Fix short-conv padding masking on transformers >=5

Update README.md

config: dtype bfloat16

cast dense head to bf16

cast weights to bf16 (lossless from training precision)

Update README.md

readme: stack GPU plot above the latency table instead of side-by-side

Update README.md

readme: split inference-speed sections (llama.cpp vs Enterprise GPU)

readme: add Enterprise GPU serving section (image + p50/p95/p99 table)

Update README.md

Update README.md

readme: speed table — remove bold and add 32q/256d setup note

readme: add inference-speed table (llama.cpp, M4 Max, fp16)

Update README.md

Update README.md

Update README.md

Update README.md

Update README.md

Create LICENSE

Update README.md

Update README.md

Update README.md

Update README.md

fix: accept input_embeds/inputs_embeds kwarg rename across transformers versions

readme: drop trust_remote_code callout

readme: enrich How-to-run with the original LFM2-ColBERT-350M prose + step-comments

Update README.md

Update README.md

Update README.md

Update README.md

Update README.md

readme: add sales contact link

readme: add LFM2 tech report citation

readme: label second benchmark as MKQA Recall@20

readme: add NanoBEIR Multilingual Extended benchmark tables (NDCG@10 + second metric)

readme: swap banner image to 2b08LKpev0DNEk6DlnWkY.png

readme: move use cases below the specs table and module repr

readme: drop 'this model' qualifiers; move RAG use cases under Model details

readme: shorter intro + side-by-side specs table; drop arch/training (will live in blog)

readme: intro both models; mark shared architecture/training sections

readme: full model card with bidi/LFM2.5 details and PyLate usage

readme: update repo id to LFM2.5-ColBERT-350M

Add FA2-capable modeling file but disable flash_attention_2 via config (incompatible with PyLate query expansion; ~10 ndcg@10 loss)

Update weights to checkpoint train_1811379_colbert-ft-s2/best (multilingual NanoBEIR ndcg@10 0.5854 -> 0.6044)

initial release: LFM2 ColBERT (bidi-patched), from train_1369809/best

initial commit