Update max seq len

Signed-off-by: Michal Guzek <[email protected]>

Author: moraxu

tests/integration/defs/accuracy/references/gsm8k.yaml

@@ -171,6 +171,6 @@ microsoft/Phi-4-multimodal-instruct:
 microsoft/Phi-4-multimodal-instruct-long-rope:
   - accuracy: 75.85
 microsoft/Phi-4-mini-instruct:
-  - accuracy: 75.00
+  - accuracy: 82.30
 LGAI-EXAONE/EXAONE-4.0-32B:
   - accuracy: 88.36

tests/integration/defs/accuracy/references/mmlu.yaml

@@ -226,7 +226,7 @@ nvidia/Nemotron-H-56B-Base-8K:
     kv_cache_quant_algo: FP8
     accuracy: 83.82
 microsoft/Phi-4-mini-instruct:
-  - accuracy: 65.00
+  - accuracy: 68.98
 # Created a dummy accuracy to track tp_size=2 for phi4-mini model.
 # TODO: update once https://nvbugs/5393849 is fixed.
 microsoft/Phi-4-mini-instruct-tp2:

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -2303,7 +2303,7 @@ class TestPhi4MiniInstruct(LlmapiAccuracyTestHarness):
     MODEL_PATH = f"{llm_models_root()}/Phi-4-mini-instruct"
 
     def test_auto_dtype(self):
-        with LLM(self.MODEL_PATH) as llm:
+        with LLM(self.MODEL_PATH, max_seq_len=4096) as llm:
             task = CnnDailymail(self.MODEL_NAME)
             task.evaluate(llm)
             task = MMLU(self.MODEL_NAME)