diff --git a/cutedsl/bridge.py b/cutedsl/bridge.py index 06464bc2..dfbdd521 100644 --- a/cutedsl/bridge.py +++ b/cutedsl/bridge.py @@ -163,7 +163,10 @@ def quantize_activation_nvfp4(x_bf16, global_scale, block_size=SF_VEC_SIZE): sf_shape = list(x_bf16.shape[:-1]) + [n_blocks] block_scale = block_scale.reshape(sf_shape) - return x_fp4, block_scaledef quantize_weight_to_nvfp4(w_bf16, block_size=SF_VEC_SIZE): + return x_fp4, block_scale + + +def quantize_weight_to_nvfp4(w_bf16, block_size=SF_VEC_SIZE): """Quantize BF16 weight matrix to NVFP4. The weight is (K, N) where K is the input dim (packed dimension).