A production investigation of DeepSeek V4.1 disk-backed Engram on four DGX Sparks: how syscall, page-cache, and component profiling found a Python scheduling bottleneck, how RWF_NOWAIT fixed it, and what the production evidence does and does not prove.
Llm-inference
All Posts
- ebpf (40)
- systems (39)
- notes (21)
- draft (13)
- ai (7)
- tutorial (7)
- gpt (6)
- others (4)
- webassembly (3)
- llm-inference (2)
- linux (2)
- gis (1)
- docker (1)
- claude-code (1)
- containers (1)
- deepseek (1)
- vllm (1)
- storage (1)
- profiling (1)
- 未来技术 (1)
- 性学研究 (1)
- 数字亲密 (1)
- 人工智能 (1)
- 生物技术 (1)
- 虚拟现实 (1)
- 伦理学 (1)
- 人际关系 (1)
- fantasy (1)
- sci-fi (1)
- literature (1)
- society (1)
- gpu (1)
- thunderbolt (1)
- science-fiction (1)
- technology (1)
- future (1)
A complete experiment report on running Qwen 27B with TP2 and native speculative decoding across two RTX 5090 hosts, finding a Thunderbolt transport bottleneck, and validating an opt-in Linux TX-E2E fix.