more test variants for qwen3.5

Signed-off-by: Anthony Chang <27950904+rosenrodt@users.noreply.github.com>

Author: rosenrodt

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -5831,20 +5831,24 @@ class TestQwen3_5_35B_A3B(LlmapiAccuracyTestHarness):
     )
 
     @pytest.mark.parametrize("moe_backend", ["CUTLASS", "TRTLLM"])
-    def test_bf16(self, moe_backend):
+    @pytest.mark.parametrize(
+        "tp_size",
+        [1, pytest.param(2, marks=pytest.mark.skip_less_device(2))],
+        ids=["tp1", "tp2"],
+    )
+    def test_bf16(self, moe_backend, tp_size):
         if moe_backend == "TRTLLM" and get_sm_version() not in (100, 103):
             pytest.skip(f"{moe_backend} backend supports SM 100 and 103 only")
 
-        world_size = 1
         kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.8,
                                         enable_block_reuse=False)
         cuda_graph_config = CudaGraphConfig(
             enable_padding=True, batch_sizes=[1, 2, 4, 8, 16, 32, 64, 128])
         moe_config = MoeConfig(backend=moe_backend)
 
         with LLM(self.MODEL_PATH,
-                 tensor_parallel_size=world_size,
-                 moe_expert_parallel_size=world_size,
+                 tensor_parallel_size=tp_size,
+                 moe_expert_parallel_size=1,
                  max_seq_len=4096,
                  max_num_tokens=4096,
                  max_batch_size=128,
@@ -5856,20 +5860,24 @@ def test_bf16(self, moe_backend):
             task.evaluate(llm,
                           extra_evaluator_kwargs=self.EXTRA_EVALUATOR_KWARGS)
 
-    def test_fp8(self):
+    @pytest.mark.parametrize(
+        "tp_size",
+        [1, pytest.param(2, marks=pytest.mark.skip_less_device(2))],
+        ids=["tp1", "tp2"],
+    )
+    def test_fp8(self, tp_size):
         model_dir = f"{self.MODEL_PATH}-FP8"
         # Model is being added to CI. Skip at the moment.
         if not os.path.exists(model_dir):
             pytest.skip(f"Model directory {model_dir} does not exist")
 
-        world_size = 1
         kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.8,
                                         enable_block_reuse=False)
         moe_config = MoeConfig(backend='DEEPGEMM')
 
         with LLM(model_dir,
-                 tensor_parallel_size=world_size,
-                 moe_expert_parallel_size=world_size,
+                 tensor_parallel_size=tp_size,
+                 moe_expert_parallel_size=1,
                  max_seq_len=4096,
                  enable_chunked_prefill=True,
                  kv_cache_config=kv_cache_config,

tests/integration/test_lists/qa/llm_function_core.txt

@@ -186,8 +186,9 @@ accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_cu
 accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm]
 accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm_attention_dp]
 accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4_4gpus[latency_moe_trtllm_eagle3]
-accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16
-accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8
+accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16[tp1-CUTLASS]
+accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16[tp1-TRTLLM]
+accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8[tp1]
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_1gpu[v1_kv_cache-True-True-cutlass-auto]
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_1gpu[v1_kv_cache-True-True-cutlass-fp8]
 accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_1gpu[v1_kv_cache-True-True-triton-auto]

tests/integration/test_lists/qa/llm_function_core_sanity.txt

@@ -175,8 +175,10 @@ accuracy/test_llm_api_pytorch.py::TestQwen3_4B::test_eagle3
 accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_fp8_block_scales[latency]
 accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_w4a8_mxfp4[fp8-latency]
 accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_w4a8_mxfp4[mxfp8-latency]
-accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16
-accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8
+accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16[tp1-CUTLASS]
+accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16[tp1-TRTLLM]
+accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8[tp1]
+accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8[tp2]
 
 # disaggregated serving accuracy test
 accuracy/test_disaggregated_serving.py::TestDeepSeekV3Lite::test_auto_dtype[mtp_nextn=0-overlap_scheduler=False]

tests/integration/test_lists/test-db/l0_b200.yml

@@ -72,8 +72,12 @@ l0_b200:
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a16_mxfp4[latency-TRTLLM]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.5]
   - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.9]
-  - accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16
-  - accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8
+  - accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16[tp1-CUTLASS]
+  - accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16[tp1-TRTLLM]
+  - accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16[tp2-CUTLASS]
+  - accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_bf16[tp2-TRTLLM]
+  - accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8[tp1]
+  - accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8[tp2]
   - accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp1-cutlass]
   - disaggregated/test_workers.py::test_workers_kv_cache_aware_router_eviction[TinyLlama-1.1B-Chat-v1.0] # nvbugs 5300551
   - test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-8B-NVFP4-nvfp4-quantized/Meta-Llama-3.1-8B]