Model techniques map
Techniquesinference & servinginference quantization

specific method · filed under inference & serving

FP4 KV-cache quantization

Stores the main key-value cache in FP4 to reduce cache storage, with one cited implementation using E2M1 values and per-16-channel E4M3 scales.

Also called FP4 main KV cache, FP4 KV caching, FP4 main KV caching, FP4-quantized main key-value cache, 4-bit floating-point key-value caching, FP4 main KV-cache compression.

sources
2
model
1
lab adopt it
1
strongest
core

How sources treat it

One count per evidence span, weakest treatment to strongest.

used 1core 4

Documented in

Evidence

5 spans quoted from the sources, strongest treatment first.

we use FP4 global KV caches during training with only marginal performance degradation. Together, CSA2 and FP4 KV caching reduce global KV cache storage to approximately 1/4 of that of DeepSeek-V4-Flash.

coremodel architecturein DeepSeek-V4.1-FlashDeepSeek

We now extend QAT to the main KV cache, where FP4 reduces storage rather than accelerates matrix multiplication.

coreunclearin DeepSeek-V4.1-FlashDeepSeek

cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) and FP4 KV caching reduce its global KV cache footprint

coreinference servingin DeepSeek-V4.1-FlashDeepSeek

Combined with FP4 main KV caching (E2M1 format, one E4M3 scale per 16 channels), these designs reduce the global KV cache footprint to 890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash.

coremodel architecturein DeepSeek-V4.1-FlashDeepSeek

we compress the main KV cache to FP4 to further reduce storage overhead.

usedinference servingin DeepSeek-V4.1-FlashDeepSeek

Filed alongside

Other methods under inference & serving :: inference quantization.