Let reward model take text inputs instead of message lists (#1907)

Co-authored-by: Kyle Corbitt <kyle@corbt.com>
2024-11-03 13:27:12 -08:00
parent 793b79dbe9
commit 2ce32db6fb
12 changed files with 43 additions and 58 deletions
--- a/python/sglang/srt/models/qwen2_vl.py
+++ b/python/sglang/srt/models/qwen2_vl.py
@@ -618,7 +618,7 @@ class Qwen2VLForConditionalGeneration(nn.Module):
            extend_start_loc_cpu = forward_batch.extend_start_loc.cpu().numpy()
            prefix_lens_cpu = forward_batch.extend_prefix_lens.cpu().numpy()
            for i, image in enumerate(forward_batch.image_inputs):
-                if image == None:
+                if image is None:
                    continue
                start_idx = extend_start_loc_cpu[i]
                prefix_len = prefix_lens_cpu[i]