diff --git a/load_test.py b/load_test.py
new file mode 100644
index 0000000..9f1cfdf
--- /dev/null
+++ b/load_test.py
@@ -0,0 +1,163 @@
+#!/usr/bin/env python3
+"""并发压力测试 llmage /v1/chat/completions，统计 TTFB / 完成时间 / QPM"""
+
+import asyncio
+import aiohttp
+import time
+import json
+import sys
+import statistics
+from dataclasses import dataclass, field
+from typing import List
+
+URL = "https://token.opencomputing.cn/llmage/v1/chat/completions"
+TOKEN = "V9J41PngWBUU6gdHWJWDJ"
+MODEL = "qwen3.6-35b-a3b"
+DURATION = 180  # 3 分钟
+CONCURRENCIES = [10, 50, 100, 200]
+
+
+@dataclass
+class ReqStat:
+    prompt_idx: int
+    start_ts: float
+    first_byte_ts: float | None = None
+    end_ts: float | None = None
+
+
+async def worker(session: aiohttp.ClientSession, idx: int, stats_out: list):
+    """单个请求：发送 stream 请求，记录首字时间和完成时间"""
+    prompt = f"请用一句话介绍你自己，编号{idx}"
+    payload = {
+        "model": MODEL,
+        "stream": True,
+        "messages": [{"role": "user", "content": prompt}],
+    }
+    stat = ReqStat(prompt_idx=idx, start_ts=time.monotonic())
+    try:
+        async with session.post(
+            URL,
+            json=payload,
+            headers={
+                "Content-Type": "application/json",
+                "Authorization": f"Bearer {TOKEN}",
+            },
+            timeout=aiohttp.ClientTimeout(total=120),
+        ) as resp:
+            first = True
+            async for line in resp.content:
+                if first:
+                    stat.first_byte_ts = time.monotonic()
+                    first = False
+                # 读完所有 chunk 才算完成
+        stat.end_ts = time.monotonic()
+    except Exception as e:
+        # 异常请求也记录（TTFB=None 表示失败）
+        stat.end_ts = time.monotonic()
+    stats_out.append(stat)
+
+
+async def run_concurrency(concurrency: int):
+    """以固定并发运行 DURATION 秒，持续发起新请求"""
+    stats: List[ReqStat] = []
+    idx = 0
+    stop_at = time.monotonic() + DURATION
+
+    connector = aiohttp.TCPConnector(limit=concurrency + 20, force_close=True)
+    async with aiohttp.ClientSession(connector=connector) as session:
+        tasks: list[asyncio.Task] = []
+
+        while time.monotonic() < stop_at:
+            # 保持并发数：补满到 concurrency
+            while len(tasks) < concurrency and time.monotonic() < stop_at:
+                idx += 1
+                tasks.append(
+                    asyncio.create_task(worker(session, idx, stats))
+                )
+
+            if not tasks:
+                break
+
+            # 等待任意一个完成，腾出槽位
+            done, tasks = await asyncio.wait(
+                tasks, return_when=asyncio.FIRST_COMPLETED, timeout=0.5
+            )
+            # 清理已完成的
+            tasks = list(tasks)
+
+        # 时间到，等待所有进行中的请求完成
+        if tasks:
+            await asyncio.wait(tasks)
+
+    return stats
+
+
+def analyze(name: str, stats: List[ReqStat]):
+    """分析并打印统计"""
+    ttfb_list = [s.first_byte_ts - s.start_ts for s in stats if s.first_byte_ts]
+    total_list = [s.end_ts - s.start_ts for s in stats if s.end_ts and s.first_byte_ts]
+    failed = sum(1 for s in stats if s.first_byte_ts is None)
+    total_req = len(stats)
+    elapsed = DURATION
+    qpm = total_req / (elapsed / 60)
+
+    print(f"\n{'='*60}")
+    print(f"  并发={name} | 运行{DURATION}s | 总请求={total_req} | 失败={failed}")
+    print(f"{'='*60}")
+    if ttfb_list:
+        print(f"  TTFB (s):  min={min(ttfb_list):.3f}  avg={statistics.mean(ttfb_list):.3f}  "
+              f"p50={statistics.median(ttfb_list):.3f}  p95={_pct(ttfb_list, 95):.3f}  p99={_pct(ttfb_list, 99):.3f}")
+    if total_list:
+        print(f"  完成 (s):  min={min(total_list):.3f}  avg={statistics.mean(total_list):.3f}  "
+              f"p50={statistics.median(total_list):.3f}  p95={_pct(total_list, 95):.3f}  p99={_pct(total_list, 99):.3f}")
