From ca5eb12f98e4943ce06242bbb9c52415ab2df166 Mon Sep 17 00:00:00 2001 From: Zheng QiHang Date: Tue, 20 Oct 2026 12:24:51 +0000 Subject: [PATCH 1/3] docs(readme): announce TileFoundry 0.0.3 and the October examples --- README.md | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index ea3724878..61bf6de92 100644 --- a/README.md +++ b/README.md @@ -31,14 +31,17 @@ ## Latest News +- 10/2026 🚀: **TileFoundry 0.0.3 is on PyPI** — scheduled HIR lowers end to end to TIR, analysis models staged memory allocation, and the CLI starts without loading torch. +- 10/2026 ⚡: [Qwen3.8-27B prefill](https://github.com/tile-ai/TileFoundry/tree/main/examples/qwen3_8_27b_prefill-tilelang) — a 64-layer gated-delta-net / attention hybrid, **faster than SGLang** at 1024 / 4096 / 8192 tokens, and [**1.88x–1.03x faster** at 128–1024 tokens](https://github.com/tile-ai/TileFoundry/tree/main/examples/qwen3_8_27b_prefill_short-tilelang), across the roofline ridge. +- 10/2026 🍎: [Qwen prefill on Apple AMX](https://github.com/tile-ai/TileFoundry/tree/main/examples/qwen_prefill-amx) — Qwen3-1.7B and Qwen3.5-4B text prefill on an M2 Pro, with the large projections as FP16 AMX outer products. - 09/2026 📝: [AI Compilers in the Agentic Era](https://tile-ai.github.io/TileFoundry.github.io/blog/AI-compilers-in-the-agentic-era/) — where is the right boundary between coding agents and AI compilers? - 09/2026 🔧: **TileFoundry 0.0.2 is on PyPI** — bug fixes and refactoring across the analysis, parser, and IR layers. -- 08/2026 ⚡: [Nemotron-3.5-Lightning-30B-A3B](https://github.com/tile-ai/TileFoundry/tree/main/examples/nemotron_3_5_lightning_30b_a3b-tilelang) — a 52-layer Mamba2 / attention / MoE hybrid in a **single mega decode kernel**, one launch a step with no CUDA graph, reaching **97.5% of SGLang** at short context and **83.2%** at 262144. -- 08/2026 📦: Four [worked examples](https://github.com/tile-ai/TileFoundry/tree/main/examples) added — Qwen3-1.7B (tilelang), Qwen3.5-35B-A3B (tilelang), MiniCPM3-4B (CuTeDSL) and granite-4.0-h-small (CUDA C) — each one a real agent run kept whole, with the decode throughput it measured.
Earlier +- 08/2026 ⚡: [Nemotron-3.5-Lightning-30B-A3B](https://github.com/tile-ai/TileFoundry/tree/main/examples/nemotron_3_5_lightning_30b_a3b-tilelang) — a 52-layer Mamba2 / attention / MoE hybrid in a **single mega decode kernel**, one launch a step with no CUDA graph, reaching **97.5% of SGLang** at short context and **83.2%** at 262144. +- 08/2026 📦: Four [worked examples](https://github.com/tile-ai/TileFoundry/tree/main/examples) added — Qwen3-1.7B (tilelang), Qwen3.5-35B-A3B (tilelang), MiniCPM3-4B (CuTeDSL) and granite-4.0-h-small (CUDA C) — each one a real agent run kept whole, with the decode throughput it measured. - 08/2026 🎉: **TileFoundry 0.0.1 is on PyPI** — the first public release.
From b9211b0ea1cbef9431a627a5a8e97bd72fd7a2de Mon Sep 17 00:00:00 2001 From: Zheng QiHang Date: Tue, 20 Oct 2026 12:29:40 +0000 Subject: [PATCH 2/3] docs(readme): describe what 0.0.3 lets an agent do --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 61bf6de92..3c8b0dd15 100644 --- a/README.md +++ b/README.md @@ -31,7 +31,7 @@ ## Latest News -- 10/2026 🚀: **TileFoundry 0.0.3 is on PyPI** — scheduled HIR lowers end to end to TIR, analysis models staged memory allocation, and the CLI starts without loading torch. +- 10/2026 🚀: **TileFoundry 0.0.3 is on PyPI** — `tilefoundry schedule` shows which hardware instructions can implement an HIR program and lowers the chosen schedule to TIR, down to block-scaled FP8 GEMM on SM90 WGMMA; `analyze` now models staged memory allocation and reuse. - 10/2026 ⚡: [Qwen3.8-27B prefill](https://github.com/tile-ai/TileFoundry/tree/main/examples/qwen3_8_27b_prefill-tilelang) — a 64-layer gated-delta-net / attention hybrid, **faster than SGLang** at 1024 / 4096 / 8192 tokens, and [**1.88x–1.03x faster** at 128–1024 tokens](https://github.com/tile-ai/TileFoundry/tree/main/examples/qwen3_8_27b_prefill_short-tilelang), across the roofline ridge. - 10/2026 🍎: [Qwen prefill on Apple AMX](https://github.com/tile-ai/TileFoundry/tree/main/examples/qwen_prefill-amx) — Qwen3-1.7B and Qwen3.5-4B text prefill on an M2 Pro, with the large projections as FP16 AMX outer products. - 09/2026 📝: [AI Compilers in the Agentic Era](https://tile-ai.github.io/TileFoundry.github.io/blog/AI-compilers-in-the-agentic-era/) — where is the right boundary between coding agents and AI compilers? From c5b9dd8939f100a9e59a7e0aec739ca2776ab9ab Mon Sep 17 00:00:00 2001 From: Zheng QiHang Date: Tue, 20 Oct 2026 12:32:54 +0000 Subject: [PATCH 3/3] docs(readme): shorten the 0.0.3 entry --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 3c8b0dd15..c92a22af1 100644 --- a/README.md +++ b/README.md @@ -31,7 +31,7 @@ ## Latest News -- 10/2026 🚀: **TileFoundry 0.0.3 is on PyPI** — `tilefoundry schedule` shows which hardware instructions can implement an HIR program and lowers the chosen schedule to TIR, down to block-scaled FP8 GEMM on SM90 WGMMA; `analyze` now models staged memory allocation and reuse. +- 10/2026 🚀: **TileFoundry 0.0.3 is on PyPI** — a new instruction-level schedule, a restructured analysis, and prefill examples on H200 and Apple AMX. - 10/2026 ⚡: [Qwen3.8-27B prefill](https://github.com/tile-ai/TileFoundry/tree/main/examples/qwen3_8_27b_prefill-tilelang) — a 64-layer gated-delta-net / attention hybrid, **faster than SGLang** at 1024 / 4096 / 8192 tokens, and [**1.88x–1.03x faster** at 128–1024 tokens](https://github.com/tile-ai/TileFoundry/tree/main/examples/qwen3_8_27b_prefill_short-tilelang), across the roofline ridge. - 10/2026 🍎: [Qwen prefill on Apple AMX](https://github.com/tile-ai/TileFoundry/tree/main/examples/qwen_prefill-amx) — Qwen3-1.7B and Qwen3.5-4B text prefill on an M2 Pro, with the large projections as FP16 AMX outer products. - 09/2026 📝: [AI Compilers in the Agentic Era](https://tile-ai.github.io/TileFoundry.github.io/blog/AI-compilers-in-the-agentic-era/) — where is the right boundary between coding agents and AI compilers?