Pipeline failure fixes : Fixing Lint failures on test, example and base file

Author: Aishwarya-Tonpe

examples/benchmarks/pytorch_deterministic_example.py

@@ -37,35 +37,33 @@
 ]
 
 DEFAULT_PARAMS = {
-    "bert-large": "--batch_size 1 --seq_len 128 --num_warmup 1 --num_steps 300 --precision float32 "
+    "bert-large":
+    "--batch_size 1 --seq_len 128 --num_warmup 1 --num_steps 300 --precision float32 "
     "--model_action train --deterministic --deterministic_seed 42 --check_frequency 20",
-    "gpt2-small": "--batch_size 1 --num_steps 300 --num_warmup 1 --seq_len 128 --precision float32 "
+    "gpt2-small":
+    "--batch_size 1 --num_steps 300 --num_warmup 1 --seq_len 128 --precision float32 "
     "--model_action train --deterministic --deterministic_seed 42 --check_frequency 20",
-    "llama2-7b": "--batch_size 1 --num_steps 300 --num_warmup 1 --seq_len 512 --precision float32 --model_action train "
+    "llama2-7b":
+    "--batch_size 1 --num_steps 300 --num_warmup 1 --seq_len 512 --precision float32 --model_action train "
     "--deterministic --deterministic_seed 42 --check_frequency 20",
-    "mixtral-8x7b": "--hidden_size=4096 --num_hidden_layers=32 --num_attention_heads=32 --intermediate_size=14336 "
+    "mixtral-8x7b":
+    "--hidden_size=4096 --num_hidden_layers=32 --num_attention_heads=32 --intermediate_size=14336 "
     "--num_key_value_heads=8 --max_position_embeddings=32768 --router_aux_loss_coef=0.02 "
     "--deterministic --deterministic_seed 42 --check_frequency 20",
-    "resnet101": "--batch_size 192 --precision float32 float32 --num_warmup 64 --num_steps 512 --sample_count 8192 "
+    "resnet101":
+    "--batch_size 192 --precision float32 float32 --num_warmup 64 --num_steps 512 --sample_count 8192 "
     "--pin_memory --model_action train --deterministic --deterministic_seed 42 --check_frequency 20",
-    "lstm": "--batch_size 1 --num_steps 300 --num_warmup 1 --seq_len 256 --precision float16 "
+    "lstm":
+    "--batch_size 1 --num_steps 300 --num_warmup 1 --seq_len 256 --precision float16 "
     "--model_action train --deterministic --deterministic_seed 42 --check_frequency 20",
 }
 
 
 def main():
-    parser = argparse.ArgumentParser(
-        description="Unified PyTorch deterministic training example."
-    )
-    parser.add_argument(
-        "--model", type=str, choices=MODEL_CHOICES, required=True, help="Model to run."
-    )
-    parser.add_argument(
-        "--generate-log", action="store_true", help="Enable fingerprint log generation."
-    )
-    parser.add_argument(
-        "--log-path", type=str, default=None, help="Path to save fingerprint log."
-    )
+    parser = argparse.ArgumentParser(description="Unified PyTorch deterministic training example.")
+    parser.add_argument("--model", type=str, choices=MODEL_CHOICES, required=True, help="Model to run.")
+    parser.add_argument("--generate-log", action="store_true", help="Enable fingerprint log generation.")
+    parser.add_argument("--log-path", type=str, default=None, help="Path to save fingerprint log.")
     parser.add_argument(
         "--compare-log",
         type=str,
@@ -92,9 +90,7 @@ def main():
         parameters += f" --compare-log {args.compare_log}"
 
     print(f"Running {args.model} with parameters: {parameters}")
-    context = BenchmarkRegistry.create_benchmark_context(
-        args.model, parameters=parameters, framework=Framework.PYTORCH
-    )
+    context = BenchmarkRegistry.create_benchmark_context(args.model, parameters=parameters, framework=Framework.PYTORCH)
     benchmark = BenchmarkRegistry.launch_benchmark(context)
     print(f"Benchmark finished. Return code: {benchmark.return_code}")
     if hasattr(benchmark, "_model_run_metadata"):

superbench/benchmarks/model_benchmarks/pytorch_base.py

@@ -99,9 +99,7 @@ def _assign_model_run_metadata(self, precision, extra_keys=None):
         # Common metadata keys
         metadata = {
             "model_name": self._name,
-            "precision": (
-                precision.value if hasattr(precision, "value") else str(precision)
-            ),
+            "precision": (precision.value if hasattr(precision, "value") else str(precision)),
             "seed": getattr(self._args, "deterministic_seed", None),
             "batch_size": getattr(self._args, "batch_size", None),
             "seq_len": getattr(self._args, "seq_len", None),
@@ -126,9 +124,7 @@ def _assign_model_run_metadata(self, precision, extra_keys=None):
         self._model_run_metadata = metadata
         return None
 
-    def record_determinism_fingerprint(
-        self, curr_step, loss, logits, periodic, check_frequency
-    ):
+    def record_determinism_fingerprint(self, curr_step, loss, logits, periodic, check_frequency):
         """Centralized logic for recording per-step loss and periodic fingerprints for deterministic runs.
 
