NeuronORPOTrainer

benchmark/text-generation-inference/performance/generate_csv.py

@@ -3,7 +3,6 @@
 import os
 
 import pandas as pd
-
 from guidellm.core import GuidanceReport, TextGenerationBenchmark
 
 
@@ -16,11 +15,7 @@ def _benchmark_rate_id(benchmark: TextGenerationBenchmark) -> str:
     :return: A string representing the benchmark rate ID.
     :rtype: str
     """
-    rate_id = (
-        f"{benchmark.mode}@{benchmark.rate:.2f} req/sec"
-        if benchmark.rate
-        else f"{benchmark.mode}"
-    )
+    rate_id = f"{benchmark.mode}@{benchmark.rate:.2f} req/sec" if benchmark.rate else f"{benchmark.mode}"
     return rate_id
 
 
@@ -38,20 +33,20 @@ def main():
     for path in paths:
         filename = os.path.basename(path)
         # Extract model_id
-        model_id, date = filename.replace(suffix, '').split('#')
+        model_id, date = filename.replace(suffix, "").split("#")
         with open(path) as f:
             report = GuidanceReport.from_json(f.read())
             for benchmark in report.benchmarks:
                 for b in benchmark.benchmarks_sorted:
                     d = {
-                    "model_id": model_id,
-                    "Date": date,
-                    "Input type": _benchmark_rate_id(b),
-                    "Requests per Second": b.completed_request_rate,
-                    "Request Latency (s)": b.request_latency,
-                    "Time-to-first-token (ms)": b.time_to_first_token,
-                    "Inter Token Latency (ms)": b.inter_token_latency,
-                    "Output Token Throughput (t/s)": b.output_token_throughput,
+                        "model_id": model_id,
+                        "Date": date,
+                        "Input type": _benchmark_rate_id(b),
+                        "Requests per Second": b.completed_request_rate,
+                        "Request Latency (s)": b.request_latency,
+                        "Time-to-first-token (ms)": b.time_to_first_token,
+                        "Inter Token Latency (ms)": b.inter_token_latency,
+                        "Output Token Throughput (t/s)": b.output_token_throughput,
                     }
                     results.append(pd.DataFrame.from_dict(d, orient="index").transpose())
 

optimum/neuron/__init__.py

@@ -27,7 +27,7 @@
 
 _import_structure = {
     "hf_argparser": ["NeuronHfArgumentParser"],
-    "trainers": ["NeuronTrainer", "Seq2SeqNeuronTrainer", "NeuronSFTTrainer"],
+    "trainers": ["NeuronTrainer", "Seq2SeqNeuronTrainer", "NeuronSFTTrainer", "NeuronORPOTrainer"],
     "training_args": ["NeuronTrainingArguments", "Seq2SeqNeuronTrainingArguments"],
     "modeling_traced": ["NeuronTracedModel"],
     "modeling": [
@@ -69,7 +69,7 @@
         "ModelParallelismPlugin",
     ],
     "pipelines": ["pipeline"],
-    "utils": ["NeuronSFTConfig", "get_peft_model"],
+    "utils": ["NeuronSFTConfig", "NeuronORPOConfig", "get_peft_model"],
 }
 
 if TYPE_CHECKING:
@@ -109,9 +109,9 @@
     from .modeling_seq2seq import NeuronModelForSeq2SeqLM
     from .modeling_traced import NeuronTracedModel
     from .pipelines import pipeline
-    from .trainers import NeuronSFTTrainer, NeuronTrainer, Seq2SeqNeuronTrainer
+    from .trainers import NeuronORPOTrainer, NeuronSFTTrainer, NeuronTrainer, Seq2SeqNeuronTrainer
     from .training_args import NeuronTrainingArguments, Seq2SeqNeuronTrainingArguments
-    from .utils import NeuronSFTConfig, get_peft_model
+    from .utils import NeuronORPOConfig, NeuronSFTConfig, get_peft_model
 
 else:
     import sys