zhengjia

wonder how to generate the zipped imageNet labels, e.g. train_map.txt ?

> 问题好像是在predict_system.py中 StructureSystem 的__call__中if region['label'] == 'table' ，因为不同的版面分析lable不一样，需要改成region['label'] == 'table' or region['label'].lower() == 'table'才能正确的走到表格模型输出带html的res。不知道这个是不是需要提pr. @Sunting78 not working ...

looks the bugging is coming from **handel_table(self, html, doc)** tmp fix by: ```py while docx_cell.text != "": # Skip the merged cell cell_col += 1 if cell_col + colspan -1...

> 你好， > > 1.2. streetsurf 内部唯一定死接口的只有相机坐标系，我们会要求所有不同数据集的相机坐标系统一转到 opencv 格式，目前已经开放的版本中，像 waymo 的、colmap 的、megaNeRF 的、dtu / neus 的，这些不同的数据集的接口都是支持读取的。这些数据集的接口都会 将 c2w 统一转换成 opencv 的 c2w，包括 waymo 也是先转换到 openCV 的 c2w / c2v 来读取的。如果你的相机已经是...

or maybe possible newer mi300 kind of?

> For half, we have `unsafeAtomicAdd` instead of `atomicAdd`. > > https://github.com/ROCm/clr/blob/aa6d07518fdb211c49fd617ee9f69408f1acddfd/hipamd/include/hip/amd_detail/amd_hip_fp16.h#L1511 is there any risk concern for `unsafeAtomicAdd`, just wonder in which way it's `unsafe`

hi, @tcgu-amd , I tried rocm/torch images with rocm6.2.3, rocm6.2.4, neither has `unsafeAtomicAdd` yet. anther thing may need you clarify, looks in `/opt/rocm/include`, there is no `clr` headers, only found:...

related to https://github.com/ROCm/composable_kernel/issues/419 ?

also got errors with bs=64, inOut: 2048,2048; 128:2048; 20488,128. rebuilt engine with ```sh --gemm_plugin float16 \ --gpt_attention_plugin float16 \ --context_fmha enable \ --remove_input_padding disable \ --paged_kv_cache disable \ --max_batch_size 64...

hi, @schung-amd , you can try on this repo: https://github.com/ZJLi2013/grouped_gemm/blob/rocm/csrc/permute.hip