fix mllama inference without image (#2947)

* fix

* add hf token in ut

* debug

* fix

.github/workflows/unit-test.yml

@@ -39,6 +39,7 @@ jobs:
       options: "--gpus=all --ipc=host --user root -e PIP_CACHE_DIR=/root/.cache/pip -e CUDA_VISIBLE_DEVICES=2,3 --pull never"
       volumes:
         - /nvme/share_data/github-actions/pip-cache:/root/.cache/pip
+        - /nvme/share_data/github-actions/hf_home:/root/.cache/huggingface
         - /nvme/share_data/github-actions/packages:/root/packages
         - /usr/share/zoneinfo/Asia/Shanghai:/etc/localtime:ro
     steps:
@@ -90,6 +91,7 @@ jobs:
           echo "TODO"
       - name: Test lmdeploy python UT
         run: |
+          huggingface-cli login --token ${{ secrets.HF_TOKEN }}
           coverage run --branch --source lmdeploy -m pytest -rsE tests
           coverage xml
           coverage report -m

lmdeploy/pytorch/models/mllama.py

@@ -1289,8 +1289,11 @@ def prepare_inputs_for_generation(
         attn_metadata = context.attn_metadata
         cross_attn_metadata = context.cross_attn_metadata
 
-        if int(cross_attn_metadata.kv_seqlens.sum()) == 0:
+        # cross_attn_metadata is None when inputs without image
+        if cross_attn_metadata is not None and int(
+                cross_attn_metadata.kv_seqlens.sum()) == 0:
             cross_attn_metadata.kv_seqlens = None
+
         device = input_ids.device
 
         # process image input