Huang Xin

I'm getting this error after pulling the Docker image with CUDA 11.8 and installation of vLLM: from vllm import LLM, SamplingParams Traceback (most recent call last): File "", line 1,...

> ```shell > pip install vllm > ``` Thanks, it resolved my issue

Simply add following code after allocation of optimizer in `optimizers.py` support the gradient accumulation: ``` if config.accumulate_gradient_steps > 1: optimizer = optax.MultiSteps(optimizer, config.accumulate_gradient_steps) ```

Hi @A9isha , I found two bugs in your conversion code, and I have fixed it and validated the weights converted from maxtext version of llama3-8b with the HF one....

> If anyone has any ideas, I can try it out on a RTX 5090. > > Apparently: > > * I guess you need https://github.com/vllm-project/vllm/pull/12702/files > * Adapt the...

I have added conversion script for gemma-3 and examined converted parameter weights and it matches with official huggingface gemma-3 weight.

I got the exact same problem with my RTX 5090 setup

You can check my commit here for the conversion script [here](https://github.com/AI-Hypercomputer/maxtext/blob/f6ebc1662cb944bd7748fb350bba164b13479b68/MaxText/gemma2_orbax_to_hf.py)

> [@hxssgaa](https://github.com/hxssgaa) I made a quick test of the script, trying to convert a 2B Gemma2 model. However, I am seeing this error: `ValueError: Requested shape: (2048,) is not compatible...

@peregilk, no need to save the ckpt locally, you can just point the maxtext_checkpoint to the google bucket checkpoint location. Sorry tor the confusion here, I think I have changed...