collect the configs for rocblas miopen and tensile

onnxruntime/python/tools/transformers/models/llama2/llama-v2.py

@@ -290,7 +290,7 @@ def run_generate(args, local_rank):
         _run_gen(args, torch_model, tokenizer, input_ids, attention_mask, "Torch")
 
 
-def func_benchmark(fn, warm=5, steps=10):
+def func_benchmark(fn, warm=1, steps=1):
     for _ in range(warm):
         torch.cuda.nvtx.range_push("gen warmup")
         fn()
@@ -322,9 +322,9 @@ def _run_bmk(args, model, name):
     # print(tokenizer.pad_token_id, tokenizer.eos_token_id)
 
     batch = 1
-    prompt_lens = [32, 64, 128, 256, 512, 1024, 2048]
+    prompt_lens = [32, 64, 128, 256, 512, 1024, 2048, 3072, 4096]
     # prompt_lens = [32]
-    generate_lens = [1, 33]
+    generate_lens = [1, 2]
 
     if args.custom_gen:
         print_out("[benchmark] Using custom_generate")

onnxruntime/python/tools/transformers/models/llama2/llama2_config.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+N_GPU=1
+
+bash sample_run.sh $N_GPU --export --optimize --merge
+
+MODEL_NAME=ort-llama2-7b-${N_GPU}gpu
+LOGFILE=$MODEL_NAME.log
+
+export ROCBLAS_LAYER=2
+bash sample_run.sh $N_GPU --optimize --merge --custom-gen --benchmark --ort 2>&1 | tee ${MODEL_NAME}_rocblas_configs.log
+sed -i '/\[[0-9]+,[0-9]+\]<stderr>:/d' ${MODEL_NAME}_rocblas_configs.log
+grep -o "rocblas-bench.*" ${MODEL_NAME}_rocblas_configs.log | sort -u &> unique_rocblas_configs_$MODEL_NAME.log
+unset ROCBLAS_LAYER
+
+export MIOPEN_ENABLE_LOGGING_CMD=1
+bash sample_run.sh $N_GPU --optimize --merge --custom-gen --benchmark --ort 2>&1 | tee ${MODEL_NAME}_miopen_configs.log
+sed -i '/\[[0-9]+,[0-9]+\]<stderr>:/d' ${MODEL_NAME}_miopen_configs.log
+grep "MIOpenDriver " ${MODEL_NAME}_miopen_configs.log | sed -e 's/.*]//' | sort -u &> unique_miopen_configs_$MODEL_NAME.log
+unset MIOPEN_ENABLE_LOGGING_CMD
+
+export TENSILE_DB=0x8000  # dump Tensile kernel names
+bash sample_run.sh $N_GPU --optimize --merge --custom-gen --benchmark --ort 2>&1 | tee ${MODEL_NAME}_tensile_configs.log
+sed -i '/\[[0-9]+,[0-9]+\]<stderr>:/d' 
+grep "Running kernel: " ${MODEL_NAME}_tensile_configs.log | sort -u &> unique_kernel_names_$MODEL_NAME.log
+unset TENSILE_DB
+
+# export HIPBLASLT_LOG_LEVEL=2
+# unset HIPBLASLT_LOG_LEVEL
+
+echo "========================Math lib profiling done"
+
+export RCCL_MSCCL_ENABLE=0
+NCCL_LOGFILE=$MODEL_NAME-NCCL.log
+HSA_FORCE_FINE_GRAIN_PCIE=1 NCCL_DEBUG=INFO NCCL_DEBUG_SUBSYS=INIT,COLL ./sample_run.sh $N_GPU --optimize --merge --custom-gen --benchmark --ort 2>&1 | tee $NCCL_LOGFILE
+python /workspace/nccl-rccl-parser/rccl_nccl_parser.py --nccl-debug-log $NCCL_LOGFILE --output-script-name unique_nccl_configs_$MODEL_NAME.log --unique

onnxruntime/python/tools/transformers/models/llama2/parse_log.py

@@ -0,0 +1,43 @@
+import os
+import glob
+
+logList = glob.glob('ort-llama2*gpu.log')
+print(logList)
+
+for logFile in logList[:1]:
+    print("============================")
+    print("processing", logFile)
+    rocblasLines=set()
+    miopenLines=set()
+    tensileLines=set()
+
+    with open(logFile) as file:
+        for line in file:
+            # print(line)
+            if "rocblas-bench" in line:
+                rocblasLines.add(line[14:])
+            if "MIOpenDriver" in line:
+                miopenLines.add(line[14:])
+            if "Running kernel" in line:
+                tensileLines.add(line[14:])
+
+    print(len(rocblasLines))
+    print(len(miopenLines))
+    print(len(tensileLines))
+
+    rocblasLines = list(rocblasLines)
+    print(rocblasLines)
+    miopenLines = list(miopenLines)
+    tensileLines = list(tensileLines)
+
+    rocblasFileName = "unique_rocblas_configs_" + logFile
+    miopenFileName = "unique_miopen_configs_" + logFile
+    tensileFileName = "unique_tensile_configs_" + logFile
+
+    # outFiles = [rocblasFileName, miopenFileName, tensileFileName]
+    # outLines = [rocblasLines, miopenLines, tensileLines]
+
+    # for outFileName, outLine in zip(outFiles, outLines):
+    #     with open(outFileName, 'w') as f:
+    #         for line in outLine:
+    #             f.write(f"{line}")

onnxruntime/python/tools/transformers/models/llama2/sample_run.sh

@@ -12,10 +12,10 @@ MPI="mpirun --allow-run-as-root
 MPI+=" -x TRANSFORMERS_CACHE=/hf_cache"
 MPI+=" -x NCCL_DEBUG=VERSION"
 
-MODEL_NAME=${LLAMA2_MODEL_NAME:-"Llama-2-70b-hf"}
-OUTPUT=$MODEL_NAME
+MODEL_NAME=${LLAMA2_MODEL_NAME:-"llama-2-7b-chat-hf"}
+OUTPUT=/data/onnx/$MODEL_NAME
 
-CMD="$MPI python llama-v2.py --model=meta-llama/$MODEL_NAME --output-name=$OUTPUT ${@:2}"
+CMD="$MPI python llama-v2.py --model=/data/$MODEL_NAME --output-name=$OUTPUT ${@:2}"
 
 set -x
 $CMD