Initial: Phi-4-mini-reasoning WebGPU
Browse files
README.md
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
base_model: microsoft/Phi-4-mini-reasoning
|
| 4 |
+
tags:
|
| 5 |
+
- phi-4
|
| 6 |
+
- webgpu
|
| 7 |
+
- browser-inference
|
| 8 |
+
- strix-halo
|
| 9 |
+
- amd
|
| 10 |
+
- unified-memory
|
| 11 |
+
- reasoning
|
| 12 |
+
- math
|
| 13 |
+
- chain-of-thought
|
| 14 |
+
pipeline_tag: text-generation
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Phi-4-mini-reasoning on WebGPU
|
| 18 |
+
|
| 19 |
+
First WebGPU package for Microsoft's Phi-4-mini-reasoning model.
|
| 20 |
+
|
| 21 |
+
The reasoning variant (not instruct) trained on DeepSeek-R1 chain-of-thought distillation. 3.8B params, 2.4 GB Q4_K_M. Runs entirely in browser via WebGPU + wllama.
|
| 22 |
+
|
| 23 |
+
## Quick Start
|
| 24 |
+
|
| 25 |
+
1. Download Q4_K_M GGUF from bartowski
|
| 26 |
+
2. Place in model_splits/ (single file, no splitting needed)
|
| 27 |
+
3. node serve.js (port 8190)
|
| 28 |
+
4. Open http://localhost:8190 in Chrome
|
| 29 |
+
|
| 30 |
+
## Hardware
|
| 31 |
+
|
| 32 |
+
Tested on GMKTEC EVO-X2 (AMD Strix Halo). Works on any WebGPU-capable device with 3+ GB available memory.
|
| 33 |
+
|
| 34 |
+
## Credits
|
| 35 |
+
|
| 36 |
+
Built by Joshua (LJTSG) and Claude.
|
| 37 |
+
Co-Authored-By: Claude <noreply@anthropic.com>
|