fix: detect vision patch size for unsplit (HF-format) Qwen3-VL (#1811)

This commit is contained in:
fszontagh
2026-07-23 16:22:21 +02:00
committed by GitHub
parent 8a51eb9284
commit 5114672c48

View File

@@ -200,7 +200,11 @@ namespace LLM {
config.vision.in_channels = tensor_storage.ne[2];
config.vision.hidden_size = tensor_storage.ne[3];
}
if (contains(name, "visual.patch_embed.bias")) {
// HF-format checkpoints keep the patch embed unsplit under a single name.
if (contains(name, "visual.patch_embed.proj.weight")) {
config.vision.patch_size = static_cast<int>(tensor_storage.ne[0]);
}
if (contains(name, "visual.patch_embed.bias") || contains(name, "visual.patch_embed.proj.bias")) {
config.vision.hidden_size = tensor_storage.ne[0];
}
if (contains(name, "visual.pos_embed.weight")) {