Align Qwen235 attention profile sequence limit

2026-07-19 06:19:08 +08:00
parent d563c30b42
commit ac5061fa5d
1 changed files with 1 additions and 1 deletions
--- a/runs/frontier-fidelity-envelope-v1/run_qwen235_v020_profiles.sh
+++ b/runs/frontier-fidelity-envelope-v1/run_qwen235_v020_profiles.sh
@@ -35,7 +35,7 @@ common_profile() {
  env CUDA_VISIBLE_DEVICES="${gpu}" PYTHONPATH="${FRONTIER_SOURCE}" \
    "${VENV_ROOT}/bin/python" -m frontier.profiling.attention.main \
      --disable_ray --num_gpus 1 --device h20 --output_dir "${root}" --models "${MODEL}" \
-      --num_tensor_parallel_workers 4 8 --max_model_len 40960 --max_seq_len 8192 \
+      --num_tensor_parallel_workers 4 8 --max_model_len 40960 --max_seq_len 40960 \
      --batch_size_list "${BATCHES[@]}" --decode_kv_cache_size_list "${KV[@]}" \
      --fixed_chunked_prefill_size 8192 --attention_backend FLASHINFER \
      --profile_method "${method}" --yes