highperfocused/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-04-08 11:58:07 +02:00
Code Issues Packages Projects Releases Wiki Activity

count gemma3 vision tensors

Browse Source
This commit is contained in:
Michael Yang 2025-03-12 14:18:06 -07:00
parent 6b45b1d6b4
commit 033cec232a
2 changed files with 10 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
fs/ggml/ggml.go
View File

@ -611,6 +611,14 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
embeddingLength*numPatches*maxNumTiles +
9*embeddingLength*numPaddedPatches*maxNumTiles +
numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
case "gemma3":
for name, layer := range llm.Tensors().GroupLayers() {
if strings.HasPrefix(name, "v.") {
for _, tensor := range layer {
weights += tensor.Size()
}
}
}
}
return weights, graphSize
}

4
llm/memory.go
View File

@ -218,8 +218,8 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
layerSize = blk.Size()
layerSize += kv / f.KV().BlockCount()
memoryWeights += blk.Size()
}
memoryWeights += layerSize
if opts.NumGPU >= 0 && layerCount >= opts.NumGPU {
// Stop allocating on GPU(s) once we hit the users target NumGPU
@ -376,7 +376,7 @@ func (m MemoryEstimate) LogValue() slog.Value {
// memory of the weights
"total", format.HumanBytes2(m.memoryWeights),
// memory of repeating layers
"repeating", format.HumanBytes2(m.memoryWeights-m.memoryLayerOutput),
"repeating", format.HumanBytes2(m.memoryWeights),
// memory of non-repeating layers
"nonrepeating", format.HumanBytes2(m.memoryLayerOutput),
),