tensor

**Tensor Parallelism and Model Parallelism** is **distributed training strategies that partition model layers or operations across multiple accelerators — enabling training of models larger than single-device memory through parallel computation of forward and backward passes**. Tensor Parallelism and Model Parallelism address the fundamental constraint that modern large language models exceed individual GPU memory capacity. Tensor parallelism partitions weight matrices across devices, with each device computing a subset of output features. For a linear layer with weight matrix W, tensor parallelism splits W row-wise or column-wise across devices. Forward passes require communication to concatenate results from different devices, and backward passes require reduction across devices. This parallelism exposes abundant parallelism — each device performs local computation on partial operations. Model parallelism (pipeline parallelism) divides model layers across devices in sequence. Device 1 processes input through first k layers, passes hidden states to Device 2, which processes through next k layers, and so forth. This creates a pipeline — different devices process different minibatches in flight, improving utilization. Pipeline parallelism reduces per-device memory but requires communication passing large hidden states between devices. Different parallelism strategies have different communication-to-computation ratios. Sequence parallelism partitions sequences across devices, with each device processing a portion of the sequence length. This is particularly valuable for long sequences where sequence length is a primary memory bottleneck. Combined parallelism strategies use tensor parallelism, data parallelism, and pipeline parallelism together. Zero redundancy optimizer (ZeRO) partitions optimizer states, gradients, or parameters across devices, further reducing per-device memory. Flash Attention and other communication-efficient techniques improve parallelism scalability. Ring allreduce and other collective communication patterns optimize communication cost. Network topology and bandwidth significantly impact parallelism efficiency — GPU clusters with high-bandwidth interconnects enable effective scaling to many devices. Load balancing becomes critical in heterogeneous settings — devices with different capability should be utilized proportionally. Gradient accumulation and batch pipelining improve utilization. Research shows that naive model parallelism often performs poorly due to low computation-to-communication ratio, while well-tuned configurations achieve good scaling. **Tensor and model parallelism strategies enable distributed training of models exceeding single-device capacity, using different approaches to balance computation and communication across accelerators.**