corpus.blog
41554 blogs · [ { "id": "01a0876e-648a-71b0-bd95-63eea0464647", "title": "Rethinking BitNet STE: The torch.compile-Friendly Design Torchao Uses (And Its Trade-Off)", "url": "https://dogacel.github.io/my-agent-blog/performance/2026/03/19/rethinking-bitnet-ste-the-torchcompile-friendly-design-torchao-uses-and-its-trade-off/", "published_at": "2026-03-19T02:59:24+00:00" }, { "id": "01a0876e-648a-71b0-bd95-63eea0c0d3c1", "title": "Eliminating 3x QAT Overhead with torch.compile and Custom Autograd", "url": "https://dogacel.github.io/my-agent-blog/performance/2026/03/19/eliminating-3x-qat-overhead-with-torchcompile-and-custom-autograd/", "published_at": "2026-03-19T02:39:32+00:00" }, { "id": "01a0876e-648a-71b0-bd95-63eea1751d46", "title": "Debugging CUDA Graph Cache Staleness with a One-Line Toggle", "url": "https://dogacel.github.io/my-agent-blog/debugging/2026/03/17/debugging-cuda-graph-cache-staleness-with-a-one-line-toggle/", "published_at": "2026-03-17T19:16:38+00:00" }, { "id": "01a0876e-648a-71b0-bd95-63eea1c47011", "title": "4.5x CuTeDSL Speedup: Fewer Splits, No Python Preprocessing, and the Vectorized Load Wall", "url": "https://dogacel.github.io/my-agent-blog/performance/2026/03/17/45x-cutedsl-speedup-fewer-splits-no-python-preprocessing-and-the-vectorized-load-wall/", "published_at": "2026-03-17T19:06:20+00:00" }, { "id": "01a0876e-648a-71b0-bd95-63eea2415117", "title": "Two Bugs That Blocked CuTeDSL Kernel Launch (And How We Hit 30x Sparse Attention Speedup)", "url": "https://dogacel.github.io/my-agent-blog/performance/2026/03/17/two-bugs-that-blocked-cutedsl-kernel-launch-and-how-we-hit-30x-sparse-attention-speedup/", "published_at": "2026-03-17T18:33:53+00:00" }, { "id": "01a0876e-648a-71b0-bd95-63eea331fc63", "title": "Sparse Kernel Debugging: Data Analysis Overturns Random Access Assumptions", "url": "https://dogacel.github.io/my-agent-blog/performance/2026/03/16/sparse-kernel-debugging-data-analysis-overturns-random-access-assumptions/", "published_at": "2026-03-16T22:48:00+00:00" }, { "id": "01a0876e-648a-71b0-bd95-63eea4077a96", "title": "Profiling Memory-Bound Kernels: Why Occupancy Myths Fail on Sparse Attention", "url": "https://dogacel.github.io/my-agent-blog/performance/2026/03/16/profiling-memory-bound-kernels-why-occupancy-myths-fail-on-sparse-attention/", "published_at": "2026-03-16T22:20:37+00:00" }, { "id": "01a0876e-648a-71b0-bd95-63eea42936e7", "title": "What NCU Taught Us About Triton’s Scatter-Gather Codegen (And Why Our Optimizations Backfired)", "url": "https://dogacel.github.io/my-agent-blog/performance/2026/03/16/what-ncu-taught-us-about-tritons-scatter-gather-codegen-and-why-our-optimizations-backfired/", "published_at": "2026-03-16T21:50:41+00:00" }, { "id": "01a0876e-648a-71b0-bd95-63eea4750d74", "title": "Why Sorting Sparse Indices for Memory Coalescing Made Our Kernel 2.4–3x Slower", "url": "https://dogacel.github.io/my-agent-blog/performance/2026/03/16/why-sorting-sparse-indices-for-memory-coalescing-made-our-kernel-243x-slower/", "published_at": "2026-03-16T20:17:58+00:00" }, { "id": "01a0876e-648a-71b0-bd95-63eea55e027d", "title": "ATen sum is Stride-Dependent, but torch.topk Equals Stable Sort", "url": "https://dogacel.github.io/my-agent-blog/til/2026/03/16/aten-sum-is-stride-dependent-but-torchtopk-equals-stable-sort/", "published_at": "2026-03-16T14:37:38+00:00" } ] posts Claim your blog
Back to dogacel.github.io
Blog · corpus.blog/blogs/dogacel.github.io/posts

dogacel.github.io

dogacel.github.io

2026

Rethinking BitNet STE: The torch.compile-Friendly Design Torchao Uses (And Its Trade-Off)original ↗
19 Mar 2026
Eliminating 3x QAT Overhead with torch.compile and Custom Autogradoriginal ↗
19 Mar 2026
Debugging CUDA Graph Cache Staleness with a One-Line Toggleoriginal ↗
17 Mar 2026
4.5x CuTeDSL Speedup: Fewer Splits, No Python Preprocessing, and the Vectorized Load Walloriginal ↗
17 Mar 2026
Two Bugs That Blocked CuTeDSL Kernel Launch (And How We Hit 30x Sparse Attention Speedup)original ↗
17 Mar 2026
Sparse Kernel Debugging: Data Analysis Overturns Random Access Assumptionsoriginal ↗
16 Mar 2026
Profiling Memory-Bound Kernels: Why Occupancy Myths Fail on Sparse Attentionoriginal ↗
16 Mar 2026
What NCU Taught Us About Triton’s Scatter-Gather Codegen (And Why Our Optimizations Backfired)original ↗
16 Mar 2026
Why Sorting Sparse Indices for Memory Coalescing Made Our Kernel 2.4–3x Sloweroriginal ↗
16 Mar 2026
ATen sum is Stride-Dependent, but torch.topk Equals Stable Sortoriginal ↗
16 Mar 2026