disable_saturate=True, ) bias_quant = bias_quantizer.quantize(bias, name=bias.name) bias_quant = bias_quantizer.quantize(bias, name=bias_name) updated_tensors[bias_quant.name] = bias_quant # Spot check that things look sane. bias_dequant = bias_quant.unpack().dequant() bias_diff = ...
bias_after_scale=True, ) qk_matmul_op = OpConfig( "matmul_v2", inputs={"X": ["scale_out"], "Y": ["transpose2_2_out"]}, outputs={"Out": ["qk_matmul_out"]}, trans_x=False, trans_y=False, ) qk_softmax_op = OpConfig( "softmax", inputs={"X": ["qk_matmul_out"]}...
Fine-tuning RWKV-4 Pile models:use 'prepare-data.py' inhttps://github.com/BlinkDL/RWKV-v2-RNN-Pile/tree/main/RWKV-v3to tokenize .txt into train.npy data. Then set EXPRESS_PILE_MODE to True in train.py, and run it. Read the inference code in src/model.py and try using the fin...
🐛 Describe the bug torch._transform_bias_rescale_qkv causes FPE with specific input. Test code: import torch qkv = torch.full((11, 0, 4, 0, 0, 5, 6, 8, 0, 10, 0, 0, 10, 0, 0, 3, 12, 15, 0, 11,), -1.5e+300, dtype=torch.float64, requires_g...