         Args:
@@ -144,9 +140,7 @@ def record_determinism_fingerprint(
         except Exception:
             v = None
         # Periodic fingerprint logging
-        if getattr(self._args, "deterministic", False) and (
-            curr_step % check_frequency == 0
-        ):
+        if getattr(self._args, "deterministic", False) and (curr_step % check_frequency == 0):
             # 1) Loss fingerprint (only at fingerprinting frequency)
             try:
                 if "loss" in periodic and v is not None:
@@ -160,8 +154,7 @@ def record_determinism_fingerprint(
                 if logits is not None:
                     act_mean = (
                         float(logits[0].detach().float().mean().item())
-                        if hasattr(logits[0], "detach")
-                        else float(logits[0])
+                        if hasattr(logits[0], "detach") else float(logits[0])
                     )
                     logger.info(f"ActMean at step {curr_step}: {act_mean}")
                     periodic["act_mean"].append(act_mean)
@@ -286,9 +279,7 @@ def _preprocess(self):
                 has_cmp = getattr(self._args, "compare_log", None)
                 if not has_gen and not has_cmp:
                     setattr(self._args, "generate_log", True)
-                    logger.info(
-                        "Deterministic run detected with no log options; defaulting to --generate-log."
-                    )
+                    logger.info("Deterministic run detected with no log options; defaulting to --generate-log.")
         except Exception:
             # Never fail preprocessing due to optional defaulting
             pass
@@ -611,19 +602,11 @@ def _process_info(self, model_action, precision, info):
                 "float64": "fp64",
                 "bfloat16": "bf16",
             }
-            prec_value = (
-                precision.value if hasattr(precision, "value") else str(precision)
-            )
+            prec_value = (precision.value if hasattr(precision, "value") else str(precision))
             prefix = precision_metric.get(prec_value, prec_value)
             metric_loss = f"{prefix}_{model_action}_loss"
-            if (
-                "loss" in info
-                and isinstance(info["loss"], list)
-                and len(info["loss"]) > 0
-            ):
-                self._result.add_raw_data(
-                    metric_loss, info["loss"], self._args.log_raw_data
-                )
+            if ("loss" in info and isinstance(info["loss"], list) and len(info["loss"]) > 0):
+                self._result.add_raw_data(metric_loss, info["loss"], self._args.log_raw_data)
         except Exception as e:
             logger.error(
                 f"Exception in _process_info: {e}\n"

tests/benchmarks/model_benchmarks/test_pytorch_determinism_all.py

@@ -9,7 +9,6 @@
 import pytest
 from superbench.benchmarks import BenchmarkRegistry, Platform, Framework, ReturnCode
 
-
 os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
 
 
@@ -92,9 +91,7 @@ def test_pytorch_model_determinism(model_name, params):
 
     # Run with compare-log for success
     extra_args = f"--compare-log {log_path} --check_frequency 10"
-    benchmark_compare, _ = run_deterministic_benchmark(
-        model_name, params, log_path, extra_args
-    )
+    benchmark_compare, _ = run_deterministic_benchmark(model_name, params, log_path, extra_args)
     assert benchmark_compare and benchmark_compare.return_code == ReturnCode.SUCCESS
 
     os.remove(log_path)
@@ -138,33 +135,19 @@ def test_pytorch_model_nondeterministoc_default(model_name, params):
     )
 
     benchmark = BenchmarkRegistry.launch_benchmark(context)
-    assert (
-        benchmark and benchmark.return_code == ReturnCode.SUCCESS
-    ), "Benchmark did not run successfully."
+    assert (benchmark and benchmark.return_code == ReturnCode.SUCCESS), "Benchmark did not run successfully."
     args = benchmark._args
     assert args.deterministic is False, "Expected deterministic to be False by default."
-    assert (
-        getattr(args, "generate_log", False) is False
-    ), "Expected generate_log to be False by default."
-    assert (
-        getattr(args, "log_path", None) is None
-    ), "Expected log_path to be None by default."
-    assert (
-        getattr(args, "compare_log", None) is None
-    ), "Expected compare_log to be None by default."
-    assert (
-        getattr(args, "check_frequency", None) == 100
-    ), "Expected check_frequency to be 100 by default."
+    assert (getattr(args, "generate_log", False) is False), "Expected generate_log to be False by default."
+    assert (getattr(args, "log_path", None) is None), "Expected log_path to be None by default."
+    assert (getattr(args, "compare_log", None) is None), "Expected compare_log to be None by default."
+    assert (getattr(args, "check_frequency", None) == 100), "Expected check_frequency to be 100 by default."
 
     # Periodic fingerprints exist but are empty when not deterministic
-    assert hasattr(
-        benchmark, "_model_run_periodic"
-    ), "Benchmark missing _model_run_periodic attribute."
+    assert hasattr(benchmark, "_model_run_periodic"), "Benchmark missing _model_run_periodic attribute."
     periodic = benchmark._model_run_periodic
     assert isinstance(periodic, dict), "_model_run_periodic should be a dict."
     for key in ("loss", "act_mean", "step"):
         assert key in periodic, f"Key '{key}' missing in _model_run_periodic."
-        assert (
-            len(periodic[key]) == 0
-        ), f"Expected empty list for periodic['{key}'], got {periodic[key]}."
+        assert (len(periodic[key]) == 0), f"Expected empty list for periodic['{key}'], got {periodic[key]}."
     pass