Fix _riscv_supports_rvv_vlen128() to detect RVV on hardware without zvl flags (#43179)

Signed-off-by: liuyudong <liuyudong@iscas.ac.cn>
Co-authored-by: YuanSheng <yuansheng@isrc.iscas.ac.cn>

csrc/cpu/cpu_attn.cpp

@@ -11,6 +11,17 @@ static inline cpu_attention::Fp8KVCacheDataType parse_fp8_kv_dtype(
   return cpu_attention::Fp8KVCacheDataType::kAuto;
 }
 
+bool cpu_attn_has_isa(const std::string& isa) {
+  if (isa == "rvv") {
+#if defined(__riscv) && defined(__riscv_v_min_vlen) && __riscv_v_min_vlen == 128
+    return true;
+#else
+    return false;
+#endif
+  }
+  return false;
+}
+
 torch::Tensor get_scheduler_metadata(
     const int64_t num_req, const int64_t num_heads_q,
     const int64_t num_heads_kv, const int64_t head_dim,

csrc/cpu/torch_bindings.cpp

@@ -146,6 +146,8 @@ at::Tensor causal_conv1d_update_cpu(
 void activation_lut_bf16(torch::Tensor& out, torch::Tensor& input,
                          const std::string& activation);
 
+bool cpu_attn_has_isa(const std::string& isa);
+
 torch::Tensor get_scheduler_metadata(
     const int64_t num_req, const int64_t num_heads_q,
     const int64_t num_heads_kv, const int64_t head_dim,
@@ -497,6 +499,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("fused_gdn_gating_cpu", torch::kCPU, &fused_gdn_gating_cpu);
 
   // CPU attention kernels
+  ops.def("cpu_attn_has_isa(str isa) -> bool", &cpu_attn_has_isa);
   ops.def(
       "get_scheduler_metadata(int num_req, int num_heads_q, int num_heads_kv, "
       "int head_dim, Tensor seq_lens, ScalarType dtype, Tensor "

vllm/v1/attention/backends/cpu_attn.py

@@ -438,9 +438,22 @@ def _riscv_supports_rvv() -> bool:
             cpuinfo = f.read()
     except OSError:
         return False
-    return any(f"zvl{n}b" in cpuinfo for n in (128, 256)) and all(
-        f"zvl{n}b" not in cpuinfo for n in (512, 1024)
-    )
+    # If VLEN >= 512 is detected, the RVV kernel was not compiled.
+    if any(f"zvl{n}b" in cpuinfo for n in (512, 1024)):
+        return False
+
+    # zvl128b or zvl256b explicitly advertised -> RVV kernel available.
+    if any(f"zvl{n}b" in cpuinfo for n in (128, 256)):
+        return True
+
+    # No zvl<N>b flag at all (e.g. some hardware reports zve* without
+    # a VLEN hint).  Delegate to the C++ compile-time check instead.
+    try:
+        import torch
+
+        return torch.ops._C.cpu_attn_has_isa("rvv")
+    except Exception:
+        return False
 
 
 def _get_attn_isa(