v0.3.3

Changelog

• Qwen 3.6 MTP Conversion Fixes: Normalized model.mtp_layer, model.mtp_layers, mtp_layer, and mtp_layers tensor prefixes to mtp. during HF-to-GGUF conversion. Registered Qwen 3.6 architectures (Qwen3_6ForConditionalGeneration, Qwen3_6ForCausalLM, Qwen3_6MoeForConditionalGeneration, and Qwen3_6MoeForCausalLM) to enable seamless, out-of-the-box conversion.
• Speculative, Quantization, and Concurrency Fixes:
  • Modified ggml-cuda/cross-ring-interleave.cu to clamp cross_len to prevent GPU buffer OOB writes, and corrected a peer-to-peer capability check.
  • Corrected block-wise quantization/dot-product logic for Q8_K64 CPU activations in ggml-iq2-bn.c.
  • Added GGML_TYPE_TURBO2_0 to token-tracking InnerQ calibration checks in set-rows.cu.
  • Resolved MSVC warning C4319 in tests/test-gguf.cpp and demoted checkpoint matching logs to LOG_DBG in prompt matching loops.
  • Dynamically scaled loop guard constraints and fixed suffix tree dangling references during alternate sequence removal.

macOS:

• macOS Apple Silicon

Linux:

• Ubuntu x64 CPU
• Ubuntu arm64 CPU
• Ubuntu x64 CUDA 12.4
• Ubuntu x64 CUDA 13.1
• Ubuntu x64 Vulkan
• Ubuntu x64 ROCm 7.2
• Ubuntu x64 SYCL

Windows:

• Windows x64 CPU
• Windows x64 SYCL
• Windows x64 CUDA 12.4 - DLLs
• Windows x64 CUDA 13.1 - DLLs
• Windows x64 HIP

Docker:

• CPU - docker pull ghcr.io/atomicmilkshake/godzilla-llama.cpp:server-cpu-v0.3.3
• CUDA - docker pull ghcr.io/atomicmilkshake/godzilla-llama.cpp:server-cuda-v0.3.3
• CUDA 12 - docker pull ghcr.io/atomicmilkshake/godzilla-llama.cpp:server-cuda12-v0.3.3
• CUDA 13 - docker pull ghcr.io/atomicmilkshake/godzilla-llama.cpp:server-cuda13-v0.3.3
• ROCm - docker pull ghcr.io/atomicmilkshake/godzilla-llama.cpp:server-rocm-v0.3.3
• Vulkan - docker pull ghcr.io/atomicmilkshake/godzilla-llama.cpp:server-vulkan-v0.3.3
• SYCL - docker pull ghcr.io/atomicmilkshake/godzilla-llama.cpp:server-sycl-v0.3.3

v0.3.2

Full Changelog: kv-god-20260629...v0.3.2