
Scheme	`FP8_DYNAMIC` — weights static per-channel FP8, activations dynamic per-token FP8
Calibration	none needed; dynamic activation scales are computed at inference time
Left in bf16	`lm_head`, and the token embeddings
Weights before	7.49 GiB
Weights after	4.83 GiB (35% smaller)

The score is read from the yes/no logits of the model head, which is left in bf16. Quantising it would put error directly into the number documents are ranked by.

Quality gate

WikipediaRerankingMultilingual from MTEB — reranking Wikipedia passages in 16 languages, scored by map_at_1000.

language	bf16	FP8	Δ
de	0.9600	0.9594	-0.0006
en	0.9715	0.9726	+0.0011
it	0.9710	0.9702	-0.0008
mean	0.9675	0.9674	-0.0001

Languages evaluated: de, en, it. Tolerance: 0.0100 map_at_1000 per language.

PASSED — no language lost more than 0.0100 map_at_1000.

Why

Serving Qwen/Qwen3-Reranker-4B in bf16 leaves little room for KV cache on a small GPU: the weights take what the cache needs, and the context length has to be cut until it fits. Halving the weights gives that memory back — the same card serves a longer context without any other change.

Serving

bash

vllm serve DCC-BS/Qwen3-Reranker-4B-FP8-Dynamic \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.85

The checkpoint is in compressed-tensors format, so vLLM detects the quantisation from config.json; no extra flag is required.

FP8 arithmetic is native on Ada and Hopper (compute capability 8.9+). On Ampere it runs through Marlin: the memory saving still applies, the speed is roughly unchanged.