[Tests][LM Eval] Fix test seeding for consistent results (#2395)

SUMMARY:
- Enables consistent test results before runs

Test Run:
https://github.com/neuralmagic/llm-compressor-testing/actions/runs/22371360237

---------

Signed-off-by: Dipika Sikka <dipikasikka1@gmail.com>

Author: dsikka

tests/lmeval/configs/w4a16_awq_sym.yaml

@@ -4,10 +4,11 @@ scheme: W4A16_awq_sym
 recipe: tests/e2e/vLLM/recipes/WNA16/recipe_w4a16_awq_sym.yaml
 dataset_id: HuggingFaceH4/ultrachat_200k
 dataset_split: train_sft
+seed: 100
 lmeval:
   recovery_threshold:
-    exact_match,strict-match: 0.91
-    exact_match,flexible-extract: 0.91
+    exact_match,strict-match: 0.92
+    exact_match,flexible-extract: 0.92
   metrics:
     exact_match,flexible-extract: 0.70
     exact_match,strict-match: 0.70

tests/lmeval/configs/w4a4_nvfp4.yaml

@@ -3,6 +3,7 @@ model: meta-llama/Llama-3.1-8B-Instruct
 scheme: NVFP4
 dataset_id: HuggingFaceH4/ultrachat_200k
 dataset_split: train_sft
+seed: 100
 num_calibration_samples: 20
 lmeval:
   # NVFP4 (4-bit weights + 4-bit activations) has lower recovery than FP8/INT8

tests/lmeval/test_lmeval.py

@@ -100,12 +100,18 @@ def set_up(self, test_data_file: str):
         self.recipe = eval_config.get("recipe")
         self.quant_type = eval_config.get("quant_type")
         self.save_dir = eval_config.get("save_dir")
-        self.seed = eval_config.get("seed", None)
+        self.seed = eval_config.get("seed", 42)
 
-        if self.seed is not None:
-            random.seed(self.seed)
-            numpy.random.seed(self.seed)
-            torch.manual_seed(self.seed)
+        random.seed(self.seed)
+        numpy.random.seed(self.seed)
+        torch.manual_seed(self.seed)
+
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(self.seed)
+            torch.backends.cudnn.deterministic = True
+            torch.backends.cudnn.benchmark = False
+
+        logger.info(f"Seed set to {self.seed} with deterministic mode enabled")
 
         logger.info("========== RUNNING ==============")
         logger.info(self.scheme)
@@ -187,6 +193,11 @@ def _eval_model(self, model: str) -> dict:
             limit=self.lmeval.limit,
             apply_chat_template=self.lmeval.apply_chat_template,
             batch_size=self.lmeval.batch_size,
+            # Pass seeds to lm_eval for deterministic evaluation
+            random_seed=self.seed,
+            numpy_random_seed=self.seed,
+            torch_random_seed=self.seed,
+            fewshot_random_seed=self.seed,
         )
 
         return results