修改后的函数：计算模型预测速度的对数概率。 保持输入参数不变。

保存 EMA 参数到文件

Load lines from a file. First tries to load from `path` directly, and if that doesn't exist, searches the `flow_grpo/assets` directory for a

从文件加载 EMA 参数

跨 GPU 生成一个共享的 timestep，均匀分布在 [tmin, 1]。 Args: accelerator: Accelerator 对象 Returns: timestep: 标量 tensor，所有 GPU 上

为每个 group 生成共享的噪声，跨 GPU 同步。 Args: x0: [batch_size, C, H, W] 用于获取形状 group_info: 预计算的 group 信息，包含 inverse_indices, num_gro

Return every nn.Linear module path in `model`.

Give position of A relative to B, factoring in object dimensions

Singleton-cached HPSv3 scorer (Qwen2-VL-7B based, expensive to construct). `device=None` defers GPU placement to the first `__call__` (loads on C

Calculate the proportion of unique prompts whose reward standard deviation is zero. Args: prompts: List of prompts. gath

AllReduce实现的梯度等价版本

Predict the sample from the previous timestep by reversing the SDE. This function propagates the flow process from the learned model outputs

Evaluate given image using detected objects on the global metadata specifications. Assumptions: * Metadata combines 'include'

Evaluate given image using detected objects on the global metadata specifications. Assumptions: * Metadata combines 'include'

跨 GPU 生成共享的采样 timesteps。 Args: accelerator: Accelerator 对象 train_timesteps: 可采样的 timestep 列表 M: 要采样的数量

用 prompt_ids 预计算 group 信息（更快）

Save script copy, resolved config JSON, argv, optional --config source file, and git metadata.

Build the wide all-linear `tdm` LoRA and, when `config.train.lora_path` is set, remap a narrow (8 attn-projection) TDM LoRA EMA into it. Retur

Calculate OCR reward :param images: List of input images (PIL or numpy format) :param prompts: Corresponding target text list

OCR reward calculator :param use_gpu: Whether to use GPU acceleration for PaddleOCR

Submits images to DeQA and computes a reward.