+    print(f"  QPM:       {qpm:.1f}")
+    print(f"  QPS:       {total_req / elapsed:.1f}")
+
+    # 按分钟分段统计
+    for minute in range(int(elapsed / 60)):
+        win_start = minute * 60
+        win_end = (minute + 1) * 60
+        cnt = sum(1 for s in stats if s.end_ts and (s.end_ts - s.start_ts) >= 0
+                  and win_start <= (s.start_ts - stats[0].start_ts) < win_end)
+        print(f"  第{minute+1}分钟完成请求数: {cnt}")
+
+    return {
+        "concurrency": name,
+        "total": total_req,
+        "failed": failed,
+        "ttfb_avg": statistics.mean(ttfb_list) if ttfb_list else None,
+        "ttfb_p50": statistics.median(ttfb_list) if ttfb_list else None,
+        "ttfb_p95": _pct(ttfb_list, 95) if ttfb_list else None,
+        "total_avg": statistics.mean(total_list) if total_list else None,
+        "total_p50": statistics.median(total_list) if total_list else None,
+        "qpm": qpm,
+    }
+
+
+def _pct(data, p):
+    return sorted(data)[int(len(data) * p / 100)]
+
+
+async def main():
+    results = []
+    for c in CONCURRENCIES:
+        print(f"\n>>> 开始测试 并发={c} ...")
+        stats = await run_concurrency(c)
+        r = analyze(str(c), stats)
+        results.append(r)
+
+    # 汇总表格
+    print(f"\n{'='*60}")
+    print("  汇总对比")
+    print(f"{'='*60}")
+    print(f"  {'并发':>6} {'总请求':>8} {'失败':>5} {'TTFB_avg':>9} {'TTFB_p50':>9} {'TTFB_p95':>9} {'完成_avg':>9} {'QPM':>8}")
+    for r in results:
+        print(f"  {r['concurrency']:>6} {r['total']:>8} {r['failed']:>5} "
+              f"{r['ttfb_avg']:.3f}s" if r['ttfb_avg'] else "N/A".rjust(9) + " "
+              f"{(r['ttfb_p50'] or 0):.3f}s".rjust(9) + " "
+              f"{(r['ttfb_p95'] or 0):.3f}s".rjust(9) + " "
+              f"{r['total_avg']:.3f}s".rjust(9) if r['total_avg'] else "N/A".rjust(9))
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/wwwroot/v1/audio/speech/index.dspy b/wwwroot/v1/audio/speech/index.dspy
index c9487b9..6ee1a33 100644
--- a/wwwroot/v1/audio/speech/index.dspy
+++ b/wwwroot/v1/audio/speech/index.dspy
@@ -41,22 +41,11 @@ if not params_kw.prompt:
 lctype = params_kw.catelogid
 
 env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-    # Look up llm by model name and catalog type through llm_api_map
-    sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${lctype}$ OR b.name = ${lctype}$)
-    and a.name=${model}$
-    and a.status = 'published'"""
-    recs = await sor.sqlExe(sql, {
-        'lctype': lctype,
-        'model': params_kw.model
-    })
-    if len(recs) == 0:
-        debug(f'{params_kw.model=} not found for catalog {lctype}')
-        return openai_400()
-    params_kw.llmid = recs[0].id
+llmid = await env.get_llmid_cached(env, params_kw.model, lctype)
+if not llmid:
+	debug(f'{params_kw.model=} not found for catalog {lctype}')
+	return openai_400()
+params_kw.llmid = llmid
     params_kw.llmcatelogid = lctype
 
 debug(f'{params_kw.llmid=}')
diff --git a/wwwroot/v1/audio/transcriptions/index.dspy b/wwwroot/v1/audio/transcriptions/index.dspy
index 42f6664..203ce99 100644
--- a/wwwroot/v1/audio/transcriptions/index.dspy
+++ b/wwwroot/v1/audio/transcriptions/index.dspy
@@ -38,22 +38,11 @@ if not params_kw.audio_file:
 lctype = params_kw.catelogid
 
 env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-    # Look up llm by model name and catalog type through llm_api_map
-    sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${lctype}$ OR b.name = ${lctype}$)
-    and a.name=${model}$
-    and a.status = 'published'"""
-    recs = await sor.sqlExe(sql, {
-        'lctype': lctype,
-        'model': params_kw.model
-    })
-    if len(recs) == 0:
-        debug(f'{params_kw.model=} not found for catalog {lctype}')
-        return openai_400()
-    params_kw.llmid = recs[0].id
