H-Simpson123

bug: Memory usage is increased with each request

10

### Describe the bug I'm trying to create embeddings for some documents with langchain and openllm. With each request the GPU RAM consumption increases by some hundred MB until OpenLLM...

MoE finetuning extreme slow

3

The finetuning of Qwen2-57B-A14B-Instruct is extremely slow compared to finetuning of Qwen2-72B-Instruct. Here are the runtimes: **Qwen/Qwen2-7B-Instruct:** {'train_runtime': 100.8509, 'train_samples_per_second': 5.652, 'train_steps_per_second': 0.099, 'train_loss': 0.751581035554409, 'epoch': 10.0} **Qwen/Qwen2-72B-Instruct:** {'train_runtime': 483.8572,...