From ee8c47961b927320400f2684029945761bdb72aa Mon Sep 17 00:00:00 2001 From: SeongJun Lee Date: Fri, 16 Oct 2026 05:19:09 +0900 Subject: [PATCH] [CI] Use an ungated model as the attention test config default (#61732) Signed-off-by: lesj0610 Co-authored-by: Wei Zhao <51183510+wzhao18@users.noreply.github.com> --- tests/v1/attention/test_attention_backends.py | 12 ++++++------ tests/v1/attention/test_deepseek_v4_swa_visible.py | 2 +- tests/v1/attention/utils.py | 2 +- 3 files changed, 8 insertions(+), 8 deletions(-) diff --git a/tests/v1/attention/test_attention_backends.py b/tests/v1/attention/test_attention_backends.py index b537a17823..abe13137f7 100644 --- a/tests/v1/attention/test_attention_backends.py +++ b/tests/v1/attention/test_attention_backends.py @@ -763,7 +763,7 @@ def test_flashinfer_cross_layer_layout( _test_backend_correctness( batch_spec=BATCH_SPECS[batch_spec_name], - model="meta-llama/Meta-Llama-3-8B", + model="Qwen/Qwen3-0.6B", backend_to_test=[AttentionBackendEnum.FLASHINFER], mask_mod=causal_mask_mod, kv_cache_dtype=kv_cache_dtype, @@ -786,7 +786,7 @@ def test_flashinfer_cross_layer_layout( "single_prefill", ], ) -@pytest.mark.parametrize("model", ["meta-llama/Meta-Llama-3-8B"]) +@pytest.mark.parametrize("model", ["Qwen/Qwen3-0.6B"]) @pytest.mark.parametrize("tensor_parallel_size", [1, 2, 4]) @pytest.mark.parametrize("kv_cache_dtype", ["auto", "fp8", "fp8_e4m3"]) def test_causal_backend_correctness( @@ -1582,7 +1582,7 @@ def test_flashinfer_native_prefill_with_sinks(default_vllm_config): _test_backend_correctness( BATCH_SPECS["small_prefill"], - "meta-llama/Meta-Llama-3-8B", + "Qwen/Qwen3-0.6B", [AttentionBackendEnum.FLASHINFER], causal_mask_mod, use_sinks=True, @@ -1629,7 +1629,7 @@ def test_flashinfer_xqa_decode_correctness(default_vllm_config): batch_spec = BATCH_SPECS["small_decode"] vllm_config = create_vllm_config( - model_name="meta-llama/Meta-Llama-3-8B", + model_name="Qwen/Qwen3-0.6B", max_model_len=max(batch_spec.seq_lens), block_size=16, ) @@ -1691,7 +1691,7 @@ def test_flashinfer_xqa_decode_correctness(default_vllm_config): _test_backend_correctness( batch_spec, - "meta-llama/Meta-Llama-3-8B", + "Qwen/Qwen3-0.6B", [AttentionBackendEnum.FLASHINFER], causal_mask_mod, ) @@ -1994,7 +1994,7 @@ if current_platform.is_rocm(): "mixed_small", ], ) -@pytest.mark.parametrize("model", ["meta-llama/Meta-Llama-3-8B"]) +@pytest.mark.parametrize("model", ["Qwen/Qwen3-0.6B"]) def test_non_causal_backend_correctness( default_vllm_config, batch_spec_name: str, diff --git a/tests/v1/attention/test_deepseek_v4_swa_visible.py b/tests/v1/attention/test_deepseek_v4_swa_visible.py index 5dcea926f0..a46495f790 100644 --- a/tests/v1/attention/test_deepseek_v4_swa_visible.py +++ b/tests/v1/attention/test_deepseek_v4_swa_visible.py @@ -529,7 +529,7 @@ def make_builder( window: int = WINDOW, max_image_tokens: int = MAX_IMG, max_num_batched_tokens: int = 64, - model_name: str = "meta-llama/Meta-Llama-3-8B", + model_name: str = "Qwen/Qwen3-0.6B", builder_cls: type[ DeepseekSparseSWAMetadataBuilder ] = DeepseekSparseSWAMetadataBuilder, diff --git a/tests/v1/attention/utils.py b/tests/v1/attention/utils.py index abb29a03c4..a0a24fd176 100644 --- a/tests/v1/attention/utils.py +++ b/tests/v1/attention/utils.py @@ -180,7 +180,7 @@ def create_standard_kv_cache_spec( def create_vllm_config( - model_name: str = "meta-llama/Meta-Llama-3-8B", + model_name: str = "Qwen/Qwen3-0.6B", tensor_parallel_size: int = 1, max_model_len: int = 1024, dtype: ModelDType | torch.dtype = "auto",