Allow long-context SSE events in exact replay

runs/frontier-s3-real-v0/qwen30_exact_trace_client.py

@@ -13,6 +13,7 @@ from typing import Any
 
 
 TARGET_PASS_RATE = 0.95
+SSE_READ_BUFFER_BYTES = 8 * 1024 * 1024
 
 
 def parse_args() -> argparse.Namespace:
@@ -177,6 +178,10 @@ async def replay(args: argparse.Namespace, rows: list[dict[str, Any]]) -> list[d
         base_url=f"http://{args.host}:{args.port}",
         timeout=timeout,
         connector=connector,
+        # return_token_ids can put the full long prompt token vector in one
+        # SSE event. aiohttp's readline limit is 2x read_bufsize; the default
+        # 512 KiB limit rejects valid 100k+ context responses.
+        read_bufsize=SSE_READ_BUFFER_BYTES,
     ) as session:
         tasks = [
             asyncio.create_task(
@@ -230,6 +235,7 @@ def main() -> None:
             "input_output_prompt": "exact_source_values",
             "served_model_alias": args.served_model,
             "http_connection_reuse": False,
+            "sse_read_buffer_bytes": SSE_READ_BUFFER_BYTES,
             "ttft_slo": "1000ms + 1000ms * input_tokens / 8000",
             "tpot_slo_ms": args.tpot_slo_ms,
             "target_pass_rate": TARGET_PASS_RATE,