fix: bound LLM inference memory for batch screening

chore(docker): 补齐容器构建的忽略规则与运行依赖
feat(app): 新增批量筛选，结果与模式可跨刷新恢复
2026-10-05 15:33:12 +08:00 · 2026-09-23 10:11:41 +08:00 · 2026-09-16 06:11:46 +00:00 · 2026-09-05 17:15:37 +00:00
29 changed files with 8433 additions and 133 deletions
--- a/.dockerignore
+++ b/.dockerignore
@ -69,6 +69,16 @@ logs/
 models/finetune_cv/
 models/benchmark/
 models/mpnn/
+models/qwen2.5-7b-instruct/
+models/cv5_sample/
+models/molt5-base/
+models/nested_cv/
+models/pretrain/
+MolE_ckpt/
+MolE_repo/
+models/abl/
+models/biot5-plus-base/
+models/abl_full/
 models/*.pt
 models/*.json
 !models/final/
--- a/5
+++ b/5
@ -36,8 +36,8 @@ COPY app/ ./app/
 # 安装项目包
 RUN pip install -e .

-# 复制模型文件
-COPY models/final/ ./models/final/
+# 模型、Qwen 权重和 RAG 数据在运行时以只读 volume 挂载。
+# 不把数 GB 权重烘进镜像，也避免误把 Git LFS pointer 当成模型。

 # ============ API 服务 ============
 FROM base AS api
@ -60,4 +60,3 @@ ENV STREAMLIT_SERVER_PORT=8501 \
    STREAMLIT_BROWSER_GATHER_USAGE_STATS=false

 CMD ["streamlit", "run", "app/app.py"]
-
--- a/SERVER_DEPLOY.md
+++ b/SERVER_DEPLOY.md
@ -0,0 +1,59 @@
+# GPU 服务器测试与部署
+
+这个代码包不包含大模型权重。解压后，请先在项目根目录放好以下运行时文件：
+
+```text
+models/final/model.pt
+models/mpnn/all_amine_split_for_LiON/cv_0/fold_0/model_0/model.pt
+models/mpnn/all_amine_split_for_LiON/cv_1/fold_0/model_0/model.pt
+models/mpnn/all_amine_split_for_LiON/cv_2/fold_0/model_0/model.pt
+models/mpnn/all_amine_split_for_LiON/cv_3/fold_0/model_0/model.pt
+models/mpnn/all_amine_split_for_LiON/cv_4/fold_0/model_0/model.pt
+models/qwen2.5-7b-instruct/config.json
+models/qwen2.5-7b-instruct/                 # 其余 tokenizer/权重文件
+data/interim/internal.csv
+```
+
+`.pt` 必须是真实二进制权重，不能是 100 多字节的 Git LFS pointer。
+
+## 1. 服务器预检
+
+```bash
+nvidia-smi
+docker compose -f docker-compose-gpu.yml config
+bash server_gpu_test.sh preflight
+```
+
+## 2. 构建并启动测试服务
+
+```bash
+LLM_INFERENCE_BATCH_SIZE=4 bash server_gpu_test.sh start
+```
+
+对 24GB 显卡可将微批设为 `2`；48GB/80GB 显卡通常可使用 `4`。这个参数只控制 Qwen 内部微批，不会限制 API 一次提交的条数。
+
+## 3. 单条与 32 条 LLM 冒烟测试
+
+```bash
+bash server_gpu_test.sh smoke
+```
+
+另开一个终端观察峰值：
+
+```bash
+watch -n 0.5 nvidia-smi
+```
+
+预期：两个请求都返回 HTTP 200，日志中出现 `LLM inference micro-batching`，32 条请求不再把 32 条同时送入 Qwen。
+
+## 4. 确认后上线
+
+如果当前机器就是正式机，`start` 启动的已是 Compose 中的正式服务。检查：
+
+```bash
+docker compose -f docker-compose-gpu.yml ps
+docker compose -f docker-compose-gpu.yml logs --tail=200 api
+curl -fsS http://127.0.0.1:18000/
+```
+
+回滚时保留旧代码目录，然后在旧目录里重新执行 `docker compose up -d --build`。
--- a/app/api.py
+++ b/app/api.py
@ -5,32 +5,30 @@ FastAPI 配方优化 API
    uvicorn app.api:app --host 0.0.0.0 --port 8000 --reload
 """

+from contextlib import asynccontextmanager
 import os
 from pathlib import Path
-from typing import List, Dict, Optional
-from contextlib import asynccontextmanager
+from typing import Dict, List, Optional

-import torch
 from fastapi import FastAPI, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
-from pydantic import BaseModel, Field
 from loguru import logger
+from pydantic import BaseModel, Field
+import torch

-from lnp_ml.config import MODELS_DIR
-from lnp_ml.modeling.predict import load_model
 from app.optimize import (
-    optimize,
-    format_results,
    AVAILABLE_ORGANS,
+    HELPER_LIPID_OPTIONS,
+    ROUTE_OPTIONS,
    TARGET_BIODIST,
    CompRanges,
    ScoringWeights,
-    HELPER_LIPID_OPTIONS,
-    ROUTE_OPTIONS,
    create_dataframe_from_formulations,
+    optimize,
    predict_all,
 )
-
+from lnp_ml.config import MODELS_DIR
+from lnp_ml.modeling.predict import load_model

 # ============ Pydantic Models ============

@ -220,6 +218,37 @@ class PredictResponse(BaseModel):
        }


+class BatchPredictItemError(BaseModel):
+    """批量请求中单个非法条目的错误信息"""
+    index: int
+    detail: str
+
+
+class BatchPredictRequest(BaseModel):
+    """批量配方预测请求"""
+    items: List[PredictRequest] = Field(
+        ..., min_length=1, max_length=500, description="配方列表，单次最多 500 条"
+    )
+    batch_size: int = Field(
+        default=32,
+        ge=1,
+        le=256,
+        description="外层模型前向批大小；7B LLM 会按显存安全值内部微批",
+    )
+    use_llm: bool = Field(
+        default=True,
+        description="是否启用 LLM 分支；关闭可大幅提速，但会损失 delivery 精度",
+    )
+
+
+class BatchPredictResponse(BaseModel):
+    """批量配方预测响应。单条非法不中断整批，汇总在 errors 中返回"""
+    n_requested: int
+    n_succeeded: int
+    predictions: List[PredictResponse]
+    errors: List[BatchPredictItemError] = []
+
+    
 # ============ Global State ============

 class ModelState:
@ -266,6 +295,7 @@ async def lifespan(app: FastAPI):
        if _llm is not None and getattr(_llm, "use_rag", False):
            import numpy as np
            import pandas as pd
+
            from lnp_ml.dataset import LNPDataset, process_dataframe
            from lnp_ml.modeling.nested_cv_optuna import _build_rag_pool

@ -409,6 +439,112 @@ async def predict_formulation(request: PredictRequest):
    )


+@app.post("/predict/batch", response_model=BatchPredictResponse)
+async def predict_formulations_batch(request: BatchPredictRequest):
+    """批量配方预测：一次请求多条，模型侧批量前向，用于虚拟筛选。"""
+    import pandas as pd
+    from rdkit import Chem, RDLogger
+
+    if state.model is None:
+        raise HTTPException(status_code=503, detail="Model not loaded")
+
+    RDLogger.DisableLog("rdApp.*")
+
+    valid_idx: List[int] = []
+    frames = []
+    errors: List[BatchPredictItemError] = []
+
+    for i, item in enumerate(request.items):
+        if item.helper_lipid not in HELPER_LIPID_OPTIONS:
+            errors.append(BatchPredictItemError(
+                index=i, detail=f"Invalid helper_lipid: {item.helper_lipid}"))
+            continue
+        if item.route not in ROUTE_OPTIONS:
+            errors.append(BatchPredictItemError(
+                index=i, detail=f"Invalid route: {item.route}"))
+            continue
+        mol_sum = (
+            item.cationic_lipid_mol_ratio + item.phospholipid_mol_ratio
+            + item.cholesterol_mol_ratio + item.peg_lipid_mol_ratio
+        )
+        if abs(mol_sum - 100.0) > 0.5:
+            errors.append(BatchPredictItemError(
+                index=i, detail=f"四项 mol 比例之和须为 100（当前 {mol_sum:.2f}）"))
+            continue
+        if Chem.MolFromSmiles(item.smiles) is None:
+            errors.append(BatchPredictItemError(
+                index=i, detail=f"无法解析的 SMILES: {item.smiles[:80]}"))
+            continue
+
+        frames.append(create_dataframe_from_formulations(
+            item.smiles,
+            [(
+                item.cationic_lipid_to_mrna_ratio,
+                item.cationic_lipid_mol_ratio,
+                item.phospholipid_mol_ratio,
+                item.cholesterol_mol_ratio,
+                item.peg_lipid_mol_ratio,
+            )],
+            [item.helper_lipid],
+            [item.route],
+        ))
+        valid_idx.append(i)
+
+    if not frames:
+        raise HTTPException(
+            status_code=400,
+            detail=f"无合法条目；首个错误: {errors[0].detail if errors else 'unknown'}",
+        )
+
+    logger.info(f"Batch predict: {len(frames)}/{len(request.items)} valid, "
+                f"use_llm={request.use_llm}, batch_size={request.batch_size}")
+
+    df = pd.concat(frames, ignore_index=True)
+    _has_llm = getattr(state.model, "llm_prompt", None) is not None
+
+    try:
+        if _has_llm:
+            state.model.set_llm_enabled(request.use_llm)
+        df = predict_all(state.model, df, state.device, batch_size=request.batch_size)
+    except Exception as e:
+        logger.error(f"Batch prediction failed: {e}")
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        raise HTTPException(status_code=500, detail=str(e))
+    finally:
+        # /optimize 与本接口共享同一模型实例，用完必须复位，避免污染后续请求
+        if _has_llm:
+            state.model.set_llm_enabled(True)
+
+    predictions: List[PredictResponse] = []
+    for j, i in enumerate(valid_idx):
+        row = df.iloc[j]
+        item = request.items[i]
+        _size, _unnorm = row.get("pred_size"), row.get("pred_unnorm_delivery")
+        predictions.append(PredictResponse(
+            smiles=item.smiles,
+            helper_lipid=item.helper_lipid,
+            route=item.route,
+            biodist={
+                c.replace("Biodistribution_", ""): float(row[f"pred_{c}"])
+                for c in TARGET_BIODIST
+            },
+            size=float(_size) if pd.notna(_size) else None,
+            quantified_delivery=float(row["pred_delivery"]),
+            unnormalized_delivery=float(_unnorm) if pd.notna(_unnorm) else None,
+            pdi_class=int(row["pred_pdi_class"]),
+            ee_class=int(row["pred_ee_class"]),
+            toxic_class=int(row["pred_toxic_class"]),
+        ))
+
+    return BatchPredictResponse(
+        n_requested=len(request.items),
+        n_succeeded=len(predictions),
+        predictions=predictions,
+        errors=errors,
+    )
+
+    
@app.post("/optimize", response_model=OptimizeResponse)
 async def optimize_formulation(request: OptimizeRequest):
    """
@ -481,7 +617,7 @@ async def optimize_formulation(request: OptimizeRequest):
        )
        
        # 用于计算综合评分的权重
-        from app.optimize import compute_formulation_score, DEFAULT_SCORING_WEIGHTS
+        from app.optimize import DEFAULT_SCORING_WEIGHTS, compute_formulation_score
        actual_scoring_weights = scoring_weights if scoring_weights is not None else DEFAULT_SCORING_WEIGHTS
        
        # 转换结果
@ -537,4 +673,3 @@ if __name__ == "__main__":
        port=8000,
        reload=True,
    )
-
--- a/app/app.py
+++ b/app/app.py
@ -6,50 +6,32 @@ Streamlit 配方优化交互界面

 Docker 环境变量:
    API_URL: API 服务地址 (默认: http://localhost:8000)
+    APP_PASSWORD: 访问口令，设置后进入页面需先输入；不设则不启用
 """

-import io
-import os
 from datetime import datetime

 import httpx
 import pandas as pd
 import streamlit as st

-# ============ 配置 ============
-
-# 从环境变量读取 API 地址，支持 Docker 环境
-API_URL = os.environ.get("API_URL", "http://localhost:8000")
-
-AVAILABLE_ORGANS = [
-    "liver",
-    "spleen", 
-    "lung",
-    "heart",
-    "kidney",
-    "muscle",
-    "lymph_nodes",
-]
-
-ORGAN_LABELS = {
-    "liver": "肝脏 (Liver)",
-    "spleen": "脾脏 (Spleen)",
-    "lung": "肺 (Lung)",
-    "heart": "心脏 (Heart)",
-    "kidney": "肾脏 (Kidney)",
-    "muscle": "肌肉 (Muscle)",
-    "lymph_nodes": "淋巴结 (Lymph Nodes)",
-}
-
-AVAILABLE_ROUTES = [
-    "intravenous",
-    "intramuscular",
-]
-
-ROUTE_LABELS = {
-    "intravenous": "静脉注射 (Intravenous)",
-    "intramuscular": "肌肉注射 (Intramuscular)",
-}
+import batch_screening
+from ui_common import (
+    API_URL,
+    AVAILABLE_ORGANS,
+    AVAILABLE_ROUTES,
+    EE_CLASS_LABELS,
+    MODE_LABELS,
+    MODE_OPTIMIZE,
+    MODE_SCREEN,
+    ORGAN_LABELS,
+    PDI_CLASS_LABELS,
+    ROUTE_LABELS,
+    TOXIC_CLASS_LABELS,
+    check_api_status,
+    render_api_status,
+    require_auth,
+)

 # ============ 页面配置 ============

@ -60,6 +42,10 @@ st.set_page_config(
    initial_sidebar_state="expanded",
 )

+# ============ 访问口令 ============
+
+require_auth()
+
 # ============ 自定义样式 ============

 st.markdown("""
@ -127,15 +113,6 @@ st.markdown("""

