From 75441c56f3b31df4a47b60220e815f76e4f4a55b Mon Sep 17 00:00:00 2001 From: Bruce MacDonald Date: Fri, 2 May 2025 15:16:22 -0700 Subject: [PATCH] add comment explaining rope theta --- convert/convert_qwen25vl.go | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/convert/convert_qwen25vl.go b/convert/convert_qwen25vl.go index c71a096aa..12d061e55 100644 --- a/convert/convert_qwen25vl.go +++ b/convert/convert_qwen25vl.go @@ -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