The systems behind useful intelligence.
Compute / Reliability / Scale
Paper uses expert correlations to reduce MoE training communication
The submitted research reports lower all-to-all communication time in tested Megatron-LM MoE pretraining configurations without changing model routing decisions or expert parameters.
2 min read
NVIDIA outlines governed digital-twin workflow for AI factory changes
NVIDIA’s implementation guide describes using DSX Air simulations, Brev GPU resources and bounded agents to validate supported infrastructure changes before they reach physical systems or production.
4 min read
Ollama v0.40.0-rc1 updates MLX tokenizer handling to match publisher semantics
The release candidate adds tokenizer compatibility fixes and shared reference cases intended to catch differences between Ollama’s MLX path and published tokenizers.
1 min read
PyTorch adds per-parameter mixed precision policy to FSDP2
The change supports mixed parameter dtypes in distributed training while retaining a single collective where dtypes allow it, according to PyTorch’s reported benchmarks.
1 min read
PyTorch adds in-place MPS reductions for supported strided tensors
The new FlatStrided path removes a costly contiguous-copy step for eligible full reductions, though performance still depends on tensor layout.
2 min read
PyTorch narrows NVFP4 autotuning for Blackwell decode workloads
The Inductor update keeps shape-specific GEMM tactics while cutting the default ranked candidate pool, which PyTorch says reduced tuning time in its tests.
2 min read
PyTorch folds NVFP4 output scaling into NVGEMM candidates for decode workloads
The Inductor change can remove a separate scale launch before QKV fan-out while retaining fallback and protected tensor-parallel paths.
2 min read
NVIDIA Dynamo-Triton Adds Multi-GPU TensorRT Serving Through One Model Endpoint
NVIDIA’s eight-GPU Cosmos 3 Nano test cut mean generation latency from 156.595 seconds to 34.183 seconds, but it measured neither throughput nor deployment economics.
3 min read