 # ============ 辅助函数 ============

-def check_api_status() -> bool:
-    """检查 API 状态"""
-    try:
-        with httpx.Client(timeout=5) as client:
-            response = client.get(f"{API_URL}/")
-            return response.status_code == 200
-    except:
-        return False
-

 def call_optimize_api(
    smiles: str,
@ -174,26 +151,6 @@ def call_optimize_api(
        return response.json()


-# PDI 分类标签
-PDI_CLASS_LABELS = {
-    0: "<0.2 (优)",
-    1: "≥0.2 (欠佳)",
-}
-
-# EE 分类标签
-EE_CLASS_LABELS = {
-    0: "<50% (低)",
-    1: "50-80% (中)",
-    2: ">80% (高)",
-}
-
-# 毒性分类标签
-TOXIC_CLASS_LABELS = {
-    0: "无毒 ✓",
-    1: "有毒 ⚠",
-}
-
-
 def format_results_dataframe(results: dict, smiles_label: str = None) -> pd.DataFrame:
    """将 API 结果转换为 DataFrame"""
    formulations = results["formulations"]
@ -265,26 +222,32 @@ def main():
    # 检查 API 状态
    api_online = check_api_status()
    
+    # ========== 功能选择 ==========
+    _modes = [MODE_OPTIMIZE, MODE_SCREEN]
+    if "mode" not in st.session_state:
+        _from_url = st.query_params.get("mode")
+        st.session_state["mode"] = _from_url if _from_url in _modes else MODE_OPTIMIZE
+
+    mode = st.sidebar.radio(
+        "功能",
+        options=_modes,
+        format_func=lambda x: MODE_LABELS[x],
+        label_visibility="collapsed",
+        key="mode",
+    )
+    if st.query_params.get("mode") != mode:
+        st.query_params["mode"] = mode
+    with st.sidebar:
+        render_api_status(api_online)
+
+    if mode == MODE_SCREEN:
+        batch_screening.render(api_online)
+        return
+
    # ========== 侧边栏 ==========
    with st.sidebar:
        # st.header("⚙️ 参数设置")
        
-        # API 状态
-        if api_online:
-            st.success("🟢 API 服务在线")
-            try:
-                with httpx.Client(timeout=5) as _c:
-                    _info = _c.get(f"{API_URL}/").json()
-                _caps = [n for n, k in (("MoE", "use_moe"), ("LLM", "use_llm"), ("RAG", "use_rag"))
-                        if _info.get(k)]
-                st.caption(f"模型: {' + '.join(_caps) if _caps else '仅 backbone'}｜{_info.get('device', '?')}")
-            except Exception:
-                pass
-        else:
-            st.error("🔴 API 服务离线")
-            st.info(f"请先启动 API 服务:\n```\nuvicorn app.api:app --port 8010\n```\n"
-                    f"当前 API_URL: {API_URL}")
-        
        # st.divider()
        
        # SMILES 输入
@ -579,7 +542,7 @@ def main():
        
        # 优化按钮
        optimize_button = st.button(
-            "🚀 开始配方优选",
+            "开始配方优选",
            type="primary",
            use_container_width=True,
            disabled=not api_online or not smiles_input.strip() or not selected_routes,
@ -636,7 +599,7 @@ def main():
                except httpx.HTTPStatusError as e:
                    try:
                        error_detail = e.response.json().get("detail", str(e))
-                    except:
+                    except ValueError:
                        error_detail = str(e)
                    errors.append(f"SMILES {idx + 1}: {error_detail}")
                except httpx.RequestError as e:
@ -740,7 +703,7 @@ def main():
                target_organ = results["target_organ"]
            
            st.download_button(
-                label="📥 导出 CSV",
+                label="导出 CSV",
                data=csv_content,
                file_name=f"lnp_optimization_{target_organ}_{datetime.now().strftime('%Y%m%d_%H%M%S')}.csv",
                mime="text/csv",
--- a/app/batch_screening.py
+++ b/app/batch_screening.py
@ -0,0 +1,466 @@
+"""批量筛选页面。
+
+在一组固定的标准配方下，对上传的 SMILES 列表批量调用 /predict/batch 并排序，
+用于从大量候选脂质中初筛。本页不做配方搜索——搜索请用主页的配方优选。
+"""
+
+import io
+import json
+import os
+from pathlib import Path
+import time
+from typing import Optional
+import uuid
+
+import httpx
+import pandas as pd
+import streamlit as st
+
+from ui_common import (
+    API_URL,
+    AVAILABLE_ORGANS,
+    AVAILABLE_ROUTES,
+    EE_CLASS_LABELS,
+    HELPER_LIPID_OPTIONS,
+    ORGAN_LABELS,
+    PDI_CLASS_LABELS,
+    ROUTE_LABELS,
+    TOXIC_CLASS_LABELS,
+)
+
+CHUNK_SIZE = 100
+
+RUNS_DIR = Path(
+    os.environ.get("SCREENING_RUNS_DIR", Path.home() / ".cache" / "lnp_screening_runs")
+)
+RUN_TTL_HOURS = 48
+
+# 标准配方默认值，取自 data/interim/internal.csv 的中位配方
+DEFAULT_FORMULATION = {
+    "weight_ratio": 10.0,
+    "cationic_mol": 35.0,
+    "phospholipid_mol": 16.0,
+    "cholesterol_mol": 46.5,
+    "peg_mol": 2.5,
+}
+
+MOL_SUM_TOLERANCE = 0.5
+_SMILES_MIN_LEN = 10
+_SMILES_COL_RATIO = 0.5
+
+LAYOUT_LONG = "long"
+LAYOUT_MATRIX = "matrix"
+
+LAYOUT_LABELS = {
+    LAYOUT_LONG: "长表：某一列是 SMILES，一行一个分子",
+    LAYOUT_MATRIX: "矩阵表：行列各代表一个结构维度，单元格是 SMILES",
+}
+
+
+def looks_like_smiles(value) -> bool:
+    """判断一个单元格是否为 SMILES。"""
+    text = str(value).strip()
+    if len(text) < _SMILES_MIN_LEN or " " in text:
+        return False
+    return any(atom in text for atom in "CNOcno")
+
+
+def smiles_like_columns(df: pd.DataFrame) -> list:
+    """返回过半单元格看起来像 SMILES 的列。"""
+    return [
+        col
+        for col in df.columns
+        if df[col].map(looks_like_smiles).mean() > _SMILES_COL_RATIO
+    ]
+
+
+def guess_layout(df: pd.DataFrame) -> str:
+    """两列以上都装着 SMILES，基本只能是矩阵表。"""
+    return LAYOUT_MATRIX if len(smiles_like_columns(df)) >= 2 else LAYOUT_LONG
+
+
+def melt_matrix(
+    df: pd.DataFrame,
+    index_col,
+    value_cols: list,
+    row_name: str,
+    col_name: str,
+) -> pd.DataFrame:
+    """把矩阵表展开为 [row_name, col_name, SMILES] 三列的长表。"""
+    long_df = df.melt(
+        id_vars=[index_col],
+        value_vars=value_cols,
+        var_name=col_name,
+        value_name="SMILES",
+    ).rename(columns={index_col: row_name})
+    long_df["SMILES"] = long_df["SMILES"].astype(str).str.strip()
+    long_df[row_name] = long_df[row_name].astype(str).str.strip()
+    return long_df[long_df["SMILES"].map(looks_like_smiles)].reset_index(drop=True)
+
+
+def save_run(payload: dict) -> str:
+    """把一次筛选结果写盘，返回可放进 URL 的 token。"""
+    RUNS_DIR.mkdir(parents=True, exist_ok=True)
+    os.chmod(RUNS_DIR, 0o700)
+    token = uuid.uuid4().hex[:12]
+    # 标签列可能来自 Excel 的数值列，取出来是 numpy 标量，json 无法直接序列化，
+    # 故用 default=str 兜底。
+    (RUNS_DIR / f"{token}.json").write_text(
+        json.dumps(payload, ensure_ascii=False, default=str), encoding="utf-8"
+    )
+    return token
+
+
+def load_run(token: str) -> Optional[dict]:
+    """按 token 读回结果。token 来自 URL，属于外部输入，必须校验以防目录穿越。"""
+    if not token or not token.isalnum():
+        return None
+    path = RUNS_DIR / f"{token}.json"
+    if not path.is_file():
+        return None
+    try:
+        return json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        return None
+
+
+def purge_old_runs() -> None:
+    """清掉过期结果，避免长期堆积。"""
+    cutoff = time.time() - RUN_TTL_HOURS * 3600
+    for path in RUNS_DIR.glob("*.json"):
+        try:
+            if path.stat().st_mtime < cutoff:
+                path.unlink()
+        except OSError:
+            continue
+
+
+def render_results(run: dict) -> None:
+    """渲染一次筛选的结果表与导出按钮。run 可能来自 session_state，也可能来自磁盘。"""
+    df = pd.DataFrame(run["rows"])
+
+    # 排序是纯事后操作，放在这里而不是运行前，改排序就不必重跑整批
+    sort_options = ["delivery"] + AVAILABLE_ORGANS
+    stored = run.get("sort_key", "delivery")
+    sort_key = st.selectbox(
+        "结果排序依据",
+        options=sort_options,
+        index=sort_options.index(stored) if stored in sort_options else 0,
+        format_func=lambda x: (
+            "递送效率 (delivery)" if x == "delivery" else f"{ORGAN_LABELS[x]} 分布占比"
+        ),
+        key="_result_sort",
+    )
+    sort_col = "delivery" if sort_key == "delivery" else f"分布_{ORGAN_LABELS[sort_key]}"
+
+    df = df.sort_values(sort_col, ascending=False).reset_index(drop=True)
+    df.insert(0, "排名", range(1, len(df) + 1))
+
+    st.subheader(f"筛选结果（成功 {len(df)} / {run['n_requested']} 条）")
+    st.caption(f"运行于 {run['finished_at']}｜{run['meta_line']}")
+
+    if run["failures"]:
+        with st.expander(f"⚠️ {len(run['failures'])} 条失败，点击查看"):
+            st.code("\n".join(run["failures"][:200]))
+
+    st.dataframe(
+        df,
+        use_container_width=True,
+        hide_index=True,
+        column_config={"SMILES": st.column_config.TextColumn("SMILES", width="medium")},
+    )
+
+    buffer = io.StringIO()
+    buffer.write(f"# {run['meta_line']}, sorted_by={sort_col}\n")
+    df.to_csv(buffer, index=False)
+    st.download_button(
+        "下载完整结果 CSV",
+        data=buffer.getvalue().encode("utf-8-sig"),
+        file_name=f"screening_{len(df)}_lipids.csv",
+        mime="text/csv",
+        use_container_width=True,
+    )
+
+    if st.button("清除结果，开始新一批", use_container_width=True):
+        st.session_state.pop("screen_run", None)
+        st.query_params.pop("run", None)
+        st.rerun()
+        
+def render(api_online: bool) -> None:
+    """渲染批量筛选界面。由 app.py 在「批量筛选」模式下调用。"""
+    st.subheader("批量筛选")
+    st.caption(
+        "在同一组标准配方下批量预测多个脂质的性能并排序，用于从大量候选中初筛。"
+        "初筛出的前几十条可切回「配方优选」做配方搜索。"
+    )
+
+    if not api_online:
+        st.error(f"API 服务离线，无法筛选。当前 API_URL: {API_URL}")
+        st.stop()
+
+    if "screen_run" not in st.session_state:
+        purge_old_runs()
+        restored = load_run(st.query_params.get("run", ""))
+        if restored:
+            st.session_state["screen_run"] = restored
+
+
+    # ============ 1. 上传与解析 ============
+
+    uploaded = st.file_uploader(
+        "脂质库文件",
+        type=["xlsx", "xls", "csv"],
+        help="支持两种布局：一列 SMILES 的长表，或行列各代表一个结构维度的矩阵表。",
+    )
+    if uploaded is None:
+        # 刷新后文件对象丢失，但结果已从磁盘还原，此处直接展示，不必重新上传
+        if st.session_state.get("screen_run"):
+            st.info("已恢复上一次的筛选结果。要跑新一批，重新上传文件即可。")
+            render_results(st.session_state["screen_run"])
+        st.stop()
+
+    try:
+        if uploaded.name.lower().endswith(".csv"):
+            df_in = pd.read_csv(uploaded)
+        else:
+            df_in = pd.read_excel(uploaded)
+    except (ValueError, pd.errors.ParserError, OSError) as exc:
+        st.error(f"读取文件失败: {exc}")
+        st.stop()
+
+    if df_in.empty:
+        st.error("文件内容为空")
+        st.stop()
+
+    st.success(f"已读取 {len(df_in)} 行、{len(df_in.columns)} 列")
+    with st.expander("预览前 5 行"):
+        st.dataframe(df_in.head(), use_container_width=True)
+
+    _columns = list(df_in.columns)
+    _smiles_cols = smiles_like_columns(df_in)
+    _layouts = [LAYOUT_LONG, LAYOUT_MATRIX]
+
+    layout = st.radio(
+        "表格布局",
+        options=_layouts,
+        index=_layouts.index(guess_layout(df_in)),
+        format_func=lambda x: LAYOUT_LABELS[x],
+        help="已按文件内容自动判断，判断错误时可手动切换。",
+    )
+
+    if layout == LAYOUT_MATRIX:
+        col_left, col_right = st.columns(2)
+        with col_left:
+            index_col = st.selectbox(
+                "行标签所在列",
+                options=_columns,
+                index=0,
+                help="通常是最左侧那一列，例如头部基团编号。",
+            )
+            row_name = st.text_input("行维度名称", value="头部基团")
+        with col_right:
+            value_cols = st.multiselect(
+                "SMILES 单元格所在列",
+                options=[c for c in _columns if c != index_col],
+                default=[c for c in _smiles_cols if c != index_col],
+            )
+            col_name = st.text_input("列维度名称", value="尾链")
+
+        if not value_cols:
+            st.error("请至少选择一列 SMILES 单元格")
+            st.stop()
+
+        parsed = melt_matrix(df_in, index_col, value_cols, row_name, col_name)
+        meta_cols = [row_name, col_name]
+    else:
+        _guess = next(
+            iter(_smiles_cols),
+            next((c for c in _columns if "smiles" in str(c).lower()), _columns[0]),
+        )
+        smiles_col = st.selectbox(
+            "SMILES 所在列", options=_columns, index=_columns.index(_guess)
+        )
+        meta_cols = st.multiselect(
+            "一并带入结果的标签列",
+            options=[c for c in _columns if c != smiles_col],
+            help="例如分子名称、编号、来源。这些列不参与预测，只原样写入结果表。",
+        )
+        parsed = df_in[[smiles_col] + meta_cols].rename(columns={smiles_col: "SMILES"})
+        parsed["SMILES"] = parsed["SMILES"].astype(str).str.strip()
+        parsed = parsed[parsed["SMILES"].map(looks_like_smiles)].reset_index(drop=True)
+
+    if parsed.empty:
+        st.error("没有解析出任何有效的 SMILES，请检查布局设置与列选择")
+        st.stop()
+
+    n_parsed = len(parsed)
+    if st.checkbox("去除重复 SMILES", value=True):
+        parsed = parsed.drop_duplicates(subset="SMILES").reset_index(drop=True)
+
+    records = parsed.to_dict("records")
+    _dedup_note = f"（原始 {n_parsed} 条，已去重）" if len(records) != n_parsed else ""
+    st.write(f"待筛选：**{len(records)}** 条{_dedup_note}")
+
+    with st.expander("确认解析结果（前 10 条）"):
+        st.dataframe(parsed.head(10), use_container_width=True, hide_index=True)
+
+
+    # ============ 2. 标准配方 ============
+
+    st.subheader("统一使用的标准配方")
+    st.caption("所有 SMILES 都在同一组配方下预测，便于横向比较。默认值取自内部数据的中位配方。")
+
+    col_a, col_b, col_c = st.columns(3)
+    with col_a:
+        weight_ratio = st.number_input(
+            "脂质/mRNA 重量比", 0.1, 50.0, DEFAULT_FORMULATION["weight_ratio"], 0.5
+        )
+        cationic_mol = st.number_input(
+            "阳离子脂质 mol%", 0.0, 100.0, DEFAULT_FORMULATION["cationic_mol"], 0.5
+        )
+    with col_b:
+        phospholipid_mol = st.number_input(
+            "磷脂 mol%", 0.0, 100.0, DEFAULT_FORMULATION["phospholipid_mol"], 0.5
+        )
+        cholesterol_mol = st.number_input(
+            "胆固醇 mol%", 0.0, 100.0, DEFAULT_FORMULATION["cholesterol_mol"], 0.5
+        )
+    with col_c:
+        peg_mol = st.number_input(
+            "PEG 脂质 mol%", 0.0, 20.0, DEFAULT_FORMULATION["peg_mol"], 0.1
+        )
+        helper_lipid = st.selectbox("辅助脂质", HELPER_LIPID_OPTIONS)
+
+    route = st.selectbox(
+        "给药途径", AVAILABLE_ROUTES, format_func=lambda x: ROUTE_LABELS[x]
+    )
+
+    mol_sum = cationic_mol + phospholipid_mol + cholesterol_mol + peg_mol
+    if abs(mol_sum - 100.0) > MOL_SUM_TOLERANCE:
+        st.error(f"四项 mol 比例之和须为 100%，当前为 {mol_sum:.1f}%")
+        st.stop()
+    st.caption(f"mol 比例合计 {mol_sum:.1f}% ✓")
+
+
+    # ============ 3. 运行选项 ============
+
+    use_llm = st.checkbox(
+        "启用 LLM 分支（更准，但明显更慢）",
+        value=True,
+        help="关闭后 delivery 精度会下降，但吞吐大幅提升。条数很多时可先关闭做粗筛。",
+    )
+
+    if st.button("开始批量筛选", type="primary", use_container_width=True):
+
+        # ============ 4. 分块调用 /predict/batch ============
+
+        def build_item(smiles: str) -> dict:
+            """构造 /predict/batch 的单个 items 元素。"""
+            return {
+                "smiles": smiles,
+                "cationic_lipid_to_mrna_ratio": weight_ratio,
+                "cationic_lipid_mol_ratio": cationic_mol,
+                "phospholipid_mol_ratio": phospholipid_mol,
+                "cholesterol_mol_ratio": cholesterol_mol,
+                "peg_lipid_mol_ratio": peg_mol,
+                "helper_lipid": helper_lipid,
+                "route": route,
+            }
+
+        def flatten_prediction(pred: dict) -> dict:
+            """把 PredictResponse 展平为一行表格数据。"""
+            row = {
+                "SMILES": pred["smiles"],
+                "delivery": pred["quantified_delivery"],
+                "delivery_原始值": pred.get("unnormalized_delivery"),
+                "粒径_nm": pred.get("size"),
+                "PDI": PDI_CLASS_LABELS.get(pred.get("pdi_class")),
+                "EE": EE_CLASS_LABELS.get(pred.get("ee_class")),
+                "毒性": TOXIC_CLASS_LABELS.get(pred.get("toxic_class")),
+            }
+            for organ, value in pred["biodist"].items():
+                row[f"分布_{ORGAN_LABELS.get(organ, organ)}"] = value
+            return row
+
+        chunks = [records[i:i + CHUNK_SIZE] for i in range(0, len(records), CHUNK_SIZE)]
+        progress = st.progress(0.0)
+        status = st.empty()
+        rows: list[dict] = []
+        failures: list[str] = []
+        started = time.time()
+
+        with httpx.Client(timeout=1800) as client:
+            for idx, chunk in enumerate(chunks):
+                done = idx * CHUNK_SIZE
+                eta = ""
+                if done:
+                    per_item = (time.time() - started) / done
+                    eta = f"，预计剩余 {per_item * (len(records) - done) / 60:.1f} 分钟"
+                status.text(
+                    f"批次 {idx + 1}/{len(chunks)}"
+                    f"（已完成 {done}/{len(records)} 条）{eta}"
+                )
+
+                try:
+                    response = client.post(
+                        f"{API_URL}/predict/batch",
+                        json={
+                            "items": [build_item(r["SMILES"]) for r in chunk],
+                            "batch_size": 32,
+                            "use_llm": use_llm,
+                        },
+                    )
+                    response.raise_for_status()
+                    payload = response.json()
+                except httpx.HTTPError as exc:
+                    failures.append(f"批次 {idx + 1} 整体失败: {exc}")
+                    progress.progress((idx + 1) / len(chunks))
+                    continue
+
+                error_idx = {err["index"] for err in payload.get("errors", [])}
+                for err in payload.get("errors", []):
+                    failures.append(
+                        f"{chunk[err['index']]['SMILES'][:60]}: {err['detail']}"
+                    )
+
+                valid = [r for j, r in enumerate(chunk) if j not in error_idx]
+                for record, pred in zip(valid, payload["predictions"]):
+                    row = {col: record.get(col) for col in meta_cols}
+                    row.update(flatten_prediction(pred))
+                    rows.append(row)
+
+                progress.progress((idx + 1) / len(chunks))
+
+        progress.progress(1.0)
+        status.text(f"完成，用时 {(time.time() - started) / 60:.1f} 分钟")
+
+
+        # ============ 5. 结果入库 ============
+
+        if not rows:
+            st.error("没有任何成功的预测结果")
+            if failures:
+                st.code("\n".join(failures[:50]))
+            st.stop()
+
+        run = {
+            "rows": rows,
+            "failures": failures,
+            "n_requested": len(records),
+            "sort_key": "delivery",
+            "finished_at": time.strftime("%Y-%m-%d %H:%M"),
+            "meta_line": (
+                f"标准配方: weight_ratio={weight_ratio}, cationic={cationic_mol}, "
+                f"phospholipid={phospholipid_mol}, cholesterol={cholesterol_mol}, "
+                f"peg={peg_mol}, helper_lipid={helper_lipid}, route={route}, "
+                f"use_llm={use_llm}"
+            ),
+        }
+        st.session_state["screen_run"] = run
+        st.query_params["run"] = save_run(run)
+
+
+    # ============ 6. 结果展示 ============
+
+    if st.session_state.get("screen_run"):
+        render_results(st.session_state["screen_run"])
--- a/app/optimize.py
+++ b/app/optimize.py
@ -7,30 +7,24 @@
    python -m app.optimize --smiles "CC(C)..." --organ liver
 """

