Issues

LoRAMLPv6.apply has 17 positional parameters and two call sites have already silently broken
Forge-researchlab/forge-kernels#19 · opened on Aug 14, 2026
ForgeRoPEv3 requires a power-of-two head_dim, excluding models with head_dim 80 or 96
Forge-researchlab/forge-kernels#18 · opened on Aug 14, 2026
GPU CI: run the real test suites and catch benchmark regressions
Forge-researchlab/forge-kernels#17 · opened on Aug 14, 2026
Publish forge to PyPI once it is installable
Forge-researchlab/forge-kernels#16 · opened on Aug 14, 2026
1
Benchmark the kernels on hardware other than an A100-80GB
Forge-researchlab/forge-kernels#14 · opened on Aug 14, 2026
53 MB of LoRA adapter weights are tracked in git
Forge-researchlab/forge-kernels#13 · opened on Aug 14, 2026
GeGLU: fp64 gradcheck is deferred
Forge-researchlab/forge-kernels#11 · opened on Aug 14, 2026
1
Expose fused linear cross-entropy chunk_size through the public API and the patch layer
Forge-researchlab/forge-kernels#10 · opened on Aug 14, 2026
Embedding: dispatch to F.embedding below a dimension threshold
Forge-researchlab/forge-kernels#8 · opened on Aug 14, 2026
1
ForgeRMSNorm does not accept Gemma's offset, so forge.patch silently skips RMSNorm on Gemma
Forge-researchlab/forge-kernels#5 · opened on Aug 14, 2026
No two kernel test suites can share a pytest process (experiments/ name collision)
Forge-researchlab/forge-kernels#3 · opened on Aug 14, 2026