fix(evals): preserve single-node DP attention metadata / 修复单节点评测的 DP attention 元数据

.github/workflows/test-changelog-gate.yml

@@ -29,6 +29,7 @@ on:
       - "utils/test_collect_eval_results.py"
       - "utils/evals/validate_scores.py"
       - "utils/evals/test_batched_eval.py"
+      - "utils/evals/test_run_eval_dispatch.py"
       - "utils/prepare_perf_changelog_merge.py"
       - "utils/recover_failed_ingest.py"
       - "utils/changelog_gate_tests/test_prepare_perf_changelog_merge.py"
@@ -83,4 +84,5 @@ jobs:
             utils/test_process_changelog.py \
             utils/test_collect_eval_results.py \
             utils/evals/test_batched_eval.py \
+            utils/evals/test_run_eval_dispatch.py \
             -v

benchmarks/benchmark_lib.sh

@@ -2267,7 +2267,12 @@ _write_lm_eval_meta_json() {
     local batch_metadata="${2:-}"
     local metadata_conc="${3:-${CONC:-1}}"
 
-    bridge_disagg_eval_metadata
+    # Single-node jobs already export TP/EP/DP_ATTENTION. The disaggregated
+    # bridge defaults missing per-phase DP flags to false, so applying it to
+    # single-node jobs would overwrite their actual DP-attention setting.
+    if [ "${IS_MULTINODE:-false}" = "true" ]; then
+        bridge_disagg_eval_metadata
+    fi
 
     local model_name="${MODEL_NAME:-$MODEL}"
     local is_multinode_json="false"

docs/results-and-ingestion.md

@@ -102,6 +102,16 @@ InferenceX-app treats routing fields as columns or config dimensions and stores
 
 Each eval upload is named `eval_<EXP_NAME>_<RESULT_FILENAME>`. Its current allowed payload includes `meta_env.json`, `results*.json`, sample JSONL, predictions, SWE-bench reports, and trajectory files. The collector uses only the metadata and lm-eval result JSON for aggregate rows.
 
+The shared eval metadata writer preserves single-node `DP_ATTENTION` and uses it
+as the default for both `prefill_dp_attention` and `decode_dp_attention`. Only
+`IS_MULTINODE=true` jobs bridge the separate prefill/decode environment variables;
+those jobs may have different DP-attention settings on each side. The collector
+does not reconstruct topology from artifact names or server logs. Older artifacts
+affected by the unconditional bridge can report `false` for a single-node
+DP-attention eval. Fixing the writer does not repair those artifacts or existing
+database rows: verify the original job configuration and server logs before
+correcting metadata, regenerating aggregates, and re-ingesting affected results.
+
 [`utils/collect_eval_results.py`](../utils/collect_eval_results.py) applies these rules:
 
 1. An eval set is a root or immediate child directory containing `meta_env.json`.

docs/results-and-ingestion_zh.md

@@ -102,6 +102,15 @@ InferenceX-app 将路由字段作为列或配置维度，并把数值测量存
 
 每个评测上传名为 `eval_<EXP_NAME>_<RESULT_FILENAME>`。当前允许的载荷包括 `meta_env.json`、`results*.json`、样本 JSONL、预测、SWE-bench 报告和轨迹文件。收集器只使用元数据和 lm-eval 结果 JSON 来生成聚合记录。
 
+共享评测元数据写入器保留单节点的 `DP_ATTENTION`，并将其作为
+`prefill_dp_attention` 和 `decode_dp_attention` 的默认值。只有
+`IS_MULTINODE=true` 的任务才会转换独立的 prefill/decode 环境变量；
+这类任务两侧的 DP attention 设置可以不同。收集器不会根据工件名称或
+服务端日志重建拓扑。受旧版无条件转换逻辑影响的历史工件，可能将实际启用
+DP attention 的单节点评测记录为 `false`。修复写入器不会修复这些工件或
+已有数据库记录：应先核实原始任务配置和服务端日志，再更正元数据、重新生成
+聚合结果并重新摄取受影响的数据。
+
 [`utils/collect_eval_results.py`](../utils/collect_eval_results.py) 执行以下规则：
 
 1. 评测集是包含 `meta_env.json` 的根目录或一级子目录。

utils/evals/test_run_eval_dispatch.py

@@ -1489,13 +1489,46 @@ def _summary_metadata(tmp_path: Path, **overrides: str) -> dict:
         "CONC": "7",
         "KV_OFFLOADING": "none",
     }
-    for key in ("EVAL_COMPLETED_SUITE", "EVAL_SUITE", "EVAL_TASKS_DIR"):
+    for key in (
+        "EVAL_COMPLETED_SUITE", "EVAL_SUITE", "EVAL_TASKS_DIR",
+        "IS_MULTINODE", "DP_ATTENTION",
+        "PREFILL_DP_ATTN", "PREFILL_DP_ATTENTION", "PREFILL_ENABLE_DP",
+        "DECODE_DP_ATTN", "DECODE_DP_ATTENTION", "DECODE_ENABLE_DP",
+    ):
         env.pop(key, None)
     env.update(overrides)
     subprocess.run(["bash", "-c", script], env=env, check=True)
     return json.loads((work_dir / "meta_env.json").read_text())
 
 
+@pytest.mark.parametrize("dp_attention, expected", [("true", True), ("false", False)])
+def test_summary_preserves_single_node_dp_attention(
+    tmp_path: Path, dp_attention: str, expected: bool,
+) -> None:
+    meta = _summary_metadata(
+        tmp_path, IS_MULTINODE="false", TP="8", EP_SIZE="8",
+        DP_ATTENTION=dp_attention,
+    )
+    assert meta["dp_attention"] is expected
+    assert meta["prefill_dp_attention"] is expected
+    assert meta["decode_dp_attention"] is expected
+    assert meta["tp"] == 8
+    assert meta["ep"] == 8
+
+
+def test_summary_preserves_asymmetric_multinode_dp_attention(tmp_path: Path) -> None:
+    meta = _summary_metadata(
+        tmp_path, IS_MULTINODE="true", DP_ATTENTION="false",
+        PREFILL_TP="4", PREFILL_EP="4", DECODE_TP="8", DECODE_EP="8",
+        PREFILL_DP_ATTN="true", DECODE_DP_ATTN="false",
+    )
+    assert meta["dp_attention"] is True
+    assert meta["prefill_dp_attention"] is True
+    assert meta["decode_dp_attention"] is False
+    assert meta["prefill_tp"] == 4
+    assert meta["decode_tp"] == 8
+
+
 def test_summary_stages_bfcl_upstream_archive_before_cleanup(tmp_path: Path) -> None:
     work_dir = tmp_path / "work"
     results_dir = tmp_path / "results"