[TRTLLM-15078][test] Remove all DeepSeek-R1-Distill-* tests

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

@@ -202,8 +202,6 @@ deepseek-ai/DeepSeek-R1-0528:
   - quant_algo: FP8_BLOCK_SCALES
     kv_cache_quant_algo: FP8
     accuracy: 84.722
-deepseek-ai/DeepSeek-R1-Distill-Llama-70B:
-  - accuracy: 78.19
 deepseek-ai/DeepSeek-V3.2-Exp:
   - quant_algo: FP8_BLOCK_SCALES
     accuracy: 88.2

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -3306,20 +3306,6 @@ def test_fp8_blockscale_chunked_prefill(self, tp_size, pp_size, ep_size,
             task.evaluate(llm)
 
 
-class TestDeepSeekR1DistillLlama70B(LlmapiAccuracyTestHarness):
-    MODEL_NAME = "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
-    MODEL_PATH = f"{llm_models_root()}/DeepSeek-R1/DeepSeek-R1-Distill-Llama-70B"
-
-    @skip_pre_hopper
-    @pytest.mark.skip_less_mpi_world_size(2)
-    def test_auto_dtype_tp2(self):
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.4)
-        _run_multinode_accuracy(self.MODEL_PATH,
-                                self.MODEL_NAME,
-                                benchmarks=["mmlu"],
-                                kv_cache_config=kv_cache_config)
-
-
 @pytest.mark.timeout(14400)
 @pytest.mark.skip_less_mpi_world_size(8)
 class TestDeepSeekV3(LlmapiAccuracyTestHarness):

tests/integration/defs/perf/_model_paths.py

@@ -26,8 +26,6 @@
     "llama_v4_scout_17b_16e_instruct": "llama4-models/Llama-4-Scout-17B-16E-Instruct",
     "llama_v4_scout_17b_16e_instruct_fp8": "llama4-models/Llama-4-Scout-17B-16E-Instruct-FP8",
     "llama_v4_scout_17b_16e_instruct_fp4": "llama4-models/Llama-4-Scout-17B-16E-Instruct-FP4",
-    "deepseek_r1_distill_qwen_32b": "DeepSeek-R1/DeepSeek-R1-Distill-Qwen-32B",
-    "deepseek_r1_distill_llama_70b": "DeepSeek-R1/DeepSeek-R1-Distill-Llama-70B/",
     "gemma_3_27b_it": "gemma/gemma-3-27b-it",
     "gemma_3_27b_it_fp8": "gemma/gemma-3-27b-it-fp8",
     "gemma_3_27b_it_fp4": "gemma/gemma-3-27b-it-FP4",

