Fix missing scale attributes for GPTJ (#3256)

Co-authored-by: NOlatunji Ruwase <olruwase@microsoft.com>
Co-authored-by: NMichael Wyatt <michaelwyatt@microsoft.com>

csrc/transformer/inference/csrc/pt_binding.cpp

@@ -462,9 +462,9 @@ std::vector<at::Tensor> ds_softmax_context(at::Tensor& query_key_value,
 
     T* workspace = (T*)InferenceContext::Instance().GetWorkSpace();
     size_t buf_size = bsz * seq_len * hidden_dim;
-    auto output = torch::from_blob(workspace + 3 * buf_size, {bsz, seq_len, hidden_dim}, options);
+    auto output = torch::from_blob(workspace + 4 * buf_size, {bsz, seq_len, hidden_dim}, options);
 
-    auto query_cont = workspace + 4 * buf_size;
+    auto query_cont = workspace + 5 * buf_size;
     size_t offset =
         10 * (hidden_dim * bsz * InferenceContext::Instance().GetMaxTokenLenght()) +
         layer_id * 2 * bsz * InferenceContext::Instance().GetMaxTokenLenght() * hidden_dim;

deepspeed/ops/transformer/inference/op_binding/gelu_gemm.py

@@ -23,7 +23,9 @@ class GELUGemmOp(BaseOp):
                 bias: torch.Tensor,
                 weight_out: torch.Tensor,
                 async_op: bool = False):
-        output = self.fused_gemm_gelu(input, weight, weight.scale, bias, weight_out, weight_out.scale,
+        output = self.fused_gemm_gelu(input, weight, weight.scale if hasattr(weight, "scale") else torch.empty(1),
+                                      bias, weight_out,
+                                      weight_out.scale if hasattr(weight_out, "scale") else torch.empty(1),
                                       self.config.epsilon, self.config.pre_layer_norm, self.config.q_int8, async_op,
                                       self.config.transposed_mode)
         return output