TenStrip

/

LTX2.3-10Eros_Version-Testing

Like 128

Beta7 W4A8 conservative quant by cicalooo, made with cicalooo's h3quant v2 script (his default profile conservative-w4a8, group size 16), uploaded on his behalf by sandpies while he's away.

• int8 ConvRot (g256) on all block linears, W4A8 g16 codebook on mlp.fc1 of blocks 10-37 (172 int8 / 28 W4A8 layers); everything else kept byte-identical bf16
• 19.08 GB (19,079,401,912 bytes), aimed at RTX 3090 / 24 GB cards
• Script's --verify passed: int8 rel err ~0.009-0.012, W4 fc1 ~0.073
• Run on Windows with I/O-only portability shims; quant math unchanged

All credit for the quant method to cicalooo.