perf: llmid缓存(启动预热+内存查),v1/chat/completions去3表JOIN

llmage/init.py

@@ -23,6 +23,9 @@ from .utils import (
 	invalidate_uapi_cache,
 	get_llmusage_by_id,
 	read_ioinfo_content,
+	_warm_llmid_cache,
+	get_llmid_cached,
+	invalidate_llmid_cache,
 )
 
 from .llmclient import (
@@ -122,10 +125,11 @@ order by lc.name, a.name"""
 
 
 def _on_hot_reload(data=None):
-	"""Event handler for hot_reload — wraps invalidate_uapi_cache to accept dispatcher's data arg."""
+	"""Event handler for hot_reload — invalidate caches."""
 	from appPublic.log import debug
-	debug(f'[llmage] on_hot_reload called, invalidating uapi cache (data={data})')
+	debug(f'[llmage] on_hot_reload called, invalidating caches (data={data})')
 	invalidate_uapi_cache()
+	invalidate_llmid_cache()
 
 
 def load_llmage():
@@ -148,6 +152,10 @@ def load_llmage():
 	env.get_tpac_balance = get_tpac_balance
 	env.inference_generator = inference_generator
 	env.get_llms_by_catelog = get_llms_by_catelog
+	env.get_llmid_cached = get_llmid_cached
+	env.invalidate_llmid_cache = invalidate_llmid_cache
+	# 启动时预热 llmid 缓存
+	asyncio.ensure_future(_warm_llmid_cache(env))
 	env.get_llmcatelogs = get_llmcatelogs
 	env.checkCustomerBalance = checkCustomerBalance
 	env.get_llmproviders = get_llmproviders

llmage/utils.py

@@ -63,6 +63,55 @@ def invalidate_uapi_cache(upappid=None, apiname=None):
         _uapiio_cache.clear()
 
 
+# =============================================================
+# Process-level cache for llmid lookup (model+catelogid -> llmid)
+# =============================================================
+_llmid_cache = {}  # key: "model_name:catelogid" -> llmid
+
+
+async def _warm_llmid_cache(env):
+    """启动时全量加载 model+catelogid -> llmid 映射"""
+    global _llmid_cache
+    try:
+        async with get_sor_context(env, 'llmage') as sor:
+            sql = """SELECT a.name, b.id as catelogid, m.llmid
+            FROM llm_api_map m
+            JOIN llm a ON a.id = m.llmid AND a.status = 'published'
+            JOIN llmcatelog b ON b.id = m.llmcatelogid"""
+            recs = await sor.sqlExe(sql, {})
+            for r in recs:
+                key = f"{r.name}:{r.catelogid}"
+                _llmid_cache[key] = r.llmid
+        debug(f'[llmage] llmid cache warmed: {len(_llmid_cache)} entries')
+    except Exception as e:
+        exception(f'[llmage] llmid cache warm failed: {e}')
+        _llmid_cache = {}
+
+
+async def get_llmid_cached(env, model_name, catelogid):
+    """从缓存获取 llmid，未命中则查 DB 并缓存"""
+    global _llmid_cache
+    key = f"{model_name}:{catelogid}"
+    if key in _llmid_cache:
+        return _llmid_cache[key]
+    # 缓存未命中，查 DB（兼容模型在缓存预热后新增的场景）
+    async with get_sor_context(env, 'llmage') as sor:
+        sql = """SELECT m.llmid
+        FROM llm_api_map m
+        JOIN llm a ON a.id = m.llmid AND a.name = ${model}$ AND a.status = 'published'
+        JOIN llmcatelog b ON b.id = m.llmcatelogid AND (b.id = ${catelogid}$ OR b.name = ${catelogid}$)"""
+        recs = await sor.sqlExe(sql, {'model': model_name, 'catelogid': catelogid})
+        llmid = recs[0].llmid if recs else None
+        if llmid:
+            _llmid_cache[key] = llmid
+        return llmid
+
+
+def invalidate_llmid_cache():
+    global _llmid_cache
+    _llmid_cache.clear()
+
+
 async def update_llmusage(ns):
 	env = ServerEnv()
 	async with get_sor_context(env, 'llmage') as sor:

wwwroot/v1/chat/completions/index.dspy

@@ -29,21 +29,11 @@ if not params_kw.prompt and not params_kw.messages:
 	d = return_error('Missing need data(prompt or messages)')
 	return json_response(d, status=400)
 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.name=${model}$
-	and a.status = 'published'"""
-	recs = await sor.sqlExe(sql, {
-		'catelogid': catelogid,
-		'model': params_kw.model or 'qwen3-max'
-	})
-	if len(recs) == 0:
-		debug(f'{params_kw.model=} not found')
-		return openai_400()
-	params_kw.llmid = recs[0].id
+llmid = await env.get_llmid_cached(env, params_kw.model or 'qwen3-max', catelogid)
+if not llmid:
+	debug(f'{params_kw.model=} not found for catelogid={catelogid}')
+	return openai_400()
+params_kw.llmid = llmid
 	params_kw.llmcatelogid = catelogid
 
 debug(f'{params_kw.llmid=}')