Kimi K3 uses FlashKDA to make chunkwise parallel context computation work
nrehiew_ · x · 2026-07-29
The thread explains the KDA chunkwise-parallel implementation used in Kimi K3.
- FlashKDA implements a chunkwise parallel form of KDA.
- Tensor parallelism is less useful here because KDA has fewer heads, so context parallelism is done across chunks.
- The key challenge is that the delta update rule depends on the previous state, which prevents independent computation.
- The solution splits the computation into two terms: one local delta-rule factor that can be computed independently, and another term that applies the delta rule after a single all-gather.
- The flow is summarized as: compute local delta-rule factor, compute local state, all-gather once, then apply the delta rule.
More from Infra
- Reflection AI open-sources Beam under Apache 2.0 with 40+ ecosystem partners — zhyncs42 · 2026-10-16
- NVIDIA launches RTX Spark for local AI on Windows PCs — ostrisai · 2026-10-16
- SkyRL ships Day-0 RL support for Beam with +62% LoRA decode throughput in vLLM — vllm_project · 2026-10-16
- DFlash2 delivers up to 3.03x decode-only speedup for Beam across five benchmarks — vllm_project · 2026-10-16
- vLLM's Beam serving kernels cut GB200 forward time from 8.15 to 6.47 ms at TP8 batch 128 — vllm_project · 2026-10-16
- Reflection's 501B MoE Beam hits 463 tok/s on 8x GB200 with NVFP4 at day 0 — zhyncs42 · 2026-10-16