BestHub
Sign in
Architect
Architect
Apr 25, 2026 · Artificial Intelligence

DeepSeek V4: 1M‑Token Context’s Impact on Model, Inference, Cache & Agents

The DeepSeek V4 technical report shows how a 1 million‑token context forces a redesign of attention, KV‑cache, optimizer, quantization and inference budgeting, turning long‑context capability from a costly showcase into a production‑ready feature for agents, search and Chinese professional tasks.

1M contextDeepSeekKV cache
0 likes · 28 min read
DeepSeek V4: 1M‑Token Context’s Impact on Model, Inference, Cache & Agents
DaTaobao Tech
DaTaobao Tech
Sep 27, 2023 · Artificial Intelligence

FlashAttention-2: Efficient Attention Algorithm for Transformer Acceleration and AIGC Applications

FlashAttention‑2 is an IO‑aware exact attention algorithm that cuts GPU HBM traffic through tiling and recomputation, optimizes non‑matmul FLOPs, expands sequence‑parallelism and warp‑level work distribution, delivering up to 2× speedup over FlashAttention, near‑GEMM efficiency, and enabling longer‑context Transformer training and inference for AIGC with fastunet and negligible accuracy loss.

AIGCFlashAttention-2GPU
0 likes · 20 min read
FlashAttention-2: Efficient Attention Algorithm for Transformer Acceleration and AIGC Applications