Y

プロ野球（楽天ファン）/basketball/SDVエンジニア

[Anshul @anshulcreates]

these 10 GPU optimization papers are why your model runs fast today

the engineers who actually read these are the ones who'll be writing your kernels in 2 years.

links to the papers below! 👇

1. FlashAttention (2022)
• IO-aware exact attention: tiles Q/K/V into on-chip SRAM and uses online softmax to avoid materializing the quadratic attention matrix in HBM.
• an incredible example of how optimizing data movement can unlock longer contexts and faster Transformers.

2. FlashAttention-2 (2023)
• redesigns work partitioning across thread blocks and warps, improving occupancy while cutting non-matmul FLOPs and shared-memory communication.
• makes attention substantially better at actually using the GPU's compute capacity. around 2× faster than the original in the paper's benchmarks.

3. Stream-K (2023)
• partitions aggregate GEMM inner-loop work into balanced chunks, reducing the wave-quantization inefficiency of fixed output-tile scheduling.
• especially elegant for matrix shapes that would otherwise leave a partially occupied final wave of thread blocks.

4. PagedAttention (2023)
• brings virtual-memory-style paging to KV-cache management, with noncontiguous physical blocks and flexible cache sharing.
• foundational to vLLM: reduces memory waste so serving systems can fit larger batches.

5. FlashAttention-3 (2024)
• exploits Hopper's asynchronous Tensor Cores and TMA through warp specialization, overlapping data movement with computation and interleaving matmul with softmax.
• adds an FP8 path with block quantization and incoherent processing. serious hardware-algorithm co-design.

6. Tiled GEMM / Tensor Core dataflow (foundational technique)
• stages matrix tiles through shared memory and registers, reusing operands across Tensor Core multiply-accumulate operations.
• the foundation underneath so much of this list. a family of techniques, so this entry has no single canonical paper.

7. Online normalizer calculation for softmax (2018)
• updates the running maximum and normalization sum as elements stream through, reducing memory passes while preserving numerical stability.
• a beautifully compact building block for memory-efficient attention.

8. MARLIN (2024)
• combines low-bit weight storage with higher-precision computation, using asynchronous loads and carefully pipelined dequantization to feed Tensor Cores efficiently.
• makes weight-only quantization useful for batched inference, with substantial speedups extending beyond single-user decoding.

9. Ring Attention (2023)
• distributes sequence blocks across devices and circulates KV blocks in a ring, overlapping communication with blockwise attention computation.
• scales context capacity across GPUs while preserving exact attention.

10. FlashInfer (2025)
• combines customizable attention kernels with load-balanced scheduling for dynamic serving workloads and heterogeneous KV-cache layouts.
• tackles the messy inference conditions: variable sequence lengths and changing request batches.

links to the papers, same order:

1. FlashAttention
https://t.co/wkRORf0tRf

2. FlashAttention-2
https://t.co/v23kmebrGN

3. Stream-K
https://t.co/V8dldoTDZd

4. PagedAttention
https://t.co/mmtorgjXd8

5. FlashAttention-3
https://t.co/YCJDZJaDpS

6. Tiled GEMM / Tensor Core dataflow
no single paper. FlashAttention above is one concrete example of IO-aware tiling.

7. Online normalizer calculation for softmax
https://t.co/3zl1ZunPDJ

8. MARLIN
https://t.co/hvd8geyy8K

9. Ring Attention
https://t.co/eQfdn3EVhl

10. FlashInfer
https://t.co/bOLhjdPcvD

[Atsushi Sakai @Atsushi_twi]

ウーブンの自動運転の人達は🤔

[Little V @LittleVMills]

🔥ウ ウ・ワ・ワ・ウワ 🔥

#ちいかわ #Chiikawa

[たむお☘️ @JaySmooth_0]

日本対韓国のバスケ決勝これにつきる

[狩野英孝 @kano9x]

しつこいなw
こんなに撮影したっけ？？

[ちいかわ💫アニメ火金 @ngnchiikawa]

⭐️

[AB @iga_2012]

中国の製薬企業、製造だけでなく研究開発も三交代で進めてるらしい。日本が勝てなくなる未来しか見えない

[湘南ベルマーレウィメン @bellmare_women]

💚Welcome💙

この度、#岡村來佳 選手が新加入することとなりました。

「海外挑戦を目指していますが、それまでの間、精一杯湘南のためにプレーしたいと思います。よろしくお願いします！」

岡村選手、ようこそ湘南へ🤝

✅コメント全文はこちら
https://t.co/IzM2jGuIcc

[シャパリントン @syaparinton]

…これだッ！！

[ゼロ次郎 @zerojirou]

クマが出るかチンポが出るか、二択の様相を呈している

[高岸の夢がパンパン @yumegapanpan]

本日の動画公開されました！
「高岸をバレずにハーランドにしてみた」
https://t.co/HbPVVQa3p6

[数え続ける世界のナベアツbot @3_aho_bot]

ｻｧﾝｼﾞｭｰﾏﾝwww!? (300000)

[タマゴケ @s5ml]

社員「Git！VScode！python！」
会社「危険、利用禁止」
社員「ChatGPT！Claude！」
会社「危険、利用禁止」
Microsoft「M365にCopilotついてます」
会社「素晴らしい！これで社員の悩みは全て解決だ！」

[Bleacher Report @BleacherReport]

BREAKING: Giannis has been traded to the Heat per @ShamsCharania 🚨😱

OMG. 🤯🤯🤯

[Basketball Forever @bballforever_]

BREAKING: Cavaliers star James Harden was arrested early Saturday morning in Texas and charged with the unlawful carrying of weapons 🚨

Harden was caught with a handgun in his vehicle, and it was allegedly in plain sight and not being carried in a holster. It’s a misdemeanor charge.

Harden was released on bond, and he has a court date scheduled for June 22nd.

(via TMZ)

[Y @tgt0916]

自動運転の認識屋として転職したが、未経験なためプロジェクトに入れてもらえないので自分で勉強の日々（Turingとかで働いてみたい）

[猪瀬直樹 【作家・参議院議員、日本維新の会 参議院幹事長】 @inosenaoki]

日本の自動車産業と消費者も、こうした進化を素直に認めること、そうでないととても巻き返しなどできない。

[仙台89ERS @89ERS_Official]

いつも黄援いただき、ありがとうございます。
このたび、シェーファーアヴィ幸樹選手と2026-27シーズンの契約が合意に至ったことをお知らせいたします。

▼詳細とコメントはこちら
https://t.co/wMgE2LTswD

@avi_schafer

#シェーファーアヴィ幸樹
#89ERS

[スポニチ楽天担当2026 @SponichiEagles]

【球団発表】
ゴンザレス選手のウエーバー公示申請の手続きを行いました。
#RakutenEagles