root
/
ollama
mirror of https://github.com/ollama/ollama.git
1
0
Fork 0
Code Issues Actions Packages Projects Releases Wiki Activity

ggml: update qwen25vl vision size estimate (#10711)

This commit is contained in:
Bruce MacDonald 2025-05-14 16:42:30 -07:00 committed by GitHub
parent ff80718e9c
commit bd68d3ae50
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
fs/ggml/ggml.go
View File

@ -6,7 +6,6 @@ import (
"fmt" "fmt"
"io" "io"
"log/slog" "log/slog"
"math"
"slices" "slices"
"strings" "strings"
@ -653,24 +652,15 @@ func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
numPatches*numPatches*headCount) numPatches*numPatches*headCount)
case "qwen25vl": case "qwen25vl":
maxPixels := uint64(llm.KV().Uint("vision.max_pixels", 28*28*1280)) maxPixels := uint64(llm.KV().Uint("vision.max_pixels", 28*28*1280))
mergeSize := uint64(llm.KV().Uint("vision.spatial_merge_size", 2))
temporalPatchSize := uint64(2)
// Calculate max possible patches based on max_pixels numPatches := maxPixels / (patchSize * patchSize)
maxHeight := uint64(math.Sqrt(float64(maxPixels)))
maxWidth := maxPixels / maxHeight
maxGridHeight := maxHeight / patchSize
maxGridWidth := maxWidth / patchSize
// Account for merged patches (2x2 grid)
numPatches := (maxGridHeight * maxGridWidth) / (mergeSize * mergeSize)
// Calculate graph size based on typical operations in ProcessImage and createPatches
graphSize = 4 * (maxPixels*numChannels + // Original image storage graphSize = 4 * (maxPixels*numChannels + // Original image storage
// Normalized pixels // Normalized pixels
maxPixels*numChannels + maxPixels*numChannels +
// Patches storage (numPatches * channels * temporalPatchSize * patchSize^2) // Patches storage (numPatches * channels * patchSize^2)
numPatches*numChannels*temporalPatchSize*patchSize*patchSize + numPatches*numChannels*patchSize*patchSize +
// Self-attention calculations (similar to other architectures) // Self-attention calculations
numPatches*numPatches*headCount + numPatches*numPatches*headCount +
// Additional buffer for processing // Additional buffer for processing
embeddingLength*numPatches) embeddingLength*numPatches)