Agneya Tharun

Reproduced the paper's benchmarks on one NVIDIA L40S GPU. Matched the paper's ViT-B Sintel depth loss and stayed competitive with the 1B-parameter VGGT model on Sintel PC L1 geometric-mean loss (1.635 vs 1.582) at 7× fewer parameters.

Engineered a Pi3/LoGeR-style 5-layer pose head with 9D-rotation SVD orthogonalization to fix the pose-estimation bottleneck, cutting ATE by 53% and beating the published ViT-B baseline on camera pose by 26%.

Vectorized query generation (256× speedup), pre-resized fp16 dataset caches (35× less disk), AMP mixed-precision, and CUDA-allocator tuning enabled full ViT-B training in 46 GB on one consumer GPU. 50 clips/sec at fp16, 3.5 GB peak for inferences.

A three-process system: a phone sender streams frames over secure WebSockets to a FastAPI server with a bounded, drop-on-overflow queue, and a three.js viewer renders the cloud as it grows.

Frames run through Pi3X in 16-frame chunks with 6-frame overlap; the previous chunk's poses, depth, and rays are injected as a prior, and each new chunk is aligned into the global frame by a masked Umeyama SVD.

SAM3 segments prompted objects on keyframes; the 2D masks are lifted into 3D through Pi3X's dense point maps and fitted to oriented bounding boxes, deduplicated across frames and streamed to the viewer.

A block of 128 weights stores one fp16 scale plus 32 bytes of 2-bit codes. Each code maps to a ternary value, so 1.58-bit models pack into roughly 2.1 bits per weight on the GPU.

Five GLSL shaders cover every Vulkan path, wired in at ~32 registration points. A 959-line CPU simulator re-implements each shader bit-for-bit and validates it against the ggml reference.

End to end on an Intel Arc GPU, an 8B ternary model offloads fully to Vulkan and generates coherent text at 31.5 tokens/sec generation, 52 tokens/sec prompt. Merged as PR #32, +1,128 lines.