1
How are you gonna be a revolutionary if you’re such a traditionalist?

This is a line from my favorite movie, La La Land. I hope I can try more bravely this semester and step outside the usual patterns.

26 Spring Record

0

I feel that this record still has a reason to exist, so I will write it.

1

  • It is predictable that OS this semester will again be a mess. I wanted to choose track 1 and hand-roll an OS kernel combining page attention and radix attention. That sounds interesting, but is it really meaningful? If a hand-rolled kernel is better, why do vLLM and SGLang exist? I do not understand and I do not know enough. Track 2 is indeed simpler, and rCore can also teach some OS fundamentals, which feels fine. I am confused. Update: I ended up choosing track 2. Better not create extra trouble for myself.
  • This semester I started running and working out, and it seems quite effective. I had dinner with a junior on the weekend. Next weekend I will go watch F1. I will probably return to Nanjing during Qingming. The shell around my phone camera broke, and the Apple Store in Shanghai is really far away. I also feel like I need to spend a weekend in Nanjing to handle a visa.

2

  • Going to the Shanghai race should count as my one-year anniversary of getting into F1. I was so stupid. I thought I probably would not need my camera before Qingming, so I left the Z30 at home. Thanks to my bestie’s Z6 + 24-200 lens and post-processing.

Why am I so handsome? Ordinary-confident girl energy.
  • I signed up for driving school. The earliest estimate is that I can get my license around May Day. I went full power in the question-practice area and cleared subject one in one day. My memory really is unbeatable.
  • Recently I vibed an interesting project with @Parsifal1986. After we optimize and polish it better, we will open-source it. Stay tuned. PS: this project is limited by us being broke and unable to buy APIs. The compute resources of a certain model are scarce beyond imagination.
  • I am now an honorable Gemini Pro user.

3

  • I passed all codegen tests. Debugging with reimu is quite fast. Next: register allocation. I plan to try vibing it, but that can wait. I will not touch compiler again within the next 10 days.
  • I started lab rotation. There was not much to hesitate about. I added the advisor on WeChat and scheduled a talk for next week.
  • Recently I read some papers and talked with several people. Later I will also schedule meals with a few people. I have always felt that systems work never ends, whether it is systems for LLM, systems for diffusion, or even possible future systems for world models. As long as ML exists, systems will be needed to support it, and many optimization methods can carry over, though adaptive innovation is also needed. Why not be the person selling shovels?
  • I fell into xzjt anxiety. I hope the last attempt can produce a satisfactory result.

4

  • Qingming and the university anniversary holiday basically exhausted me. I worked intensely on xzjt while also having fun. Sichuan Airlines’ 773 was really average. There was also a Jiaxing elderly tour group on the same flight. Hard to judge. During summer break I may go to Beijing again with Guimi, with museums and street photography as the main quests. But thinking more carefully, summer break seems even busier. Anyway, we will see.
  • For xzjt, a dean-level leader sat in and questioned us. It was intense. Four student forums have ended, and there seem to be many regrets. I was never fully satisfied with the PPT until the end. Speaking without a script was hard, and there were still moments when I was incoherent. Compared with some students who did very well, the gap was large. Although I rehearsed three times, I still felt underprepared. The road is still long.
  • I should write a little about this. Recently I have been repeatedly pressured by the parent of a next-cohort student. I do not understand why some parents, after their child has entered university, still care so much about exam scores and rankings. Caring about semester grades to avoid academic warning is one thing, but what does caring about machine-test scores even mean? It feels like high school all over again: not allowing the child to get distracted in class, let alone skip class, or play games after class, and creating inexplicable anxiety for a long time. Sigh. The old saying goes that family shame should not be aired in public. But is it really right to belittle a younger student, who is also their own child, in front of a senior? Even coming to find me was done secretly, like guerrilla warfare. Many times, children’s thinking is much better than their parents’, but parents not only fail to realize this, they also love being self-righteous, as if they never went to university and were never socialized.
  • Sister Yuki brought me to a meal with people from ZhenFund, and I unexpectedly developed some interest in investment and quantitative finance. I had always thought I would never be interested in investment or finance in my life. Naturally, I scheduled more chats. Communicating with excellent peers from other universities and outstanding investors and founders in industry is very valuable to me.

5

  • I will face a rather busy period around May Day. There are still many classes to catch up on. This semester, as usual, I rely entirely on self-study and do not attend many classes. It is worth mentioning that the OS class in the neighboring section is decent and taught me some things. This semester feels unusually fast. I have done nothing, and yet more than half has already passed. The biggest problem this semester is unprecedented financial pressure. Copilot’s series of money-grabbing actions caused me some losses, and I had no choice but to switch to relay services and buy Opus quota at a high price. My domain expires next month and will cost more to renew. It is safe to say that my current balance may barely last until the next living allowance. Paying out of pocket to go to school, truly.
  • I went to Hangzhou on the weekend, revisited West Lake, and attended a concert. Although I had not heard most of the songs, some of them accompanied me through the hardest period of the pandemic. It felt like a pure prayer to the sky. Every time I listen, it feels like rain falling inside my heart. I was still very moved.
  • I had a successful coffee chat. I feel coffee chats are more efficient than meals and cost much less. They will probably become a common self-improvement method for me in the near future.

6

  • It has been a long time since I wrote anything. Unfortunately, the attempt to commercialize the relay service is almost dead, but acmapi was apparently imitated by SAI next door. I mean, acmapi is much more watered down than joypiggy.
  • I completed the ML course project. Because of dataset limitations, the model training result did not reach the original paper’s baseline, with a small gap, but there were indeed some findings. I realized how difficult it is to do systems research alone, and not having SPEC2017 access created a huge obstacle for my experiments. At first, when I saw the PF-LLM paper, I thought AI4Sys was extremely interesting, so in a rush of excitement I searched for many similar works. My sudden obsession with AI4Sys was more about my doubts over applying LLMs to traditional systems optimization. Taking prefetchers as an example, existing sandbox prefetching already feels like a very good strategy. In fact, in experiments LLMs often went all in on sandbox. Meanwhile, with some old-school ML methods, such as naive logistic regression or DNN, I achieved the same results as Qwen2.5 on some benchmarks. Personally, I think this is not very worth it. But anyway, it is indeed an interesting idea, and aiming for ASPLOS BP feels reasonable. Recently I have been curious about ML Compiler, and I also chose it as the topic for Advanced Compiler. It feels like another thing I will become obsessed with.
  • The day after tomorrow I will send Parsifal to the airport. For the details, P chose 5.21. See you in seven months, or maybe in three months, since I am considering going to Hawaii for vacation during summer break. I also want to experience the long-awaited ANA Flying Honu, and I have always wanted to visit the Hawaii Botanical Garden and Kualoa Ranch, since I am a die-hard Jurassic Park fan and love it too much.
  • I discovered that I have quite a talent for karting. Since my first day at driving school, I felt a natural familiarity with cars. When driving, it is easy for me to enter a flow state. I have seriously decided to visit every karting track in Shanghai before going abroad.

7

  • News: I join SJTU EPCC Lab and work with Prof. Jingwen Leng and Dr. Xinhao Luo.

8

  • As of today, 2026.6.21, Parsifal has been in the United States for one month. A 15-hour time difference seems to have passed quite easily, and the progress bar has reached 1/7. Also as of today, I have worked in the lab for a full month. The good news is that I believe I am doing something correct and meaningful. I can say for sure that my condition has never been better, both academically and in research. Although this month was hard, it feels like the month with the highest meaningful-output power since I entered university. Balancing coursework and research is of course a challenge, but I think I did it well. Once I truly started research, many worries about the future seemed to fade, replaced by steadiness and love for being immersed in the work. I choose to take it one step at a time.
  • Finals week is unprecedentedly easy. I only need to take Maoism and Algorithms, which really means only Algorithms. The only remaining course project is OS, which I will leave until after the algorithms exam. I think when I have time later, I will write a blog post about my Advanced Compiler course project. Honestly, I did not do an innovative research project. I merely migrated an optimization for an ML compiler from the original paper’s Triton implementation to TileLang. For me, the more important part was learning TileLang and ML compiler itself. As for RCompiler, my code is stuck somewhere between O0 and O1/O2. It seems to have reached a serious bottleneck. It may be due to the test cases themselves, no idea. I have already added more than a dozen optimizations and truly have no more tricks.
  • Recently I really like drinking Rio Light, and I also like going to Lawson at night to buy Baileys and coconut water to make a simple drink. Slightly tipsy!
  • I decided that once OS CR ends, I will immediately return to Nanjing to drive karts. I cannot endure it anymore. If, as a child, I had chosen karting instead of studying, I would probably be an excellent professional racing driver now.