docs: add proper YAML metadata, description, and tags for HF Space card
Browse files
README.md
CHANGED
|
@@ -1,14 +1,56 @@
|
|
| 1 |
---
|
| 2 |
-
title:
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
sdk: gradio
|
| 7 |
sdk_version: 6.14.0
|
| 8 |
python_version: '3.13'
|
| 9 |
app_file: app.py
|
| 10 |
pinned: false
|
| 11 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
---
|
| 13 |
|
| 14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
title: REPOMIND
|
| 3 |
+
emoji: 🧠
|
| 4 |
+
colorFrom: indigo
|
| 5 |
+
colorTo: red
|
| 6 |
sdk: gradio
|
| 7 |
sdk_version: 6.14.0
|
| 8 |
python_version: '3.13'
|
| 9 |
app_file: app.py
|
| 10 |
pinned: false
|
| 11 |
license: mit
|
| 12 |
+
short_description: Cursor for self-hosters — 256K context on AMD MI300X
|
| 13 |
+
tags:
|
| 14 |
+
- amd-developer-hackathon
|
| 15 |
+
- agents
|
| 16 |
+
- coding-agent
|
| 17 |
+
- long-context
|
| 18 |
+
- rocm
|
| 19 |
+
- mi300x
|
| 20 |
+
- qwen3-coder
|
| 21 |
+
- vllm
|
| 22 |
---
|
| 23 |
|
| 24 |
+
# REPOMIND
|
| 25 |
+
|
| 26 |
+
> Open-source Cursor for self-hosters. Ingest an entire git repo (256K tokens, FP8) and reason across it on a single AMD MI300X — what NVIDIA H100 80GB physically cannot do.
|
| 27 |
+
|
| 28 |
+
**Built for the [AMD Developer Hackathon 2026](https://lablab.ai/ai-hackathons/amd-developer)** · MIT License · [GitHub source](https://github.com/SRKRZ23/repomind)
|
| 29 |
+
|
| 30 |
+
## Why MI300X?
|
| 31 |
+
|
| 32 |
+
- Qwen3-Coder-Next-FP8 weights ≈ 80 GB
|
| 33 |
+
- 256K KV cache @ FP8 ≈ 38 GB
|
| 34 |
+
- activations ≈ 25 GB → **~143 GB total on a single GPU**
|
| 35 |
+
- NVIDIA H100 80GB physically OOMs. AMD MI300X 192GB just runs it.
|
| 36 |
+
|
| 37 |
+
This is a memory-architecture story, not a CUDA-vs-ROCm one.
|
| 38 |
+
|
| 39 |
+
## Stack
|
| 40 |
+
|
| 41 |
+
- **Model**: `Qwen/Qwen3-Coder-Next-FP8` — 80B params, 3B active (MoE)
|
| 42 |
+
- **Inference**: vLLM ROCm 7 with `qwen3_coder` tool-call parser
|
| 43 |
+
- **Agent loop**: SC-TIR style (PLAN → CALL TOOL → OBSERVE → THINK → ANSWER)
|
| 44 |
+
- **Tools**: `read_file` · `grep_codebase` · `execute_code` (sandboxed) · `run_tests` · `git_log`
|
| 45 |
+
|
| 46 |
+
## Status
|
| 47 |
+
|
| 48 |
+
This Space runs on CPU-basic with the **mock LLM backend** for testing the agent loop without GPU credits. The `vllm` backend wires up automatically once the AMD MI300X endpoint comes online (AMD Cloud credits incoming).
|
| 49 |
+
|
| 50 |
+
If the MI300X memory-architecture pitch resonates, **a like on this Space helps us with the Hugging Face Special Prize judging** 🤗
|
| 51 |
+
|
| 52 |
+
## Author
|
| 53 |
+
|
| 54 |
+
[Sardor Razikov](https://lablab.ai/u/@Sardor_R) — independent ML engineer · Tashkent 🇺🇿
|
| 55 |
+
- Kaggle SPR 2026 #7/371 (Top 1.9%) · S6E3 #23/4,142 · AIMO3 39/50 (XTX $2.2M)
|
| 56 |
+
- [Epistemic Curie Benchmark](https://doi.org/10.5281/zenodo.19791329)
|