vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

convert: update comment of MOE tensors mapping

Browse source
This commit is contained in:
Pierrick HYMBERT 2024-04-12 22:15:11 +02:00
parent 03bdc36e8b
commit 8e6758f2f4
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
convert-hf-to-gguf.py
View file

@ -1470,9 +1470,9 @@ class DbrxModel(Model):
# But llama.cpp moe graph works differently
# AND the dimensions in ggml are typically in the reverse order of the pytorch dimensions
# so (n_expert, n_ff, n_embd) in pytorch is {n_embd, n_ff, n_expert} in ggml_tensor
exp_tensor_names = {"ffn.experts.mlp.v1": None, # LLM_TENSOR_FFN_GATE_EXPS ggml_tensor->ne{n_embd, n_ff, n_expert}
exp_tensor_names = {"ffn.experts.mlp.w1": None, # LLM_TENSOR_FFN_GATE_EXPS ggml_tensor->ne{n_embd, n_ff, n_expert}
"ffn.experts.mlp.w2": (0, 2, 1), # LLM_TENSOR_FFN_DOWN_EXPS ggml_tensor->ne{n_ff, n_embd, n_expert}
"ffn.experts.mlp.w1": None} # LLM_TENSOR_FFN_UP_EXPS ggml_tensor->ne{n_embd, n_ff, n_expert}
"ffn.experts.mlp.v1": None} # LLM_TENSOR_FFN_UP_EXPS ggml_tensor->ne{n_embd, n_ff, n_expert}
experts = False
for exp_tensor_name in exp_tensor_names.keys():
if name.find(exp_tensor_name) != -1 and name.find(".weight") == -1: