Vikram Rajpurohit (@imvikramraj) header
Vikram Rajpurohit (@imvikramraj) profile picture

Tech Enthusiast. Idea Guy using AI to bring them in reality https://xitter.cf/t.co/l06dC5AcxF

मुंबई, भारत
Joined December 2010
Time is linearly cyclical or Cyclically Linear.
1
2
#Gemini created this deck of card design
12
Can we have concept of interchangeable equipment with heroes who are worthy of it, inspired by Captain America Thor Hammer style? @ClashofClans
9
Penned my thoughts about the Next Gen Social Media imvikramraj.substack.com/p/n…
9
More we are seeing AI slop taking over content consumption more we will value human made content. Youtube will have a switch in few years it will say do you want to disable AI slop.
12
Project Idea: An website which scan all the e-commerce platforms for an specific search product and then lists the price, offers, delivery timeline and other details for better shopping experience.
1
14
I am trying to find resources which can help make AI coding better with resources like Code review, Graph Mapping for code base and other such worthy GitHub projects. Will be sharing such projects coming days.
17
Day 2 of Figuring out how to Code using AI, still understanding how to prompt LLM with specific technical depth for code as expectation.
1
1
16
Project Ideas that should need AI accelerated assistance for app development for Linux Open Source : Codex or Claude Work like IDE with Coding Harness which is Open Source. (Opencode doesn't have all the bells & whistle)
1
1
50
Vikram Rajpurohit retweeted
Day 3/30 of Inference Engineering - read about why CUDA Graphs are especially useful for LLM inference, particularly during the decode phase, where repeated small GPU operations can become limited by CPU launch overhead. - looked into how CUDA Graphs work with KV cache and PagedAttention, and how inference systems deal with CUDA Graph constraints like fixed memory addresses while the KV cache keeps changing. - also read on how vLLM and SGLang use CUDA Graphs internally, including capturing graphs for different batch sizes, reusing static buffers, padding to captured sizes, and falling back to eager execution when needed and more - LLaMA-2 inference benchmarks and profiler timelines to understand where the actual speedup comes from, and how reducing CPU dispatch overhead keeps the GPU busier during generation. I've also updated my site and added a video section for all my videos, so now u can watch on X, youtube, my website Part 1 was the basics & intro of cuda graphs, now Part 2 of my CUDA Graph series will focus on CUDA Graphs specifically in LLM inference and how they’re used in real inference systems. link for both the reads + my website below
Part 1: CUDA Graphs - Intro & Basics - how normal CPU ←→ GPU execution works - what a kernel launch is and where CPU launch overhead comes from - why launching many small kernels individually can become a bottleneck - what CUDA Graphs are and how they reduce CPU launch overhead - common benefits of using CUDA Graphs - the two ways to create CUDA Graphs: a. Explicit Graph Construction b. Stream Capture - where these graph creation methods are used - the CUDA Graph lifecycle: capture → instantiate/prepare → replay this is the first ever content that ive recorded and taught, its genuinely so much harder to record than i thought lmao, feedback's are appreciated and hopefully it just gets better from here youtube link shared below
8
15
1
191
14,064
Day 1 of Learning how to code with AI : 1) Start by Creating a Implementation Road Map 2) Then Delegate Task to Cheaper Model with More Specific Technical Details for more precise implementation 3) Using Skills like Code Review to understand Bugs in Code. Will go in more details in coming days about how to prompt & learning how to specify for precise execution.
19
The Evolution of how the AI coding prompt have evolved & trajectory : First came the Idea of using loops like Ralph Mode Now people are talking of using Graph based Loop for Coding session. Next, people will be using Graph with loop along with LLM as verifying. Future would be Human Feedback with Verifying Checklist with Graph loop. Lets see how the evolution unfolds. #AIprompting
30
I have been following AI trends for a quite a while now, haven't come across any one who can dissect & tell how to write reproducible code using AI which shows how to build complex software using AI assisted coding. One shot prompt App & Websites are fun to watch but they can't be deployed on production environment.
15
Vikram Rajpurohit retweeted
We're close to hitting 10K on X. Should celebrate. Giving away 3 Pro subs STEPS 1. Follow @CommandCodeAI 2. Repost this post 3. Reply with what you're building with Command Code We'll pick randomly soon as we hit 10K. LFG!
175
170
10
272
19,902
Local LLM have just started heating up the game AI with one device learning and Agentic Operation. With Research like QTP and many more techniques to come they will only improve for better
1
46
Self Evolving Agents with Personal local LLM needs still has a very large scope of improvement in terms of Security Contextual Awareness and learning
22
Self Evolving LLM over enterprise generated data will cater to needs better solving real value problems
8
Teams with evolving framework for non deterministic nature of LLM for code changes effecting on large codebase will build reliable and scalable solutions
9