-import itertools
+from dataclasses import dataclass, field
 import json
 from pathlib import Path
 from typing import Dict, List, Optional, Tuple
-from dataclasses import dataclass, field

+from loguru import logger
 import numpy as np
 import pandas as pd
 import torch
 from torch.utils.data import DataLoader
-from loguru import logger
-from tqdm import tqdm
 import typer

 from lnp_ml.config import MODELS_DIR
 from lnp_ml.dataset import (
-    LNPDataset,
-    LNPDatasetConfig,
-    collate_fn,
    SMILES_COL,
-    COMP_COLS,
-    HELP_COLS,
    TARGET_BIODIST,
-    get_phys_cols,
+    LNPDataset,
+    collate_fn,
    get_exp_cols,
 )
 from lnp_ml.modeling.predict import load_model
@ -577,7 +571,9 @@ def predict_all(
    all_ee_preds = []
    all_toxic_preds = []
    
-    with torch.no_grad():
+    # inference_mode 比 no_grad 还会关闭 version counter/view tracking，
+    # 对大批量 LLM 推理可再减少一部分临时内存。
+    with torch.inference_mode():
        for batch in dataloader:
            smiles = batch["smiles"]
            tabular = {k: v.to(device) for k, v in batch["tabular"].items()}
@ -1151,4 +1147,3 @@ def main(

 if __name__ == "__main__":
    app()
-
--- a/app/ui_common.py
+++ b/app/ui_common.py
@ -0,0 +1,141 @@
+"""Streamlit 前端共享配置与工具。
+"""
+
+import os
+
+import httpx
+import streamlit as st
+
+# ============ API ============
+
+API_URL = os.environ.get("API_URL", "http://localhost:8000")
+
+
+# ============ 选项与标签 ============
+
+AVAILABLE_ORGANS = [
+    "liver",
+    "spleen",
+    "lung",
+    "heart",
+    "kidney",
+    "muscle",
+    "lymph_nodes",
+]
+
+ORGAN_LABELS = {
+    "liver": "肝脏 (Liver)",
+    "spleen": "脾脏 (Spleen)",
+    "lung": "肺 (Lung)",
+    "heart": "心脏 (Heart)",
+    "kidney": "肾脏 (Kidney)",
+    "muscle": "肌肉 (Muscle)",
+    "lymph_nodes": "淋巴结 (Lymph Nodes)",
+}
+
+AVAILABLE_ROUTES = [
+    "intravenous",
+    "intramuscular",
+]
+
+ROUTE_LABELS = {
+    "intravenous": "静脉注射 (Intravenous)",
+    "intramuscular": "肌肉注射 (Intramuscular)",
+}
+
+HELPER_LIPID_OPTIONS = ["DOPE", "DSPC"]
+
+# PDI 分类标签
+PDI_CLASS_LABELS = {
+    0: "<0.2 (优)",
+    1: "≥0.2 (欠佳)",
+}
+
+EE_CLASS_LABELS = {
+    0: "<50% (低)",
+    1: "50-80% (中)",
+    2: ">80% (高)",
+}
+
+TOXIC_CLASS_LABELS = {
+    0: "无毒 ✓",
+    1: "有毒 ⚠",
+}
+
+
+# ============ 功能模式 ============
+
+MODE_OPTIMIZE = "optimize"
+MODE_SCREEN = "screen"
+
+MODE_LABELS = {
+    MODE_OPTIMIZE: "配方优选",
+    MODE_SCREEN: "批量筛选",
+}
+
+# ============ 访问口令 ============
+
+def check_password() -> bool:
+    """校验访问口令，未设置 APP_PASSWORD 时直接放行，内网访问不受影响。"""
+    expected = os.environ.get("APP_PASSWORD")
+    if not expected:
+        return True
+    if st.session_state.get("_authed"):
+        return True
+
+    gate = st.empty()
+    with gate.container():
+        st.markdown("### LNP 配方优化")
+        pwd = st.text_input("🔒 访问口令", type="password")
+        if pwd and pwd != expected:
+            st.error("口令不正确")
+
+    if pwd == expected:
+        st.session_state["_authed"] = True
+        gate.empty()
+        return True
+    return False
+
+
+def require_auth() -> None:
+    """口令未通过时终止页面渲染。应用只有 app.py 一个入口，在其顶部调用一次即可。"""
+    if not check_password():
+        st.stop()
+
+
+# ============ API 客户端 ============
+
+def check_api_status(timeout: float = 5.0) -> bool:
+    """检查 API 是否在线。"""
+    try:
+        with httpx.Client(timeout=timeout) as client:
+            return client.get(f"{API_URL}/").status_code == 200
+    except httpx.HTTPError:
+        return False
+
+
+def render_api_status(api_online: bool) -> None:
+    """在侧边栏渲染 API 状态与模型能力，两种功能模式共用。"""
+    if not api_online:
+        st.error("🔴 API 服务离线")
+        st.info(
+            f"请先启动 API 服务:\n```\nuvicorn app.api:app --port 8010\n```\n"
+            f"当前 API_URL: {API_URL}"
+        )
+        return
+
+    st.success("🟢 API 服务在线")
+    try:
+        with httpx.Client(timeout=5) as client:
+            info = client.get(f"{API_URL}/").json()
+    except httpx.HTTPError:
+        return
+
+    caps = [
+        name
+        for name, key in (("MoE", "use_moe"), ("LLM", "use_llm"), ("RAG", "use_rag"))
+        if info.get(key)
+    ]
+    st.caption(
+        f"模型: {' + '.join(caps) if caps else '仅 backbone'}｜{info.get('device', '?')}"
+    )
--- a/docker-compose-gpu.yml
+++ b/docker-compose-gpu.yml
@ -6,12 +6,20 @@ services:
      dockerfile: Dockerfile
      target: api
    container_name: lnp-api
+    ports:
+      # 仅绑定本机，供服务器上的冒烟测试使用，不直接暴露到公网。
+      - "127.0.0.1:${API_PORT:-18000}:8000"
    environment:
      - MODEL_PATH=/app/models/final/model.pt
+      - RAG_POOL_CSV=/app/data/interim/internal.csv
+      # 7B LLM 内部微批；外层批量筛选仍可一次提交更多条。
+      - LLM_INFERENCE_BATCH_SIZE=${LLM_INFERENCE_BATCH_SIZE:-4}
    volumes:
-      # 挂载模型目录以便更新模型
+      # 权重和 RAG 数据不进镜像，服务器上需预先放到这些路径。
      - ./models/final:/app/models/final:ro
      - ./models/mpnn:/app/models/mpnn:ro
+      - ./models/qwen2.5-7b-instruct:/app/models/qwen2.5-7b-instruct:ro
+      - ./data/interim/internal.csv:/app/data/interim/internal.csv:ro
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8000/"]
--- a/lnp_ml/modeling/layers/llm_prompt.py
+++ b/lnp_ml/modeling/layers/llm_prompt.py
@ -9,6 +9,16 @@ import torch.nn as nn

