v1.4.1

1.4.1 (2026-04-04)

Bug Fixes

• kv_cache: pass both key/value states to lazy_initialization for transformers 5.x (#68) (bee2071), closes #66
• quantizer: add dimension guards for clear errors on GQA shape mismatch (#69) (4271915), closes #67
• test: add create=True to flash_attn_varlen_func mocks for vllm 0.19 (006d057)
• verify: handle num_kv_shared_layers in cache layer iteration (#71) (006d057), closes #70
• verify: pass HF_TOKEN to from_pretrained for gated repos (#65) (9d19b87)

Documentation

• readme: update for v1.4.0 — asymmetric K/V, 8-model quality matrix, serving benchmarks (#63) (6af1b0b)