tests/integration/defs/perf/base_perf_pytorch.csv

@@ -3,7 +3,3 @@ network_name,perf_case_name,test_name,threshold,absolute_threshold,metric_type,p
 "llama_v3.1_8b_instruct-bench-pytorch-float16-maxbs:512-maxnt:2048-input_output_len:128,128-reqs:8192","H100_PCIe-PyTorch-Perf-1/perf/test_perf.py::test_perf_metric_seq_throughput[llama_v3.1_8b_instruct-bench-pytorch-float16-maxbs:512-maxnt:2048-input_output_len:128,128-reqs:8192]","test_perf_metric_seq_throughput[llama_v3.1_8b_instruct-bench-pytorch-float16-maxbs:512-maxnt:2048-input_output_len:128,128-reqs:8192]",-0.20,5,SEQ_THROUGHPUT,76.45,
 "llama_v3.1_8b_instruct-bench-pytorch-float16-maxbs:512-maxnt:2048-input_output_len:128,128-reqs:8192","H100_PCIe-PyTorch-Perf-1/perf/test_perf.py::test_perf_metric_total_output_throughput[llama_v3.1_8b_instruct-bench-pytorch-float16-maxbs:512-maxnt:2048-input_output_len:128,128-reqs:8192]","test_perf_metric_total_output_throughput[llama_v3.1_8b_instruct-bench-pytorch-float16-maxbs:512-maxnt:2048-input_output_len:128,128-reqs:8192]",-0.20,500,TOTAL_OUTPUT_THROUGHPUT,9785.75,
 "llama_v3.1_8b_instruct-bench-pytorch-float16-maxbs:512-maxnt:2048-input_output_len:128,128-reqs:8192","H100_PCIe-PyTorch-Perf-1/perf/test_perf.py::test_perf_metric_kv_cache_size[llama_v3.1_8b_instruct-bench-pytorch-float16-maxbs:512-maxnt:2048-input_output_len:128,128-reqs:8192]","test_perf_metric_kv_cache_size[llama_v3.1_8b_instruct-bench-pytorch-float16-maxbs:512-maxnt:2048-input_output_len:128,128-reqs:8192]",0.20,2,KV_CACHE_SIZE,55.64,
-"deepseek_r1_distill_qwen_32b-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024","H100_PCIe-PyTorch-Perf-1/perf/test_perf.py::test_perf_metric_inference_time[deepseek_r1_distill_qwen_32b-subtype:H100_PCIe-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024]","test_perf_metric_inference_time[deepseek_r1_distill_qwen_32b-subtype:H100_PCIe-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024]",0.1,50,INFERENCE_TIME,1359184.5059,H100_PCIe
-"deepseek_r1_distill_qwen_32b-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024","H100_PCIe-PyTorch-Perf-1/perf/test_perf.py::test_perf_metric_kv_cache_size[deepseek_r1_distill_qwen_32b-subtype:H100_PCIe-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024]","test_perf_metric_kv_cache_size[deepseek_r1_distill_qwen_32b-subtype:H100_PCIe-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024]",-0.1,50,KV_CACHE_SIZE,10.92,H100_PCIe
-"deepseek_r1_distill_qwen_32b-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024","H100_PCIe-PyTorch-Perf-1/perf/test_perf.py::test_perf_metric_seq_throughput[deepseek_r1_distill_qwen_32b-subtype:H100_PCIe-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024]","test_perf_metric_seq_throughput[deepseek_r1_distill_qwen_32b-subtype:H100_PCIe-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024]",-0.1,10,SEQ_THROUGHPUT,0.3767,H100_PCIe
-"deepseek_r1_distill_qwen_32b-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024","H100_PCIe-PyTorch-Perf-1/perf/test_perf.py::test_perf_metric_total_output_throughput[deepseek_r1_distill_qwen_32b-subtype:H100_PCIe-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024]","test_perf_metric_total_output_throughput[deepseek_r1_distill_qwen_32b-subtype:H100_PCIe-bench-_autodeploy-float16-maxbs:512-maxnt:2048-kv_frac:0.8-input_output_len:1024,1024]",-0.1,10,TOTAL_OUTPUT_THROUGHPUT,385.7372,H100_PCIe

tests/integration/defs/perf/pytorch_model_config.py

@@ -291,16 +291,6 @@ def get_model_yaml_config(model_label: str,
                 },
             }
         },
-        # Model-specific cases with attention_dp disabled to prevent hangs
-        {
-            'patterns': [
-                'deepseek_r1_distill_llama_70b',
-            ],
-            'config': {
-                # True causes hang, needs model-specific fix.
-                'enable_attention_dp': False,
-            }
-        },
         # Qwen3 models with fp4 quantization on B200 and fp8 quantization on H200/H20
         {
             'patterns': [

tests/integration/defs/test_e2e.py

@@ -41,13 +41,13 @@
 
 
 @pytest.mark.parametrize("model_name,model_path", [
-    ("DeepSeek-R1-Distill-Qwen-1.5B", "DeepSeek-R1-Distill-Qwen-1.5B"),
+    ("Qwen3/Qwen3-0.6B", "Qwen3/Qwen3-0.6B"),
 ])
 def test_qwen_e2e_cpprunner_large_new_tokens(model_name, model_path, llm_venv):
     """RCCA: https://nvbugs/5238105 - none of n>1 sequences may be empty."""
     from tensorrt_llm import LLM, SamplingParams
 
