Fix qwen vl (#3096)

* Fixing qwen2.5 VL.

* Fixing the CI.

Cargo.lock

@@ -4754,7 +4754,7 @@ dependencies = [
  "async-trait",
  "bindgen 0.71.1",
  "clap 4.5.30",
- "hf-hub 0.3.2",
+ "hf-hub 0.4.2",
  "num_cpus",
  "pkg-config",
  "text-generation-router",

server/text_generation_server/models/custom_modeling/qwen2_5_vl.py

@@ -313,7 +313,6 @@ class Qwen2_5_VLVisionConfig(PretrainedConfig):
 
 
 class Qwen2_5_VLConfig(PretrainedConfig):
-
     def __init__(
         self,
         vocab_size=152064,
@@ -702,7 +701,6 @@ class Qwen2_5VisionModel(nn.Module):
         pixel_values: torch.Tensor,
         grid_thw: Optional[torch.LongTensor] = None,
     ) -> torch.Tensor:
-
         # reshape the input tensor for processing
         shape = (
             -1,
@@ -967,6 +965,7 @@ class Qwen2_5VLForConditionalGeneration(nn.Module):
             max_s=max_s,
             true_max_s=max_s,
             prefill_cache_indices=prefill_cache_indices,
+            adapter_data=adapter_data,
         )
         if lm_head_indices is not None:
             hidden_states = hidden_states[lm_head_indices]