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

add comment explaining rope theta

Browse Source
This commit is contained in:
Bruce MacDonald 2025-05-02 15:16:22 -07:00
parent 45f96e898d
commit 75441c56f3
1 changed files with 1 additions and 1 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)
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.fullatt_block_indexes"] = q.VisionModel.FullAttentionBlocks
kv["qwen25vl.vision.temporal_patch_size"] = q.VisionModel.TemporalPatchSize