Strata 0.2.19 Mais recente

Downloads

Notas da versão

# Strata 0.2.19

The app and the Terminal are 0.2.19. The engines (strata-engine, the new strata-inkling shape engine, and the updated Linux SYCL strata-hy4 package) are 0.1.65, and so is Control, the program that installs and updates them. The update feeds are sequence 131 (shared) and 127 (linux-aarch64).

What changes for you

  • Native Thinking Machines Lab Inkling 2 Mini support (`strata-inkling` 0.1.65). Run Thinking Machines Lab's 26.4B-parameter looped Mixture-of-Experts model (3.4B active per token / 5.4B looped-active across 48 virtual visits, up to 1M context, text + image inputs, OpenAI-compatible tool calling, direct non-thinking answers) locally across Windows, Linux, macOS Apple Silicon, Intel Arc, and NVIDIA DGX Spark (linux-aarch64).
  • Three Strata Quantization packs ([`strata-engine/Inkling-2-Mini-Strata`](https://huggingface.co/strata-engine/Inkling-2-Mini-Strata)), starting from a 4 GiB graphics card:
  • SQ-10 (10.47 GB / 9.75 GiB): fits 100% in VRAM on 12 GiB cards, 16 GiB Macs, and Radeon 780M / Strix Point iGPUs (and runs in hybrid RAM+VRAM mode from 4 GiB GPUs).
  • SQ-15 (14.35 GB / 13.36 GiB): fits 100% in VRAM on 16 GiB cards and 24 GiB Macs (96.68% top-1 agreement with the original weights, 0.00889 test KL divergence).
  • SQ-20 (19.35 GB / 18.02 GiB): bit-exact lossless native SF4+MX8 build for 24 GiB+ GPUs and 32 GiB+ unified memory (100.00% top-1 agreement and 0.00000 KL divergence on held-out test prompts and WikiText-2).
  • Vision (`Inkling-2-Mini-Vision.strata`, `0.64 GB` lossless BF16) and built-in `W=4` Multi-Branch On-GPU Markov Beam Tree `DSpark` speculative decoding work out of the box in both the Desktop app and strata launch.
  • Direct `safetensors`, `.gguf`, and Dual-GPU EP support: strata-inkling also loads official thinkingmachines/Inkling-2-Mini safetensors (+ mtp.safetensors) and community .gguf files directly, with --gpus 2 Dual-GPU Expert Parallelism and CUDA-Graph mapped-memory doorbell NUMA offload.
  • 15 ready-made `strata-inkling` 0.1.65 packages in shape_engines, fetched automatically on demand when you choose Inkling 2 Mini:
  • Windows `x86_64`: cuda-cuda13, cuda-cuda12, hip-strix (gfx1151), hip-rx9070 (gfx1200/gfx1201), hip-rdna3 (gfx1100–gfx1150, including Radeon 780M & Strix Point iGPUs), hip-rdna2 (gfx1030–gfx1032).
  • Linux `x86_64`: cuda-cuda13, cuda-cuda12, hip-halo (gfx1151), hip-rx9070, hip-rx7900 (gfx1100–gfx1150), hip-rx6000, and unified fat sycl (Intel Arc A-series & Battlemage B-series).
  • macOS `arm64`: metal (Metal 3 & Metal 4 simdgroup_matrix + barrier-free GEMV, signed & Apple-notarized).
  • Linux `aarch64`: cuda-cuda13 (NVIDIA Grace Blackwell / DGX Spark GB10).
  • Faster Qwen3.8-Flash-Next out of the box on NVIDIA Blackwell, RTX 40, RTX 30, and AMD RDNA3 (`strata-engine` 0.1.65, zero environment variables needed).
  • Update as usual. Your existing Qwen, Hy4-Flash, and DeepSeek models, chats, and settings stay untouched.

