matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-10 15:57:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #9703 from ollama/mxyng/gemma3-memory

Browse Source 
count gemma3 vision tensors
This commit is contained in:
Michael Yang
2025-03-13 16:56:34 -07:00
committed by GitHub
parent 74b44fdf8f 8d76fa23ef
commit 4ea4d2b189
2 changed files with 39 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
fs/ggml/ggml.go
View File

@@ -583,39 +583,52 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
}
func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
if llm.KV().Uint("vision.block_count") == 0 {
return
}
for name, layer := range llm.Tensors().GroupLayers() {
if name == "v" || strings.HasPrefix(name, "v.") {
for _, tensor := range layer {
weights += tensor.Size()
}
}
}
imageSize := uint64(llm.KV().Uint("vision.image_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)
if _, ok := llm.Tensors().GroupLayers()["v"]["class_embd"]; ok {
numPatches++
}
headCount := uint64(llm.KV().Uint("vision.attention.head_count"))
embeddingLength := uint64(llm.KV().Uint("vision.embedding_length"))
switch llm.KV().Architecture() {
case "mllama":
for _, layer := range llm.Tensors().GroupLayers()["v"] {
weights += layer.Size()
}
kv := func(n string) uint64 {
if v, ok := llm.KV()["mllama.vision."+n].(uint32); ok {
return uint64(v)
}
return 0
}
imageSize := kv("image_size")
maxNumTiles := kv("max_num_tiles")
embeddingLength := kv("embedding_length")
headCount := kv("attention.head_count")
numPatches := (imageSize / kv("patch_size")) * (imageSize / kv("patch_size"))
if _, ok := llm.Tensors().GroupLayers()["v"]["class_embd"]; ok {
numPatches++
}
numPaddedPatches := numPatches + 8 - (numPatches%8)%8
maxNumTiles := uint64(llm.KV().Uint("vision.max_num_tiles"))
graphSize = 4 * (8 +
imageSize*imageSize*kv("num_channels")*maxNumTiles +
imageSize*imageSize*numChannels*maxNumTiles +
embeddingLength*numPatches*maxNumTiles +
9*embeddingLength*numPaddedPatches*maxNumTiles +
numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
case "gemma3":
graphSize = 4 * (imageSize*imageSize*numChannels +
embeddingLength*patchSize +
numPatches*numPatches*headCount)
}
return weights, graphSize
}