Training LLMs with Fault Tolerant HSDP on 100,000 GPUs
This addresses a critical scalability bottleneck for organizations training massive AI models, though it is an incremental improvement on existing parallelism methods.
The paper tackles the problem of low efficiency in large-scale LLM training due to frequent GPU failures and long recovery times in synchronous systems, proposing Fault Tolerant Hybrid-Shared Data Parallelism (FT-HSDP) to reduce stall time from 10 minutes to 3 minutes and increase effective training time from 44% to 80% on 100,000 GPUs without accuracy degradation.
Large-scale training systems typically use synchronous training, requiring all GPUs to be healthy simultaneously. In our experience training on O(100K) GPUs, synchronous training results in a low efficiency due to frequent failures and long recovery time. To address this problem, we propose a novel training paradigm, Fault Tolerant Hybrid-Shared Data Parallelism (FT-HSDP). FT-HSDP uses data parallel replicas as units of fault tolerance. When failures occur, only a single data-parallel replica containing the failed GPU or server is taken offline and restarted, while the other replicas continue training. To realize this idea at scale, FT-HSDP incorporates several techniques: 1) We introduce a Fault Tolerant All Reduce (FTAR) protocol for gradient exchange across data parallel replicas. FTAR relies on the CPU to drive the complex control logic for tasks like adding or removing participants dynamically, and relies on GPU to perform data transfer for best performance. 2) We introduce a non-blocking catch-up protocol, allowing a recovering replica to join training with minimal stall. Compared with fully synchronous training at O(100K) GPUs, FT-HSDP can reduce the stall time due to failure recovery from 10 minutes to 3 minutes, increasing effective training time from 44\% to 80\%. We further demonstrate that FT-HSDP's asynchronous recovery does not bring any meaning degradation to the accuracy of the result model.