diff --git a/tests/models/multimodal/generation/test_common.py b/tests/models/multimodal/generation/test_common.py index 6640e1ff9474d..299f57f6c4f86 100644 --- a/tests/models/multimodal/generation/test_common.py +++ b/tests/models/multimodal/generation/test_common.py @@ -513,6 +513,7 @@ VLM_TEST_SETTINGS = { max_model_len=8192, use_tokenizer_eos=True, patch_hf_runner=model_utils.internvl_patch_hf_runner, + num_logprobs=10 if current_platform.is_rocm() else 5, ), "intern_vl-hf": VLMTestInfo( models=["OpenGVLab/InternVL3-1B-hf"], diff --git a/vllm/model_executor/models/siglip2navit.py b/vllm/model_executor/models/siglip2navit.py index efdee255ab5eb..15d0ff30ed9bb 100644 --- a/vllm/model_executor/models/siglip2navit.py +++ b/vllm/model_executor/models/siglip2navit.py @@ -163,8 +163,10 @@ def apply_rotary_pos_emb( enable_fp32_compute=True, ) - if is_flash_attn_backend and not current_platform.is_cuda(): + if is_flash_attn_backend and current_platform.is_cuda(): apply_rotary_emb_func = apply_rotary_emb.forward_cuda + elif is_flash_attn_backend and current_platform.is_rocm(): + apply_rotary_emb_func = apply_rotary_emb.forward_hip else: apply_rotary_emb_func = apply_rotary_emb.forward_native