[qwen3-omni] Add Qwen3-Omni moe thinker

2025-10-09 17:51:14 +08:00
parent bc57e2ef60
commit 24fab12b2f
8 changed files with 1543 additions and 37 deletions
--- a/vllm/attention/layer.py
+++ b/vllm/attention/layer.py
@@ -54,6 +54,14 @@ def check_xformers_availability():
    return USE_XFORMERS_OPS


+def check_upstream_fa_availability(dtype: torch.dtype):
+    if dtype in (torch.float16, torch.bfloat16) and current_platform.is_cuda(
+    ) and current_platform.has_device_capability(80):
+        from transformers.utils import is_flash_attn_2_available
+        return is_flash_attn_2_available()
+    return False
+
+
 class Attention(nn.Module):
    """Attention layer.