Fix MUSA detection in compiled prefill path (#39061)

python/sglang/srt/layers/attention/dsa/utils.py

@@ -115,13 +115,14 @@ def should_use_dsa_fused_topk(seed_dsa_topk_from_draft_extend: bool) -> bool:
 
 
 def is_dsa_enable_prefill_cp():
+    if get_parallel().attn_cp_size <= 1:
+        return False
+
     if is_hip() or is_npu() or is_musa():
         return False
 
     # Generic prefill CP derives activation from the runtime topology and model
     # architecture.
-    if get_parallel().attn_cp_size <= 1:
-        return False
     from sglang.srt.configs.model_config import is_deepseek_dsa, is_deepseek_v4
 
     hf_config = process_model_config().hf_config

python/sglang/srt/utils/common.py

@@ -221,13 +221,17 @@ def is_cpu() -> bool:
     return os.getenv("SGLANG_USE_CPU_ENGINE", "0") == "1" and is_host_cpu_supported
 
 
+try:
+    import torchada  # noqa: F401
+except ImportError:
+    _IS_MUSA = False
+else:
+    _IS_MUSA = hasattr(torch.version, "musa") and torch.version.musa is not None
+
+
 @lru_cache(maxsize=1)
 def is_musa() -> bool:
-    try:
-        import torchada  # noqa: F401
-    except ImportError:
-        return False
-    return hasattr(torch.version, "musa") and torch.version.musa is not None
+    return _IS_MUSA
 
 
 @lru_cache(maxsize=1)