+llmid = await env.get_llmid_cached(env, params_kw.model, lctype)
+if not llmid:
+	debug(f'{params_kw.model=} not found for catalog {lctype}')
+	return openai_400()
+params_kw.llmid = llmid
     params_kw.llmcatelogid = lctype
 
 debug(f'{params_kw.llmid=}')
diff --git a/wwwroot/v1/image/generations/index.dspy b/wwwroot/v1/image/generations/index.dspy
index bf2300f..2c4695d 100644
--- a/wwwroot/v1/image/generations/index.dspy
+++ b/wwwroot/v1/image/generations/index.dspy
@@ -39,22 +39,11 @@ if not params_kw.prompt:
 lctype = params_kw.catelogid
 
 env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-    # Look up llm by model name and catalog type through llm_api_map
-    sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${lctype}$ OR b.name = ${lctype}$)
-    and a.name=${model}$
-    and a.status = 'published'"""
-    recs = await sor.sqlExe(sql, {
-        'lctype': lctype,
-        'model': params_kw.model
-    })
-    if len(recs) == 0:
-        debug(f'{params_kw.model=} not found for catalog {lctype}')
-        return openai_400()
-    params_kw.llmid = recs[0].id
+llmid = await env.get_llmid_cached(env, params_kw.model, lctype)
+if not llmid:
+	debug(f'{params_kw.model=} not found for catalog {lctype}')
+	return openai_400()
+params_kw.llmid = llmid
     params_kw.llmcatelogid = lctype
 
 debug(f'{params_kw.llmid=}')
diff --git a/wwwroot/v1/media/asr-transcribe/index.dspy b/wwwroot/v1/media/asr-transcribe/index.dspy
index 357b373..37c859a 100644
--- a/wwwroot/v1/media/asr-transcribe/index.dspy
+++ b/wwwroot/v1/media/asr-transcribe/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - asr-transcribe
-# /v1/media/asr-transcribe
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'asr-transcribe'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - asr-transcribe
+2|# /v1/media/asr-transcribe
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'asr-transcribe'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/demucs-separate/index.dspy b/wwwroot/v1/media/demucs-separate/index.dspy
index c1514b3..0119edb 100644
--- a/wwwroot/v1/media/demucs-separate/index.dspy
+++ b/wwwroot/v1/media/demucs-separate/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - demucs-separate
-# /v1/media/demucs-separate
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'demucs-separate'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - demucs-separate
+2|# /v1/media/demucs-separate
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'demucs-separate'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/face-compare/index.dspy b/wwwroot/v1/media/face-compare/index.dspy
index 3379346..a2a4960 100644
--- a/wwwroot/v1/media/face-compare/index.dspy
+++ b/wwwroot/v1/media/face-compare/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - face-compare
-# /v1/media/face-compare
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'face-compare'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - face-compare
+2|# /v1/media/face-compare
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'face-compare'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/face-detect/index.dspy b/wwwroot/v1/media/face-detect/index.dspy
index 2d982e5..c999e71 100644
--- a/wwwroot/v1/media/face-detect/index.dspy
+++ b/wwwroot/v1/media/face-detect/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - face-detect
-# /v1/media/face-detect
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'face-detect'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - face-detect
+2|# /v1/media/face-detect
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'face-detect'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/face-recognize/index.dspy b/wwwroot/v1/media/face-recognize/index.dspy
index 28a06b9..e081dfd 100644
