Jiading Gai

Yongjun He, Shuai Zhang, Jiading Gai et al. · amazon-science

As large language models (LLMs) continue to scale and new GPUs are released even more frequently, there is an increasing demand for LLM post-training in heterogeneous environments to fully leverage underutilized mid-range or previous-generation GPUs and alleviate the shortage of homogeneous high-end GPUs within a single availability zone. However, achieving high-performance reinforcement learning (RL) training for LLMs on such computing resources remains challenging because the workflow involves multiple models and tasks with complex computation and data dependencies. In this paper, we present HetRL, a distributed system for efficient RL training in infrastructures with heterogeneous GPUs and networks. HetRL formulates the scheduling of RL training in heterogeneous environments as a constrained joint optimization problem and provides two complementary approaches for addressing this problem: (1) a hybrid scheduling algorithm that efficiently identifies near-optimal solutions, and (2) an integer linear programming (ILP)-based scheduling algorithm that obtains optimal solutions, enabling flexible trade-offs between solution optimality and efficiency. Our extensive evaluation, consuming 20,000 GPU-hours, shows that HetRL achieves up to 9.17x the throughput of state-of-the-art systems, and 3.17x on average, across a wide range of workloads and settings.

Jiading Gai, Shuai Zhang, Xiang Song et al.

Modern RL post-training methods such as GRPO and DAPO train on $N$ response sequences of $R$ tokens sampled from a shared prompt of $P$ tokens, but standard FlashAttention replicates all $P$ prompt tokens $N$ times across both forward and backward passes -- duplicating compute and memory on identical hidden states. In large-rollout, long-context RL training ($N{\geq}16$, $P{\geq}8\text{K}$), this redundancy dominates the policy update cost. We observe that in decoder-only models, causal masking makes prompt representations invariant across sequences at every layer, so all per-token operations (norms, projections, MLP) and attention can process the prompt once -- a property not yet exploited at the kernel level for training. We propose \textbf{DualKV}, the first FlashAttention kernel variant that eliminates shared-prompt replication during RL training, via (1)~fused CUDA forward and backward kernels that iterate over two disjoint KV regions -- shared context and per-sequence response -- in a single kernel launch, and (2)~a data-pipeline redesign in veRL that repacks $N(P{+}R)$ tokens into $P{+}NR$ tokens per micro-batch, extending the token reduction from attention to the entire model by a factor $ρ= N(P{+}R)/(P{+}NR)$. DualKV is mathematically equivalent to standard attention and introduces no approximation. On Qwen3-8B GRPO training with 8$\times$H100 GPUs ($N{=}32$, 8K-context), DualKV achieves $1.63$--$2.09\times$ policy-update speedup, enables $2\times$ larger micro-batches, and raises MFU from $36\%$ to $76\%$. Similar gains hold for DAPO ($2.47\times$ speedup, $77\%$ MFU). At 30B MoE scale on 16$\times$H100, DualKV achieves $3.82\times$ policy-update and $3.38\times$ end-to-end step speedup over FlashAttention (which requires 4-way Ulysses sequence parallelism to avoid OOM).

Hongbin Zheng, Sejong Oh, Huiqing Wang et al.

Deep learning (DL) workloads are moving towards accelerators for faster processing and lower cost. Modern DL accelerators are good at handling the large-scale multiply-accumulate operations that dominate DL workloads; however, it is challenging to make full use of the compute power of an accelerator since the data must be properly staged in a software-managed scratchpad memory. Failing to do so can result in significant performance loss. This paper proposes a systematic approach which leverages the polyhedral model to analyze all operators of a DL model together to minimize the number of memory accesses. Experiments show that our approach can substantially reduce the impact of memory accesses required by common neural-network models on a homegrown AWS machine-learning inference chip named Inferentia, which is available through Amazon EC2 Inf1 instances.