Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.
Python ai-filmmakingai-videoai-video-generatorapple-siliconhailuoimage-to-videolip-synclocal-ailora-trainingltx-2ltx-2-5ltx-videomacosminimax-h3mlxmusic-generationoffline-aipinokiotext-to-videovideo-generation
loras: the probe stops shouting at the person it is meant to help
mrbizarro69ee0b4f08ce
diff --git a/lora_compat.py b/lora_compat.pyindex 29a3bd8..12888ac 100644--- a/lora_compat.py+++ b/lora_compat.py@@ -378,9 +378,18 @@ def measure_adapter_effect(continueif a.shape[0] != b.shape[1]:continue # rank disagreement: a shape problem, not a size one- gram = b.T @ b- cov = a @ a.T- frob = float(np.sqrt(max(float((gram * cov.T).sum()), 0.0)))+ # Accelerate's BLAS raises spurious FPE flags on Apple Silicon —+ # numpy turns them into "divide by zero"/"overflow encountered in+ # matmul" on inputs that are entirely finite. Measured: identical+ # results with and without. A diagnostic that shouts three+ # warnings before its verdict is a diagnostic nobody trusts.+ with np.errstate(all="ignore"):+ gram = b.T @ b+ cov = a @ a.T+ total = float((gram * cov.T).sum())+ if not (total == total): # NaN: a product we cannot stand behind+ continue+ frob = float(np.sqrt(max(total, 0.0)))entries = float(a.shape[1]) * float(b.shape[0])value = frob / (entries**0.5) if entries else 0.0rms.append(value)