[https://nvbugs/5575913][fix] Use separate thresholds for 2120b/20b gptoss

Signed-off-by: Dongfeng Yu <[email protected]>

Author: dongfengy

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

@@ -212,7 +212,7 @@ GPT-OSS/BF16:
   - accuracy: 90.3
   - kv_cache_quant_algo: FP8
     accuracy: 90.3
-GPT-OSS/MXFP4:
+GPT-OSS/120B-MXFP4:
   - accuracy: 90.3
   - quant_algo: W4A8_MXFP4_MXFP8
     accuracy: 90.3
@@ -229,6 +229,18 @@ GPT-OSS/MXFP4:
   - quant_algo: W4A16_MXFP4
     kv_cache_quant_algo: FP8
     accuracy: 90.3
+GPT-OSS/20B-MXFP4:
+  - accuracy: 85.0
+  - quant_algo: W4A8_MXFP4_MXFP8
+    accuracy: 85.0
+  - quant_algo: W4A8_MXFP4_MXFP8
+    kv_cache_quant_algo: FP8
+    accuracy: 85.0
+  - quant_algo: W4A16_MXFP4
+    accuracy: 85.0
+  - quant_algo: W4A16_MXFP4
+    kv_cache_quant_algo: FP8
+    accuracy: 85.0
 LGAI-EXAONE/EXAONE-4.0-32B:
   - accuracy: 88.36
 ByteDance-Seed/Seed-OSS-36B-Instruct:

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -3450,7 +3450,7 @@ def test_w4_1gpu(self, kv_cache_dtype, moe_backend, cuda_graph,
                   moe_config=MoeConfig(backend=moe_backend))
 
         with llm:
-            model_name = "GPT-OSS/MXFP4"
+            model_name = "GPT-OSS/20B-MXFP4"
             task = GSM8K(model_name)
             task.evaluate(llm,
                           extra_evaluator_kwargs=self.extra_evaluator_kwargs)
@@ -3461,7 +3461,7 @@ def test_dummy_load_format(self):
             load_format="dummy",
         )
         with llm:
-            model_name = "GPT-OSS/MXFP4"
+            model_name = "GPT-OSS/20B-MXFP4"
             task = GSM8K(model_name)
             task.evaluate(llm, is_integration_test=True)
 
@@ -3508,7 +3508,7 @@ def test_w4_4gpus(self, kv_cache_dtype, moe_backend, tp_size, pp_size,
                   moe_config=MoeConfig(backend=moe_backend))
 
         with llm:
-            model_name = "GPT-OSS/MXFP4"
+            model_name = "GPT-OSS/120B-MXFP4"
             task = GSM8K(model_name)
             task.evaluate(llm,
                           extra_evaluator_kwargs=self.extra_evaluator_kwargs)
@@ -3595,7 +3595,7 @@ def test_w4_2gpus(self, kv_cache_dtype, moe_backend, tp_size, pp_size,
                   moe_config=MoeConfig(backend=moe_backend))
 
         with llm:
-            model_name = "GPT-OSS/MXFP4"
+            model_name = "GPT-OSS/20B-MXFP4"
             task = GSM8K(model_name)
             mocker.patch.object(GSM8K, "MAX_OUTPUT_LEN", 8192)
             mocker.patch.dict(GSM8K.EVALUATE_KWARGS,
@@ -3622,7 +3622,7 @@ def test_w4_chunked_prefill(self, kv_cache_dtype, moe_backend, mocker):
         kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6,
                                         dtype=kv_cache_dtype)
 
-        model_name = "GPT-OSS/MXFP4"
+        model_name = "GPT-OSS/120B-MXFP4"
         with LLM(self.MODEL_PATH,
                  tensor_parallel_size=4,
                  pipeline_parallel_size=1,