[Spec][trtllm] use decode kernel for draft extend (#24566)

python/sglang/srt/layers/attention/trtllm_mha_backend.py

@@ -848,7 +848,10 @@ class TRTLLMHAAttnBackend(FlashInferAttnBackend):
 
         page_table = self._get_layer_page_table(layer, forward_batch)
 
-        if forward_batch.forward_mode.is_target_verify():
+        if (
+            forward_batch.forward_mode.is_target_verify()
+            or forward_batch.forward_mode.is_draft_extend_v2()
+        ):
             o = flashinfer.decode.trtllm_batch_decode_with_kv_cache(
                 query=q,
                 kv_cache=kv_cache,

python/sglang/test/bench_one_batch_server_internal.py

@@ -464,6 +464,18 @@ def _warmup_cache(
     print("Cache warmup completed")
 
 
+def _flush_cache_with_retry(url: str, endpoint: str, max_retries: int = 3):
+    """Post to a cache flush endpoint with retries on failure."""
+    for attempt in range(max_retries):
+        response = requests.post(url + endpoint, timeout=DEFAULT_TIMEOUT)
+        if response.status_code == 200:
+            return
+        if attempt < max_retries - 1:
+            time.sleep(2)
+        else:
+            response.raise_for_status()
+
+
 def run_one_case(
     url: str,
     batch_size: int,
@@ -500,11 +512,9 @@ def run_one_case(
 ):
     if backend == "vllm":
         # You need to have export VLLM_SERVER_DEV_MODE=1 in your environment to use this endpoint.
-        response = requests.post(url + "/reset_prefix_cache", timeout=DEFAULT_TIMEOUT)
-        response.raise_for_status()
+        _flush_cache_with_retry(url, "/reset_prefix_cache")
     else:
-        response = requests.post(url + "/flush_cache", timeout=DEFAULT_TIMEOUT)
-        response.raise_for_status()
+        _flush_cache_with_retry(url, "/flush_cache")
 
     # Load input token ids via bench_serving.get_dataset
     supported_datasets = ("random", "mmmu", "generated-shared-prefix")