license: mit
base_model:
- intfloat/multilingual-e5-large-instruct
tags:
- llmcompressor
- compressed-tensors
- quantized
- nvfp4
- vllm
pipeline_tag: feature-extraction
library_name: vllm
multilingual-e5-large-instruct-nvfp4
NVFP4 quantization of intfloat/multilingual-e5-large-instruct, produced with
llm-compressor 0.13.0 and measured end to end on
NVIDIA H100 80 GB with vLLM 0.26.0.
Only the Linear layers of the encoder are quantized; the vocabulary, position and token-type embeddings, the pooler and any classification head stay in bf16.
Measured
| variant | size | SciFact nDCG@10 | STS-B de ρ | STS17 en-de ρ | texts/s (bs=256) |
|---|---|---|---|---|---|
| bf16 source | 1068 MiB | 0.7204 | 0.8347 | 0.8525 | 1488 |
| FP8 | 802 MiB | 0.7164 | 0.8349 | 0.8526 | 1511 |
| NVFP4 | 675 MiB | 0.7036 | 0.8371 | 0.8440 | – |
Retrieval costs 1.7 pp nDCG@10 against bf16; STS is unchanged. Size is −37%.
SciFact: 300 queries over 5183 documents, nDCG@10. STS-B de: mteb/stsb_multi_mt (1379 pairs). STS17 en-de: mteb/sts17-crosslingual-sts (250 pairs). Reranking: 1 positive + 19 sampled negatives per query. Throughput over 4096 multilingual texts through the vLLM OpenAI endpoint.
A bug you should know about if you quantize BERT-family models yourself
The first NVFP4 build of this model scored 0.042 nDCG@10 in vLLM — random-level — while the same weights scored 0.66 under plain transformers. The weights were fine; the checkpoint was not.
For NVFP4, every linear carries its own weight_global_scale. vLLM fuses query/key/value into one
qkv_proj and then collapses their scales with weight_global_scale.max()
(compressed_tensors_w4a4_nvfp4.py:111), silently rescaling two of the three projections. It warns
("the weight global scale is different for parallel layers") but continues.
llm-compressor is supposed to prevent this: fuse_weight_observers() links observers across fused
groups so they share a global scale. But its FUSED_LAYER_NAMES
(llmcompressor/observers/helpers.py:162) only lists decoder naming — q_proj/k_proj/v_proj,
gate_proj/up_proj, w1/w3. BERT and RoBERTa call them query/key/value, so the group is
never matched and each projection keeps its own scale.
This checkpoint was built with ("query", "key", "value") appended to that list. Layer 0 q/k/v
global scales went from {query: 5344, key: 6048, value: 7904} to a shared 5344, vLLM's warning
disappeared, and nDCG@10 went from 0.042 to 0.704.
Serving
vllm serve mbehr90/multilingual-e5-large-instruct-nvfp4 --runner pooling