--- a/wwwroot/v1/media/face-recognize/index.dspy
+++ b/wwwroot/v1/media/face-recognize/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - face-recognize
-# /v1/media/face-recognize
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'face-recognize'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - face-recognize
+2|# /v1/media/face-recognize
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'face-recognize'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/merge-video/index.dspy b/wwwroot/v1/media/merge-video/index.dspy
index 2bba72f..f981cb4 100644
--- a/wwwroot/v1/media/merge-video/index.dspy
+++ b/wwwroot/v1/media/merge-video/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - merge-video
-# /v1/media/merge-video
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'merge-video'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - merge-video
+2|# /v1/media/merge-video
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'merge-video'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/realesrgan-upscale/index.dspy b/wwwroot/v1/media/realesrgan-upscale/index.dspy
index 6ae1cd8..1d76ae4 100644
--- a/wwwroot/v1/media/realesrgan-upscale/index.dspy
+++ b/wwwroot/v1/media/realesrgan-upscale/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - realesrgan-upscale
-# /v1/media/realesrgan-upscale
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'realesrgan-upscale'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - realesrgan-upscale
+2|# /v1/media/realesrgan-upscale
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'realesrgan-upscale'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/rvc-convert/index.dspy b/wwwroot/v1/media/rvc-convert/index.dspy
index 505622a..6fb4492 100644
--- a/wwwroot/v1/media/rvc-convert/index.dspy
+++ b/wwwroot/v1/media/rvc-convert/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - rvc-convert
-# /v1/media/rvc-convert
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'rvc-convert'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - rvc-convert
+2|# /v1/media/rvc-convert
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'rvc-convert'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/songrate-evaluate/index.dspy b/wwwroot/v1/media/songrate-evaluate/index.dspy
index 85fbf83..6dc7ac2 100644
--- a/wwwroot/v1/media/songrate-evaluate/index.dspy
+++ b/wwwroot/v1/media/songrate-evaluate/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - songrate-evaluate
-# /v1/media/songrate-evaluate
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'songrate-evaluate'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - songrate-evaluate
+2|# /v1/media/songrate-evaluate
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'songrate-evaluate'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/subtitle-render/index.dspy b/wwwroot/v1/media/subtitle-render/index.dspy
index 4fd9949..940fb4a 100644
--- a/wwwroot/v1/media/subtitle-render/index.dspy
+++ b/wwwroot/v1/media/subtitle-render/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - subtitle-render
-# /v1/media/subtitle-render
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'subtitle-render'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - subtitle-render
+2|# /v1/media/subtitle-render
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'subtitle-render'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/synth-generate/index.dspy b/wwwroot/v1/media/synth-generate/index.dspy
index 470c79f..0fdf7a9 100644
--- a/wwwroot/v1/media/synth-generate/index.dspy
+++ b/wwwroot/v1/media/synth-generate/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - synth-generate
-# /v1/media/synth-generate
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'synth-generate'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - synth-generate
