fix(agentic): filter Kimi MI355X replay context

Author: cquil11

benchmarks/single_node/agentic/kimik2.5_fp4_mi355x.sh

@@ -22,6 +22,10 @@ MAX_DELAY=${MAX_DELAY:-60}
 ADVANCE_MIN=${ADVANCE_MIN:-0.0}
 ADVANCE_MAX=${ADVANCE_MAX:-0.7}
 EP_SIZE=${EP_SIZE:-1}
+# Kimi-K2.5 advertises a 262144-token context window in vLLM 0.21.0.
+# Keep the benchmark loader's trace filter aligned with the server so
+# prompt+max_tokens overflows are removed before replay.
+MAX_MODEL_LEN=${MAX_MODEL_LEN:-262144}
 
 if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
@@ -512,6 +516,7 @@ VLLM_CMD=(
     --gpu-memory-utilization 0.90
     --block-size=1
     --trust-remote-code
+    --max-model-len "$MAX_MODEL_LEN"
     --max-num-seqs "$CONC"
     --mm-encoder-tp-mode data
     "${PREFIX_CACHE_ARGS[@]}"