Commit Graph

58 Commits

Author SHA1 Message Date
aee72e65ec Improved gemma training 2026-01-11 16:25:40 +01:00
4e50d1f7c4 Use Vulkan 2026-01-10 19:52:41 +01:00
5821391309 Fix continue from prefix 2026-01-10 13:50:57 +01:00
39e0c81ef1 Better stop token handling 2026-01-10 13:41:32 +01:00
b89cb9cfe1 Switch to Vulkan 2026-01-07 19:32:50 +01:00
ab349993ba Larger context and WiP fix to prefix handling 2025-05-22 20:24:29 +00:00
c22917e3c9 Fixes for 12B 2025-05-22 18:51:07 +00:00
Niels Geens
19dd9fa2ee Fixes for saving the gemma tokenizer 2025-05-22 18:22:15 +02:00
Niels Geens
f2c70cd05d Gemma training script 2025-05-20 20:46:16 +02:00
d4a4902b94 New web interface, move llm engine to separate process 2025-05-20 09:43:17 +02:00
f8365ef698 Fixes from runpod 2025-05-06 20:35:30 +02:00
c41a86a00a Gemma inference with llama.cpp and logits processor 2025-05-06 20:31:42 +02:00
f61c9ab15f Add inference in notebook 2025-05-02 19:32:03 +00:00
f1f1a42fb3 Fixes after local testing 2025-05-02 18:28:07 +02:00
e78fca8233 Gemma notebook working on runpod 2025-05-02 17:05:28 +02:00
Niels Geens
ff6a6ab135 Add bitsandbytes to gemma notebook 2025-05-01 10:46:33 +02:00
Niels Geens
95546208e8 Gemma notebook 2025-04-30 13:36:24 +02:00
8544f12d41 Added example notebook 2025-04-29 14:58:27 +00:00
36d718d326 Test train qwen 3 2025-04-29 10:17:03 +00:00
e2e346d134 Implement 4-bit quant for train script and inference 2025-04-23 16:43:55 +00:00
41c336165f Updated notebooks to support 4-bit quantization 2025-04-23 16:37:10 +00:00
d7c340a937 Updated notebooks wip 2025-04-23 10:32:33 +00:00
3e471eced3 Use Unsloth for QwQ inference 2025-04-21 10:42:35 +00:00
64fc621d44 Fix error print in dataset 2025-04-18 12:42:24 +00:00
5e47200ece Fixes for QwQ llm engine 2025-04-18 12:30:47 +00:00
Niels Geens
faf78b7948 Lock vllm version to avoid bug 2025-04-07 12:17:56 +02:00
0e2cca2e4f Fixed QwQ tokenizer 2025-04-04 08:54:06 +00:00
156d3aa7d3 Bootstrapping QwQ works!! 2025-04-02 16:37:17 +00:00
Niels Geens
e573b9ebe7 Removed flash-attn as dependency because it's hard to install properly 2025-04-02 14:42:35 +02:00
130afbdd92 Updates after testing 2025-04-01 11:49:21 +00:00
Niels Geens
6f3d414d17 Converted QwQ notebooks to .py files 2025-04-01 10:42:54 +02:00
642c5181bb Fixed inference notebook 2025-03-30 08:52:17 +00:00
290c718187 Added qwq inference notebook 2025-03-30 09:26:18 +02:00
30eb5c419d Updated QwQ train notebook 2025-03-29 11:03:53 +00:00
56242496bb Add support for jupyter notebooks for training 2025-03-28 16:10:43 +00:00
8975a54d58 WIP QwQ train 2025-03-24 15:46:54 +00:00
1ef32ed33e WIP manual implementation of QwQ finetune 2025-03-24 12:05:43 +00:00
3594c8150a WIP Simplify training 2025-03-24 10:53:51 +00:00
9a171ab622 Fixes for QwQ training 2025-03-20 09:41:02 +01:00
Niels Geens
3ea5ba87a9 Fix venv add to path 2025-03-18 09:46:16 +01:00
Niels Geens
2e66020f8e Replaced deepseek with qwq 2025-03-14 11:22:30 +01:00
Niels Geens
3ea3239a9b wip deepseek train 2025-03-03 16:57:04 +01:00
b64f8d7d33 wip deepseek r1 2025-03-02 22:01:24 +01:00
b7e95d7398 Added restart script and fixed dockerfile and run to match docs 2025-03-01 13:29:17 +01:00
Niels Geens
4fcb32e6a1 Added tool for finetuning mistral model based on training config 2025-01-24 11:05:23 +01:00
3058820668 Merge branch 'master' of https://git.nielsgeens.be/llm/SIA 2025-01-16 14:37:10 +01:00
fffdefcd92 More updates to itb (wip) 2025-01-16 14:36:37 +01:00
Niels Geens
6669b7a682 Fixed timestamp timezone for entry id's 2025-01-02 16:52:43 +01:00
e0bc7a7159 Update to itb_screenshot and friends 2024-12-27 17:07:19 +01:00
fe72969b4f Cleanup of itb readme 2024-12-26 18:19:27 +01:00