+2|# /v1/media/synth-generate
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'synth-generate'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/media/video-eval-evaluate/index.dspy b/wwwroot/v1/media/video-eval-evaluate/index.dspy
index 4dae516..ea63d21 100644
--- a/wwwroot/v1/media/video-eval-evaluate/index.dspy
+++ b/wwwroot/v1/media/video-eval-evaluate/index.dspy
@@ -1,34 +1,30 @@
-# KTV Media API - video-eval-evaluate
-# /v1/media/video-eval-evaluate
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	return openai_403()
-
-catelogid = 'ktv_pipeline'
-model_name = 'video-eval-evaluate'
-params_kw.model = model_name
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.model=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {'catelogid': catelogid, 'model': model_name})
-	if len(recs) == 0:
-		return openai_400()
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	return openai_429()
-
-return await inference(request, env=env)
+1|# KTV Media API - video-eval-evaluate
+2|# /v1/media/video-eval-evaluate
+3|
+4|debug_params('params_kw', params_kw)
+5|
+6|userid = await get_user()
+7|userorgid = await get_userorgid()
+8|if userid is None:
+9|	return openai_403()
+10|
+11|catelogid = 'ktv_pipeline'
+12|model_name = 'video-eval-evaluate'
+13|params_kw.model = model_name
+14|params_kw.catelogid = catelogid
+15|
+16|env = request._run_ns
+17|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return openai_400()
+params_kw.llmid = llmid
+28|	params_kw.llmcatelogid = catelogid
+29|
+30|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+31|if not f:
+32|	return openai_429()
+33|
+34|return await inference(request, env=env)
+35|
\ No newline at end of file
diff --git a/wwwroot/v1/music/generations/index.dspy b/wwwroot/v1/music/generations/index.dspy
index 160b238..79e364d 100644
--- a/wwwroot/v1/music/generations/index.dspy
+++ b/wwwroot/v1/music/generations/index.dspy
@@ -47,22 +47,11 @@ if not params_kw.lyrics:
 lctype = params_kw.catelogid
 
 env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-    # Look up llm by model name and catalog type through llm_api_map
-    sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${lctype}$ OR b.name = ${lctype}$)
-    and a.name=${model}$
-    and a.status = 'published'"""
-    recs = await sor.sqlExe(sql, {
-        'lctype': lctype,
-        'model': params_kw.model
-    })
-    if len(recs) == 0:
-        debug(f'{params_kw.model=} not found for catalog {lctype}')
-        return openai_400()
-    params_kw.llmid = recs[0].id
+llmid = await env.get_llmid_cached(env, params_kw.model, lctype)
+if not llmid:
+	debug(f'{params_kw.model=} not found for catalog {lctype}')
+	return openai_400()
+params_kw.llmid = llmid
     params_kw.llmcatelogid = lctype
 
 debug(f'{params_kw.llmid=}')
diff --git a/wwwroot/v1/pipeline/submit/index.dspy b/wwwroot/v1/pipeline/submit/index.dspy
index a339d2c..525ceb7 100644
--- a/wwwroot/v1/pipeline/submit/index.dspy
+++ b/wwwroot/v1/pipeline/submit/index.dspy
@@ -1,74 +1,64 @@
-# KTV Pipeline / GPU Service Inference API
-# POST /v1/pipeline/submit
-# catelogid 固定为 ktv_pipeline，无需传 catelogid
-#
-# Required params:
-#   model: string - 模型名称，如 "ky-asr-transcribe"
-#
-# 各服务特定参数见 API 文档
-#
-# 同步服务直接返回结果：
-# {
-#   "taskid": "luid_xxx",
-#   "taskstatus": "SUCCEEDED",
-#   ...服务特定结果字段...,
-#   "usage": {...}
-# }
-#
-# 异步服务返回任务信息：
-# {
-#   "taskid": "luid_xxx",
-#   "taskstatus": "PENDING"
-# }
-# 异步结果通过 /v1/tasks?taskid=xxx 查询
-
-debug_params('params_kw', params_kw)
-
-userid = await get_user()
-userorgid = await get_userorgid()
-if userid is None:
-	debug('need login')
-	return openai_403()
-
-# Validate required parameters
-if not params_kw.model:
-	d = return_error('Missing required parameter: model')
-	return json_response(d, status=400)
-
