ollama/convert/convert_llama.go

214 lines
6.1 KiB
Go
Raw Normal View History

2024-06-01 03:00:49 +00:00
package convert
import (
"cmp"
"fmt"
2024-07-29 21:53:02 +00:00
"math"
2024-06-01 03:00:49 +00:00
"strings"
"github.com/pdevine/tensor"
"github.com/pdevine/tensor/native"
2024-08-01 21:52:15 +00:00
"github.com/ollama/ollama/llm"
2024-06-01 03:00:49 +00:00
)
type llama struct {
Parameters
NLayers uint32 `json:"n_layers"`
NumHiddenLayers uint32 `json:"num_hidden_layers"`
NLayer uint32 `json:"n_layer"`
MaxPositionEmbeddings uint32 `json:"max_position_embeddings"`
NCtx uint32 `json:"n_ctx"`
HiddenSize uint32 `json:"hidden_size"`
NEmbd uint32 `json:"n_embd"`
IntermediateSize uint32 `json:"intermediate_size"`
NInner uint32 `json:"n_inner"`
NumAttentionHeads uint32 `json:"num_attention_heads"`
NHead uint32 `json:"n_head"`
NumKeyValueHeads uint32 `json:"num_key_value_heads"`
RopeTheta float32 `json:"rope_theta"`
RopeScaling struct {
2024-07-29 21:53:02 +00:00
Type string `json:"type"`
RopeType string `json:"rope_type"`
Factor float32 `json:"factor"`
LowFrequencyFactor float32 `json:"low_freq_factor"`
HighFrequencyFactor float32 `json:"high_freq_factor"`
OriginalMaxPositionalEmbeddings uint32 `json:"original_max_positional_embeddings"`
factors ropeFactor
2024-06-01 03:00:49 +00:00
} `json:"rope_scaling"`
RMSNormEPS float32 `json:"rms_norm_eps"`
LayerNormEPS float32 `json:"layer_norm_eps"`
LayerNormEpsilon float32 `json:"layer_norm_epsilon"`
NormEpsilon float32 `json:"norm_epsilon"`
HeadDim uint32 `json:"head_dim"`
}
var _ Converter = (*llama)(nil)
func (p *llama) KV(t *Tokenizer) llm.KV {
kv := p.Parameters.KV(t)
kv["general.architecture"] = "llama"
kv["llama.vocab_size"] = p.VocabSize
kv["llama.block_count"] = cmp.Or(p.NLayers, p.NumHiddenLayers, p.NLayer)
if contextLength := cmp.Or(p.MaxPositionEmbeddings, p.NCtx); contextLength > 0 {
kv["llama.context_length"] = contextLength
}
if embeddingLength := cmp.Or(p.HiddenSize, p.NEmbd); embeddingLength > 0 {
kv["llama.embedding_length"] = cmp.Or(p.HiddenSize, p.NEmbd)
}
if feedForwardLength := cmp.Or(p.IntermediateSize, p.NInner); feedForwardLength > 0 {
kv["llama.feed_forward_length"] = cmp.Or(p.IntermediateSize, p.NInner)
}
if headCount := cmp.Or(p.NumAttentionHeads, p.NHead); headCount > 0 {
kv["llama.attention.head_count"] = cmp.Or(p.NumAttentionHeads, p.NHead)
kv["llama.rope.dimension_count"] = p.HiddenSize / headCount
}
if p.RopeTheta > 0 {
kv["llama.rope.freq_base"] = p.RopeTheta
}
if p.RopeScaling.Type == "linear" {
kv["llama.rope.scaling.type"] = p.RopeScaling.Type
kv["llama.rope.scaling.factor"] = p.RopeScaling.Factor
2024-07-29 21:53:02 +00:00
} else if p.RopeScaling.RopeType == "llama3" {
dim := p.HiddenSize / p.NumAttentionHeads
for i := uint32(0); i < dim; i += 2 {
factor := cmp.Or(p.RopeScaling.Factor, 8.0)
factorLow := cmp.Or(p.RopeScaling.LowFrequencyFactor, 1.0)
factorHigh := cmp.Or(p.RopeScaling.HighFrequencyFactor, 4.0)
original := cmp.Or(p.RopeScaling.OriginalMaxPositionalEmbeddings, 8192)
lambdaLow := float32(original) / factorLow
lambdaHigh := float32(original) / factorHigh
lambda := 2 * math.Pi * math.Pow(float64(p.RopeTheta), float64(i)/float64(dim))
