Report issues regarding the architecture-agnostic branch.

[win10ogod]

When performing passthrough, the model will be severely shrunk. For example, the original model bf16 7GB will have 1G left.

[win10ogod]

merge_method: passthrough
slices:
- sources:
  - layer_range: [0, 3]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [2, 5]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [4, 7]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [6, 9]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [8, 11]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [10, 13]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [12, 15]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [14, 17]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [16, 19]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [18, 21]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [20, 23]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [22, 25]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [24, 27]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- sources:
  - layer_range: [26, 28]
    model: deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct

[cg123]

Thanks for giving it a try. The architecture-agnostic branch is work in progress at the moment and it's expected that this sort of merge in particular won't work yet.

Merges that don't use layer slicing or the tokenizer functionality should all work though.

[win10ogod]

Thanks for giving it a try. The architecture-agnostic branch is work in progress at the moment and it's expected that this sort of merge in particular won't work yet.

Merges that don't use layer slicing or the tokenizer functionality should all work though.

Looking forward to a better version.