tmp: bert npu run

furiosa-ai · Jul 22, 2024 · 034ae10 · 034ae10
1 parent 238106e
commit 034ae10
Show file tree

Hide file tree

Showing 6 changed files with 17 additions and 7 deletions.
diff --git a/language/bert/RNGD_SUT.py b/language/bert/RNGD_SUT.py
@@ -92,6 +92,8 @@ def __init__(self, args):
                 args.quant_param_path,
                 args.quant_format_path,
             )
+            for n, p in self.model.named_parameters():
+                print(n, p)
         else:
             self.model = self.model.trace()
 
@@ -141,11 +143,15 @@ def process_sample(self, sample_input, query_id=None):
             self.dump.update({"input": query})
 
         with torch.no_grad():
+            print(f"{input_ids=}")
+            print(f"{input_mask=}")
+            print(f"{segment_ids=}")
             model_output = self.encoder.encode(
                 input_ids=torch.LongTensor(input_ids).unsqueeze(0).to(self.dev),
                 attention_mask=torch.LongTensor(input_mask).unsqueeze(0).to(self.dev),
                 token_type_ids=torch.LongTensor(segment_ids).unsqueeze(0).to(self.dev),
             )
+            print("model_output", model_output)
             if self.dump_path:
                 assert len(model_output) == 1
                 self.dump.update({"output": {"output_ids": model_output[0].tolist()}})

diff --git a/language/bert/RNGD_encoder.py b/language/bert/RNGD_encoder.py
@@ -59,9 +59,9 @@ def encode(
             for arg in self.model.concrete_args:
                 if arg in model_kwargs:
                     del model_kwargs[arg]
-
+        print(model_kwargs)
         logits = self.model(**model_kwargs)
-
+        print(logits)
         outputs = []
         for batch_index, target_location in enumerate(packed_target_locations):
             for single_target_location in target_location:

diff --git a/language/bert/run.py b/language/bert/run.py
@@ -30,7 +30,7 @@
 def get_args():
     parser = argparse.ArgumentParser()
     parser.add_argument(
-            "--backend", choices=["tf", "pytorch", "onnxruntime", "tf_estimator", "ray", "rngd"], default="tf", help="Backend")
+            "--backend", choices=["tf", "pytorch", "onnxruntime", "tf_estimator", "ray", "rngd", "rngd-npu"], default="tf", help="Backend")
     parser.add_argument("--scenario", choices=["SingleStream", "Offline",
                                                "Server", "MultiStream"], default="Offline", help="Scenario")
     parser.add_argument("--accuracy", action="store_true",
@@ -102,6 +102,9 @@ def main():
             assert not args.profile, "Profiling is only supported by onnxruntime backend!"
             from RNGD_SUT import get_rngd_sut
             sut = get_rngd_sut(args)
+        elif args.backend == "rngd-npu":
+            from RNGD_NPU_SUT import get_rngd_npu_sut
+            sut = get_rngd_npu_sut(args)
         else:
             raise ValueError("Unknown backend: {:}".format(args.backend))
 

diff --git a/scripts/build_rngd_env.sh b/scripts/build_rngd_env.sh
@@ -108,7 +108,7 @@ install_mlperf_loadgen() {
 install_pip_packages() {
     echo "Installing DVC with S3 support..."
     pip install 'dvc[s3]'
-    pip install datasets==2.18.0 nltk==3.8.1 evaluate==0.4.1 absl-py==2.1.0 rouge_score==0.1.2
+    pip install datasets==2.18.0 nltk==3.8.1 evaluate==0.4.1 absl-py==2.1.0 rouge_score==0.1.2 tokenization==1.0.7
 }
 
 # https://www.notion.so/furiosa/RNGD-SW-Stack-runtime-7df73fb4d92241e09a2721612ebd9c3d?pvs=4#847dd27d9857405a8b3c26b4538c3656

diff --git a/scripts/eval_mlperf_gptj.sh b/scripts/eval_mlperf_gptj.sh
@@ -98,11 +98,11 @@ run_multi_device_eval() {
     echo -e "\nAll evaluations completed.\n"
 
     if [ "$N_DEVICES" == "$N_PARTITIONS" ]; then
-        python "$work_dir/gather_log_accuracy.py" --log-dir="$LOG_PATH"
+        python "$work_dir/gather_log.py" "$LOG_PATH" -p "$N_PARTITIONS" -n "$N_COUNT"
         MLPERF_ACCURACY_FILE="$LOG_PATH/merged_mlperf_log_accuracy.json"
 
         if [ "$DO_DUMP" = true ]; then
-            python "$work_dir/gather_generator_dump.py" --log-dir="$LOG_PATH"
+            python "$work_dir/gather_log.py" "$LOG_PATH" -p "$N_PARTITIONS" -n "$N_COUNT"
         fi
     else
         SKIP_VERIFY_ACCURACY=true
@@ -162,6 +162,7 @@ LOG_PATH=${LOG_PATH:="$log_dir/$model_name/$SCENARIO/$(date +%Y%m%d_%H%M%S%Z)"}
 DEVICE=${DEVICE:="npu"}
 DEVICE_NUM=${DEVICE_NUM:="0"}
 DEVICES=${DEVICES:="$DEVICE:$DEVICE_NUM:0-3,$DEVICE:$DEVICE_NUM:4-7"}
+DEVICES="npu:1:0-3,npu:1:4-7"
 N_COUNT=${N_COUNT:="13368"}
 N_DEVICES=${N_DEVICES:="1"}
 N_PARTITIONS="${N_PARTITIONS:-$N_DEVICES}"

diff --git a/scripts/eval_qbert.sh b/scripts/eval_qbert.sh
@@ -74,7 +74,7 @@ python $work_dir/accuracy-squad.py --vocab_file=$VOCAB_PATH \
                          --out_file=$LOG_PATH/predictions.json \
                          --max_examples=$N_COUNT \
                          &> $LOG_PATH/accuracy_result.log
-
+cat $LOG_PATH/accuracy_result.log
 printf "Save evaluation log to $LOG_PATH"
 
 unset LOG_PATH