HWresearch

/

GNN4Colliders

like 0

Follow

Haichen Wang Research Group 11

2.0 -> 2.1:

• torch.sparse now includes prototype support for semi-structured (2:4) sparsity on NVIDIA® GPUs
• New CPU performance features include inductor improvements (e.g. bfloat16 support and dynamic shapes), AVX512 kernel support, and scaled-dot-product-attention kernels
• torch.compile can now compile NumPy operations via translating them into PyTorch-equivalent operations

2.1->2.2

• scaled_dot_product_attention (SDPA) now supports FlashAttention-2, yielding around 2x speedups compared to previous versions

2.2->2.3

• Tensor Parallelism improves the experience for training Large Language Models using native PyTorch function
• semi-structured sparsity implements semi-structured sparsity as a Tensor subclass, with observed speedups of up to 1.6 over dense matrix multiplication.

2.3->2.4

• Introduced a new default server backend for TCPStore built with libuv which should introduce significantly lower initialization times and better scalability
• Pytorch users can now experience improved quality and performance gains with the beta BF16 symbolic shape support

Based on a small setup with 4 trainings for each of torch 2.0 vs. torch 2.4, it looks like about 8% decrease in training time.

using 600k events total, and 1024 batch size, the speed up is less dramatic but still noticeable (about 3%)

Torch 2.4
Batch: 809.5060 s
Train: 417.8602 s
Eval: 790.6683 s

Torch 2.0
Batch: 829.1964 s
Train: 432.0877 s
Eval: 811.7211 s