gahow/agentic-kvc
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix retry syntax: async generator can't use return, use break+try/finally

Browse Source 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Gahow Wang
2026-05-23 22:37:32 +08:00
parent daeb95eca0
commit a7514fc3d5
1 changed files with 17 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
scripts/cache_aware_proxy.py
View File

@@ -446,28 +446,24 @@ async def _handle_combined(api, req_data, token_ids, input_length, session_id, h
async def generate(): async def generate():
prefill_done = False prefill_done = False
last_err = None try:
for attempt in range(MAX_STREAM_RETRIES): for attempt in range(MAX_STREAM_RETRIES):
try: try:
async with inst.client.stream("POST", api, json=req_data, headers=headers) as resp: async with inst.client.stream("POST", api, json=req_data, headers=headers) as resp:
resp.raise_for_status() resp.raise_for_status()
async for chunk in resp.aiter_bytes(): async for chunk in resp.aiter_bytes():
if not prefill_done: if not prefill_done:
inst.pending_prefill_tokens -= estimated_new inst.pending_prefill_tokens -= estimated_new
inst.ongoing_decode_tokens += input_length inst.ongoing_decode_tokens += input_length
breakdown["t_first_token"] = _time.monotonic() breakdown["t_first_token"] = _time.monotonic()
prefill_done = True prefill_done = True
yield chunk yield chunk
inst.record_prefix(token_ids) inst.record_prefix(token_ids)
return break
except (httpx.ConnectError, httpx.RemoteProtocolError) as e: except (httpx.ConnectError, httpx.RemoteProtocolError):
last_err = e if prefill_done or attempt >= MAX_STREAM_RETRIES - 1:
if prefill_done: raise
raise
if attempt < MAX_STREAM_RETRIES - 1:
await asyncio.sleep(RETRY_DELAY_S) await asyncio.sleep(RETRY_DELAY_S)
if last_err:
raise last_err
finally: finally:
if not prefill_done: if not prefill_done:
inst.pending_prefill_tokens -= estimated_new inst.pending_prefill_tokens -= estimated_new