agent/toolchain.md

Choosing a toolchain

Triton is the default reflex and is often not the right answer. Pick deliberately.

workload	first choice	why
elementwise / reduction fusion	Triton	fast to write, autotunes well, this is its sweet spot
standard GEMM shapes	cuBLAS / CUTLASS	already at the hardware limit; do not rewrite it
GEMM + unusual epilogue or prologue	CUTLASS / CuTe DSL	epilogue fusion with full layout control
quantised GEMM (fp8/int4/mxfp4/nvfp4)	CUTLASS / CUDA	the dequant must fuse into the MMA pipeline; no library does it for you
attention variants (masks, sparsity, sinks)	CUDA or Triton	no library kernel applies as-is; Triton is viable if the mask is regular
persistent kernel / grid-wide sync	CUDA	Triton can do it (atomics + spin) but it is awkward and fragile
whole-model fusion (megakernel)	CUDA	needs instruction scheduling, async weight staging, warp specialization
sub-byte packing, bit tricks	CUDA	lop3, prmt, byte permutes have no Triton expression
variable-length / ragged batching	Triton or CUDA	both fine; Triton's masking is convenient

Where Triton specifically loses

• Fine-grained warp specialization — producer/consumer warp roles are not directly expressible.
• TMA multicast, cluster/DSMEM (sm_90+), tcgen05 (sm_100+) — either unavailable or awkward.
• Custom epilogues with awkward layouts — you get what the compiler decides.
• Sub-byte dtypes — 4-bit packing needs bit manipulation Triton does not express well.
• Register pressure control — no direct control; spills appear and are hard to fix.
• Persistent kernels with device-wide barriers — possible via tl.atomic_add + a spin loop, and it works, but the grid must stay co-resident or the spinning blocks deadlock. Verify block count against multi_processor_count before relying on it.

Where Triton wins

Iteration speed, autotuning over num_warps/num_stages/tiles, automatic masking for ragged shapes, and readable reduction code. For a bandwidth-bound fusion it will usually reach near-peak with far less effort than CUDA.

The pragmatic path

Prototype in Triton to establish correctness and a baseline number, dump its PTX/SASS to see what it generated, then decide whether hand-written CUDA can beat it. Frequently the Triton version is already at bandwidth and the answer is to stop; for MMA-heavy work it usually is not.