modify cpp_benchmark.sh

benchmarks/gpt/cpp_benchmark.sh

@@ -33,20 +33,21 @@ echo -e "| model size | Batch Size | Input length | Output length | Decode value
 echo -e "|:----------:|:----------:|:------------:|:-------------:|:------------:|:---------:|:---------------:|" >> $all_log
 
 cat /proc/cpuinfo > ${logdir}/cpuinfo.txt
-nvidia-smi > ${logdir}/gpuinfo.txt
+rocm-smi > ${logdir}/gpuinfo.txt
 
-for model_size in "345m" "5b";
+# for model_size in "345m" "5b";
+for model_size in "345m";
 do
     if [ "$model_size" = "345m" ]; then
         head_num=16
         size_per_head=64
         inter_size=$(echo "scale=2; $head_num * ${size_per_head} * 4 " | bc)
         num_layer=24
-    elif [ "$model_size" = "5b" ]; then
-        head_num=32
-        size_per_head=128
-        inter_size=$(echo "scale=2; $head_num * ${size_per_head} * 4 " | bc)
-        num_layer=24
+    # elif [ "$model_size" = "5b" ]; then
+    #     head_num=32
+    #     size_per_head=128
+    #     inter_size=$(echo "scale=2; $head_num * ${size_per_head} * 4 " | bc)
+    #     num_layer=24
     fi
 
 for decode_type in "beamsearch" "sampling";
@@ -93,7 +94,7 @@ python ../examples/pytorch/gpt/utils/generate_gpt_config.py \
                                         --size_per_head ${size_per_head} \
                                         --inter_size ${inter_size} \
                                         --num_layer ${num_layer} \
-                                        -v 51200 \
+                                        -v 50304 \
                                         -d fp16 \
                                         -topk ${topk} \
                                         -topp ${topp} \