mirror of
https://github.com/ollama/ollama.git
synced 2025-03-18 05:41:43 +01:00
58245413f4
feat: add new Ollama engine using ggml through cgo This change introduces a new way to run pretrained models. It introduces 3 high level interfaces and a bunch of smaller helper interfaces to facilitate this. - `model.Model` defines the interface for a model architecture. Models such as `llama` and `mllama`, which are provided as examples, can implement the model's forward propagation in the `Forward` method. This method will be called to generate completions. This interface can be found in `model/model.go` - `ml.Backend` defines the interface for a backend tensor library, in this case `ggml`. Among other things, a Backend is responsible for loading a pretrained model into hardware (GPU, CPU, etc) and providing an interface for Models to access loaded tensors. This interface can be found in `ml/backend.go` - `ml.Tensor` defines the interface for a tensor and tensor operations This is the first implementation of the new engine. Follow up PRs will implement more features: - non-greedy sampling (#8410) - integration with Ollama and KV caching (#8301) - more model support (#9080) with more coming soon Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
77 lines
2.3 KiB
Go
77 lines
2.3 KiB
Go
package convert
|
|
|
|
import (
|
|
"cmp"
|
|
|
|
"github.com/ollama/ollama/fs/ggml"
|
|
)
|
|
|
|
type commandrModel struct {
|
|
ModelParameters
|
|
MaxPositionEmbeddings uint32 `json:"max_position_embeddings"`
|
|
HiddenSize uint32 `json:"hidden_size"`
|
|
HiddenLayers uint32 `json:"num_hidden_layers"`
|
|
IntermediateSize uint32 `json:"intermediate_size"`
|
|
NumAttentionHeads uint32 `json:"num_attention_heads"`
|
|
NumKeyValueHeads uint32 `json:"num_key_value_heads"`
|
|
LayerNormEPS float32 `json:"layer_norm_eps"`
|
|
RopeTheta float32 `json:"rope_theta"`
|
|
UseQKNorm bool `json:"use_qk_norm"`
|
|
MaxLength uint32 `json:"model_max_length"`
|
|
LogitScale float32 `json:"logit_scale"`
|
|
NCtx uint32 `json:"n_ctx"`
|
|
}
|
|
|
|
var _ ModelConverter = (*commandrModel)(nil)
|
|
|
|
func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
|
|
kv := p.ModelParameters.KV(t)
|
|
kv["general.architecture"] = "command-r"
|
|
kv["general.name"] = "command-r"
|
|
kv["command-r.context_length"] = cmp.Or(p.MaxLength, p.MaxPositionEmbeddings, p.NCtx)
|
|
kv["command-r.embedding_length"] = p.HiddenSize
|
|
kv["command-r.block_count"] = p.HiddenLayers
|
|
kv["command-r.feed_forward_length"] = p.IntermediateSize
|
|
kv["command-r.attention.head_count"] = p.NumAttentionHeads
|
|
kv["command-r.attention.head_count_kv"] = p.NumKeyValueHeads
|
|
kv["command-r.attention.layer_norm_epsilon"] = p.LayerNormEPS
|
|
kv["command-r.rope.freq_base"] = p.RopeTheta
|
|
kv["command-r.max_position_embeddings"] = cmp.Or(p.MaxLength, p.MaxPositionEmbeddings)
|
|
kv["command-r.logit_scale"] = p.LogitScale
|
|
kv["command-r.rope.scaling.type"] = "none"
|
|
|
|
return kv
|
|
}
|
|
|
|
func (p *commandrModel) Tensors(ts []Tensor) []ggml.Tensor {
|
|
var out []ggml.Tensor
|
|
for _, t := range ts {
|
|
out = append(out, ggml.Tensor{
|
|
Name: t.Name(),
|
|
Kind: t.Kind(),
|
|
Shape: t.Shape(),
|
|
WriterTo: t,
|
|
})
|
|
}
|
|
|
|
return out
|
|
}
|
|
|
|
func (p *commandrModel) Replacements() []string {
|
|
return []string{
|
|
"self_attn.q_norm", "attn_q_norm",
|
|
"self_attn.k_norm", "attn_k_norm",
|
|
"model.layers", "blk",
|
|
"input_layernorm", "attn_norm",
|
|
"mlp.down_proj", "ffn_down",
|
|
"mlp.gate_proj", "ffn_gate",
|
|
"mlp.up_proj", "ffn_up",
|
|
"self_attn.k_proj", "attn_k",
|
|
"self_attn.o_proj", "attn_output",
|
|
"self_attn.q_proj", "attn_q",
|
|
"self_attn.v_proj", "attn_v",
|
|
"model.norm", "output_norm",
|
|
"model.embed_tokens", "token_embd",
|
|
}
|
|
}
|