diff --git a/python/sglang/srt/environ.py b/python/sglang/srt/environ.py
index e874228d27bc..ec3a56041eeb 100644
--- a/python/sglang/srt/environ.py
+++ b/python/sglang/srt/environ.py
@@ -1359,7 +1359,7 @@ class Envs:
     # capacity).
     SGLANG_DISABLE_DRAFT_EXTEND_CUDA_GRAPH = EnvBool(False)
     # Use the split-KV (flash-decode) kernel for EAGLE target-verify on the
-    # Triton backend (ROCm). Only active at speculative topk == 1; falls back to
+    # Triton backend (ROCm and CUDA). Only active at speculative topk == 1; falls back to
     # extend_attention_fwd for unsupported cases or when set false (e.g. for
     # debugging). Correctness is unaffected; this only changes performance.
     SGLANG_ENABLE_SPLITKV_VERIFY = EnvBool(True)
diff --git a/python/sglang/srt/layers/attention/triton_backend.py b/python/sglang/srt/layers/attention/triton_backend.py
index 914def77f6b9..8ffa62310281 100644
--- a/python/sglang/srt/layers/attention/triton_backend.py
+++ b/python/sglang/srt/layers/attention/triton_backend.py
@@ -222,12 +222,14 @@ def __init__(
         self.target_verify_num_tokens_per_req = model_runner.decode_num_tokens_per_req()
         self.speculative_num_steps = get_spec().speculative_num_steps
         self.topk = get_spec().speculative_eagle_topk or 0
-        # Split-KV verify is bit-equivalent only for a pure-causal chain (topk==1)
-        # and is gfx95-only; else fall back to extend_attention_fwd.
+        # Split-KV verify is bit-equivalent only for a pure-causal chain (topk==1);
+        # else fall back to extend_attention_fwd. The kernel is NV-safe (HIP-only
+        # launch kwargs are gated inside verify_splitkv.py), and on CUDA the
+        # fallback extend kernel loops the whole prefix serially per (seq, head),
+        # so long-context verify degrades linearly with prefix length. Opt out
+        # with SGLANG_ENABLE_SPLITKV_VERIFY=0.
         self.use_verify_splitkv = (
-            is_gfx95_supported()
-            and envs.SGLANG_ENABLE_SPLITKV_VERIFY.get()
-            and self.topk == 1
+            envs.SGLANG_ENABLE_SPLITKV_VERIFY.get() and self.topk == 1
         )
         self.use_mla = model_runner.model_config.attention_arch == AttentionArch.MLA
         # The grouped-head verify kernel is tuned for Kimi-K3 MLA and Qwen3.5