 DEFAULT_MOLT5_PATH = os.environ.get("MOLT5_PATH", "models/molt5-base")

+
+def _positive_int_env(name: str, default: int) -> int:
+    """Read a positive integer environment variable with a safe fallback."""
+    try:
+        value = int(os.environ.get(name, str(default)))
+    except (TypeError, ValueError):
+        return default
+    return value if value > 0 else default
+
+
 # 检索池支持的额外多任务（除 delivery 外）
 _EXTRA_TASKS = ["size", "pdi", "ee", "toxic", "biodist"]

@ -58,6 +68,9 @@ class LLMPromptEncoder(nn.Module):
        self.use_rag = use_rag
        self.rag_top_k = rag_top_k
        self.use_soft_prompt = use_soft_prompt
+        # 7B 模型即使在 no_grad 下，大批量长 prompt 的临时激活也很大。
+        # 外层 DataLoader 仍可用大 batch；这里只把 LLM 前向切成小块。
+        self.inference_batch_size = _positive_int_env("LLM_INFERENCE_BATCH_SIZE", 4)

        _name_l = model_name_or_path.lower()
        _is_t5 = "t5" in _name_l
@ -145,16 +158,18 @@ class LLMPromptEncoder(nn.Module):
    def _apply_lora(self, r, alpha, dropout, prepare_kbit=False):
        from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training

-        if prepare_kbit:
+        # 4-bit/8-bit 基座只 prepare 一次。PEFT 会处理输入梯度并开启
+        # gradient checkpointing；后面再显式开一次，兼容不同 PEFT 版本。
+        is_kbit = (
+            prepare_kbit
+            or getattr(self.encoder, "is_loaded_in_4bit", False)
+            or getattr(self.encoder, "is_loaded_in_8bit", False)
+        )
+        if is_kbit:
            self.encoder = prepare_model_for_kbit_training(self.encoder)
        else:
            for p in self.encoder.parameters():
                p.requires_grad = False
-
-        # 4-bit 量化模型(QLoRA)需先 prepare，才能正确接收梯度
-        if getattr(self.encoder, "is_loaded_in_4bit", False) or getattr(self.encoder, "is_loaded_in_8bit", False):
-            from peft import prepare_model_for_kbit_training
-            self.encoder = prepare_model_for_kbit_training(self.encoder)
        for p in self.encoder.parameters():
            p.requires_grad = False
        # 不同架构注意力层命名不同：
@ -170,6 +185,24 @@ class LLMPromptEncoder(nn.Module):
                         target_modules=_targets, bias="none")
        self.encoder = get_peft_model(self.encoder, cfg)

+        # Decoder-only LLM 做表征编码不需要 KV cache。训练时保留 cache
+        # 只会额外占显存；梯度检查点则用计算换显存。
+        if hasattr(self.encoder, "config"):
+            self.encoder.config.use_cache = False
+        if hasattr(self.encoder, "gradient_checkpointing_enable"):
+            try:
+                self.encoder.gradient_checkpointing_enable(
+                    gradient_checkpointing_kwargs={"use_reentrant": False}
+                )
+            except TypeError:
+                self.encoder.gradient_checkpointing_enable()
+
+    def _encoder_forward(self, **kwargs):
+        """Forward the backbone without allocating an unused decoder KV cache."""
+        if self._is_qwen:
+            kwargs["use_cache"] = False
+        return self.encoder(**kwargs)
+
    # ---------- 检索池 ----------
    def set_retrieval_pool(self, smiles_list, labels, pool_id="train", extra_labels=None):
        """设置 RAG 检索池（防泄漏：只传训练集分子与标签）。
@ -326,14 +359,15 @@ class LLMPromptEncoder(nn.Module):
    def _rag_encode_batch(self, prompts, device):
        """编码一批 RAG prompt，取最后有效 token。grad 由调用方上下文决定。"""
        outs = []
-        for i in range(0, len(prompts), 4):
-            bt = prompts[i:i+4]
+        chunk_size = self.inference_batch_size
+        for i in range(0, len(prompts), chunk_size):
+            bt = prompts[i:i + chunk_size]
            enc = self.tokenizer(
                bt, padding=True, truncation=True,
                max_length=self.max_length, return_tensors="pt",
            ).to(device)
            self._warn_if_truncated(enc)
-            out = self.encoder(**enc).last_hidden_state          # [B,L,H]
+            out = self._encoder_forward(**enc).last_hidden_state  # [B,L,H]
            lengths = enc["attention_mask"].sum(1) - 1
            b = torch.arange(out.size(0), device=device)
            outs.append(out[b, lengths.long(), :].float())       # [B,H]
@ -359,7 +393,8 @@ class LLMPromptEncoder(nn.Module):
            return self._rag_encode_batch(prompts, device)

    # ---------- soft-prompt 编码（带梯度，不缓存特征）----------
-    def _encode_softrag(self, smiles, chem, tab, device) -> torch.Tensor:
+    def _encode_softrag_chunk(self, smiles, chem, tab, device) -> torch.Tensor:
+        """单个 LLM micro-batch 的 soft-RAG 前向。"""
        prompts = [self._get_prompt(s) for s in smiles]
        enc = self.tokenizer(
            prompts, padding=True, truncation=True,
@ -393,7 +428,7 @@ class LLMPromptEncoder(nn.Module):
        position_ids = attn_mask.long().cumsum(-1) - 1
        position_ids = position_ids.masked_fill(attn_mask == 0, 1)

-        out = self.encoder(
+        out = self._encoder_forward(
            inputs_embeds=inputs_embeds,
            attention_mask=attn_mask,
            position_ids=position_ids,
@ -407,6 +442,43 @@ class LLMPromptEncoder(nn.Module):
            feat = out[b, lengths.long(), :]
        return self.proj_down(feat.float())

+    def _encode_softrag(self, smiles, chem, tab, device) -> torch.Tensor:
+        """
+        soft-RAG 编码。推理时自动微批，避免 batch API 把 7B LLM 的
+        临时激活按请求 batch_size 线性放大。训练仍保持原 batch，
+        由 gradient checkpointing 降低激活显存。
+        """
+        batch_size = len(smiles)
+        micro_batch = self.inference_batch_size
+        should_chunk = (
+            not self.training
+            and not torch.is_grad_enabled()
+            and batch_size > micro_batch
+        )
+        if not should_chunk:
+            return self._encode_softrag_chunk(smiles, chem, tab, device)
+
+        if not getattr(self, "_microbatch_logged", False):
+            from loguru import logger
+
+            logger.info(
+                f"LLM inference micro-batching: batch={batch_size}, "
+                f"micro_batch={micro_batch} (override with LLM_INFERENCE_BATCH_SIZE)"
+            )
+            self._microbatch_logged = True
+
+        outputs = []
+        for start in range(0, batch_size, micro_batch):
+            end = min(start + micro_batch, batch_size)
+            chem_chunk = chem[start:end] if chem is not None else None
+            tab_chunk = tab[start:end] if tab is not None else None
+            outputs.append(
+                self._encode_softrag_chunk(
+                    smiles[start:end], chem_chunk, tab_chunk, device
+                )
+            )
+        return torch.cat(outputs, dim=0)
+
    # ---------- 旧路径（保留，向后兼容）----------
    def _mean_pool(self, last_hidden, mask):
        m = mask.unsqueeze(-1).float()
@ -417,12 +489,13 @@ class LLMPromptEncoder(nn.Module):
        missing = [s for s in smiles if s not in self._cache]
        if missing:
            uniq = list(dict.fromkeys(missing))
-            for i in range(0, len(uniq), 256):
-                chunk = uniq[i:i + 256]
+            chunk_size = self.inference_batch_size if self._is_qwen else 256
+            for i in range(0, len(uniq), chunk_size):
+                chunk = uniq[i:i + chunk_size]
                mols = [self._fmt_mol(s) for s in chunk]
                enc = self.tokenizer(mols, padding=True, truncation=True,
                                     max_length=self.max_length, return_tensors="pt").to(device)
-                out = self.encoder(**enc).last_hidden_state
+                out = self._encoder_forward(**enc).last_hidden_state
                pooled = self._mean_pool(out, enc["attention_mask"])
                for s, v in zip(chunk, pooled):
                    self._cache[s] = v.cpu()
@ -432,7 +505,7 @@ class LLMPromptEncoder(nn.Module):
        mols = [self._fmt_mol(s) for s in smiles]
        enc = self.tokenizer(mols, padding=True, truncation=True,
                             max_length=self.max_length, return_tensors="pt").to(device)
-        out = self.encoder(**enc).last_hidden_state
+        out = self._encoder_forward(**enc).last_hidden_state
        return self._mean_pool(out, enc["attention_mask"])

    def forward(self, smiles: List[str], chem: Optional[torch.Tensor] = None,
@ -448,4 +521,4 @@ class LLMPromptEncoder(nn.Module):

    def clear_cache(self) -> None:
        self._cache.clear()
-        self._prompt_cache.clear()
+        self._prompt_cache.clear()
--- a/lnp_ml/modeling/predict.py
+++ b/lnp_ml/modeling/predict.py
@ -118,7 +118,7 @@ def load_model(
    return model


-@torch.no_grad()
+@torch.inference_mode()
 def predict_batch(
    model: Union[LNPModel, LNPModelWithoutMPNN],
    loader: DataLoader,
--- a/pixi.toml
+++ b/pixi.toml
@ -31,5 +31,6 @@ uvicorn = ">=0.33.0, <0.34"
 optuna = ">=4.5.0, <5"
 captum = ">=0.7.0, <0.8"
 transformers = ">=4.30, <4.46"
+peft = "==0.13.2"
 sentencepiece = "*"
 protobuf = "*"
--- a/pyproject.toml
+++ b/pyproject.toml
@ -20,11 +20,11 @@ requires-python = ">=3.8"

 [tool.ruff]
 line-length = 99
-src = ["lnp_ml"]
-include = ["pyproject.toml", "lnp_ml/**/*.py"]
+src = ["lnp_ml", "app"]
+include = ["pyproject.toml", "lnp_ml/**/*.py", "app/**/*.py"]

 [tool.ruff.lint]
-extend-select = ["I"]  # Add import sorting
+select = ["E4", "E7", "E9", "F", "I"]

