藍+85CD

I fixed this by patching code (https://github.com/moeru-ai/Moeru-Llama-3-8B/commit/c65e3b1deef11f29d4a9c65b4f4cef9f11c8dc9d#diff-b135d17426f077f767e0ec29114d24b182dcaa3f6dadaee03d8ff424adcdff0bR407), the problem now is that it will Segfault in the `Starting trainer` phase: ```bash /usr/local/lib/python3.11/dist-packages/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: ''If...

> Can you provide oneAPI version, GPU driver version and pip list? We will try to reproduce this issue. I installed `mesa 24.0.5`, `intel-compute-runtime 24.09.28717.12` and `level-zero 1.16.14` on host,...

> ```shell > export ONEAPI_DEVICE_SELECTOR=level_zero:1 > ``` Hmm... Something went wrong, I tried to run `sycl-ls` inside a container and it only showed cpu. ```bash [opencl:acc:0] Intel(R) FPGA Emulation Platform...

> Hmm... Something went wrong, I tried to run `sycl-ls` inside a container and it only showed cpu. ~~This is probably a NixOS bug, after I rolled back the system,...

I think it might be useful to provide an `accelerate/default_config.yaml` reference file to avoid misconfiguration. Also, I fixed this (https://github.com/intel-analytics/ipex-llm/issues/10821#issuecomment-2071442815) by setting environments (https://github.com/intel/compute-runtime/issues/710#issuecomment-1972684586), but at the moment the Trainer...

It looks like I need to download `llama-2-7b` and try the example `lora.yml` to confirm that the current version works.

I tried `unsloth/llama-2-7b` and it was consistent with the previous behavior. ```bash [2024-04-24 17:20:25,931] [DEBUG] [axolotl.train.log:60] [PID:53] [RANK:0] loading tokenizer... /workspace/models/llama-2-7b [2024-04-24 17:20:25,986] [DEBUG] [axolotl.load_tokenizer:216] [PID:53] [RANK:0] EOS: 2 /...

Oh, sorry I missed the message before. I'm using [unsloth/llama-2-7b](https://huggingface.co/unsloth/llama-2-7b) Then I now suspect it may have something to do with the container running in the background. If I run...

Update: I fixed the iGPU display issue with `i915.enable_psr=1`, but it wasn't running Trainer in the foreground either. I also tried running `xpu-smi` inside the container, but it doesn't seem...

~~I'll probably implement block actor/acct in v0.1.2 and block instance after v0.2.0.~~ v0.3.0 The block_instance should accept a string and do an ends_with match. ``` block_instance example.com https://example.com (blocked) https://foo.example.com...