-catelogid = 'ktv_pipeline'
-params_kw.catelogid = catelogid
-
-env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	# Look up llm by model name through llm_api_map (fixed catelogid=ktv_pipeline)
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${catelogid}$ OR b.name = ${catelogid}$)
-	and a.name=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {
-		'catelogid': catelogid,
-		'model': params_kw.model
-	})
-	if len(recs) == 0:
-		debug(f'{params_kw.model=} not found under ktv_pipeline catalog')
-		d = return_error(f'Model "{params_kw.model}" not found or not available')
-		return json_response(d, status=400)
-	params_kw.llmid = recs[0].id
-	params_kw.llmcatelogid = catelogid
-
-debug(f'{params_kw.llmid=}')
-
-# Check balance
-f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
-if not f:
-	debug(f'{userid=} balance not enough')
-	return openai_429()
-
-# Generate task ID
-if not params_kw.transno:
-	params_kw.transno = getID()
-
-# Call inference
-return await inference(request, env=env)
+1|# KTV Pipeline / GPU Service Inference API
+2|# POST /v1/pipeline/submit
+3|# catelogid 固定为 ktv_pipeline，无需传 catelogid
+4|#
+5|# Required params:
+6|#   model: string - 模型名称，如 "ky-asr-transcribe"
+7|#
+8|# 各服务特定参数见 API 文档
+9|#
+10|# 同步服务直接返回结果：
+11|# {
+12|#   "taskid": "luid_xxx",
+13|#   "taskstatus": "SUCCEEDED",
+14|#   ...服务特定结果字段...,
+15|#   "usage": {...}
+16|# }
+17|#
+18|# 异步服务返回任务信息：
+19|# {
+20|#   "taskid": "luid_xxx",
+21|#   "taskstatus": "PENDING"
+22|# }
+23|# 异步结果通过 /v1/tasks?taskid=xxx 查询
+24|
+25|debug_params('params_kw', params_kw)
+26|
+27|userid = await get_user()
+28|userorgid = await get_userorgid()
+29|if userid is None:
+30|	debug('need login')
+31|	return openai_403()
+32|
+33|# Validate required parameters
+34|if not params_kw.model:
+35|	d = return_error('Missing required parameter: model')
+36|	return json_response(d, status=400)
+37|
+38|catelogid = 'ktv_pipeline'
+39|params_kw.catelogid = catelogid
+40|
+41|env = request._run_ns
+42|# llmid from cache (model+catelogid -> llmid)
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', 'ktv_pipeline')
+if not llmid:
+	debug(f'model not found: params_kw.model')
+	return {{error_response}}
+params_kw.llmid = llmid
+59|	params_kw.llmcatelogid = catelogid
+60|
+61|debug(f'{params_kw.llmid=}')
+62|
+63|# Check balance
+64|f = await checkCustomerBalance(params_kw.llmid, userid, userorgid)
+65|if not f:
+66|	debug(f'{userid=} balance not enough')
+67|	return openai_429()
+68|
+69|# Generate task ID
+70|if not params_kw.transno:
+71|	params_kw.transno = getID()
+72|
+73|# Call inference
+74|return await inference(request, env=env)
+75|
\ No newline at end of file
diff --git a/wwwroot/v1/video/generations/index.dspy b/wwwroot/v1/video/generations/index.dspy
index a24c326..b4801ec 100644
--- a/wwwroot/v1/video/generations/index.dspy
+++ b/wwwroot/v1/video/generations/index.dspy
@@ -47,22 +47,11 @@ if not params_kw.prompt:
 lctype = params_kw.catelogid
 
 env = request._run_ns
-async with get_sor_context(env, 'llmage') as sor:
-	# Look up llm by model name and catalog type through llm_api_map
-	sql = """select distinct a.* from llm a
-join llm_api_map m on a.id = m.llmid
-join llmcatelog b on m.llmcatelogid = b.id
-where (b.id = ${lctype}$ OR b.name = ${lctype}$)
-	and a.name=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {
-		'lctype': lctype,
-		'model': params_kw.model
-	})
-	if len(recs) == 0:
-		debug(f'{params_kw.model=} not found for catalog {lctype}')
-		return openai_400()
-	params_kw.llmid = recs[0].id
+llmid = await env.get_llmid_cached(env, params_kw.model, lctype)
+if not llmid:
+	debug(f'{params_kw.model=} not found for catalog {lctype}')
+	return openai_400()
+params_kw.llmid = llmid
 	params_kw.llmcatelogid = lctype
 
 debug(f'{params_kw.llmid=}')