ekuznetsov139

Fisheye lens information loss during undistortion

3

I am trying to use colmap to reconstruct a scene from a set of fisheye lens images, 2048x2048 with FOV 180 degrees (camera: RADIAL_FISHEYE 2048 2048 600 1024 1024 0...

pretrained_model_dir not implemented

5

layers/common_hparams.py mentions a hyperparameter "pretrained_model_dir": "Directory containing a checkpoint for a pretrained model. This will only be used if a new run is being started. Parameters not found in the...

All pages on https://rocm-documentation.readthedocs.io/en/latest/index.html are limited to approximately 800 pixels in width and there's no apparent way to make the text wider (I can zoom in but that just makes...

Training questions

1

I'm trying to reproduce BEV training results, and I have a few questions. 1. The source tree references a "posetrack" dataset. It is not mentioned in the paper, its official...

The following fails to compile on AMD (gfx942; have not tried on NVIDIA, the issue may be present there too): ``` @triton.jit def dot_test(p, out, N:tl.constexpr, K:tl.constexpr, M:tl.constexpr): p_block =...

[WIP] Splitting attention _fwd_grouped_kernel_stage1 to improve occupancy

1

This splits _fwd_grouped_kernel_stage1 operation in two parts, using a temporary buffer. The rationale is as follows. _fwd_grouped_kernel_stage1 becomes a significant bottleneck in DeepSeek-V3 with large context lengths. It is slow,...

Triton-fused DeepseekScalingRotaryEmbedding

1

This PR uses Triton to fuse DeepseekScalingRotaryEmbedding operation with is_neox_style=False (observed in DeepSeek-V3). It substantially reduces the number of distinct kernels in DeepSeek-V2/V3 profile (the number of kernel launches per...

BLOCK_SIZE_K fix

1

This reintroduces the bugfix from #14987 that somehow got lost during refactoring.