Lightweight proxy models enable rapid experimentation without repeatedly training frontier-scale systems, but their small kernels often leave modern accelerators underutilized. Conventional training compounds this inefficiency by scheduling the forward and backward passes as disjoint phases, so spare capacity in one cannot be filled by work from the other. We reinterpret the detachment mechanism of Forward-Forward (FF) as a scheduling primitive: given a local objective, detaching a block's output removes downstream gradient dependencies, making its backward pass ready when its forward pass finishes. ERASE launches each detached subgraph's backward pass early on a separate CUDA stream, overlapping it with subsequent forward work. Execution trace on a lightweight transformer demonstrates this overlap and its limit: a kernel that saturates the device leaves no capacity for concurrency. On a large-scale click-through-rate model, detaching six dense subarchitectures improves training throughput by up to 9.51% while keeping normalized entropy close to the baseline.
Figures & tables
Figure 1: Schematic of FF-style detachment. Each block receives a detached copy of the previous block’s activation, and contributes its own local loss. No gradient crosses a block boundary.
Figure 2: NanoChat single-batch trace with two detachment points and three streams. Backward work from one subgraph runs alongside forward work from the next.
Figure 3: NE gap between the non-blocking early-backward variant with CUDA Graphs (red curve) and the baseline (blue curve), approximately 1.38% .
#
Configuration
QPS ( p90 )
Gain vs. baseline
–
baseline
184,535.51
+0.00%
3
blocking, FUP=False
185,223.71
+0.37%
2
blocking, FUP=True
194,244.72
+5.26%
1
non-blocking + CUDA Graphs, FUP=False
198,145.28
+7.38%
0
blocking, FUP=True †
202,078.43
+9.51%
Table 1: Throughput on the CTR model, p90 queries per second
Modern industrial Deep Learning Recommendation Models typically extract user preferences through the analysis of sequential interaction histories, subsequently generating predictions based on these derived interests. The inherent heterogeneity in data characteristics frequently result in substantial under-utilization of computational resources during large-scale training, primarily due to computational bubbles caused by severe stragglers and slow blocking communications. This paper introduces FreeScale, a solution designed to (1) mitigate the straggler problem through meticulously load balanced input samples (2) minimize the blocking communication by overlapping prioritized embedding communications with computations (3) resolve the GPU resource competition during computation and communication overlapping by communicating through SM-Free techniques. Empirical evaluation demonstrates that FreeScale achieves up to 90.3% reduction in computational bubbles when applied to real-world workloads running on 256 H100 GPUs.
Modern recommendation models have increased to trillions of parameters. As cluster scales expand to O(1k), distributed training bottlenecks shift from computation and memory to data movement, especially lookup and communication latency associated with embeddings. Existing solutions either optimize only one bottleneck or improve throughput by sacrificing training consistency. This paper presents NestPipe, a large-scale decentralized embedding training framework that tackles both bottlenecks while preserving synchronous training semantics. NestPipe exploits two hierarchical sparse parallelism opportunities through nested pipelining. At the inter-batch level, Dual-Buffer Pipelining (DBP) constructs a staleness-free five-stage pipeline through dual-buffer synchronization, mitigating lookup bottlenecks without embedding staleness. At the intra-batch level, we identify the embedding freezing phenomenon, which inspires Frozen-Window Pipelining (FWP) to overlap All2All communication with dense computation via coordinated stream scheduling and key-centric sample clustering. Experiments on production GPU and NPU clusters with 1,536 workers demonstrate that NestPipe achieves up to 3.06x speedup and 94.07% scaling efficiency.
We present MegaTrain, a memory-centric system that efficiently trains 100B+ parameter large language models at full precision on a single GPU. Unlike traditional GPU-centric systems, MegaTrain stores parameters and optimizer states in host memory (CPU memory) and treats GPUs as transient compute engines. For each layer, we stream parameters in and compute gradients out, minimizing persistent device state. To battle the CPU-GPU bandwidth bottleneck, we adopt two key optimizations. 1) We introduce a pipelined double-buffered execution engine that overlaps parameter prefetching, computation, and gradient offloading across multiple CUDA streams, enabling continuous GPU execution. 2) We replace persistent autograd graphs with stateless layer templates, binding weights dynamically as they stream in, eliminating persistent graph metadata while providing flexibility in scheduling. On a single H200 GPU with 1.5TB host memory, MegaTrain reliably trains models up to 120B parameters. It also achieves 1.84× the training throughput of DeepSpeed ZeRO-3 with CPU offloading when training 14B models. MegaTrain also enables 7B model training with 512k token context on a single GH200.