MJ

Optimist

[MJ @manjitpokhrel_]

I ran a 70B-scale LLM forward pass on a $300 consumer GPU (RTX 5060, 8GB VRAM).
Measured kernel speedup: 69.83% over dense matmul. Measured dead neurons: 27.3% permanently inactive. Measured perplexity delta at 42% sparsity: +5.91%.

[MJ @manjitpokhrel_]

AI safety guardrails fail in Nepali. Same model, same harmful intent — 0% bypass in English, 73.7% in Devanagari Nepali.
I documented 3 failure modes + a novel attack vector (Vajra Morphing) across Qwen, Gemma, Llama, Gemini.

Paper: https://t.co/mEcTfUeBjS

[MJ @manjitpokhrel_]

13 mistakes over 8 hours. Most of them are undocumented for this hardware.

The model produces fluent Nepali. Factual accuracy needs more training — 200 steps is just the start.
Hardware problem: solved. Training problem: next.
Full README + code:
https://t.co/PjJzv4HjLj

[MJ @manjitpokhrel_]

Wall 3: Two undocumented Windows bugs — cp1252 encoding crash + multiprocessing failure. Fix: lock trl version + dataloader_num_workers=0

[MJ @manjitpokhrel_]

Wall 3: Two undocumented Windows bugs — cp1252 encoding crash + multiprocessing failure. Fix: lock trl version + dataloader_num_workers=0

[MJ @manjitpokhrel_]

Wall 1: https://t.co/ITTaNa0Vpr_available() = False — RTX 5060 is sm_120, stable PyTorch stops at sm_90. Fix: PyTorch nightly cu128
Wall 2: Unsloth has a GPU whitelist. Stops at sm_90. Killed it completely. Fix: dropped Unsloth, used raw HuggingFace + PEFT

[MJ @manjitpokhrel_]

Wall 1: https://t.co/ITTaNa0Vpr_available() = False — RTX 5060 is sm_120, stable PyTorch stops at sm_90. Fix: PyTorch nightly cu128
Wall 2: Unsloth has a GPU whitelist. Stops at sm_90. Killed it completely. Fix: dropped Unsloth, used raw HuggingFace + PEFT

[MJ @manjitpokhrel_]

I spent 8 hours fine-tuning a model on a GPU the ML stack couldn't even see.
RTX 5060. Blackwell. sm_120. At home. No cloud.

[MJ @manjitpokhrel_]

I spent 8 hours fine-tuning a model on a GPU the ML stack couldn't even see.
RTX 5060. Blackwell. sm_120. At home. No cloud.

[MJ @manjitpokhrel_]

I built a GPT from scratch using only NumPy. No PyTorch. No abstractions.
211K parameters.
Multi-head attention, back propagation, Adam Optimizer.

Loss: 4.21 → 1.64 in 5000 steps.

Full code: https://t.co/0wgruBonUZ

[MJ @manjitpokhrel_]

I built a tool that turns an idea into a full architecture.

Input your concept and get the following:
- Architecture blueprint
- Tech stack
- MVP roadmap
- Risk assessment

Early feedback welcome: https://t.co/V4ww8m1uNo

[MJ @manjitpokhrel_]

Day 100 of #100DaysOfCode

Journey completed! From Python basics to advanced modules, DSA, and problem solving .
This is just the beginning, excited to keep building and learning every day.

#Python #DSA

[MJ @manjitpokhrel_]

Day 99 of #100DaysOfCode

Explored Python’s itertools today : a powerful module for working with iterators. Functions like permutations, combinations, and product make solving algorithmic problems much easier.

#Python #DSA

[MJ @manjitpokhrel_]

Day 98 of #100DaysOfCode

Explored Python’s functools today.
lru_cache is a game changer for optimizing recursive functions by caching results.

#Python #DSA

[MJ @manjitpokhrel_]

Day 97 of #100DaysOfCode

Learned about Diameter of a Binary Tree.
It’s the length of the longest path between any two nodes in the tree (can pass through or avoid the root).

#Python #DSA

[MJ @manjitpokhrel_]

Day 96 of #100DaysOfCode

Explored the depth of a node in a Binary Tree.
Depth = number of edges from the root to that node.
Each step is making the tree concepts clearer.

#Python #DSA

[MJ @manjitpokhrel_]

Day 95 of #100DaysOfCode

Learned about the height of a node in a Binary Tree.
Height = number of edges on the longest path

#Python #DSA

[MJ @manjitpokhrel_]

Day 94 of #100DaysOfCode

Learned about the depth of a node in a Binary Tree.
Depth = number of edges from the root to that node.

#Python #DSA

[MJ @manjitpokhrel_]

Day 93 of #100DaysOfCode

Explored how to calculate the height of a Binary Tree.
The height is the longest path from root to a leaf.

#DSA #Python

[MJ @manjitpokhrel_]

Day 92 of #100DaysOfCode

Learned Level Order Traversal in Binary Trees using BFS.
Visited nodes level by level with the help of a queue.

#Python #DSA

[MJ @manjitpokhrel_]

Day 91 of #100DaysOfCode
Practiced tree traversals in Bi
nary Search Trees.
Inorder → sorted order
Preorder → root-first structure
Postorder → cleanup order

#Python #DSA

[MJ @manjitpokhrel_]

Day 90 of #100DaysOfCode

Explored deletion in Binary Search Trees. The tricky part: handling 3 cases: deleting a leaf, a node with one child, or a node with two children (replace with inorder successor).

#Python #DSA

[MJ @manjitpokhrel_]

Day 89 of #100DaysOfCode

Learned how to search in a Binary Search Tree. Instead of scanning all nodes, BST narrows down the search by moving left or right, making lookups much faster.

#Python #DSA