matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-11 16:26:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

add mixtral 8x7b model conversion (#3859)

Browse Source
This commit is contained in:
Patrick Devine
2024-04-23 20:17:04 -07:00
committed by GitHub
parent 4dc4f1be34
commit ce8ce82567
3 changed files with 138 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
convert/safetensors.go
View File

@@ -93,7 +93,6 @@ func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params)
}
slices.Sort(keys)
slog.Info("converting layers")
var tensors []llm.Tensor
@@ -105,7 +104,6 @@ func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params)
return nil, 0, err
}
slog.Debug(fmt.Sprintf("metadata = %#v", data))
var size uint64
var kind uint32
switch len(data.Shape) {
@@ -150,11 +148,13 @@ func (m *SafetensorFormat) readTensors(fn string, offset uint64, params *Params)
padding: 8 + jsonSize,
}
tensors = append(tensors, t)
offset += size
tensors = append(tensors, t)
}
slog.Debug(fmt.Sprintf("total tensors for file = %d", len(tensors)))
slog.Debug(fmt.Sprintf("offset = %d", offset))
return tensors, offset, nil
}
@@ -185,15 +185,19 @@ func (m *SafetensorFormat) GetLayerName(n string) (string, error) {
}
tMap := map[string]string{
"model.layers.(\\d+).input_layernorm.weight": "blk.$1.attn_norm.weight",
"model.layers.(\\d+).mlp.down_proj.weight": "blk.$1.ffn_down.weight",
"model.layers.(\\d+).mlp.gate_proj.weight": "blk.$1.ffn_gate.weight",
"model.layers.(\\d+).mlp.up_proj.weight": "blk.$1.ffn_up.weight",
"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
"model.layers.(\\d+).self_attn.k_proj.weight": "blk.$1.attn_k.weight",
"model.layers.(\\d+).self_attn.o_proj.weight": "blk.$1.attn_output.weight",
"model.layers.(\\d+).self_attn.q_proj.weight": "blk.$1.attn_q.weight",
"model.layers.(\\d+).self_attn.v_proj.weight": "blk.$1.attn_v.weight",
"model.layers.(\\d+).input_layernorm.weight": "blk.$1.attn_norm.weight",
"model.layers.(\\d+).mlp.down_proj.weight": "blk.$1.ffn_down.weight",
"model.layers.(\\d+).mlp.gate_proj.weight": "blk.$1.ffn_gate.weight",
"model.layers.(\\d+).mlp.up_proj.weight": "blk.$1.ffn_up.weight",
"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
"model.layers.(\\d+).self_attn.k_proj.weight": "blk.$1.attn_k.weight",
"model.layers.(\\d+).self_attn.o_proj.weight": "blk.$1.attn_output.weight",
"model.layers.(\\d+).self_attn.q_proj.weight": "blk.$1.attn_q.weight",
"model.layers.(\\d+).self_attn.v_proj.weight": "blk.$1.attn_v.weight",
"model.layers.(\\d+).block_sparse_moe.gate.weight": "blk.$1.ffn_gate_inp.weight",
"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w1.weight": "blk.$1.ffn_gate.$2.weight",
"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w2.weight": "blk.$1.ffn_down.$2.weight",
"model.layers.(\\d+).block_sparse_moe.experts.(\\d+).w3.weight": "blk.$1.ffn_up.$2.weight",
}
v, ok := directMap[n]
@@ -286,6 +290,15 @@ func (m *SafetensorFormat) GetModelArch(name, dirPath string, params *Params) (M
Format: m,
},
}, nil
case "MixtralForCausalLM":
return &MixtralModel{
ModelData{
Name: name,
Path: dirPath,
Params: params,
Format: m,
},
}, nil
case "GemmaForCausalLM":
return &GemmaModel{
ModelData{