Hugging Face
Models
Datasets
Spaces
Buckets
new
Docs
Enterprise
Pricing
Website
Tasks
HuggingChat
Collections
Languages
Organizations
Community
Blog
Posts
Daily Papers
Hardware
Learn
Discord
Forum
GitHub
Solutions
Team & Enterprise
Hugging Face PRO
Enterprise Support
Inference Providers
Inference Endpoints
Storage Buckets
Log In
Sign Up
๐ค
Open to Collab
Harshit Kumar Gupta
PRO
harshitkgupta
1
3
Follow
hepsie50's profile picture
nitya-gupta's profile picture
joao-silva-1's profile picture
4 followers
ยท
16 following
https://www.harshitgupta.info/
harshit_k_gupta
harshitkgupta
harshitgupta92
harshitkgupta.bsky.social
AI & ML interests
None yet
Recent Activity
reacted
to
their
post
with ๐
about 21 hours ago
Fine-tuned Qwen 2.5 (0.5B โ 3B) on real coding-agent traces, 10 controlled runs, one 16GB Mac. Compared PyTorch MPS vs. Apple MLX for local LoRA SFT โ and the honest answer is "it depends on what you're optimizing for": โข PyTorch MPS: 2.2xโ5.7x faster raw throughput, but hits a hard memory wall โ can't load a 3B model in FP16 on 16GB. โข Apple MLX: 4-bit QLoRA fits 3B+ models with almost flat memory scaling as context grows (+109 MB going from 1kโ4k tokens). โข 4-bit quantization doesn't cost you convergence โ eval loss tracks closely across backends. โข The bigger surprise: most of MLX's slowdown isn't the 4-bit dequant tax. Two of the 10 runs went unquantized to isolate it โ dequant only explains 1.07xโ1.4x of the gap. A ~4.1โ4.6x framework-level gap remains either way. All 10 LoRA adapters + Trackio logs are public so the numbers are checkable, not just claimed. Full writeup: https://huggingface.co/blog/harshitkgupta/fine-tuning-coding-agents-on-mac-pytorch-mps-mlx
posted
an
update
about 21 hours ago
Fine-tuned Qwen 2.5 (0.5B โ 3B) on real coding-agent traces, 10 controlled runs, one 16GB Mac. Compared PyTorch MPS vs. Apple MLX for local LoRA SFT โ and the honest answer is "it depends on what you're optimizing for": โข PyTorch MPS: 2.2xโ5.7x faster raw throughput, but hits a hard memory wall โ can't load a 3B model in FP16 on 16GB. โข Apple MLX: 4-bit QLoRA fits 3B+ models with almost flat memory scaling as context grows (+109 MB going from 1kโ4k tokens). โข 4-bit quantization doesn't cost you convergence โ eval loss tracks closely across backends. โข The bigger surprise: most of MLX's slowdown isn't the 4-bit dequant tax. Two of the 10 runs went unquantized to isolate it โ dequant only explains 1.07xโ1.4x of the gap. A ~4.1โ4.6x framework-level gap remains either way. All 10 LoRA adapters + Trackio logs are public so the numbers are checkable, not just claimed. Full writeup: https://huggingface.co/blog/harshitkgupta/fine-tuning-coding-agents-on-mac-pytorch-mps-mlx
upvoted
an
article
about 22 hours ago
Fine-Tuning Coding Agents on a 16GB Mac: PyTorch MPS vs. Apple MLX
View all activity
Organizations
harshitkgupta
's activity
All
Models
Datasets
Spaces
Buckets
Papers
Collections
Community
Posts
Upvotes
Likes
Articles
liked
a Space
about 23 hours ago
Running
Agents
1
Training Agents Trackio
๐ฏ
1
Show visualized I/O tracking data
liked
2 datasets
6 months ago
microsoft/Taskbench
Viewer
โข
Updated
Aug 21, 2024
โข
17.3k
โข
1.27k
โข
38
Maurus/ToolBench
Viewer
โข
Updated
Aug 17, 2023
โข
88.9k
โข
1.56k
โข
10