Using the FP16-accumulator FP8 matmul instruction with narrowed quantization ranges for P and V yields up to 3.9x speedup over FlashAttention2 with negligible accuracy loss.
Infllm: Training-free long-context extrapolation for llms with an efficient context memory
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
SageAttention2++: A More Efficient Implementation of SageAttention2
Using the FP16-accumulator FP8 matmul instruction with narrowed quantization ranges for P and V yields up to 3.9x speedup over FlashAttention2 with negligible accuracy loss.