SQLite vs vLLM

Export SQL Results to CSV and Excel Safely

A technical how-to showing how to export SQL query results to CSV or XLSX without corrupting values (leading zeros, accented characters, dates). The author demonstrates that CSV exports preserve bytes but that damage usually occurs when Excel guesses column types on open. Recommended practices include opening CSVs in Excel via Data → Get Data → From Text/CSV (set File Origin to UTF-8 and mark code columns as Text), adding a UTF-8 byte order mark (BOM) on export (e.g., sqlite3 .once --bom) to avoid accent issues, and generating real .xlsx files (e.g., with pandas.to_excel) when humans will double-click the file. The article also lists quick checks to run after export (row counts, code endpoints, an accent, a date) and details other common CSV pitfalls.

• The author tested exports on 2026-08-08 using SQLite 3.51.1 and inspected files byte-by-byte to show exports themselves are correct.
• Opening a CSV by double-clicking in Excel can silently change data types (e.g., drop leading zeros or mis-interpret character encoding), while importing via Data → Get Data → From Text/CSV lets the user set types and encoding.
• Adding a UTF-8 byte order mark (BOM) at file start (sqlite3 .once --bom results.csv) signals encoding to Excel and prevents accented-character corruption.

Trelix v3.0: Six opt-in features, audit trail added

The article documents trelix releases from v2.11.0 through v3.1.1 (last dated 2026-08-15). v3.0.0 introduced six major, opt-in feature areas: Anthropic extended thinking, a model-aware context token budget, a VS Code extension, a hash-chained append-only SQLite audit trail, OIDC SSO (resource-server mode), and query-conditioned context compression, plus an optional FTS5 declaration boost. Defaults keep all features off so a default v3.0.0 install is byte-identical to v2.12.0. v3.0.1 fixed a parser index bug and requires a full reindex; v3.1.0 and v3.1.1 addressed rendering, symlink handling and safety fixes. The audit trail is a separate opt-in SQLite DB with tamper-evident checks, and the OIDC verifier enforces asymmetric algorithms and forged-token tests.

• v3.0.0 introduces six new, opt-in feature areas: Anthropic extended thinking, model-aware context budget, VS Code extension, hash-chained append-only audit trail, OIDC SSO, and query-conditioned context compression.
• The release span v2.11.0 → v3.1.1 comprises 68 commits, 137 files changed, and +19,829/-1,211 lines; v3.1.1 dated 2026-08-15.
• The audit trail is an opt-in, separate SQLite file (default <cwd>/.trelix/audit.db) and uses sha256(prev_hash || canonical_json(content)) to produce tamper-evident hashes.

DeepSeek Open-Sources DeepInfer Inference Engine

DeepSeek open-sourced DeepInfer, the production inference engine behind its V4.1 Flash model, under Apache 2.0 during PyTorch Conference 2026. The framework supports Nvidia Blackwell (including SM100) and Huawei Ascend 950 chips, but not older Hopper, aligning with DeepSeek's strategy to reduce Nvidia dependence. It integrates previously open-sourced components like FlashMLA, DeepGEMM, and DeepEP, and includes techniques for distributed inference and speculative decoding. However, key features such as cache management, distributed execution, and load balancing remain proprietary. DeepSeek disclosed impressive metrics: 155 trillion tokens per day per cluster, 96% input cache hits, and $38 per accelerator-hour at off-peak, suggesting high margins. vLLM claims its throughput is within 6% of DeepInfer's, sparking debate. The release coincides with U.S. government actions against DeepSeek.

• DeepSeek released DeepInfer as open source under Apache 2.0.
• DeepInfer supports Nvidia Blackwell and Huawei Ascend 950, but not older Hopper.
• DeepSeek reports 155 trillion tokens per day per cluster and $38 revenue per accelerator-hour at off-peak.

MiniMax H3 on vLLM-Omni: From System-Wide Optimization to Real-Time Serving with FastVideo’s FastH3

How vLLM-Omni optimizes and scales the complete MiniMax H3 stack, then integrates FastVideo’s four-step FastH3 for generation faster than playback.

Qwen3-8B inference benchmark and FP8 on Blackwell

Independent benchmarks compare Qwen3-8B inference on an RTX PRO 6000 Blackwell (96 GB) across three serving stacks (vLLM 0.27.1, SGLang 0.5.9, and llama.cpp CUDA). At concurrency 32 using BF16, vLLM achieved 1,725 aggregate tokens/s (TTFT p50 39 ms), SGLang 1,327 tok/s (TTFT p50 42 ms), and llama.cpp 428 tok/s (TTFT p50 316 ms). Applying an FP8 checkpoint to vLLM increased throughput by ~1.5x (aggregate 1,725 -> 2,597 tok/s; single-stream 86 -> 130 tok/s) with lower latency and no detected regressions on a fixed factual check. The author documents methodology, reproductions, and an sm_120-specific kernel workaround required to run FP8 on workstation Blackwell hardware.

• GPU used: RTX PRO 6000 Blackwell, 96 GB (workstation Blackwell, sm_120).
• Model benchmarked: Qwen3-8B across vLLM 0.27.1, SGLang 0.5.9, and llama.cpp (CUDA).
• BF16, concurrency 32 aggregate throughput: vLLM 1,725 tok/s; SGLang 1,327 tok/s; llama.cpp 428 tok/s.