Lynx partitions KV cache bits into anchor and residual streams for progressive transfer, enabling speculative decoding on partial data followed by verification to match BF16 accuracy at 4-bit-like TTFT.
2025.COMET: Towards Practical W4A4KV4 LLMs Serving
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.DC 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Lynx: Progressive Speculative Quantization for accelerating KV Transfer in Long-Context Inference
Lynx partitions KV cache bits into anchor and residual streams for progressive transfer, enabling speculative decoding on partial data followed by verification to match BF16 accuracy at 4-bit-like TTFT.