Model training/serving bugs, LLM integrations, GPU/CUDA issues.
Our document tagger hit HTTP 429 constantly while sending only 40 requests a minute, far below the request limit. The token-per-minute limit was the real ceiling, inflated by an oversized max_tokens and minute-aligned bursts. How we read the headers and built a client-side limiter.
Our meeting transcripts contained sentences nobody said, usually during silence or hold music, plus the same line repeated a dozen times. Whisper conditions on its own previous output and has no idea what silence is. A VAD pre-filter and stricter decoding thresholds removed most of it.
We added cache_control to a 6,000-token system prompt and the bill went up instead of down. Every request wrote the cache and none read it, because the prompt prefix changed on every call. A stable prefix brought cache reads to 92 percent.
Our chat assistant streamed tokens smoothly on localhost but in production the whole answer appeared after 14 seconds. nginx was buffering the event stream and gzip was holding bytes. How we proved it with curl and fixed it without breaking other routes.
Serving Llama 3.1 8B with vLLM on a 24 GB GPU crashed at startup with a KV cache error. By default vLLM sizes for the model's full 128k context, and after loading weights there was room for about 45k tokens of KV cache. Setting max_model_len to what we need fixed it.
Small tenants in our RAG product got 2 or 3 chunks back instead of 10, and sometimes none. HNSW finds ef_search candidates first and applies the tenant filter afterwards. Iterative index scans in pgvector 0.8 plus partial indexes for big tenants fixed recall.
Our self-hosted LLM box produced three tokens per second on a server with an RTX 4090. ollama ps reported 100% CPU: the container had no access to the GPU. Installing the NVIDIA Container Toolkit and requesting the GPU in Compose brought it to 120 tokens per second.
Semantic search in our docs assistant took 1.9 seconds and returned odd rankings. The query sorted by L2 distance while the HNSW index was built for cosine, so Postgres scanned every row. Matching the operator to the opclass brought it to 14 ms.
After rebuilding our inference image, models silently ran on CPU on half the GPU fleet. The new PyTorch wheel targets CUDA 12, and older nodes still had a driver that only supports CUDA 11. We aligned drivers, added a startup check that fails without a GPU, and pinned images.
Our nightly embedding backfill died on the same few documents every night. A handful of chunks were far over the embedding model's token limit because the chunker split by characters and some files were minified JSON. Token-aware chunking and batching fixed it.
Our support agent started failing with a 400 on roughly one conversation in twelve. The model had begun returning several tool calls in one turn and our loop only answered the first. Answering every tool_call_id in order fixed it.
Our text classification service crashed with CUDA out of memory every few hours, always on a big batch. Logits kept for drift monitoring were holding autograd graphs alive, batches were sized by count not tokens, and the allocator was fragmented. How we measured each one and fixed it.
Switching our extraction endpoint to strict structured outputs failed with an invalid schema error. In strict mode every property must be listed in required and additionalProperties must be false. Optional fields became nullable fields and the errors stopped.
Our fraud model scored 0.91 AUC on holdout data and close to random in production. The serving service re-implemented preprocessing in another language: the scaler was refit on each batch, one-hot categories came in a different order and text was not lowercased. Shipping the pipeline with the model fixed it.
Our embedding API logged a tokenizers fork warning thousands of times a day and tokenisation got slower after we enabled Gunicorn preload. The tokenizer was used in the master process before workers forked. Loading it after fork fixed both.
Long chats in our internal assistant failed with a context length error even though we kept only the last 20 messages. One pasted log file was bigger than the whole window. We moved to token-based trimming with a reserved output budget.