 [tool.ruff.lint.isort]
 known-first-party = ["lnp_ml"]
--- a/requirements.txt
+++ b/requirements.txt
@ -21,5 +21,8 @@ uvicorn>=0.33.0,<0.34
 optuna>=4.5.0,<5
 captum>=0.7.0
 transformers>=4.30,<4.46
+peft==0.13.2
 sentencepiece
 protobuf
+bitsandbytes
+accelerate
--- a/server_gpu_test.sh
+++ b/server_gpu_test.sh
@ -0,0 +1,108 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+cd "${ROOT_DIR}"
+
+COMPOSE_FILE=${COMPOSE_FILE:-docker-compose-gpu.yml}
+API_PORT=${API_PORT:-18000}
+API_URL=${API_URL:-http://127.0.0.1:${API_PORT}}
+LLM_INFERENCE_BATCH_SIZE=${LLM_INFERENCE_BATCH_SIZE:-4}
+export API_PORT
+export LLM_INFERENCE_BATCH_SIZE
+
+fail() {
+  echo "ERROR: $*" >&2
+  exit 1
+}
+
+assert_real_file() {
+  local path=$1
+  [ -f "${path}" ] || fail "missing file: ${path}"
+  if head -n 1 "${path}" 2>/dev/null | grep -q "git-lfs.github.com/spec"; then
+    fail "${path} is a Git LFS pointer, not the real artifact"
+  fi
+}
+
+preflight() {
+  command -v docker >/dev/null || fail "docker is not installed"
+  command -v nvidia-smi >/dev/null || fail "nvidia-smi is not installed"
+  nvidia-smi >/dev/null || fail "NVIDIA driver/GPU is unavailable"
+  docker compose version >/dev/null || fail "docker compose plugin is unavailable"
+
+  assert_real_file models/final/model.pt
+  assert_real_file data/interim/internal.csv
+  assert_real_file models/qwen2.5-7b-instruct/config.json
+
+  local fold
+  for fold in 0 1 2 3 4; do
+    assert_real_file "models/mpnn/all_amine_split_for_LiON/cv_${fold}/fold_0/model_0/model.pt"
+  done
+
+  find models/qwen2.5-7b-instruct -maxdepth 1 -type f \
+    \( -name '*.safetensors' -o -name '*.bin' \) -print -quit | grep -q . \
+    || fail "Qwen weight shards (*.safetensors or *.bin) are missing"
+
+  docker compose -f "${COMPOSE_FILE}" config >/dev/null
+  echo "preflight: OK"
+}
+
+wait_for_api() {
+  local attempt
+  for attempt in $(seq 1 120); do
+    if curl -fsS "${API_URL}/" >/dev/null 2>&1; then
+      echo "API ready after ${attempt} checks"
+      return 0
+    fi
+    sleep 2
+  done
+  docker compose -f "${COMPOSE_FILE}" logs --tail=200 api || true
+  fail "API did not become healthy within 240 seconds"
+}
+
+start() {
+  preflight
+  docker compose -f "${COMPOSE_FILE}" build
+  docker compose -f "${COMPOSE_FILE}" up -d
+  wait_for_api
+  docker compose -f "${COMPOSE_FILE}" ps
+  curl -fsS "${API_URL}/"
+  echo
+}
+
+request_item() {
+  printf '%s' '{"smiles":"CC(C)NCCNC(C)C","cationic_lipid_to_mrna_ratio":10.0,"cationic_lipid_mol_ratio":35.0,"phospholipid_mol_ratio":16.0,"cholesterol_mol_ratio":46.5,"peg_lipid_mol_ratio":2.5,"helper_lipid":"DOPE","route":"intravenous"}'
+}
+
+smoke() {
+  wait_for_api
+
+  local item payload i
+  item=$(request_item)
+  curl -fsS -H 'Content-Type: application/json' \
+    -d "${item}" "${API_URL}/predict" >/tmp/lnp-single-response.json
+  echo "single prediction: OK"
+
+  payload='{"items":['
+  for i in $(seq 1 32); do
+    [ "${i}" -eq 1 ] || payload+=','
+    payload+="${item}"
+  done
+  payload+='],"batch_size":32,"use_llm":true}'
+
+  curl -fsS -H 'Content-Type: application/json' \
+    -d "${payload}" "${API_URL}/predict/batch" >/tmp/lnp-batch-response.json
+  grep -q '"n_succeeded":32' /tmp/lnp-batch-response.json \
+    || fail "batch response did not report 32 successes"
+  echo "32-item LLM prediction: OK"
+
+  docker compose -f "${COMPOSE_FILE}" logs --tail=200 api \
+    | grep -E 'LLM inference micro-batching|ERROR|CUDA out of memory' || true
+}
+
+case "${1:-}" in
+  preflight) preflight ;;
+  start) start ;;
+  smoke) smoke ;;
+  *) echo "Usage: $0 {preflight|start|smoke}" >&2; exit 2 ;;
+esac
--- a/server_test_cases/README.md
+++ b/server_test_cases/README.md
@ -0,0 +1,88 @@
+# LNP API 与 GPU 显存测试
+
+这里有两套测试：
+
+- `run_lnp_batch_stress.sh`：专门复现 LNP 批量调用的显存问题。
+- `run_cases.sh`：检查健康状态、器官列表、单条预测、批量预测、异常输入和小规模配方优化。
+
+## 1. 复现 LNP 批量显存场景
+
+推荐先运行与问题最接近的场景：
+
+```bash
+bash server_test_cases/run_lnp_batch_stress.sh repro
+```
+
+`repro` 会依次提交：
+
+1. 32 条不同 LNP 配方，`use_llm=true`、外层 `batch_size=32`。这是原先最容易从约 10GB 涨到 63GB 的场景。
+2. 500 条不同 LNP 配方，仍使用外层 `batch_size=32`，检查多轮处理后是否持续涨显存。
+
+逐级观察批量大小与显存的关系：
+
+```bash
+bash server_test_cases/run_lnp_batch_stress.sh ladder
+```
+
+完整测试（关闭 LLM 对照 + 阶梯 + 500 条持续压力）：
+
+```bash
+bash server_test_cases/run_lnp_batch_stress.sh full
+```
+
+每个场景都会打印并写入 CSV：运行前显存、采样峰值、显存增量、耗时、HTTP 状态。结果文件名形如 `vram-results-20260922-160000.csv`。
+
+可以设置显存上限，让超过阈值时测试直接失败。例如期望不超过 20GB：
+
+```bash
+MAX_GPU_MIB=20480 bash server_test_cases/run_lnp_batch_stress.sh repro
+```
+
+这里采样的是整张卡的显存；测试时请尽量不要让其他任务共用该卡。
+
+## 2. Mock 数据
+
+已附带三组可直接提交的数据：
+
+- `mock_batch_32_llm.json`：核心显存回归场景。
+- `mock_batch_500_llm.json`：最大请求量与持续处理场景。
+- `mock_batch_128_no_llm.json`：关闭 LLM 的对照组。
+
+这些数据从项目 `data/interim/internal.csv` 抽取不同 SMILES，再随机生成满足总和 100% 的组分比例。也可以自行生成：
+
+```bash
+python3 server_test_cases/generate_mock_batch.py \
+  --count 200 \
+  --batch-size 32 \
+  --use-llm \
+  --source-csv data/interim/internal.csv \
+  --output /tmp/lnp-batch-200.json
+```
+
+生成过程使用固定随机种子，便于前后版本用完全相同的数据对比。
+
+## 3. 其他功能测试
+
+服务启动后执行：
+
+```bash
+bash server_test_cases/run_cases.sh
+```
+
+如果 API 端口或 GPU 编号不同：
+
+```bash
+API_PORT=18000 GPU_INDEX=0 bash server_test_cases/run_cases.sh
+```
+
+用例包含：
+
+1. 健康检查与模型状态。
+2. 可用器官列表。
+3. 启用 LLM 的单条预测，检查输出完整性、有限数和 biodistribution 归一化。
+4. 64 条关闭 LLM 的吞吐对照。
+5. 32 条启用 LLM 的显存回归测试。
+6. 2 条合法 + 3 条非法输入，检查批量接口的局部容错。
+7. 小范围 `/optimize` 配方搜索，检查优化功能和返回结果。
+
+GPU 显存采样来自整张卡；如果还有其他进程使用同一张卡，请以 `nvidia-smi` 中 API 进程的变化为准。
--- a/server_test_cases/generate_mock_batch.py
+++ b/server_test_cases/generate_mock_batch.py
@ -0,0 +1,102 @@
+#!/usr/bin/env python3
+"""Generate reproducible /predict/batch payloads for GPU memory tests."""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import random
+from pathlib import Path
+
+
+# Used only when data/interim/internal.csv is not available.  These are valid,
+# deliberately varied SMILES; the project dataset is preferred for realistic RAG prompts.
+FALLBACK_SMILES = [
+    "CC(C)NCCNC(C)C",
+    "CCCCN(CC)CC",
+    "CCCCCCCCN(C)C",
+    "CCN(CC)CCOC(=O)CCCC",
+    "CCCCCCCCCCN(CC)CC",
+    "CC(C)(C)OC(=O)NCCN(C)C",
+    "CCCCCCCCCCCCN(C)C",
+    "CCOC(=O)CN(C)C",
+    "CCCCCCN(CC)CC",
+    "CN(C)CCOC(=O)CCCCCCCC",
+]
+
+
+def load_smiles(source: Path | None) -> list[str]:
+    if source and source.is_file():
+        with source.open(encoding="utf-8-sig", newline="") as handle:
+            rows = csv.DictReader(handle)
+            values = []
+            seen = set()
+            for row in rows:
+                value = (row.get("smiles") or "").strip()
+                if value and value.lower() != "nan" and value not in seen:
+                    seen.add(value)
+                    values.append(value)
+            if values:
+                return values
+    return FALLBACK_SMILES
+
+
+def make_composition(rng: random.Random) -> tuple[float, float, float, float]:
+    # Match the ranges used by the optimizer and guarantee an exact 100% sum.
+    for _ in range(1000):
+        peg = round(rng.uniform(1.0, 5.0), 2)
+        cationic = round(rng.uniform(25.0, 53.0), 2)
+        phospholipid = round(rng.uniform(8.0, 30.0), 2)
+        cholesterol = round(100.0 - peg - cationic - phospholipid, 2)
+        if 15.0 <= cholesterol <= 46.0:
+            return cationic, phospholipid, cholesterol, peg
+    raise RuntimeError("unable to generate a valid composition")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--count", type=int, required=True, choices=range(1, 501))
+    parser.add_argument("--batch-size", type=int, default=32, choices=range(1, 257))
+    parser.add_argument("--use-llm", action=argparse.BooleanOptionalAction, default=True)
+    parser.add_argument("--seed", type=int, default=20260922)
+    parser.add_argument("--source-csv", type=Path)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+
+    rng = random.Random(args.seed)
+    smiles_pool = load_smiles(args.source_csv)
+    rng.shuffle(smiles_pool)
+    items = []
+    for index in range(args.count):
+        cationic, phospholipid, cholesterol, peg = make_composition(rng)
+        items.append(
+            {
+                "smiles": smiles_pool[index % len(smiles_pool)],
+                "cationic_lipid_to_mrna_ratio": round(rng.uniform(7.0, 20.0), 2),
+                "cationic_lipid_mol_ratio": cationic,
+                "phospholipid_mol_ratio": phospholipid,
+                "cholesterol_mol_ratio": cholesterol,
+                "peg_lipid_mol_ratio": peg,
+                "helper_lipid": ("DOPE", "DSPC")[index % 2],
+                "route": ("intravenous", "intramuscular")[index % 2],
+            }
+        )
+
+    payload = {
+        "items": items,
+        "batch_size": args.batch_size,
+        "use_llm": args.use_llm,
+    }
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(
+        json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8"
+    )
+    print(
+        f"generated {args.output}: items={args.count}, unique_smiles={len(set(x['smiles'] for x in items))}, "
+        f"batch_size={args.batch_size}, use_llm={args.use_llm}"
+    )
+
+
+if __name__ == "__main__":
+    main()
--- a/server_test_cases/mixed_valid_invalid.json
+++ b/server_test_cases/mixed_valid_invalid.json
@ -0,0 +1,56 @@
+{
+  "items": [
+    {
+      "smiles": "CC(C)NCCNC(C)C",
+      "cationic_lipid_to_mrna_ratio": 10.0,
+      "cationic_lipid_mol_ratio": 35.0,
+      "phospholipid_mol_ratio": 16.0,
+      "cholesterol_mol_ratio": 46.5,
+      "peg_lipid_mol_ratio": 2.5,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "CCCCCCCCCCCCCCCCN",
+      "cationic_lipid_to_mrna_ratio": 12.0,
+      "cationic_lipid_mol_ratio": 40.0,
+      "phospholipid_mol_ratio": 15.0,
+      "cholesterol_mol_ratio": 42.5,
+      "peg_lipid_mol_ratio": 2.5,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "this-is-not-smiles",
+      "cationic_lipid_to_mrna_ratio": 10.0,
+      "cationic_lipid_mol_ratio": 35.0,
+      "phospholipid_mol_ratio": 16.0,
+      "cholesterol_mol_ratio": 46.5,
+      "peg_lipid_mol_ratio": 2.5,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "CCN(CC)CC",
+      "cationic_lipid_to_mrna_ratio": 10.0,
+      "cationic_lipid_mol_ratio": 35.0,
+      "phospholipid_mol_ratio": 16.0,
+      "cholesterol_mol_ratio": 30.0,
+      "peg_lipid_mol_ratio": 2.5,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "C1CCN(CC1)CCO",
+      "cationic_lipid_to_mrna_ratio": 10.0,
+      "cationic_lipid_mol_ratio": 35.0,
+      "phospholipid_mol_ratio": 16.0,
+      "cholesterol_mol_ratio": 46.5,
+      "peg_lipid_mol_ratio": 2.5,
