highperfocused/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-06-21 20:01:02 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix divide by zero

Browse Source
This commit is contained in:
Michael Yang 2025-03-13 15:05:42 -07:00
parent a422ba39c9
commit 65b88c544f
1 changed files with 14 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
fs/ggml/ggml.go
View File

@ -579,6 +579,10 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
} }
func (llm GGML) VisionGraphSize() (weights, graphSize uint64) { func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
if llm.KV().Uint("vision.block_count") == 0 {
return
}
for name, layer := range llm.Tensors().GroupLayers() { for name, layer := range llm.Tensors().GroupLayers() {
if strings.HasPrefix(name, "v.") { if strings.HasPrefix(name, "v.") {
for _, tensor := range layer { for _, tensor := range layer {
@ -589,6 +593,12 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
imageSize := uint64(llm.KV().Uint("vision.image_size")) imageSize := uint64(llm.KV().Uint("vision.image_size"))
patchSize := uint64(llm.KV().Uint("vision.patch_size")) patchSize := uint64(llm.KV().Uint("vision.patch_size"))
if patchSize == 0 {
slog.Warn("unknown patch size for vision model")
return
}
numChannels := uint64(llm.KV().Uint("vision.num_channels"))
numPatches := (imageSize / patchSize) * (imageSize / patchSize) numPatches := (imageSize / patchSize) * (imageSize / patchSize)
if _, ok := llm.Tensors().GroupLayers()["v"]["class_embd"]; ok { if _, ok := llm.Tensors().GroupLayers()["v"]["class_embd"]; ok {
@ -596,15 +606,13 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
} }
headCount := uint64(llm.KV().Uint("vision.attention.head_count")) headCount := uint64(llm.KV().Uint("vision.attention.head_count"))
embeddingLength := uint64(llm.KV().Uint("vision.embedding_length"))
switch llm.KV().Architecture() { switch llm.KV().Architecture() {
case "mllama": case "mllama":
numPaddedPatches := numPatches + 8 - (numPatches%8)%8 numPaddedPatches := numPatches + 8 - (numPatches%8)%8
maxNumTiles := uint64(llm.KV().Uint("vision.max_num_tiles")) maxNumTiles := uint64(llm.KV().Uint("vision.max_num_tiles"))
numChannels := uint64(llm.KV().Uint("vision.num_channels"))
embeddingLength := uint64(llm.KV().Uint("vision.embedding_length"))
graphSize = 4 * (8 + graphSize = 4 * (8 +
imageSize*imageSize*numChannels*maxNumTiles + imageSize*imageSize*numChannels*maxNumTiles +
@ -612,7 +620,9 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
9*embeddingLength*numPaddedPatches*maxNumTiles + 9*embeddingLength*numPaddedPatches*maxNumTiles +
numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount) numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
case "gemma3": case "gemma3":
graphSize = 4 * (numPatches * numPatches * headCount) graphSize = 4 * (imageSize*imageSize*numChannels +
embeddingLength*patchSize +
numPatches*numPatches*headCount)
} }
return weights, graphSize return weights, graphSize