if lambda < float64(lambdaHigh) {
p.RopeScaling.factors = append(p.RopeScaling.factors, 1.0)
} else if lambda > float64(lambdaLow) {
p.RopeScaling.factors = append(p.RopeScaling.factors, factor)
} else {
smooth := (float32(original)/float32(lambda) - factorLow) / (factorHigh - factorLow)
p.RopeScaling.factors = append(p.RopeScaling.factors, 1.0/((1-smooth)/factor+smooth))
}
}
2024-06-01 03:00:49 +00:00
}
if p.NumKeyValueHeads > 0 {
kv["llama.attention.head_count_kv"] = p.NumKeyValueHeads
}
if p.RMSNormEPS > 0 {
kv["llama.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
}
if layerNormEpsilon := cmp.Or(p.LayerNormEPS, p.LayerNormEpsilon, p.NormEpsilon); layerNormEpsilon > 0 {
kv["llama.attention.layer_norm_epsilon"] = layerNormEpsilon
}
if p.HeadDim > 0 {
kv["llama.attention.key_length"] = p.HeadDim
kv["llama.attention.value_length"] = p.HeadDim
}
return kv
}
2024-07-08 23:59:48 +00:00
func (p *llama) Tensors(ts []Tensor) []llm.Tensor {
var out []llm.Tensor
2024-07-29 21:53:02 +00:00
if p.RopeScaling.factors != nil {
out = append(out, llm.Tensor{
Name: "rope_freqs.weight",
Kind: 0,
Shape: []uint64{uint64(len(p.RopeScaling.factors))},
WriterTo: p.RopeScaling.factors,
})
}
2024-06-01 03:00:49 +00:00
for _, t := range ts {
2024-06-28 20:27:05 +00:00
if strings.HasSuffix(t.Name(), "attn_q.weight") ||
strings.HasSuffix(t.Name(), "attn_k.weight") {
2024-06-01 03:00:49 +00:00
t.SetRepacker(p.repack)
}
2024-07-08 23:59:48 +00:00
out = append(out, llm.Tensor{
2024-06-28 20:27:05 +00:00
Name: t.Name(),
2024-06-01 03:00:49 +00:00
Kind: t.Kind(),
Shape: t.Shape(),
WriterTo: t,
})
}
return out
}
2024-06-28 20:27:05 +00:00
func (p *llama) Replacements() []string {
return []string{
2024-06-01 03:00:49 +00:00
"lm_head", "output",
"model.embed_tokens", "token_embd",
"model.norm", "output_norm",
"model.layers", "blk",
"input_layernorm", "attn_norm",
"self_attn.q_proj", "attn_q",
"self_attn.k_proj", "attn_k",
"self_attn.v_proj", "attn_v",
"self_attn.o_proj", "attn_output",
"mlp.gate_proj", "ffn_gate",
"mlp.down_proj", "ffn_down",
"mlp.up_proj", "ffn_up",
"post_attention_layernorm", "ffn_norm",
2024-06-28 20:27:05 +00:00
}
2024-06-01 03:00:49 +00:00
}
func (p *llama) repack(name string, data []float32, shape []uint64) ([]float32, error) {
var dims []int
for _, dim := range shape {
dims = append(dims, int(dim))
}
var heads uint32
2024-06-28 20:27:05 +00:00
if strings.HasSuffix(name, "attn_q.weight") {
2024-06-01 03:00:49 +00:00
heads = p.NumAttentionHeads
2024-06-28 20:27:05 +00:00
} else if strings.HasSuffix(name, "attn_k.weight") {
2024-06-01 03:00:49 +00:00
heads = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
} else {
return nil, fmt.Errorf("unknown tensor for repack: %s", name)
}
n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
return nil, err
}
if err := n.T(0, 2, 1, 3); err != nil {
return nil, err
}
if err := n.Reshape(dims...); err != nil {
return nil, err
}
if err := n.Transpose(); err != nil {
return nil, err
}
ts, err := native.SelectF32(n, 1)
if err != nil {
return nil, err
}
var f32s []float32
for _, t := range ts {
f32s = append(f32s, t...)
}
return f32s, nil
}