KarLex AI, Inc. (@KarLexAI) profile picture
US
Joined October 2024
KarLex AI, Inc. retweeted
Where we lose: FlashQLA beats our exact Gated DeltaNet split on long single sequences (0.68-0.78x) and few-head packed documents (0.58-0.77x). Our FA3-style attention runs at 0.77-0.92x FA3, lower on some long sequences, so "auto" picks cuDNN for dense attention. H100 only.
1
1
54
KarLex AI, Inc. retweeted
End to end at 1.3B on one H100: transformer-1p3b in FP8 at 72.3% MFU, and 68.2% per GPU on 8 GPUs (FSDP). Mamba-2, Gated DeltaNet, KDA, RWKV-7 and the Kimi Linear and Qwen3-Next layouts train at 1.3B too; against flame on the same tokens: 1.04-1.67x in bf16, 1.10-1.77x in FP8.
1
1
60
KarLex AI, Inc. retweeted
Kernel speedup vs the latest upstream code, one H100, forward + backward: Mamba-2 scan 1.20-1.96x and Mamba-3 SISO 1.58-2.32x, MIMO 2.70-5.37x vs the official kernels; Gated DeltaNet 1.71-2.68x vs FLA Triton, KDA 1.47-1.78x vs FLA, GLA 1.52-1.77x, RWKV-7 1.06-1.46x.
1
1
61
KarLex AI, Inc. retweeted
jax-linear-attention 0.1.0: FLA-class linear-attention kernels for JAX in Pallas Mosaic GPU. Mamba-2/3, Gated DeltaNet, KDA, RWKV-7, GLA and FA3-style attention on H100 (fwd + bwd), plus a training stack at 72.3% MFU (1.3B transformer, FP8). Apache-2.0. github.com/jla-org/jax-linea…
2
1
1
290
KarLex AI, Inc. retweeted
Thanks to the work behind this: Mamba-2/3 (@tri_dao @_albertgu), Gated DeltaNet, GLA, FLA and flame (@SonglinYang4 @yzhang_cs @uniartisan), Kimi Linear's KDA (@Kimi_Moonshot), RWKV-7 (@BlinkDL_AI), JAX's Pallas Mosaic GPU, and many more credited in the repo.
jax-linear-attention 0.1.0: FLA-class linear-attention kernels for JAX in Pallas Mosaic GPU. Mamba-2/3, Gated DeltaNet, KDA, RWKV-7, GLA and FA3-style attention on H100 (fwd + bwd), plus a training stack at 72.3% MFU (1.3B transformer, FP8). Apache-2.0. github.com/jla-org/jax-linea…
1
2
88
KarLex AI, Inc. retweeted
With Claude Code (Opus 5.5), it ties plain grep on finding files and facts in every test we ran. On Graphify's own showcase corpus it found 99.2% of the facts vs Graphify's 87.8% (grep: 94.7%, a tie). Ties and losses included: github.com/graphicle-ai/grap…
1
1
51
KarLex AI, Inc. retweeted
Memory: lower peak memory than Graphify 0.9.91 in all 39 measurements, at 0.29× to 0.97× of its peak (pre-release build, 13 codebases × 3 operations).
1
1
39
KarLex AI, Inc. retweeted
Speed: across 13 codebases and 3 operations, it took 3.0× to 22.7× less time than Graphify 0.9.91. Kubernetes: 68 s vs about 26 min. Pre-release build, same machine, 3 CPUs each. The graphs differ slightly: node counts within 5%, edge counts within 2%.
1
1
47
KarLex AI, Inc. retweeted
Graphicle 0.1.1 is out. Code graph and memory for your coding agents, in one Rust program. brew install graphicle-ai/tap/graphicle
2
1
1
100
KarLex AI, Inc. retweeted
In any repo: graphicle setup It builds the code graph and connects Claude Code, Codex, Cursor and VS Code through one MCP server. Your agent looks up your code instead of re-reading it every session.
2
1
55
KarLex AI, Inc. retweeted
Graphify, rewritten in Rust. Kubernetes code graph: 26m 38s → 3m 18s. Same output. Plus memory for your coding agent: your docs, decisions and past sessions, with sources. brew install graphicle-ai/tap/graphicle github.com/graphicle-ai/grap…
1
1
1
100