01 Apr, 2026
2 commits
31 Mar, 2026
11 commits
30 Mar, 2026
1 commit
27 Mar, 2026
4 commits
-
coarse_rank.output_window -> 再做 SKU 选择和 title suffix -> 精排调用轻量 reranker 裁到 fine_rank.output_window -> 最终重排调用现有 reranker，并在最终融合里加入 fine_score。同时把 reranker client/provider 改成了按 service_profile 选不同 service_url，这样 fine/final 可以共用同一套服务代码，只起不同实例。
26 Mar, 2026
3 commits
25 Mar, 2026
7 commits
-
（之前因为错误将attention方法该回到TRITON_ATTN，性能相比于之前的vllm版本更差。但是那个错误是能解决的。已修复保持FLASHINFER）
-
报错），并允许通过配置或环境变量让 vLLM 自行选择 attention。 -- 临时版本
-
这两个配置、四种情况： backend: qwen3_vllm | qwen3_vllm_score instruction_format: compact | standard 调用 python scripts/benchmark_reranker_random_titles.py 100,200,400,600,800,1000 --repeat 5 产出性能测试报告 平均延迟（ms，客户端 POST /rerank 墙钟，--seed 99） backend instruction_format n=100 n=200 n=400 n=600 n=800 n=1000 qwen3_vllm compact 213.5 418.0 861.4 1263.4 1744.3 2162.2 qwen3_vllm standard 254.9 475.4 909.7 1353.2 1912.5 2406.7 qwen3_vllm_score compact 239.2 480.2 966.2 1433.5 1937.2 2428.4 qwen3_vllm_score standard 299.6 591.8 1178.9 1773.7 2341.6 2931.7 归纳： 在本机 T4、当前 vLLM 与上述 YAML（max_model_len=160、infer_batch_size=100 等）下，两种后端都是 compact 快于 standard；整体最快为 qwen3_vllm + compact（n=1000 ≈ 2.16 s），最慢为 qwen3_vllm_score + standard（≈ 2.93 s）。其他 GPU / vLLM 版本下排序可能变化。
22 Mar, 2026
1 commit
21 Mar, 2026
2 commits
20 Mar, 2026
2 commits
19 Mar, 2026
5 commits
-
- Text and image embedding are now split into separate services/processes, while still keeping a single replica as requested. The split lives in [embeddings/server.py](/data/saas-search/embeddings/server.py#L112), [config/services_config.py](/data/saas-search/config/services_config.py#L68), [providers/embedding.py](/data/saas-search/providers/embedding.py#L27), and the start scripts [scripts/start_embedding_service.sh](/data/saas-search/scripts/start_embedding_service.sh#L36), [scripts/start_embedding_text_service.sh](/data/saas-search/scripts/start_embedding_text_service.sh), [scripts/start_embedding_image_service.sh](/data/saas-search/scripts/start_embedding_image_service.sh). - Independent admission control is in place now: text and image have separate inflight limits, and image can be kept much stricter than text. The request handling, reject path, `/health`, and `/ready` are in [embeddings/server.py](/data/saas-search/embeddings/server.py#L613), [embeddings/server.py](/data/saas-search/embeddings/server.py#L786), and [embeddings/server.py](/data/saas-search/embeddings/server.py#L1028). - I checked the Redis embedding cache. It did exist, but there was a real flaw: cache keys did not distinguish `normalize=true` from `normalize=false`. I fixed that in [embeddings/cache_keys.py](/data/saas-search/embeddings/cache_keys.py#L6), and both text and image now use the same normalize-aware keying. I also added service-side BF16 cache hits that short-circuit before the model lane, so repeated requests no longer get throttled behind image inference. **What This Means** - Image pressure no longer blocks text, because they are on different ports/processes. - Repeated text/image requests now return from Redis without consuming model capacity. - Over-capacity requests are rejected quickly instead of sitting blocked. - I did not add a load balancer or multi-replica HA, per your GPU constraint. I also did not build Grafana/Prometheus dashboards in this pass, but `/health` now exposes the metrics needed to wire them. **Validation** - Tests passed: `.venv/bin/python -m pytest -q tests/test_embedding_pipeline.py tests/test_embedding_service_limits.py` -> `10 passed` - Stress test tool updates are in [scripts/perf_api_benchmark.py](/data/saas-search/scripts/perf_api_benchmark.py#L155) - Fresh benchmark on split text service `6105`: 535 requests / 3s, 100% success, `174.56 rps`, avg `88.48 ms` - Fresh benchmark on split image service `6108`: 1213 requests / 3s, 100% success, `403.32 rps`, avg `9.64 ms` - Live health after the run showed cache hits and non-zero cache-hit latency accounting: - text `avg_latency_ms=4.251` - image `avg_latency_ms=1.462`
-
The instability is very likely real overload, but `lsof -i :6005 | wc -l = 75` alone does not prove it. What does matter is the live shape of the service: it is a single `uvicorn` worker on port `6005`, and the code had one shared process handling both text and image requests, with image work serialized behind a single lock. Under bursty image traffic, requests could pile up and sit blocked with almost no useful tracing, which matches the “only blocking observed” symptom. now adds persistent log files, request IDs, per-request request/response/failure logs, text microbatch dispatch logs, health stats with active/rejected counts, and explicit overload admission control. New knobs are `TEXT_MAX_INFLIGHT`, `IMAGE_MAX_INFLIGHT`, and `EMBEDDING_OVERLOAD_STATUS_CODE`. Startup output now shows those limits and log paths in [scripts/start_embedding_service.sh](/data/saas-search/scripts/start_embedding_service.sh#L80). I also added focused tests in [tests/test_embedding_service_limits.py](/data/saas-search/tests/test_embedding_service_limits.py#L1). What this means operationally: - Text and image are still in one process, so this is not the final architecture. - But image spikes will now be rejected quickly once the image lane is full instead of sitting around and consuming the worker pool. - Logs will now show each request, each rejection, each microbatch dispatch, backend time, response time, and request ID. Verification: - Passed: `.venv/bin/python -m pytest -q tests/test_embedding_service_limits.py` - I also ran a wider test command, but 3 failures came from pre-existing drift in [tests/test_embedding_pipeline.py](/data/saas-search/tests/test_embedding_pipeline.py#L95), where the tests still monkeypatch `embeddings.text_encoder.redis.Redis` even though [embeddings/text_encoder.py](/data/saas-search/embeddings/text_encoder.py#L1) no longer imports `redis` that way. 已把 CLIP_AS_SERVICE 的默认模型切到 ViT-L-14，并把这套配置收口成可变更的统一入口了。现在默认值在 embeddings/config.py (line 29) 的 CLIP_AS_SERVICE_MODEL_NAME，当前为 CN-CLIP/ViT-L-14；scripts/start_cnclip_service.sh (line 37) 会自动读取这个配置，不再把默认模型写死在脚本里，同时支持 CNCLIP_MODEL_NAME 和 --model-name 临时覆盖。scripts/start_embedding_service.sh (line 29) 和 embeddings/server.py (line 425) 也补了模型信息输出，方便排查实际连接的配置。 文档也一起更新了，重点在 docs/CNCLIP_SERVICE说明文档.md (line 62) 和 embeddings/README.md (line 58)：现在说明的是“以配置为准、可覆盖”的机制，而不是写死某个模型名；相关总结文档和内部说明也同步改成了配置驱动表述。
-
中采用了最优T4配置：ct2_inter_threads=2、ct2_max_queued_batches=16、ct2_batch_type=examples。该设置使NLLB获得了显著更优的在线式性能，同时大致保持了大批次吞吐量不变。我没有将相同配置应用于两个Marian模型，因为聚焦式报告显示了复杂的权衡：opus-mt-zh-en 在保守默认配置下更为均衡，而 opus-mt-en-zh 虽然获得了吞吐量提升，但在 c=8 时尾延迟波动较大。 我还将部署/配置经验记录在 /data/saas-search/translation/README.md 中，并在 /data/saas-search/docs/TODO.txt 中标记了优化结果。关键实践要点现已记录如下：使用CT2 + float16，保持单worker，将NLLB的 inter_threads 设为2、max_queued_batches 设为16，在T4上避免使用 inter_threads=4（因为这会损害高批次吞吐量），除非区分在线/离线配置，否则保持Marian模型的默认配置保守。
18 Mar, 2026
2 commits
-
Implemented CTranslate2 for the three local translation models and switched the existing local_nllb / local_marian factories over to it. The new runtime lives in local_ctranslate2.py, including HF->CT2 auto-conversion, float16 compute type mapping, Marian direction handling, and NLLB target-prefix decoding. The service wiring is in service.py (line 113), and the three model configs now point at explicit ctranslate2-float16 dirs in config.yaml (line 133). I also updated the setup path so this is usable end-to-end: ctranslate2>=4.7.0 was added to requirements_translator_service.txt and requirements.txt, the download script now supports pre-conversion in download_translation_models.py (line 27), and the docs/config examples were refreshed in translation/README.md. I installed ctranslate2 into .venv-translator, pre-converted all three models, and the CT2 artifacts are now already on disk: models/translation/facebook/nllb-200-distilled-600M/ctranslate2-float16 models/translation/Helsinki-NLP/opus-mt-zh-en/ctranslate2-float16 models/translation/Helsinki-NLP/opus-mt-en-zh/ctranslate2-float16 Verification was solid. python3 -m compileall passed, direct TranslationService smoke tests ran successfully in .venv-translator, and the focused NLLB benchmark on the local GPU showed a clear win: batch_size=16: HF 0.347s/batch, 46.1 items/s vs CT2 0.130s/batch, 123.0 items/s batch_size=1: HF 0.396s/request vs CT2 0.126s/request One caveat: translation quality on some very short phrases, especially opus-mt-en-zh, still looks a bit rough in smoke tests, so I’d run your real quality set before fully cutting over. If you want, I can take the next step and update the benchmark script/report so you have a fresh full CT2 performance report for all three models.
-
batch×并发矩阵”彻底分开展示。 改动在这几处： scripts/benchmark_translation_local_models.py：新增 --suite extended，支持 batch_size=1,4,8,16,32,64、concurrency=1,2,4,8,16,64、以及 batch_size * concurrency <= 128 的组合矩阵；并且单场景模式现在只加载目标模型，load_seconds 更干净，也支持 --disable-cache。 translation/README.md：把性能章节拆成了 batch_sweep、concurrency_sweep、batch x concurrency matrix 三块，补了这次复测的参数、复现命令和摘要表。 perf_reports/20260318/translation_local_models/README.md：新增本轮补测摘要。 完整结果在 translation_local_models_extended_221846.md 和 translation_local_models_extended_221846.json。 这次补测的核心结论很明确： 在线单条请求应该看 concurrency_sweep，也就是固定 batch_size=1 的表。 离线批量吞吐应该看 batch_sweep，4 个方向的最高 raw throughput 都出现在 batch_size=64，但更均衡的默认值仍更像 batch_size=16。 当前本地 seq2seq backend 有单模型锁，提升客户端并发几乎不涨吞吐，主要是把排队时间变成更高的 p95；所以并发更像“延迟预算”问题，不是“扩容吞吐”手段。 本轮在线单条里最快的是 opus-mt-zh-en；最慢、且并发放大最明显的是 nllb-200-distilled-600m en->zh。