-    prompt = r"<｜begin▁of▁sentence｜><｜User｜>The operation $\otimes$ is defined for all nonzero numbers by $a \otimes b = \frac{a^{2}}{b}$. Determine $[(1 \otimes 2) \otimes 3] - [1 \otimes (2 \otimes 3)]$. Let's think step by step and output the final answer within \boxed{}.<｜Assistant｜>"
+    prompt = r"The operation $\otimes$ is defined for all nonzero numbers by $a \otimes b = \frac{a^{2}}{b}$. Determine $[(1 \otimes 2) \otimes 3] - [1 \otimes (2 \otimes 3)]$. Let's think step by step."
 
     sampling_params = SamplingParams(
         max_tokens=1024,
@@ -60,6 +60,14 @@ def test_qwen_e2e_cpprunner_large_new_tokens(model_name, model_path, llm_venv):
     with LLM(model=f"{llm_models_root()}/{model_path}",
              max_batch_size=8,
              max_seq_len=4224) as llm:
+        prompt = llm.tokenizer.apply_chat_template(
+            [{
+                "role": "user",
+                "content": prompt
+            }],
+            tokenize=False,
+            add_generation_prompt=True,
+        )
         outputs = llm.generate([prompt], sampling_params=sampling_params)
 
     completions = outputs[0].outputs
@@ -918,9 +926,6 @@ def test_ptp_quickstart(llm_root, llm_venv):
     pytest.param('Mistral-Nemo-12b-Base',
                  'Mistral-Nemo-Base-2407',
                  marks=skip_pre_blackwell),
-    pytest.param('DeepSeek-R1-Distill-Qwen-32B',
-                 'DeepSeek-R1/DeepSeek-R1-Distill-Qwen-32B',
-                 marks=skip_pre_blackwell),
     pytest.param('GPT-OSS-20B', 'gpt_oss/gpt-oss-20b',
                  marks=skip_pre_blackwell),
     pytest.param(
@@ -1854,23 +1859,6 @@ def test_ptp_quickstart_bert(llm_root, llm_venv, model_name, model_path,
     print("Success: HF model logits match TRTLLM logits!")
 
 
-@pytest.mark.skip_less_device_memory(80000)
-@pytest.mark.parametrize("model_name,model_path", [
-    ("DeepSeek-R1-Distill-Qwen-7B", "DeepSeek-R1/DeepSeek-R1-Distill-Qwen-7B"),
-])
-def test_ptp_scaffolding(llm_root, llm_venv, model_name, model_path):
-    print(f"Testing scaffolding {model_name}.")
-    example_root = Path(os.path.join(llm_root, "examples", "scaffolding"))
-    input_file = Path(os.path.join(example_root, "test.jsonl"))
-    llm_venv.run_cmd([
-        str(example_root / "run_majority_vote_aime24.py"),
-        "--model_dir",
-        f"{llm_models_root()}/{model_path}",
-        f"--jsonl_file={input_file}",
-        "--threshold=0.5",
-    ])
-
-
 @pytest.mark.timeout(5400)
 @pytest.mark.skip_less_device_memory(80000)
 @pytest.mark.skip_less_device(4)
@@ -1933,8 +1921,6 @@ def test_multi_nodes_eval(model_path, tp_size, pp_size, ep_size, eval_task,
                  marks=skip_pre_hopper),
     pytest.param('Qwen3/saved_models_Qwen3-235B-A22B_nvfp4_hf',
                  marks=skip_pre_blackwell),
-    pytest.param('DeepSeek-R1/DeepSeek-R1-Distill-Llama-70B',
-                 marks=skip_pre_hopper),
     pytest.param('llama4-models/Llama-4-Scout-17B-16E-Instruct-FP8',
                  marks=skip_pre_hopper),
     pytest.param('llama4-models/Llama-4-Scout-17B-16E-Instruct',