Agent skills by Krusty84

Implements request-to-token-pool copies for Triton kernels on Ascend NPUs, replacing loop-carried vector offsets with fixed-offset block loops that compile reliably.

Krusty84/triton-ascend-agent-dev-kit

Writes a Triton kernel pattern for Ascend NPU that gathers several indexed token rows per program into an on-chip buffer and stores one contiguous output tile, for MoE-style token reordering.

Krusty84/triton-ascend-agent-dev-kit

Builds Triton-Ascend kernels that pack sorted token assignments into fixed expert-capacity tensors, restore top-k outputs and compute router-weight gradients on Ascend NPUs.

Krusty84/triton-ascend-agent-dev-kit

Guides building a FlashAttention-v2-style fused forward attention kernel for Triton-Ascend on Ascend NPU, with online softmax, causal staging and float32 accumulation.

Krusty84/triton-ascend-agent-dev-kit

Builds a fused row-wise softmax kernel for Triton-Ascend that reads and writes each row once, handling padding, strides and masked loads on Ascend NPUs.

Krusty84/triton-ascend-agent-dev-kit

Build a fused forward LayerNorm kernel for Triton-Ascend with row-wise mean and variance reductions, float32 accumulation, affine weight and bias, and masked feature tiles.

Krusty84/triton-ascend-agent-dev-kit

Build Ascend-friendly MoE gather, scatter, and router-weight-gradient kernels with vector-core-sized grids, UB-aware row/column tiling, and CANN slice extensions.

Krusty84/triton-ascend-agent-dev-kit

Build padded MoE gather, scatter, and router-weight-gradient kernels for Triton-Ascend using real cumulative bins, padded cumulative bins, UB slice assembly, and vector-core-aware tiling.

Krusty84/triton-ascend-agent-dev-kit

Build a single-program Triton-Ascend matrix multiplication with fused bias using two-dimensional pointer grids and tl.dot.

Krusty84/triton-ascend-agent-dev-kit

Build and debug masked one-dimensional elementwise kernels for Triton-Ascend, including launch wrappers and PyTorch/NPU correctness checks.

Krusty84/triton-ascend-agent-dev-kit

Refactor fused causal Conv1d state updates for Triton-Ascend by replacing negative-offset cat emulation and tl.where selection with transposed UB assembly through extension.insertslice.

Krusty84/triton-ascend-agent-dev-kit

Apply Triton-Ascend maxautotune to search base kernel configurations together with Ascend compiler options for vector, cube, or mixed operators.

Krusty84/triton-ascend-agent-dev-kit

Optimize bounded discrete reads in Triton-Ascend by staging a contiguous source vector in UB and selecting elements with tl.gather.

Krusty84/triton-ascend-agent-dev-kit

Optimize irregular KV-cache access in Triton-Ascend grouped decode attention by vector-loading contiguous inner dimensions, assembling discrete outer rows with CANN slice extensions, and transposing…

Krusty84/triton-ascend-agent-dev-kit

Optimize integer vector kernels for Triton-Ascend by preferring int32 arithmetic when the value and index ranges permit it.

Krusty84/triton-ascend-agent-dev-kit

Prevent Triton-Ascend Unified Buffer overflow by streaming large logical vectors through bounded compile-time tiles with masked loop iterations.

Krusty84/triton-ascend-agent-dev-kit

Generate TTIR for Triton-Ascend kernel configurations and rank them with the Ascend costmodel backend before empirical autotuning.

Krusty84/triton-ascend-agent-dev-kit

Reorder independent Triton-Ascend loads to expose overlap with loop-carried stores and reduce dependency stalls.

Krusty84/triton-ascend-agent-dev-kit

Use Triton-Ascend tl.load carepadding=False safely to skip initialization of masked tail lanes and remove unnecessary padding work.

Krusty84/triton-ascend-agent-dev-kit

Tile two-dimensional torch.gather(dim=1) kernels for Triton-Ascend so the logical grid stays near the physical vector-core count while each program handles multiple batch rows and loops over K.

Krusty84/triton-ascend-agent-dev-kit

Add standard or advanced autotuning to Triton-Ascend kernels, define shape keys and candidate meta-parameters, and structure kernels so automatic split and tiling analysis can recognize their axes.

Krusty84/triton-ascend-agent-dev-kit

Validate Triton-Ascend kernel outputs against PyTorch references with dtype-aware tolerances, exact integer checks, bfloat16 promotion, and boolean handling.

Krusty84/triton-ascend-agent-dev-kit

Vectorize explicit integer comparisons in Triton-Ascend by casting bounded index vectors to float32 before tl.where or similar compute expressions.

Krusty84/triton-ascend-agent-dev-kit