Repository navigation
Releases
Basalt v0.3.3-prod (Windows x64, CUDA 13, sm_120)
Clean production build on upstream v0.3.3 containing strictly the verified fixes: PR jesdga95#320 (batch prefill for incremental reads >= 16 tokens) and PR jesdga95#261 (CPU pool lost-wakeup completion barrier). Includes START-HERE.bat, TUNE-HERE.bat, and bundled CUDA 13 redist.
Basalt v0.3.3-win1 (Windows x64, CUDA 13, sm_120)
Windows x64 binaries for compute capability 12.0 (RTX 50 series, RTX PRO Blackwell), built by
upstream-windows.yml from
7004c3e50aaceb902fcfc5d1d5d7cb389449780c merged with
upstream jesdga95/basalt@v0.3.3
(commit 2729e228): these binaries are upstream's v0.3.3 plus this fork's Windows work.
SHA-256 is in dist/basalt-0.3.3-win1-windows-x64-cuda13.zip.SHA256SUMS. Unpack anywhere and run START-HERE.bat; a current NVIDIA driver
is the only thing to install.
Checked by CI: MSVC compile, Windows CPU tests, AVX-512 isolation, the zip's contents, and the packaged
engine starting with no CUDA toolkit installed. Not checked by CI: there is no GPU runner, so loading a
real model, speed and PPL/KLD quality are unmeasured here.
Basalt v0.3.2-prod (PR #320 Batch Prefill + PR #261 Pool Barrier)
Production release built from clean v0.3.2 with only PR jesdga95#320 (batch prefill for incremental reads >= 16 tokens) and PR jesdga95#261 (CPU pool lost-wakeup completion barrier) merged.
Basalt v0.3.2-win1 (Windows x64, CUDA 13, sm_120)
Windows x64 binaries for compute capability 12.0 (RTX 50 series, RTX PRO Blackwell), built by
upstream-windows.yml from
c1b51a65cfdb9d2f2a926da5f7a115b35c77e4d8 merged with
upstream jesdga95/basalt@v0.3.2
(commit ffc68faa): these binaries are upstream's v0.3.2 plus this fork's Windows work.
SHA-256 is in dist/basalt-0.3.2-win1-windows-x64-cuda13.zip.SHA256SUMS. Unpack anywhere and run START-HERE.bat; a current NVIDIA driver
is the only thing to install.
Checked by CI: MSVC compile, Windows CPU tests, AVX-512 isolation, the zip's contents, and the packaged
engine starting with no CUDA toolkit installed. Not checked by CI: there is no GPU runner, so loading a
real model, speed and PPL/KLD quality are unmeasured here.
Basalt v0.3.2-prod2 (Windows x64, CUDA 13, sm_120)
Windows x64 binaries for compute capability 12.0 (RTX 50 series, RTX PRO Blackwell), built by release-windows.yml from a8c6662. Unpack anywhere, run START-HERE.bat. SHA-256 in basalt-0.3.2-prod2-windows-x64-cuda13.zip.SHA256SUMS.
Basalt v0.3.2-prod1 (Windows x64, CUDA 13, sm_120)
Basalt v0.3.2-prod1 (Windows x64, CUDA 13, sm_120)
Clean, stable production release built on upstream v0.3.2 with 5 proven performance and stability optimizations for AMD Zen 3/4 + NVIDIA RTX 5070 Ti / Blackwell architectures:
- CPU Zen 3/4 BMI2 AVX2 (aef4013): Vectorized bit unpacking and sign mask shifts behind a fast-pdep probe for CPU MoE expert execution (+15–20% CPU token rate).
- CPU Expert Pool Lost-Wakeup Fix (2c896a3): Synchronizes worker completion barrier to eliminate intermittent thread pool jitter/stalls.
- DeepSeek-V4.1 Batch Prefill for Incremental Reads >= 16 tokens (85bfa88): Accelerates multi-turn user conversation turns and tool calls from 15.5 t/s to 202.6 t/s (13x speedup).
- DeepSeek-V4.1 2048 Prefill Chunk Cap on <= 16 GB GPUs (3579e89): Caps prefill chunks to 2048 on 16 GB cards, eliminating OOMs and keeping workspace to ~738 MiB.
- DeepSeek-V4.1 Dynamic Beside Slots & 512 MiB Floor (ceee676): Reclaims freed prompt memory during decode (raising VRAM expert hit rate to 34.4%) while strictly preserving a 512 MiB Windows WDDM floor to prevent driver paging.
- Windows Launcher & Dynamic Vision Fix (b620a71): Fixes vision dynamic mode translation and adds --config support to START-HERE.bat.
Basalt PR (Cap Prefill Chunks at 2048 on 16GB GPUs)
Test build for capping prefill chunks at 2048 tokens on <=16 GiB GPUs to avoid prompt-read OOMs and driver paging.
Basalt PR #322 (Dynamic Beside Slots)
Isolated test build for PR jesdga95#322 (upstream issue jesdga95#322). Reclaim idle decode VRAM as dynamic beside slots with 512 MiB floor and warm base cache across turns.
Basalt PR #321 (BMI2 AVX2 Kernels)
Isolated test build for PR jesdga95#321 (upstream issue jesdga95#321). BMI2 sign masks and bit extraction for AVX2 i-quant rows behind fast-pdep probe.
Basalt PR #320 (Batch Prefill Tail)
Isolated test build for PR jesdga95#320 (upstream issue jesdga95#320). Batch prefill for incremental reads >= 16 tokens.