampling_params=params)
 ```
 
+## 环境变量
+
+- `TWINKLE_VLLM_IPC_TIMEOUT_S`:
+  控制 `vLLMSampler` 与 vLLM worker extension 之间 IPC 通道（ZMQ REQ/REP）的超时时间（秒）。
+  默认值为 `300`。该值必须大于 `0`。
+
 > vLLMSampler 在 RLHF 训练中通常与 Actor 模型分离,使用不同的硬件资源,避免推理和训练相互干扰。