ZLUDA 6 stack runs CUDA PyTorch on AMD GPUs on Windows
A hobbyist ZLUDA 6 build trains PyTorch on Radeon — but cuDNN's absent. Plus: OpenAI agents flooded RubyGems, and SageMaker slashes inference latency.
Copy markdown
The stack: CUDA app → ZLUDA → ROCm, no rebuild
A new Windows setup layers ZLUDA v6-preview over AMD's HIP SDK 6.4 so CUDA binaries hit a Radeon unchanged — cuBLAS, cuBLASLt, cuSPARSE and cuFFT all pass validation, and a LibTorch PPO network trained end to end on an RX 9060 XT. If you own an AMD card, that's PyTorch training with no NVIDIA GPU in the box.
The catch: no cuDNN, one GPU tested
The stable Windows HIP SDK ships no cuDNN, so convolution-heavy vision models stall, and NCCL, TensorRT and custom PTX can fail too. Validation covers exactly one GPU (gfx1200); other Radeons are candidates, not guarantees. Treat it as a real weekend path for LibTorch and RL, not a drop-in CUDA replacement.
Elsewhere: OpenAI agents flooded RubyGems with ~2,000 packages
Researchers tied an agent-driven campaign to roughly 2,000 malicious RubyGems and a remote-code-execution foothold on RubyDoc build servers, dating to May and disclosed this week. Audit your Ruby dependencies and pin versions before your next install.
Elsewhere: SageMaker cuts first-token latency up to 77%
Prefix-aware routing pins requests sharing a prompt prefix to one instance, keeping the KV cache warm: P50 time-to-first-token drops 71-77% and cache hits jump from ~25% to over 80% on long shared-prefix workloads. Turn it on with a PrefixAwareRoutingConfig flag, no container changes.