Fix parallel cpu kernel

quantization.py

@@ -89,7 +89,7 @@ class W8A16LinearCPU(torch.autograd.Function):
 default_cpu_kernel_code_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "quantization_kernels.c")
 default_cpu_kernel_code = "QlpoOTFBWSZTWXLbSoQAAgzbgERwQXxmTwAAr/ff3kABt0Q2oRVT0hpo9RtEAAAAyBEiSQ9EGjQGQAAAwANGhowjJoNGmgMEUplMTNSMJ5TQaDJpsoMyRMj8P4mZzFSVVwqSXG8GG7MlVwiToYEQwVD7noBxMhNfkeZYtYFtbgOBUSIGtIQjhNHCEnPJsadhb3yBmRIOD3TeAtNLSaU5GgvKUBWSNuuOIHmVt0YhW6rsmDMDUjeUJGJ64R1Jm5lrh0Aa0tKjhFwPdWcGogxLDSXPWQUWTM8Sd3Qz1HMYNxx3HMeiNqNo4jeRDEfZ3gUSHIcU/heomq0vEzL1Msz5KKGxH8FrNOYw3KaxdqaEmNHYMxJFgQbR0DyRknL2L4kwUSxKRdhjRpEtUqilVfggFL1klaMS3PPRDfNqbBOPWO7m4JTVGhS9QTBDDJaEbLbrUQNB+IpJSKQbG5SZZ5gkwJEhJ3aYKJipZ/i7kinChIOW2lQg"
 default_cpu_parallel_kernel_code_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "quantization_kernels_parallel.c")
-default_cpu_parallel_kernel_code = "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"
+default_cpu_parallel_kernel_code = "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"
 
 cpu_kernels = None
 

quantization_kernels_parallel.c

@@ -17,7 +17,7 @@ void compress_int4_weight(void *weight, void *out, int n, int m)
     {
         for(int j=0;j<m;j++)
         {
-            (*(unsigned char*)(out + sizeof(unsigned char) * (i * m + j))) |= ((*(unsigned char*)(weight + sizeof(unsigned char) * (i * (m << 1) + (j << 1)))) << 4);
+            (*(unsigned char*)(out + sizeof(unsigned char) * (i * m + j))) = ((*(unsigned char*)(weight + sizeof(unsigned char) * (i * (m << 1) + (j << 1)))) << 4);
             (*(unsigned char*)(out + sizeof(unsigned char) * (i * m + j))) |= (((*(unsigned char*)(weight + sizeof(unsigned char) * (i * (m << 1) + ((j << 1) | 1)))) & 15));
         }
     }