+      "helper_lipid": "INVALID_HELPER",
+      "route": "intravenous"
+    }
+  ],
+  "batch_size": 5,
+  "use_llm": false
+}
--- a/server_test_cases/mock_batch_128_no_llm.json
+++ b/server_test_cases/mock_batch_128_no_llm.json
--- a/server_test_cases/mock_batch_32_llm.json
+++ b/server_test_cases/mock_batch_32_llm.json
@ -0,0 +1,326 @@
+{
+  "items": [
+    {
+      "smiles": "O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O)CCNC(C[C@@H](CN)CC(C)C)=O",
+      "cationic_lipid_to_mrna_ratio": 7.46,
+      "cationic_lipid_mol_ratio": 37.58,
+      "phospholipid_mol_ratio": 23.5,
+      "cholesterol_mol_ratio": 36.53,
+      "peg_lipid_mol_ratio": 2.39,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCCCCCCCC)CCN(CCC(OCCCCCCCCCCCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCCCCCCCC)=O)CCC(OCCCCCCCCCCCCCC)=O)CCNC(CCC(NCC1=CC=C(B(O)O)C=C1)=O)=O",
+      "cationic_lipid_to_mrna_ratio": 16.82,
+      "cationic_lipid_mol_ratio": 50.09,
+      "phospholipid_mol_ratio": 13.01,
+      "cholesterol_mol_ratio": 33.14,
+      "peg_lipid_mol_ratio": 3.76,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "O=C(OCCCCCCN([C@H](C(NCCCN(CCCCCCOC(CCCCCCC)=O)CCCCCCOC(CCCCCCC)=O)=O)CO)CCCCCCOC(CCCCCCC)=O)CCCCCCC",
+      "cationic_lipid_to_mrna_ratio": 9.92,
+      "cationic_lipid_mol_ratio": 46.82,
+      "phospholipid_mol_ratio": 16.37,
+      "cholesterol_mol_ratio": 33.48,
+      "peg_lipid_mol_ratio": 3.33,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(OCCCCCCN([C@H](C(NCCCCCCN(CCCCCCOC(CCCCC)=O)CCCCCCOC(CCCCC)=O)=O)CO)CCCCCCOC(CCCCC)=O)CCCCC",
+      "cationic_lipid_to_mrna_ratio": 10.73,
+      "cationic_lipid_mol_ratio": 36.31,
+      "phospholipid_mol_ratio": 23.42,
+      "cholesterol_mol_ratio": 35.49,
+      "peg_lipid_mol_ratio": 4.78,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCN(CC1=NC=CS1)CC2=NC=CS2)CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O",
+      "cationic_lipid_to_mrna_ratio": 19.35,
+      "cationic_lipid_mol_ratio": 51.39,
+      "phospholipid_mol_ratio": 10.87,
+      "cholesterol_mol_ratio": 35.21,
+      "peg_lipid_mol_ratio": 2.53,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCCCCCCCC)CCN(CCC(OCCCCCCCCCCCCCC)=O)CCCCCCN(CCCCCCN(CCC(OCCCCCCCCCCCCCC)=O)CCC(OCCCCCCCCCCCCCC)=O)CCCN1CCOCC1",
+      "cationic_lipid_to_mrna_ratio": 17.21,
+      "cationic_lipid_mol_ratio": 46.17,
+      "phospholipid_mol_ratio": 19.64,
+      "cholesterol_mol_ratio": 31.96,
+      "peg_lipid_mol_ratio": 2.23,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "CCCCCCCCCCCC(OCCCCCCN(CCCCCCOC(CCCCCCCCCCC)=O)CCCOC([C@@H](N(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)CO)=O)=O",
+      "cationic_lipid_to_mrna_ratio": 10.34,
+      "cationic_lipid_mol_ratio": 44.28,
+      "phospholipid_mol_ratio": 17.74,
+      "cholesterol_mol_ratio": 34.79,
+      "peg_lipid_mol_ratio": 3.19,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(OCCCCCCN([C@H](C(NCCN(C)CCN(CCCCCCOC(CCCCCCCCC=C)=O)CCCCCCOC(CCCCCCCCC=C)=O)=O)CO)CCCCCCOC(CCCCCCCCC=C)=O)CCCCCCCCC=C",
+      "cationic_lipid_to_mrna_ratio": 9.15,
+      "cationic_lipid_mol_ratio": 42.88,
+      "phospholipid_mol_ratio": 10.3,
+      "cholesterol_mol_ratio": 43.76,
+      "peg_lipid_mol_ratio": 3.06,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "CC(CNC([C@@H](N(CCCCCCOC(CCCCCCCCC=C)=O)CCCCCCOC(CCCCCCCCC=C)=O)CO)=O)(CN(CCCCCCOC(CCCCCCCCC=C)=O)CCCCCCOC(CCCCCCCCC=C)=O)C",
+      "cationic_lipid_to_mrna_ratio": 10.12,
+      "cationic_lipid_mol_ratio": 49.74,
+      "phospholipid_mol_ratio": 12.42,
+      "cholesterol_mol_ratio": 35.63,
+      "peg_lipid_mol_ratio": 2.21,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(NCCN(CCCCN(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)CCCCN(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)CCCN",
+      "cationic_lipid_to_mrna_ratio": 19.72,
+      "cationic_lipid_mol_ratio": 32.8,
+      "phospholipid_mol_ratio": 25.26,
+      "cholesterol_mol_ratio": 38.21,
+      "peg_lipid_mol_ratio": 3.73,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCN1CC(C(OC)=O)CC1=O)CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O",
+      "cationic_lipid_to_mrna_ratio": 13.7,
+      "cationic_lipid_mol_ratio": 43.91,
+      "phospholipid_mol_ratio": 21.35,
+      "cholesterol_mol_ratio": 31.92,
+      "peg_lipid_mol_ratio": 2.82,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "CCCCCCCCCCCCOC(CCN([C@H](C(NCCCCCCN(CCC(OCCCCCCCCCCCC)=O)CCC(OCCCCCCCCCCCC)=O)=O)CO)CCC(OCCCCCCCCCCCC)=O)=O",
+      "cationic_lipid_to_mrna_ratio": 10.47,
+      "cationic_lipid_mol_ratio": 32.49,
+      "phospholipid_mol_ratio": 24.48,
+      "cholesterol_mol_ratio": 40.96,
+      "peg_lipid_mol_ratio": 2.07,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "O=C([C@H](CO)N(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)N(CC1)CCN1C2=NC=CC=N2",
+      "cationic_lipid_to_mrna_ratio": 12.86,
+      "cationic_lipid_mol_ratio": 51.8,
+      "phospholipid_mol_ratio": 19.97,
+      "cholesterol_mol_ratio": 25.07,
+      "peg_lipid_mol_ratio": 3.16,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(CN(CC(O)CCCCCCCC)CC(O)CCCCCCCC)NCCCN(CC(O)CCCCCCCC)CC(O)CCCCCCCC",
+      "cationic_lipid_to_mrna_ratio": 18.73,
+      "cationic_lipid_mol_ratio": 45.45,
+      "phospholipid_mol_ratio": 13.35,
+      "cholesterol_mol_ratio": 38.13,
+      "peg_lipid_mol_ratio": 3.07,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "CCCCCCCCCCCC(OCCCCCCN(CCN(C)CCNC([C@@H](N(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)CO)=O)CCCCCCOC(CCCCCCCCCCC)=O)=O",
+      "cationic_lipid_to_mrna_ratio": 10.09,
+      "cationic_lipid_mol_ratio": 37.43,
+      "phospholipid_mol_ratio": 11.94,
+      "cholesterol_mol_ratio": 45.78,
+      "peg_lipid_mol_ratio": 4.85,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "OC[C@@H](C(NCCCCCCN(CC(O)CCCCCCCCCC)CC(O)CCCCCCCCCC)=O)N(CC(CCCCCCCCCC)O)CC(CCCCCCCCCC)O",
+      "cationic_lipid_to_mrna_ratio": 13.05,
+      "cationic_lipid_mol_ratio": 52.84,
+      "phospholipid_mol_ratio": 8.6,
+      "cholesterol_mol_ratio": 37.06,
+      "peg_lipid_mol_ratio": 1.5,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCCCCCCCC)CCN(CCC(OCCCCCCCCCCCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCCCCCCCC)=O)CCC(OCCCCCCCCCCCCCC)=O)CCNC(C[C@@H](CN)CC(C)C)=O",
+      "cationic_lipid_to_mrna_ratio": 11.09,
+      "cationic_lipid_mol_ratio": 48.25,
+      "phospholipid_mol_ratio": 14.52,
+      "cholesterol_mol_ratio": 32.66,
+      "peg_lipid_mol_ratio": 4.57,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCC/C=C\\C/C=C\\CCCCC)CCN(CCC(OCCCCCCCC/C=C\\C/C=C\\CCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCC/C=C\\C/C=C\\CCCCC)=O)CCC(OCCCCCCCC/C=C\\C/C=C\\CCCCC)=O)CCNC(CCC(NCCC1=CC=C(O)C(O)=C1)=O)=O",
+      "cationic_lipid_to_mrna_ratio": 19.63,
+      "cationic_lipid_mol_ratio": 31.98,
+      "phospholipid_mol_ratio": 25.5,
+      "cholesterol_mol_ratio": 40.57,
+      "peg_lipid_mol_ratio": 1.95,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "O=C(NCCCCCCN(CCCCCCCCCCCC)CCCCCCCCCCCC)[C@H](C)N(CCCCCCCCCCCC)CCCCCCCCCCCC",
+      "cationic_lipid_to_mrna_ratio": 16.96,
+      "cationic_lipid_mol_ratio": 28.21,
+      "phospholipid_mol_ratio": 28.57,
+      "cholesterol_mol_ratio": 39.51,
+      "peg_lipid_mol_ratio": 3.71,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCN(CCC(F)(F)F)CCC(F)(F)F)CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O",
+      "cationic_lipid_to_mrna_ratio": 19.73,
+      "cationic_lipid_mol_ratio": 37.62,
+      "phospholipid_mol_ratio": 22.08,
+      "cholesterol_mol_ratio": 36.42,
+      "peg_lipid_mol_ratio": 3.88,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "CCCCCCCCC(CN(CCCNC([C@@H](N(CC(CCCCCCCC)O)CC(CCCCCCCC)O)CO)=O)CC(CCCCCCCC)O)O",
+      "cationic_lipid_to_mrna_ratio": 15.62,
+      "cationic_lipid_mol_ratio": 51.29,
+      "phospholipid_mol_ratio": 27.02,
+      "cholesterol_mol_ratio": 18.83,
+      "peg_lipid_mol_ratio": 2.86,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C([C@H](CO)N(CCCCCCOC(CCCCCCCCC=C)=O)CCCCCCOC(CCCCCCCCC=C)=O)N(CC1)CCN1C2=NC=CC=N2",
+      "cationic_lipid_to_mrna_ratio": 14.57,
+      "cationic_lipid_mol_ratio": 38.85,
+      "phospholipid_mol_ratio": 18.16,
+      "cholesterol_mol_ratio": 40.17,
+      "peg_lipid_mol_ratio": 2.82,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "CCCCCCCCCCCC(OCCCCCCN([C@H](C(NCCCCCCN(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)=O)CO)CCCCCCOC(CCCCCCCCCCC)=O)=O",
+      "cationic_lipid_to_mrna_ratio": 17.0,
+      "cationic_lipid_mol_ratio": 41.61,
+      "phospholipid_mol_ratio": 21.12,
+      "cholesterol_mol_ratio": 35.9,
+      "peg_lipid_mol_ratio": 1.37,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCCCCCC)CCN(CCC(OCCCCCCCCCCCC)=O)CCCCCCN(CCCCCCN(CCC(OCCCCCCCCCCCC)=O)CCC(OCCCCCCCCCCCC)=O)CCN1CCOCC1",
+      "cationic_lipid_to_mrna_ratio": 12.75,
+      "cationic_lipid_mol_ratio": 31.73,
+      "phospholipid_mol_ratio": 29.19,
+      "cholesterol_mol_ratio": 37.55,
+      "peg_lipid_mol_ratio": 1.53,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "O=C(CN(CCCCCCCCCCCCCC)CCCCCCCCCCCCCC)NCCCN(CCCCCCCCCCCCCC)CCCCCCCCCCCCCC",
+      "cationic_lipid_to_mrna_ratio": 19.09,
+      "cationic_lipid_mol_ratio": 41.63,
+      "phospholipid_mol_ratio": 16.23,
+      "cholesterol_mol_ratio": 39.43,
+      "peg_lipid_mol_ratio": 2.71,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCCCN(CCCCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O)CCNC(CCC(NCCC1=CC=C(O)C(O)=C1)=O)=O",
+      "cationic_lipid_to_mrna_ratio": 16.96,
+      "cationic_lipid_mol_ratio": 34.46,
+      "phospholipid_mol_ratio": 28.65,
+      "cholesterol_mol_ratio": 32.06,
+      "peg_lipid_mol_ratio": 4.83,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O)CCNC([C@@H](N)CCCCN)=O",
+      "cationic_lipid_to_mrna_ratio": 14.76,
+      "cationic_lipid_mol_ratio": 34.58,
+      "phospholipid_mol_ratio": 25.73,
+      "cholesterol_mol_ratio": 35.03,
+      "peg_lipid_mol_ratio": 4.66,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(CN(CCCCCCCCCCCC)CCCCCCCCCCCC)NCCCN(CCCCCCCCCCCC)CCCCCCCCCCCC",
+      "cationic_lipid_to_mrna_ratio": 7.01,
+      "cationic_lipid_mol_ratio": 42.86,
+      "phospholipid_mol_ratio": 23.71,
+      "cholesterol_mol_ratio": 30.73,
+      "peg_lipid_mol_ratio": 2.7,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCN(CC1=COC=N1)CC2=COC=N2)CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O",
+      "cationic_lipid_to_mrna_ratio": 7.11,
+      "cationic_lipid_mol_ratio": 43.3,
+      "phospholipid_mol_ratio": 22.32,
+      "cholesterol_mol_ratio": 31.55,
