Repository Issue Activity (beta)

#211 [FEATURE] H100 Lightning Indexer kernel with streaming top-k for DeepSeek-V4 CSA

#176 TVM FFI support

#199 Help wanted: Performance issue in ThunderKittens FlashAttention kernel

#196 How to support FP8 in H100

#192 [Question] Mamba 3

#194 [Question] gemm_ar: why is `store_async_read_wait()` sufficient before signaling cross-GPU barrier?

#188 FP8 shared vector support (sv_fp8e4m3 / sv_fp8e5m2/ sv_fp8e8m0)

#187 TK vs FA4 Benchmark

#191 [Question] end-to-end runnable examples with multi-GPU kernels

#182 [Question] Is there any code that make sure there is no Register Spilling before creating Register based types?

#178 [Bug] `namespace "std" has no member "copy_n"` in include/common/util.cuh when building kernels/attention/mha_h100

#180 Move + Destroy bug in TKParallelTensor

#169 [Request] microbenchmark code for ParallelKittens: TMA vs ld/st SM utilization analysis

#173 can i run these kernels for adalovelace arch?

#175 Do you support batched gemm operations in a custom kernel with custom branching?

#159 Is there a runnable TK-based Flash Attention on B200?

#174 Windows with Visual Studio and Cuda tool kit

#171 [Question] Why mxfp8 mma doesn't support trans_a or trans_b

#170 [Questions] about dispatch GEMM scheduling and barrier load semantics

#168 Will EP be supported in the future?

#166 Will parallel kittens support decoding in the future?

#167 H100: ThunderKittens/include/types/system/vmm.cuh:152 'invalid argument'

#125 Is the causal mask top-left or bottom-right?

#163 [BUG] missing symbol `make_causal`

#164 pipeline/barrier error on sm90a