jeans/ollama
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update convert

Browse Source
This commit is contained in:
Bruce MacDonald 2025-04-21 13:58:26 -07:00
parent 3fe090f447
commit 963531215e
3 changed files with 17 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
convert/convert.go
View File

@ -190,7 +190,7 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
case "Qwen2ForCausalLM":
conv = &qwen2Model{}
case "Qwen2_5_VLForConditionalGeneration":
conv = &qwen25vlModel{}
conv = &qwen25VLModel{}
case "BertModel":
conv = &bertModel{}
case "CohereForCausalLM":

42
convert/convert_qwen25vl.go
View File

@ -13,7 +13,7 @@ import (
"github.com/x448/float16"
)
type qwen25vlModel struct {
type qwen25VLModel struct {
ModelParameters
HiddenSize uint32 `json:"hidden_size"`
IntermediateSize uint32 `json:"intermediate_size"`
@ -25,18 +25,12 @@ type qwen25vlModel struct {
RMSNormEPS float32 `json:"rms_norm_eps"`
VisionModel struct {
PatchSize uint32 `json:"patch_size"`
//HeadDim uint32 `json:"num_heads"`
//RopeTheta float32 `json:"rope_theta"`
HiddenSize uint32 `json:"hidden_size"`
IntermediateSize uint32 `json:"intermediate_size"`
WindowSize uint32 `json:"window_size"`
} `json:"vision_config"`
}
var _ ModelConverter = (*qwen25vlModel)(nil)
var _ ModelConverter = (*qwen25VLModel)(nil)
func (q *qwen25vlModel) KV(t *Tokenizer) ggml.KV {
func (q *qwen25VLModel) KV(t *Tokenizer) ggml.KV {
kv := q.ModelParameters.KV(t)
kv["general.architecture"] = "qwen25vl"
kv["qwen25vl.block_count"] = q.HiddenLayers
@ -48,24 +42,18 @@ func (q *qwen25vlModel) KV(t *Tokenizer) ggml.KV {
kv["qwen25vl.rope.freq_base"] = q.RopeTheta
kv["qwen25vl.attention.layer_norm_rms_epsilon"] = q.RMSNormEPS
kv["qwen25vl.vision.embedding_length"] = q.VisionModel.HiddenSize
return kv
}
func (q *qwen25vlModel) Tensors(ts []Tensor) []ggml.Tensor {
func (q *qwen25VLModel) Tensors(ts []Tensor) []ggml.Tensor {
var out []ggml.Tensor
for _, t := range ts {
if strings.HasSuffix(t.Name(), "patch_embed.proj.weight") {
// var buf bytes.Buffer
// if _, err := t.WriteTo(&buf); err != nil {
// panic(err)
// }
// newTensors := splitPatchEmbed(buf, t.Kind(), t.Shape())
// out = append(out, newTensors...)
// } else if strings.HasPrefix(t.Name(), "v.blk.") {
// skip
var buf bytes.Buffer
t.WriteTo(&buf)
newTensors := splitPatchEmbed(buf, t.Kind(), t.Shape())
out = append(out, newTensors...)
} else {
out = append(out, ggml.Tensor{
Name: t.Name(),
@ -79,7 +67,7 @@ func (q *qwen25vlModel) Tensors(ts []Tensor) []ggml.Tensor {
return out
}
func (p *qwen25vlModel) Replacements() []string {
func (p *qwen25VLModel) Replacements() []string {
return []string{
"lm_head", "output",
"model.embed_tokens", "token_embd",
@ -151,10 +139,10 @@ func splitPatchEmbed(buf bytes.Buffer, kind uint32, shape []uint64) []ggml.Tenso
slog.Debug("first vals", "val 1", ts[0][0], "val 2", ts[0][1], "val 3", ts[0][2])
var f16s patchEmbed
for _, row := range ts {
for _, col := range row {
f16s = append(f16s, float16.Fromfloat32(col).Bits())
f16s := make(patchEmbed, shapeToSize(shape))
for r, row := range ts {
for c, col := range row {
f16s[r+c] = float16.Fromfloat32(col).Bits()
}
}
@ -163,7 +151,7 @@ func splitPatchEmbed(buf bytes.Buffer, kind uint32, shape []uint64) []ggml.Tenso
p := getDataFromSlice(f32s, intShape, []tensor.Slice{nil, nil, tensor.S(0, 1, 1), nil, nil})
newTensors = append(newTensors, ggml.Tensor{
Name: "v.patch_embed.0.weight",
Name: "v.patch_embed_0.weight",
Kind: kind,
Shape: append(shape[:2], shape[3:]...),
WriterTo: p,
@ -171,7 +159,7 @@ func splitPatchEmbed(buf bytes.Buffer, kind uint32, shape []uint64) []ggml.Tenso
p = getDataFromSlice(f32s, intShape, []tensor.Slice{nil, nil, tensor.S(1, 2, 1), nil, nil})
newTensors = append(newTensors, ggml.Tensor{
Name: "v.patch_embed.1.weight",
Name: "v.patch_embed_1.weight",
Kind: kind,
Shape: append(shape[:2], shape[3:]...),
WriterTo: p,

2
model/models/qwen25vl/model_vision.go
View File

@ -113,7 +113,7 @@ type VisionModelOptions struct {
}
type PatchEmbedding struct {
PatchConv0 *nn.Conv2D `gguf:"patch_embd_0"`
PatchConv0 *nn.Conv2D `gguf:"patch_embd_0"` // TODO: `gguf:"patch_embed_0"`
PatchConv1 *nn.Conv2D `gguf:"patch_embd_1"`
}