diff --git a/server/text_generation_server/models/custom_modeling/qwen2_vl.py b/server/text_generation_server/models/custom_modeling/qwen2_vl.py index 3553b3f9..7c07a3c3 100644 --- a/server/text_generation_server/models/custom_modeling/qwen2_vl.py +++ b/server/text_generation_server/models/custom_modeling/qwen2_vl.py @@ -516,7 +516,7 @@ class Qwen2VLForConditionalGeneration(nn.Module): prefill_cache_indices=prefill_cache_indices, ) hidden_states, _ = self.norm(hidden_states) - print("lm_head_indices", lm_head_indices) - logits, speculative_logits = self.lm_head(hidden_states) - # import ipdb; ipdb.set_trace() - return logits, speculative_logits + if lm_head_indices is not None: + hidden_states = hidden_states[lm_head_indices] + logits = self.lm_head(hidden_states) + return logits, None