Max Krasnyansky

Open-source work

Krasnyansky's pull request introducing ggml-hexagon, opened on 2025-10-13 and merged on 2025-10-22, describes an experimental llama.cpp backend for the Hexagon NPU on Snapdragon-based Android devices. It supported Hexagon versions v73 through v81, the Q4_0, Q8_0, MXFP4, and FP32 data types, and core LLM operations such as matrix multiplication, RMS normalization, RoPE, and softmax, and was tested with Llama 3, Qwen3, and gpt-oss models.¹

Later commits by Krasnyansky add flash attention, ROPE_NEOX, and further operator support to the Hexagon backend, along with reworked L2-cache handling, a buffer and DMA overhaul for 64-bit mappings, and tensor repacking for tiled and flat weights. llama.cpp's CODEOWNERS file places Krasnyansky, together with another maintainer, on the teams that own the Hexagon and OpenCL backends and the Snapdragon documentation and scripts.³²⁴

CPU and threading work

Before the Hexagon backend, Krasnyansky's llama.cpp contributions began in May 2024 with support for optimized Windows ARM64 builds using LLVM and MSVC. Commits from 2024 and 2025 improve the ggml thread pool and barrier scaling, add low scheduling-priority support, introduce chunking for flash attention and repacked matrix multiplications on the CPU, and fix thread-pool race conditions.³

Work and contributions

• llama.cpp Hexagon backend (ggml-hexagon) Open-source code

  Opened pull request #16547, "Add experimental ggml-hexagon backend for the Hexagon NPU," merged in October 2025. CODEOWNERS lists Krasnyansky in the ggml-org/ggml-hexagon and ggml-org/ggml-opencl teams. Commits authored from July to October 2026 add multi-NPU and multi-device support, DMA and L2-cache rework, MUL_MAT and FlashAttention updates, and backend sampler support. The repository is MIT-licensed.¹²³⁴

Timeline

1. May 16, 2024Added support for optimized Windows ARM64 builds of llama.cpp with LLVM and MSVC.³
2. Sep 23, 2024Improved ggml_barrier scaling for large thread counts in llama.cpp.³
3. Oct 13, 2025Opened the pull request adding the experimental ggml-hexagon backend for the Hexagon NPU.¹
4. Oct 22, 2025The ggml-hexagon backend pull request was merged into llama.cpp.¹
5. Oct 30, 2025Introduced chunking for flash attention and repacked matrix multiplications in the CPU backend.³
6. Jan 7, 2026Added f16/f32 flash attention and other operator support to the Hexagon backend.³
7. Sep 21, 2026Overhauled Hexagon buffer and DMA handling to support 64-bit mappings.³
8. Oct 16, 2026Reworked Hexagon tensor repacking to support tiled and flat quantized and floating-point weights.³

Where they work

• llama.cpp project 12 people in the wiki

Related people

• Aaron TeoAlso at llama.cpp project
• Neo ZhangAlso at llama.cpp project
• Johannes GäßlerAlso at llama.cpp project
• Jeff BolzAlso at llama.cpp project
• Daniel BeveniusAlso at llama.cpp project
• Sudeep BhojaChips and hardware · Inference and serving
• Sid ShethChips and hardware · Inference and serving
• Mitesh AgrawalChips and hardware · Inference and serving
• Jeff JohnsonChips and hardware · Inference and serving
• Ian BuckChips and hardware · Inference and serving