Decision Workspace
jammi-kernels vs candle-fused-attn vs kaio-candle
Side-by-side comparison of Rust crates
Candle CustomOp scaffolding and a feature-gated CUDA build path for fused training kernels (elementwise/normalization/attention). Leaf crate: no jammi-* dependencies.
Fused fp32 scaled-dot-product attention for candle, forward AND backward: the FlashAttention-2 algorithm in plain fp32, a bit-reproducible backward, over stock candle.
Candle bridge for KAIO — CustomOp bindings for 12 GPU ops (matmul_tc, matmul_tc_bf16, matmul_tc_async, matmul_tc_bf16_async, matmul_int4, matmul_int8, attention_tc, attention_tc_causal, attention_flash, attention_flash_causal, qkv_project_int8, qkv_project_int4). All four matmul TC variants (f16+bf16, sync+async) support backward (autograd); FlashAttention (plain+causal) supports backward via dedicated PTX kernels. Build with `cargo build --features cuda`.
Core Metrics
| jammi-kernels | candle-fused-attn | kaio-candle | |
|---|---|---|---|
| Health Score | 50 | 57 | 48 |
| Total Downloads | 301 | 60 | 65 |
| 30d Downloads | 0 | 0 | 0 |
| Dependents | 24 | 0 | 0 |
| Releases | 9 | 4 | 3 |
| Last Updated | 7d ago | 4d ago | 119d ago |
| Age | 1m | 7d | 5m |
Health Breakdown
Technical Details
| jammi-kernels | candle-fused-attn | kaio-candle | |
|---|---|---|---|
| Version | 0.54.0 | 0.4.0 | 0.2.0 |
| Stable (≥1.0) | ✗ No | ✗ No | ✗ No |
| License | Apache-2.0 | MIT OR Apache-2.0 | MIT OR Apache-2.0 |
| Dependencies | 9 | 2 | 7 |
| Crate Size | 588KB | 152KB | 42KB |
| Features | 6 | 2 | 2 |
| Yanked % | 0.0% | 0.0% | 0.0% |
| Edition | 2021 | 2024 | 2024 |
| MSRV | 1.94.0 | 1.88 | 1.94 |
| Owners | 1 | 1 | 1 |
Links
Quick Verdict
- •candle-fused-attn leads with a health score of 57/100, but none of the options score above 80.
- •jammi-kernels has the most downloads (301), suggesting wider adoption.