tensormux

2. Isolate the quantization step responsible. The quantization pipeline is: fp32 input → quantize → int8 input int8 GEMM (or other op) with int32 accumulation int32 accumulation → dequantize → fp32/fp16 output Test each boundary: - Quant-dequant roundtrip: quantize the fp32 input to INT8, then imme…