+      "peg_lipid_mol_ratio": 2.83,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCNC(CCCN)=O)CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O",
+      "cationic_lipid_to_mrna_ratio": 18.85,
+      "cationic_lipid_mol_ratio": 45.22,
+      "phospholipid_mol_ratio": 17.28,
+      "cholesterol_mol_ratio": 33.01,
+      "peg_lipid_mol_ratio": 4.49,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    },
+    {
+      "smiles": "O=C(CN(CCCCCCCCCC)CCCCCCCCCC)NCCCN(CCCCCCCCCC)CCCCCCCCCC",
+      "cationic_lipid_to_mrna_ratio": 10.74,
+      "cationic_lipid_mol_ratio": 38.47,
+      "phospholipid_mol_ratio": 25.72,
+      "cholesterol_mol_ratio": 32.81,
+      "peg_lipid_mol_ratio": 3.0,
+      "helper_lipid": "DOPE",
+      "route": "intravenous"
+    },
+    {
+      "smiles": "O=C(OCCCCCCCC/C=C\\C/C=C\\CCCCC)CCN(CCC(OCCCCCCCC/C=C\\C/C=C\\CCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCC/C=C\\C/C=C\\CCCCC)=O)CCC(OCCCCCCCC/C=C\\C/C=C\\CCCCC)=O)CCNC(CCCN)=O",
+      "cationic_lipid_to_mrna_ratio": 17.44,
+      "cationic_lipid_mol_ratio": 42.07,
+      "phospholipid_mol_ratio": 17.36,
+      "cholesterol_mol_ratio": 36.89,
+      "peg_lipid_mol_ratio": 3.68,
+      "helper_lipid": "DSPC",
+      "route": "intramuscular"
+    }
+  ],
+  "batch_size": 32,
+  "use_llm": true
+}
--- a/server_test_cases/mock_batch_500_llm.json
+++ b/server_test_cases/mock_batch_500_llm.json
--- a/server_test_cases/optimize_smoke.json
+++ b/server_test_cases/optimize_smoke.json
@ -0,0 +1,31 @@
+{
+  "smiles": "CC(C)NCCNC(C)C",
+  "organ": "liver",
+  "top_k": 3,
+  "num_seeds": 5,
+  "top_per_seed": 1,
+  "rerank_top_n": 8,
+  "step_sizes": [5],
+  "wr_step_sizes": [5],
+  "routes": ["intravenous"],
+  "comp_ranges": {
+    "weight_ratio_min": 10,
+    "weight_ratio_max": 15,
+    "cationic_mol_min": 40,
+    "cationic_mol_max": 50,
+    "phospholipid_mol_min": 10,
+    "phospholipid_mol_max": 20,
+    "cholesterol_mol_min": 28,
+    "cholesterol_mol_max": 45,
+    "peg_mol_min": 1,
+    "peg_mol_max": 5
+  },
+  "scoring_weights": {
+    "biodist_weight": 1,
+    "delivery_weight": 0,
+    "size_weight": 0,
+    "ee_class_weights": [0, 0, 0],
+    "pdi_class_weights": [0, 0],
+    "toxic_class_weights": [0, 0]
+  }
+}
--- a/server_test_cases/run_cases.sh
+++ b/server_test_cases/run_cases.sh
@ -0,0 +1,187 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+CASE_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+API_PORT=${API_PORT:-18000}
+API_URL=${API_URL:-http://127.0.0.1:${API_PORT}}
+GPU_INDEX=${GPU_INDEX:-0}
+CURL_TIMEOUT=${CURL_TIMEOUT:-1800}
+
+PASS=0
+FAIL=0
+
+run_case() {
+  local name=$1
+  shift
+  echo
+  echo "===== ${name} ====="
+  if "$@"; then
+    PASS=$((PASS + 1))
+    echo "PASS: ${name}"
+  else
+    FAIL=$((FAIL + 1))
+    echo "FAIL: ${name}" >&2
+  fi
+}
+
+validate_health() {
+  local response=/tmp/lnp-test-health.json
+  curl -fsS --max-time 30 "${API_URL}/" >"${response}"
+  python3 - "${response}" <<'PY'
+import json, sys
+data = json.load(open(sys.argv[1], encoding="utf-8"))
+assert data["status"] == "healthy", data
+assert data["model_loaded"] is True, data
+assert data["use_llm"] is True, data
+print(json.dumps(data, ensure_ascii=False, indent=2))
+PY
+}
+
+validate_organs() {
+  local response=/tmp/lnp-test-organs.json
+  curl -fsS --max-time 30 "${API_URL}/organs" >"${response}"
+  python3 - "${response}" <<'PY'
+import json, sys
+data = json.load(open(sys.argv[1], encoding="utf-8"))
+expected = {"lymph_nodes", "heart", "liver", "spleen", "lung", "kidney", "muscle"}
+assert set(data) == expected, data
+print(json.dumps(data, ensure_ascii=False))
+PY
+}
+
+validate_single() {
+  local response=/tmp/lnp-test-single.json
+  curl -fsS --max-time "${CURL_TIMEOUT}" \
+    -H 'Content-Type: application/json' \
+    --data-binary "@${CASE_DIR}/single_valid.json" \
+    "${API_URL}/predict" >"${response}"
+  python3 - "${response}" <<'PY'
+import json, math, sys
+data = json.load(open(sys.argv[1], encoding="utf-8"))
+required = ["biodist", "quantified_delivery", "pdi_class", "ee_class", "toxic_class"]
+assert all(k in data for k in required), data
+values = list(data["biodist"].values())
+assert len(values) == 7, values
+assert all(math.isfinite(float(x)) for x in values), values
+assert abs(sum(values) - 1.0) < 1e-4, sum(values)
+assert math.isfinite(float(data["quantified_delivery"]))
+print(json.dumps(data, ensure_ascii=False, indent=2))
+PY
+}
+
+make_batch_payload() {
+  local count=$1
+  local use_llm=$2
+  local batch_size=$3
+  python3 - "${CASE_DIR}/single_valid.json" "${count}" "${use_llm}" "${batch_size}" <<'PY'
+import json, sys
+item = json.load(open(sys.argv[1], encoding="utf-8"))
+print(json.dumps({
+    "items": [item for _ in range(int(sys.argv[2]))],
+    "batch_size": int(sys.argv[4]),
+    "use_llm": sys.argv[3].lower() == "true",
+}))
+PY
+}
+
+validate_batch_response() {
+  local response=$1
+  local expected=$2
+  python3 - "${response}" "${expected}" <<'PY'
+import json, math, sys
+data = json.load(open(sys.argv[1], encoding="utf-8"))
+expected = int(sys.argv[2])
+assert data["n_requested"] == expected, data
+assert data["n_succeeded"] == expected, data
+assert data["errors"] == [], data["errors"]
+assert len(data["predictions"]) == expected
+for pred in data["predictions"]:
+    values = list(pred["biodist"].values())
+    assert len(values) == 7
+    assert all(math.isfinite(float(x)) for x in values)
+    assert abs(sum(values) - 1.0) < 1e-4
+print(f"n_succeeded={data['n_succeeded']}, errors={len(data['errors'])}")
+PY
+}
+
+run_batch() {
+  local count=$1
+  local use_llm=$2
+  local batch_size=$3
+  local response="/tmp/lnp-test-batch-${count}-${use_llm}.json"
+  local payload before peak current curl_pid start_ms end_ms
+
+  payload=$(make_batch_payload "${count}" "${use_llm}" "${batch_size}")
+  before=$(nvidia-smi --id="${GPU_INDEX}" --query-gpu=memory.used \
+    --format=csv,noheader,nounits | head -n 1 | tr -d ' ')
+  peak=${before}
+  start_ms=$(date +%s%3N)
+
+  curl -fsS --max-time "${CURL_TIMEOUT}" \
+    -H 'Content-Type: application/json' \
+    -d "${payload}" "${API_URL}/predict/batch" >"${response}" &
+  curl_pid=$!
+  while kill -0 "${curl_pid}" 2>/dev/null; do
+    current=$(nvidia-smi --id="${GPU_INDEX}" --query-gpu=memory.used \
+      --format=csv,noheader,nounits | head -n 1 | tr -d ' ')
+    if [ "${current}" -gt "${peak}" ]; then peak=${current}; fi
+    sleep 0.2
+  done
+  wait "${curl_pid}"
+  end_ms=$(date +%s%3N)
+
+  validate_batch_response "${response}" "${count}"
+  echo "use_llm=${use_llm}, count=${count}, outer_batch=${batch_size}"
+  echo "GPU used before=${before} MiB, observed peak=${peak} MiB, delta=$((peak - before)) MiB"
+  echo "elapsed=$((end_ms - start_ms)) ms"
+}
+
+validate_mixed_errors() {
+  local response=/tmp/lnp-test-mixed.json
+  curl -fsS --max-time "${CURL_TIMEOUT}" \
+    -H 'Content-Type: application/json' \
+    --data-binary "@${CASE_DIR}/mixed_valid_invalid.json" \
+    "${API_URL}/predict/batch" >"${response}"
+  python3 - "${response}" <<'PY'
+import json, sys
+data = json.load(open(sys.argv[1], encoding="utf-8"))
+assert data["n_requested"] == 5, data
+assert data["n_succeeded"] == 2, data
+assert len(data["errors"]) == 3, data
+assert {x["index"] for x in data["errors"]} == {2, 3, 4}, data["errors"]
+print(json.dumps(data["errors"], ensure_ascii=False, indent=2))
+PY
+}
+
+validate_optimize() {
+  local response=/tmp/lnp-test-optimize.json
+  curl -fsS --max-time "${CURL_TIMEOUT}" \
+    -H 'Content-Type: application/json' \
+    --data-binary "@${CASE_DIR}/optimize_smoke.json" \
+    "${API_URL}/optimize" >"${response}"
+  python3 - "${response}" <<'PY'
+import json, math, sys
+data = json.load(open(sys.argv[1], encoding="utf-8"))
+assert data["target_organ"] == "liver", data
+assert 1 <= len(data["formulations"]) <= 3, data
+for row in data["formulations"]:
+    assert math.isfinite(float(row["target_biodist"])), row
+    assert abs(sum(float(x) for x in row["all_biodist"].values()) - 1.0) < 1e-4, row
+print(f"optimize returned {len(data['formulations'])} formulations")
+PY
+}
+
+run_case "health" validate_health
+run_case "available organs" validate_organs
+run_case "single prediction with LLM" validate_single
+run_case "64 items without LLM (throughput baseline)" run_batch 64 false 64
+run_case "32 items with LLM (VRAM regression)" run_batch 32 true 32
+run_case "mixed valid/invalid items" validate_mixed_errors
+run_case "small optimize search" validate_optimize
+
+echo
+echo "===== SUMMARY ====="
+echo "passed=${PASS}, failed=${FAIL}"
+if [ "${FAIL}" -ne 0 ]; then
+  exit 1
+fi
--- a/server_test_cases/run_lnp_batch_stress.sh
+++ b/server_test_cases/run_lnp_batch_stress.sh
@ -0,0 +1,146 @@
+#!/usr/bin/env bash
+set -uo pipefail
+
+CASE_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+API_PORT=${API_PORT:-18000}
+API_URL=${API_URL:-http://127.0.0.1:${API_PORT}}
+GPU_INDEX=${GPU_INDEX:-0}
+CURL_TIMEOUT=${CURL_TIMEOUT:-3600}
+MAX_GPU_MIB=${MAX_GPU_MIB:-0}
+PROFILE=${1:-repro}
+RESULT_CSV=${RESULT_CSV:-"${CASE_DIR}/vram-results-$(date +%Y%m%d-%H%M%S).csv"}
+
+if ! command -v curl >/dev/null || ! command -v python3 >/dev/null; then
+  echo "curl and python3 are required" >&2
+  exit 2
+fi
+if ! command -v nvidia-smi >/dev/null; then
+  echo "nvidia-smi is required for the GPU memory test" >&2
+  exit 2
+fi
+if ! curl -fsS --max-time 30 "${API_URL}/" >/tmp/lnp-stress-health.json; then
+  echo "API is not reachable: ${API_URL}" >&2
+  exit 2
+fi
+
+gpu_used_mib() {
+  nvidia-smi --id="${GPU_INDEX}" --query-gpu=memory.used --format=csv,noheader,nounits \
+    | head -n 1 | tr -d ' '
+}
+
+now_ms() {
+  python3 -c 'import time; print(int(time.time() * 1000))'
+}
+
+validate_response() {
+  local response=$1 expected=$2
+  python3 - "${response}" "${expected}" <<'PY'
+import json, math, sys
+data = json.load(open(sys.argv[1], encoding="utf-8"))
+expected = int(sys.argv[2])
+assert data["n_requested"] == expected, data
+assert data["n_succeeded"] == expected, data
+assert data["errors"] == [], data["errors"]
+assert len(data["predictions"]) == expected, len(data["predictions"])
+for pred in data["predictions"]:
+    values = list(pred["biodist"].values())
+    assert len(values) == 7
+    assert all(math.isfinite(float(value)) for value in values)
+    assert abs(sum(values) - 1.0) < 1e-4, sum(values)
+print(f"validated {expected} predictions")
+PY
+}
+
+run_payload() {
+  local label=$1 payload=$2 expected=$3
+  local response="/tmp/lnp-${label}-response.json"
+  local http_file="/tmp/lnp-${label}-http.txt"