Engine & performance highlights

  • Inkling 2 Mini (`strata-inkling` 0.1.65):
  • Sparse-MoE Bayes-Optimal `W=4` On-GPU Markov Beam Tree `DSpark` + Zero-Logits-D2H serving (`0.00000` KLD): reaches up to 250.3 tok/s solo (198.9 tok/s on lossless SQ-20), 11,306 tok/s prompt prefill, and 4,089.7 tok/s total output across 64 concurrent requests on RTX 5090; 154.4–214.5 tok/s on RTX PRO 5000; 131.6 tok/s (2,081.1 tok/s @ c=32) on 2x RTX PRO 6000 (--gpus 2 EP); 108.9–114.5 tok/s on Apple M5 Max 128 GB; 99.0–106.5 tok/s on Radeon AI PRO R9700 (+66%..+78%); 47.3–48.0 tok/s on Ryzen AI Max+ Strix Halo 128 GB (+46%..+64%); 74.3 tok/s (1,145.1 tok/s @ c=64) on DGX Spark GB10; and 159.2 tok/s on Intel Arc Pro B70 32 GB.
  • NVIDIA Blackwell (`sm_120`/`sm_121`), Ada (`sm_89`), & Ampere (`sm_86`) (`strata-engine` 0.1.65):
  • Native sm_120a dense_fp4_tc (192.6 TFLOPS), moe_fp4_tc v4, persistent INT8 Tensor Core dense_i8_tc, vhead_bitplane 2-stage LM head, row_dot2_80_sub16, and Q5_1 use_sub8 automatic defaults (0.000 pts force_check parity):
  • RTX 5090 (`Qwen3.8-Flash-Next`): IQ3_S prefill +76% (3,205.9 ppts, 2.91x faster TTFT) and decode 161.6 tok/s (+19%); UD-Q4_K_XL decode 115.9–117.6 tok/s (+36%..+46%); NVFP4 prefill +36% (2.08x faster TTFT) and decode 155.5 tok/s (+15%).
  • RTX PRO 5000 72 GB: IQ3_S prefill +72% (5,093–6,552 ppts, 2.96x faster TTFT) and agent8k decode 265.3 tok/s (+28%).
  • RTX 4090 (`sm_89`): IQ3_S prefill +39% / decode +20%..+31% (117.1 tok/s); UD-Q4_K_XL prefill +42% / decode +47%..+59% (87.7 tok/s).
  • RTX 3090 & 2x RTX A4000 (`sm_86`): IQ3_S / UD-Q4_K_XL / IQ2_XS prefill +12%..+26% (1.72x faster TTFT) and decode +13%..+20% (100% bit-exact).
  • AMD RDNA3 (`gfx1100`–`gfx1151`) (`strata-engine` 0.1.65):
  • NVFP4 Wave32 FP16 WMMA prefill (42–44x lower approximation error), native_h11p_kernel (PF_F16=2), discrete moe_p1a FL=3, IQ2_XS & NVFP4 single-launch bitwise-exact LDS decode (em0=8), 3.10x Wave32 WMMA DFlash2, WDDM coherent host staging, OCuLink PcieAuto, and automatic gfx1151 Strix Halo Windows defaults — delivering 1.44x–2.10x faster prefill and +17%..+22% faster decode (97.9 tok/s, peak 103.7 tok/s) on RX 7900 XT.
  • Intel Arc SYCL (`strata-hy4` 0.1.65): fixes Battlemage (Arc Pro B70) XMX SYCL dispatch on Hy4-Flash.

Community acknowledgements

We gratefully thank Baptiste Rajaut (`@baptisterajaut`, PR #6520), `@merbanan` (PR #6520 stack), `@cobra91` (PR #6506), `@mercant1982`, and Peasant Smith, OscarSheng, `alesha-pro`, and the Project Maya contributors whose MIT-licensed work on official safetensors / .gguf loading, InkDoorbell mapped-memory NUMA CPU offload, quantized KV modes, setup contracts, and --gpus 2 Dual-GPU Expert Parallelism helped shape Strata's unified strata-inkling support.

Updating

  • App: the in-app update offers 0.2.19, or download the installer below.
  • Terminal: run strata update.
  • A clone of the repository: run UPDATE.bat (Windows) or ./setup.sh --update.

Downloads

Total: 50 files, 9,999,878,199 bytes.

Unchanged strata-hy4 0.1.64 packages stay on the 0.2.18 release, the two DFlash2 drafter packs stay on the 0.2.16 release, and the frozen DeepSeek 0.1.59 engines stay on the 0.2.13 release; the signed update feed downloads them automatically from their respective releases.

| File | Bytes | |---|---| | install-manifest.txt | 545 | | install.ps1 | 16,236 | | install.sh | 12,331 | | latest.json | 2,134 | | SHA256SUMS.txt | 5,350 | | Strata-0.2.19-linux-x86_64.AppImage | 300,542,456 | | Strata-0.2.19-linux-x86_64.AppImage.sig | 428 | | Strata-0.2.19-macos-arm64.dmg | 161,211,245 | | Strata-0.2.19-windows-x64-setup.exe | 109,463,499 | | Strata-0.2.19-windows-x64-setup.exe.sig | 428 | | strata-control-0.1.65-linux-aarch64.tar.gz | 78,458,819 | | strata-control-0.1.65-linux-x86_64.tar.gz | 94,518,092 | | strata-control-0.1.65-macos-arm64.tar.gz | 45,691,587 | | strata-control-0.1.65-windows-x86_64.zip | 67,721,651 | | strata-engine-0.1.65-linux-aarch64-cuda-cuda13.tar.gz | 582,724,655 | | strata-engine-0.1.65-linux-x86_64-cuda-cuda12.tar.gz | 772,153,289 | | strata-engine-0.1.65-linux-x86_64-cuda-cuda13.tar.gz | 584,031,359 | | strata-engine-0.1.65-linux-x86_64-hip-halo.tar.gz | 265,735,598 | | strata-engine-0.1.65-linux-x86_64-hip-rx6000.tar.gz | 190,489,895 | | strata-engine-0.1.65-linux-x86_64-hip-rx7900.tar.gz | 271,188,189 | | strata-engine-0.1.65-linux-x86_64-hip-rx9070.tar.gz | 633,044,600 | | strata-engine-0.1.65-linux-x86_64-sycl.tar.gz | 188,035,148 | | strata-engine-0.1.65-macos-arm64-m

Completo no GitHub

Versão anterior: v0.2.18