[Fix] Guard kernel OOB accesses and harden runtime edge cases (#30847)
This commit is contained in:
@@ -785,9 +785,15 @@ def run_one_case(
|
||||
response.raise_for_status()
|
||||
server_info = response.json()
|
||||
internal_states = server_info.get("internal_states", [])
|
||||
internal_state = internal_states[0] if internal_states else {}
|
||||
last_gen_throughput = internal_state.get("last_gen_throughput", None) or -1
|
||||
acc_length = internal_state.get("avg_spec_accept_length", None) or -1
|
||||
acc_length = -1
|
||||
last_gen_throughput = -1
|
||||
for internal_state in internal_states:
|
||||
val_acc = internal_state.get("avg_spec_accept_length")
|
||||
if val_acc is not None:
|
||||
acc_length = val_acc
|
||||
val_thr = internal_state.get("last_gen_throughput")
|
||||
if val_thr is not None:
|
||||
last_gen_throughput = val_thr
|
||||
|
||||
# Calculate cache hit rate from before/after metrics delta
|
||||
metrics_after = get_cache_tokens_from_metrics(url)
|
||||
|
||||
Reference in New Issue
Block a user