Model techniques map
Techniquesinference & servinginference quantization

specific method · filed under inference & serving

FP8 KV-cache quantization

Stores key-value cache data in FP8; the evidence includes E4M3 and amax-scaled variants but does not establish them as a single more specific scheme.

Also called FP8 KV cache during RL rollouts, FP8 E4M3 key-value cache, fp8_e4m3 KV cache, FP8 key-value cache, FP8 KV Cache, --kv-cache-dtype fp8.

sources
13
models
10
labs adopt it
7
strongest
used

How sources treat it

One count per evidence span, weakest treatment to strongest.

optional 4used 11

Documented in

Further reading

Picked by hand, not extracted: where to read more, not evidence for anything on this page.

Evidence

15 spans quoted from the sources, strongest treatment first.

with FP8 KV cache — reaches the full 1M-token context on 8xB200.

usedinference servingin GLM-5.3vLLM

FP8 precision is applied to the remaining dimensions

usedinference servingin DeepSeek-V4DeepSeek

--kv-cache-dtype fp8

usedunclearin Qwen3.6-35B-A3BQwen

storing the KV cache in FP8 roughly doubles the number of concurrent trajectories a single replica can carry

usedinference servingin LAGUNA M.1/XS.2Poolside

The KV cache was quantized to FP8 using amax-based scaling with a calibration dataset of 128 long-contexts agentic traces.

usedpost trainingin Laguna XS.2Poolside

Lower memory per token: KV cache quantized to FP8.

usedinference servingin Laguna XS 2.1Poolside

kv-cache-dtype fp8

usedinference servingin Nemotron 3 UltraNVIDIA

KV Cache is quantized to FP8

usedinference servingin Nemotron 3 UltraNVIDIA

requires modelopt activation and FP8 KV Cache alignment

usedinference servingin Step 3.7 FlashStepFun

--kv-cache-dtype fp8_e4m3

usedinference servingin Step 3.7 FlashStepFun

--kv-cache-dtype fp8

usedinference servingin Step 3.7 FlashStepFun

--kv-cache-dtype fp8

optionalinference servingin Nemotron 3 UltraNVIDIA

--kv-cache-dtype fp8

optionalinference servingin DeepSeek-V4-Flash-Vision-ExpDeepSeek

On Blackwell you can add --kv-cache-dtype fp8 to both pools; Hopper does not support FP8 KV cache for this model and must run BF16 KV.

optionalinference servingin GLM-5.3-FlashZ.ai

--kv-cache-dtype fp8_e4m3

optionalinference servingin vLLMvLLM

Filed alongside

Other methods under inference & serving :: inference quantization.