Search headlines, summaries and articles to connect the dots.
The qwen2vl model in llama.cpp was producing incorrect results due to NULL values being passed for FFN weight scales, preventing per-tensor NVFP4 scales from being applied. This issue has been resolved by correctly passing the scales to the build_ffn function, consistent with how qwen2 handles them.
The latest release of llama.cpp introduces support for Inkling 2, including text and vision capabilities. This update includes text inference implementation, GGUF conversion for text models, vision multimodal projection support, still-image encoding, flash attention with folded bias and sinks, and improvements to tensor mapping and testing. The changes also address model conversion, checkpoint handling, and compatibility across multiple platforms, with support for CPU, GPU (CUDA, ROCm, Vulkan), and specialized hardware like Snapdragon and OpenVINO.
The latest release of llama.cpp (b12362) includes optimizations for grammar matching through left factoring, improved handling of character class boundaries, prevention of infinite expansion, and enhanced code clarity with ASCII comments and deduplicated logic. It reverts prior changes related to AST refactoring and alternative limits, while introducing shared immutable rules and removing performance benchmarks. The release supports multiple platforms including macOS, Linux, Windows, Android, and iOS, with various backend options such as CUDA, Vulkan, ROCm, OpenVINO, and SYCL.
The update refactors DFlash-based speculative decoding methods in llama.cpp by restructuring code organization, sharing components across models, splitting decoder graphs per head, and introducing dedicated classes for DFlash2 and LiliCorr drafts. It also adds support for new draft types, improves documentation, and removes unused code.
Imaging black holes is challenging because they are distant and appear tiny in the sky—for example, M87*’s observed ring spans just 42 microarcseconds. The...
Human motion offers roboticists a great starting point for teaching humanoid robots how to move. A human demonstration captures the coordination and timing...
The update requires the position embedding tensor for HunyuanVL to be explicitly provided as F32 during model loading to prevent buffer overflow issues when using F64 tensors, ensuring compatibility and stability on first image requests.
TL;DR Helion now uses compile-time heuristics that analyze a kernel’s IR and dataflow to pick fast settings without slow autotuning. It classifies the kernel as matmul, reduction, or pointwise and... The post Fast by Default: Compile-Time Heuristics for Helion Kernels appeared first on PyTorch.
The llama.cpp project has released version 1.0.0, marking a significant milestone with numerous improvements across multiple platforms including CUDA, Vulkan, Metal, SYCL, OpenCL, and WebGPU. Key updates include enhanced support for quantized models (Q5_0, Q5_1, FP8), optimizations for flash attention and MoE caching, improved server security features like CORS protection and SSRF mitigation, UI enhancements such as sidebar reorganization and live-preview markdown editing, and better cross-platform compatibility. The release also includes bug fixes, CI/CD improvements, and updated dependencies.
Running a frontier-scale model locally means keeping all its weights in GPU memory; the largest models no longer fit on one machine. Two NVIDIA DGX Stations...
On one NVIDIA Vera Rubin NVL72 rack, vLLM serves DeepSeek-V4.1-Flash on SemiAnalysis AgentX at 1.7M tokens/s per GPU at a P90 of 387 tokens/s per user, 58x GB300 NVL72 at the same interactivity.
H3-Image on Vera Rubin: 0.87-second 4K generation or 6,115 images/hour on four GPUs, with projected API revenue and support across consumer and data-center hardware.
vLLM supports Thinking Machines' Inkling 2 Small from day 0: attention residuals, a new decode kernel for relative-bias attention, all-reduce fused with the ops around it, and DSpark speculative decoding.
The update introduces graceful failure handling for unreachable RPC endpoints and failed handshakes in ggml-rpc, replacing fatal aborts with reportable errors. Functions now return appropriate values (e.g., false, nullptr) on connection failures, allowing host applications to handle issues without crashing. A regression test was added for unreachable endpoints, and low-value comments were removed. Mid-stream failures remain unchanged to prevent silent corruption.
On one NVIDIA Vera Rubin NVL72 rack, vLLM serves DeepSeek-V4.1-Flash on SemiAnalysis AgentX at 1.7M tokens/s per GPU at a P90 of 387 tokens/s per user, 58x GB300 NVL72 at the same interactivity.
Recent advances in vLLM encoder disaggregation: metadata-only multimodal requests, dedicated encoder execution, remote embedding transfer, cross-encoder caching, and lower-overhead EPD routing.
TL;DR We serve Meta’s Muse Spark Open, an 816B-parameter mixture-of-experts model, on a single TPU v7 host of four chips, with SGLang running on the TorchTPU PyTorch backend. At the... The post Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU appeared first on PyTorch.
AI agents spend most of their time idle, waiting on a human prompt, a tool result or a model response. Kubernetes wasn’t built for that. Running one Pod per...
On Qualcomm Adreno GPUs, scalar flash attention was limited to 8 rows per workgroup when head size exceeds 128 to prevent register spilling and performance degradation. This change improved performance significantly, reducing execution time from several seconds to milliseconds and enabling completion of the test-backend-ops case.
Picture guiding a robot arm by hand while its motors supply just enough torque to support its weight and keep the robot upright. This simple interaction, a...
