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

Revert "ropeTheta should be 1e5"

Browse Source 
This reverts commit cc1638b26763eae7daddd44e3975a885671ef9d3.

This reverts commit
b32385591307e2d33a8f43ce1626b529d2dac83e.
This commit is contained in:
Bruce MacDonald 2025-05-02 15:42:35 -07:00
parent 75441c56f3
commit 16b13e0cfc
2 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
convert/convert_qwen25vl.go
View File

@ -48,7 +48,7 @@ func (q *qwen25VLModel) KV(t *Tokenizer) ggml.KV {
kv["qwen25vl.vision.spatial_patch_size"] = q.VisionModel.SpatialPatchSize
kv["qwen25vl.vision.window_size"] = q.VisionModel.WindowSize
kv["qwen25vl.vision.attention.layer_norm_epsilon"] = cmp.Or(q.VisionModel.RMSNormEps, 1e-6)
kv["qwen25vl.vision.rope.freq_base"] = cmp.Or(q.VisionModel.RopeTheta, 1e5) // note: other implementations use 1e4, but we have seen better results with 1e5
kv["qwen25vl.vision.rope.freq_base"] = cmp.Or(q.VisionModel.RopeTheta, 1e4)
kv["qwen25vl.vision.fullatt_block_indexes"] = q.VisionModel.FullAttentionBlocks
kv["qwen25vl.vision.temporal_patch_size"] = q.VisionModel.TemporalPatchSize

4
model/models/qwen25vl/model_vision.go
View File

@ -360,7 +360,7 @@ func newVisionModel(c fs.Config) *VisionModel {
numHeads := int(c.Uint("vision.attention.head_count", 16))
numChannels := int(c.Uint("vision.num_channels", 3))
eps := c.Float("vision.attention.layer_norm_epsilon", 1e-6)
ropeTheta := c.Float("vision.rope.freq_base", 100000.0)
ropeTheta := c.Float("vision.rope.freq_base", 10000.0)
spatialMergeSize := int(c.Uint("vision.spatial_merge_size", 2))
windowSize := int(c.Uint("vision.window_size", 112))
fullAttnBlocks := c.Ints("qwen25vl.vision.fullatt_block_indexes", []int32{7, 15, 23, 31})
@ -383,7 +383,7 @@ func newVisionModel(c fs.Config) *VisionModel {
}
for i := range fullAttnBlocks {
// full attention block indexes have to be converted to int for use with the slices package
// full attention block indexes have to be converted to int for use with the slices package
model.fullAttnBlocks = append(model.fullAttnBlocks, int(fullAttnBlocks[i]))
}