Chris de Vries's picture

1 13

Chris de Vries

cmdevries

·

AI & ML interests

None yet

Recent Activity

liked a model 17 days ago

microsoft/VibeVoice-Realtime-0.5B

reacted to flozi00's post with ❤️ 29 days ago

When models get too large for a single GPU, simply stacking layers vertically (Pipeline Parallelism) isn't always the answer. Sometimes, you need to slice the matrices themselves. My latest guide breaks down the hardware mechanics of Tensor Parallelism (TP). We look at how to shard individual operations across devices to make a cluster function as one massive accelerator. This isn't high-level theory—it is a look at the bare metal implementation. Here is what is covered in the deep dive: The Strategies: Column vs. Row Parallelism We analyze how to split weight matrices (W) and inputs (X). Column-Linear: Splits weights by columns. Requires an All-Gather to reconstruct the output. Row-Linear: Splits weights by rows. Requires an All-Reduce to sum partial results. The "Megatron-LM" Optimization Efficiency comes from minimizing communication. By sandwiching the non-linearity (GeLU) between a Column-Parallel layer and a Row-Parallel layer, we can skip synchronization entirely during the activation phase. This cuts communication events by 50% per block. The Hardware Reality: The Bandwidth Wall In TP, the dist.all_reduce operation sits on the critical path. The CUDA cores effectively stall while waiting for the ring-reduce to finish. Intra-Node: Works well because NVLink provides enough bandwidth to hide this latency. Inter-Node: Fails at scale. Standard networking (Ethernet/InfiniBand) is too slow for the high-frequency syncs required by TP. The article includes a raw PyTorch implementation using torch.distributed primitives to show exactly where the data moves and where the bottlenecks sit. Read the full hardware-centric guide here: https://flozi.net/en/guides/ai/scaling/tensor_parallel

liked a Space about 1 month ago

HuggingFaceTB/smol-training-playbook

View all activity

Organizations

cmdevries 's datasets

None public yet