v1.2.0

RustingStrata v1.2.0

Prebuilt binaries for Linux (x86_64, glibc 2.35+) and Windows (x86_64) are attached below. They are built for
x86-64-v3 (AVX2, any CPU from roughly 2015 on). On Windows, double-click strata-windows-x86_64.exe. On Linux,
run chmod +x strata-linux-x86_64 && ./strata-linux-x86_64. Either way, a chat page opens in the browser, and you
can download a model from it.

New

• Chat page in the binary. Plain strata starts the server on http://127.0.0.1:8080 and opens the page. It
  keeps a chat list in the browser, downloads models from Hugging Face (HF_TOKEN for gated repos), switches
  between the models in the models folder, and shows context use, speed and memory use.
• Several GPUs. --device 0,1 splits the layers over the cards in proportion to their free VRAM, so more
  experts stay resident. Mixed cards and backends without peer access work. MTP drafting is off on a split model.
• Qwen3.8-Flash-Next (qwen4exp) on CUDA and Vulkan, including QSA sparse attention, hyper-connections, PLE
  and split GGUF files (pass the -00001-of-0000N file).
• More GGUF types: Q4_0, Q4_1, Q5_0 and Q5_1 (also in the experts), F16 and Q2_0.
• Anthropic Messages API (POST /v1/messages), so Claude Code can use the server. OpenCode and other agents
  use the OpenAI address.
• Agent sessions prefill less. --preserve-thinking is on by default, so a long session prefills only what was
  added. A side request (a title or a summary) parks the conversation's state in host memory (--cache-ram,
  4096 MB), and the next request takes it back. The state after the system prompt and tools is kept too, so a
  compacted conversation restarts from there instead of from token 0.
• strata perplexity computes perplexity like llama-perplexity, and with --kld-base it writes a logits
  file that llama-perplexity's --kl-divergence can read.

AMD and Vulkan fixes

• RADV device-coherent memory is set before Vulkan opens.
• Prefill submits each layer's work as it is recorded, because amdgpu resets a submission that runs over 10 s.
• A MoE combine stops waiting for a slow CPU worker after 1 s.
• A load that would leave less than the VRAM margin free fails with a hint.
• The device, driver and features are printed at load. docs/AMD.md lists the known issues and what to send in a
  report.

Faster

• Prefill: expert copies overlap the GEMMs, and small expert tiles have their own kernels.
• IQ grids and tables are in shared memory on CUDA and Vulkan.
• Faster IQ2_S and Q2_0 dots on the CPU.

Other

• GGUF files with malformed metadata are refused instead of panicking (#2, thanks @bettercallcaleb).
• STRATA_NAN_TRACE=1 prints per-layer magnitudes, to help locate NaNs on hardware we do not have.
• Docs: multi-GPU, AMD clocks in a split, page cache and ZFS