-
LFM2.5-Audio-1.5B ONNX Vulkan deck profile (Tier-1): q4 lm_head decoder + exact-bf16 host tables
Derived from LiquidAI/LFM2.5-Audio-1.5B-ONNX @ 62318d95 (byte-pinned stock files). Tier-1 changes: (1) decoder_q4 lm_head f32 table 536,870,912 B -> symmetric q4 MatMulNBits 83,886,080 B; (2) host embedding tables f32 -> native bfloat16 (exact). Validation: release-convention check, CPU ORT greedy parity 8/8 prompts bit-identical vs stock, instruct 21/21 unchanged, GPU showcase 14/14 ALL-PASS. LFM Open License v1.0 (inherited).