+  local before peak current curl_pid curl_status http_code start_ms end_ms elapsed delta status
+
+  echo
+  echo "===== ${label} ====="
+  before=$(gpu_used_mib)
+  peak=${before}
+  start_ms=$(now_ms)
+
+  curl -sS --max-time "${CURL_TIMEOUT}" -o "${response}" -w '%{http_code}' \
+    -H 'Content-Type: application/json' --data-binary "@${payload}" \
+    "${API_URL}/predict/batch" >"${http_file}" &
+  curl_pid=$!
+  while kill -0 "${curl_pid}" 2>/dev/null; do
+    current=$(gpu_used_mib 2>/dev/null || echo 0)
+    if [[ "${current}" =~ ^[0-9]+$ ]] && [ "${current}" -gt "${peak}" ]; then
+      peak=${current}
+    fi
+    sleep 0.1
+  done
+  wait "${curl_pid}"
+  curl_status=$?
+  end_ms=$(now_ms)
+  http_code=$(cat "${http_file}" 2>/dev/null || echo 000)
+  elapsed=$((end_ms - start_ms))
+  delta=$((peak - before))
+  status=PASS
+
+  if [ "${curl_status}" -ne 0 ] || [ "${http_code}" != "200" ]; then
+    status=FAIL
+    echo "request failed: curl_status=${curl_status}, HTTP=${http_code}" >&2
+    sed -n '1,30p' "${response}" 2>/dev/null || true
+  elif ! validate_response "${response}" "${expected}"; then
+    status=FAIL
+  elif [ "${MAX_GPU_MIB}" -gt 0 ] && [ "${peak}" -gt "${MAX_GPU_MIB}" ]; then
+    status=FAIL
+    echo "peak ${peak} MiB exceeds MAX_GPU_MIB=${MAX_GPU_MIB}" >&2
+  fi
+
+  echo "${label}: status=${status}, before=${before} MiB, peak=${peak} MiB, delta=${delta} MiB, elapsed=${elapsed} ms"
+  printf '%s,%s,%s,%s,%s,%s,%s\n' \
+    "${label}" "${status}" "${before}" "${peak}" "${delta}" "${elapsed}" "${http_code}" >>"${RESULT_CSV}"
+  [ "${status}" = PASS ]
+}
+
+make_payload() {
+  local count=$1 outer_batch=$2 use_llm=$3 output=$4
+  local llm_flag=--use-llm
+  [ "${use_llm}" = false ] && llm_flag=--no-use-llm
+  python3 "${CASE_DIR}/generate_mock_batch.py" \
+    --count "${count}" --batch-size "${outer_batch}" "${llm_flag}" \
+    --source-csv "${CASE_DIR}/../data/interim/internal.csv" --output "${output}"
+}
+
+printf 'case,status,before_mib,peak_mib,delta_mib,elapsed_ms,http_code\n' >"${RESULT_CSV}"
+FAILED=0
+
+case "${PROFILE}" in
+  repro)
+    run_payload "llm-32-outer32" "${CASE_DIR}/mock_batch_32_llm.json" 32 || FAILED=1
+    run_payload "llm-500-outer32" "${CASE_DIR}/mock_batch_500_llm.json" 500 || FAILED=1
+    ;;
+  ladder)
+    for count in 1 8 16 32 64 128; do
+      payload="/tmp/lnp-ladder-${count}.json"
+      make_payload "${count}" "${count}" true "${payload}"
+      run_payload "llm-${count}-outer${count}" "${payload}" "${count}" || FAILED=1
+      curl -fsS --max-time 30 "${API_URL}/" >/dev/null || break
+    done
+    ;;
+  full)
+    run_payload "baseline-no-llm-128" "${CASE_DIR}/mock_batch_128_no_llm.json" 128 || FAILED=1
+    for count in 1 8 16 32 64 128; do
+      payload="/tmp/lnp-ladder-${count}.json"
+      make_payload "${count}" "${count}" true "${payload}"
+      run_payload "llm-${count}-outer${count}" "${payload}" "${count}" || FAILED=1
+      curl -fsS --max-time 30 "${API_URL}/" >/dev/null || break
+    done
+    curl -fsS --max-time 30 "${API_URL}/" >/dev/null \
+      && run_payload "llm-500-outer32" "${CASE_DIR}/mock_batch_500_llm.json" 500 || FAILED=1
+    ;;
+  *)
+    echo "Usage: $0 [repro|ladder|full]" >&2
+    exit 2
+    ;;
+esac
+
+echo
+echo "Results: ${RESULT_CSV}"
+exit "${FAILED}"
--- a/server_test_cases/single_valid.json
+++ b/server_test_cases/single_valid.json
@ -0,0 +1,10 @@
+{
+  "smiles": "CC(C)NCCNC(C)C",
+  "cationic_lipid_to_mrna_ratio": 10.0,
+  "cationic_lipid_mol_ratio": 35.0,
+  "phospholipid_mol_ratio": 16.0,
+  "cholesterol_mol_ratio": 46.5,
+  "peg_lipid_mol_ratio": 2.5,
+  "helper_lipid": "DOPE",
+  "route": "intravenous"
+}
--- a/server_test_cases/ui_batch_screening_32.csv
+++ b/server_test_cases/ui_batch_screening_32.csv
@ -0,0 +1,33 @@
+name,SMILES
+LNP-001,O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O)CCNC(C[C@@H](CN)CC(C)C)=O
+LNP-002,O=C(OCCCCCCCCCCCCCC)CCN(CCC(OCCCCCCCCCCCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCCCCCCCC)=O)CCC(OCCCCCCCCCCCCCC)=O)CCNC(CCC(NCC1=CC=C(B(O)O)C=C1)=O)=O
+LNP-003,O=C(OCCCCCCN([C@H](C(NCCCN(CCCCCCOC(CCCCCCC)=O)CCCCCCOC(CCCCCCC)=O)=O)CO)CCCCCCOC(CCCCCCC)=O)CCCCCCC
+LNP-004,O=C(OCCCCCCN([C@H](C(NCCCCCCN(CCCCCCOC(CCCCC)=O)CCCCCCOC(CCCCC)=O)=O)CO)CCCCCCOC(CCCCC)=O)CCCCC
+LNP-005,O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCN(CC1=NC=CS1)CC2=NC=CS2)CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O
+LNP-006,O=C(OCCCCCCCCCCCCCC)CCN(CCC(OCCCCCCCCCCCCCC)=O)CCCCCCN(CCCCCCN(CCC(OCCCCCCCCCCCCCC)=O)CCC(OCCCCCCCCCCCCCC)=O)CCCN1CCOCC1
+LNP-007,CCCCCCCCCCCC(OCCCCCCN(CCCCCCOC(CCCCCCCCCCC)=O)CCCOC([C@@H](N(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)CO)=O)=O
+LNP-008,O=C(OCCCCCCN([C@H](C(NCCN(C)CCN(CCCCCCOC(CCCCCCCCC=C)=O)CCCCCCOC(CCCCCCCCC=C)=O)=O)CO)CCCCCCOC(CCCCCCCCC=C)=O)CCCCCCCCC=C
+LNP-009,CC(CNC([C@@H](N(CCCCCCOC(CCCCCCCCC=C)=O)CCCCCCOC(CCCCCCCCC=C)=O)CO)=O)(CN(CCCCCCOC(CCCCCCCCC=C)=O)CCCCCCOC(CCCCCCCCC=C)=O)C
+LNP-010,O=C(NCCN(CCCCN(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)CCCCN(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)CCCN
+LNP-011,O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCN1CC(C(OC)=O)CC1=O)CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O
+LNP-012,CCCCCCCCCCCCOC(CCN([C@H](C(NCCCCCCN(CCC(OCCCCCCCCCCCC)=O)CCC(OCCCCCCCCCCCC)=O)=O)CO)CCC(OCCCCCCCCCCCC)=O)=O
+LNP-013,O=C([C@H](CO)N(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)N(CC1)CCN1C2=NC=CC=N2
+LNP-014,O=C(CN(CC(O)CCCCCCCC)CC(O)CCCCCCCC)NCCCN(CC(O)CCCCCCCC)CC(O)CCCCCCCC
+LNP-015,CCCCCCCCCCCC(OCCCCCCN(CCN(C)CCNC([C@@H](N(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)CO)=O)CCCCCCOC(CCCCCCCCCCC)=O)=O
+LNP-016,OC[C@@H](C(NCCCCCCN(CC(O)CCCCCCCCCC)CC(O)CCCCCCCCCC)=O)N(CC(CCCCCCCCCC)O)CC(CCCCCCCCCC)O
+LNP-017,O=C(OCCCCCCCCCCCCCC)CCN(CCC(OCCCCCCCCCCCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCCCCCCCC)=O)CCC(OCCCCCCCCCCCCCC)=O)CCNC(C[C@@H](CN)CC(C)C)=O
+LNP-018,O=C(OCCCCCCCC/C=C\C/C=C\CCCCC)CCN(CCC(OCCCCCCCC/C=C\C/C=C\CCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCC/C=C\C/C=C\CCCCC)=O)CCC(OCCCCCCCC/C=C\C/C=C\CCCCC)=O)CCNC(CCC(NCCC1=CC=C(O)C(O)=C1)=O)=O
+LNP-019,O=C(NCCCCCCN(CCCCCCCCCCCC)CCCCCCCCCCCC)[C@H](C)N(CCCCCCCCCCCC)CCCCCCCCCCCC
+LNP-020,O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCN(CCC(F)(F)F)CCC(F)(F)F)CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O
+LNP-021,CCCCCCCCC(CN(CCCNC([C@@H](N(CC(CCCCCCCC)O)CC(CCCCCCCC)O)CO)=O)CC(CCCCCCCC)O)O
+LNP-022,O=C([C@H](CO)N(CCCCCCOC(CCCCCCCCC=C)=O)CCCCCCOC(CCCCCCCCC=C)=O)N(CC1)CCN1C2=NC=CC=N2
+LNP-023,CCCCCCCCCCCC(OCCCCCCN([C@H](C(NCCCCCCN(CCCCCCOC(CCCCCCCCCCC)=O)CCCCCCOC(CCCCCCCCCCC)=O)=O)CO)CCCCCCOC(CCCCCCCCCCC)=O)=O
+LNP-024,O=C(OCCCCCCCCCCCC)CCN(CCC(OCCCCCCCCCCCC)=O)CCCCCCN(CCCCCCN(CCC(OCCCCCCCCCCCC)=O)CCC(OCCCCCCCCCCCC)=O)CCN1CCOCC1
+LNP-025,O=C(CN(CCCCCCCCCCCCCC)CCCCCCCCCCCCCC)NCCCN(CCCCCCCCCCCCCC)CCCCCCCCCCCCCC
+LNP-026,O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCCCN(CCCCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O)CCNC(CCC(NCCC1=CC=C(O)C(O)=C1)=O)=O
+LNP-027,O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O)CCNC([C@@H](N)CCCCN)=O
+LNP-028,O=C(CN(CCCCCCCCCCCC)CCCCCCCCCCCC)NCCCN(CCCCCCCCCCCC)CCCCCCCCCCCC
+LNP-029,O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCN(CC1=COC=N1)CC2=COC=N2)CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O
+LNP-030,O=C(OCCCCCCCC)CCN(CCC(OCCCCCCCC)=O)CCCCN(CCNC(CCCN)=O)CCCCN(CCC(OCCCCCCCC)=O)CCC(OCCCCCCCC)=O
+LNP-031,O=C(CN(CCCCCCCCCC)CCCCCCCCCC)NCCCN(CCCCCCCCCC)CCCCCCCCCC
+LNP-032,O=C(OCCCCCCCC/C=C\C/C=C\CCCCC)CCN(CCC(OCCCCCCCC/C=C\C/C=C\CCCCC)=O)CCCCN(CCCCN(CCC(OCCCCCCCC/C=C\C/C=C\CCCCC)=O)CCC(OCCCCCCCC/C=C\C/C=C\CCCCC)=O)CCNC(CCCN)=O
--- a/server_test_cases/ui_batch_screening_matrix.csv
+++ b/server_test_cases/ui_batch_screening_matrix.csv
@ -0,0 +1,3 @@
+head,Tail-A,Tail-B
+H-01,CC(C)NCCNC(C)C,CCCCCCCCCCN(C)C
+H-02,CCN(CC)CCOC(=O)CCCC,CCCCCCCCCCCCN(C)C
--- a/server_test_cases/ui_batch_screening_mixed.csv
+++ b/server_test_cases/ui_batch_screening_mixed.csv
@ -0,0 +1,7 @@
+name,SMILES
+valid-1,CC(C)NCCNC(C)C
+valid-2,CCCCCCCCCCN(C)C
+invalid-backend-1,C1INVALIDCCCC
+filtered-short,abc
+duplicate-valid-1,CC(C)NCCNC(C)C
+invalid-backend-2,NOT_A_VALID_CCCC
--- a/tests/test_llm_memory.py
+++ b/tests/test_llm_memory.py
@ -0,0 +1,58 @@
+"""LLM 显存保护逻辑的轻量测试（不需要下载模型）。"""
+
+from types import MethodType, SimpleNamespace
+from unittest.mock import patch
+
+import torch
+
+from lnp_ml.modeling.layers.llm_prompt import LLMPromptEncoder, _positive_int_env
+
+
+def _empty_encoder(micro_batch: int = 2) -> LLMPromptEncoder:
+    encoder = LLMPromptEncoder.__new__(LLMPromptEncoder)
+    torch.nn.Module.__init__(encoder)
+    encoder.inference_batch_size = micro_batch
+    return encoder
+
+
+def test_invalid_inference_batch_env_falls_back() -> None:
+    with patch.dict("os.environ", {"LLM_INFERENCE_BATCH_SIZE": "invalid"}):
+        assert _positive_int_env("LLM_INFERENCE_BATCH_SIZE", 4) == 4
+    with patch.dict("os.environ", {"LLM_INFERENCE_BATCH_SIZE": "0"}):
+        assert _positive_int_env("LLM_INFERENCE_BATCH_SIZE", 4) == 4
+
+
+def test_softrag_inference_is_microbatched() -> None:
+    encoder = _empty_encoder(micro_batch=2).eval()
+    calls = []
+
+    def fake_chunk(self, smiles, chem, tab, device):
+        calls.append((list(smiles), chem.clone(), tab.clone()))
+        return chem[:, 0, :] + tab[:, 0, :]
+
+    encoder._encode_softrag_chunk = MethodType(fake_chunk, encoder)
+    chem = torch.arange(5 * 2 * 3, dtype=torch.float32).reshape(5, 2, 3)
+    tab = torch.ones(5, 1, 3)
+
+    with torch.no_grad():
+        output = encoder._encode_softrag(
+            [f"SMILES-{i}" for i in range(5)], chem, tab, torch.device("cpu")
+        )
+
+    assert [len(smiles) for smiles, _, _ in calls] == [2, 2, 1]
+    assert torch.equal(output, chem[:, 0, :] + tab[:, 0, :])
+
+
+def test_qwen_forward_disables_kv_cache() -> None:
+    encoder = _empty_encoder()
+    encoder._is_qwen = True
+    received = {}
+
+    class FakeBackbone(torch.nn.Module):
+        def forward(self, **kwargs):
+            received.update(kwargs)
+            return SimpleNamespace(last_hidden_state=torch.zeros(1, 1, 1))
+
+    encoder.encoder = FakeBackbone()
+    encoder._encoder_forward(input_ids=torch.ones(1, 1, dtype=torch.long))
+    assert received["use_cache"] is False
Author	SHA1	Message	Date
楚毓	c0d2c008b0	fix: bound LLM inference memory for batch screening	2026-09-23 10:11:41 +08:00
Michelle0574	a7c7d54c61	chore(docker): 补齐容器构建的忽略规则与运行依赖	2026-09-16 06:11:46 +00:00
Michelle0574	a812694500	feat(app): 新增批量筛选，结果与模式可跨刷新恢复	2026-09-05 17:15:37 +00:00