[Bugfix] Correct adapter usage for cohere and jamba (#8292)

vllm/model_executor/models/commandr.py

@@ -47,6 +47,8 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.sequence import IntermediateTensors
 
+from .interfaces import SupportsLoRA
+
 
 @torch.compile
 def layer_norm_func(hidden_states, weight, variance_epsilon):
@@ -292,8 +294,7 @@ class CohereModel(nn.Module):
         return hidden_states
 
 
-class CohereForCausalLM(nn.Module):
-
+class CohereForCausalLM(nn.Module, SupportsLoRA):
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",

vllm/model_executor/models/jamba.py

@@ -38,6 +38,8 @@ from vllm.sequence import IntermediateTensors
 from vllm.worker.model_runner import (_BATCH_SIZES_TO_CAPTURE,
                                       _get_graph_batch_size)
 
+from .interfaces import SupportsLoRA
+
 KVCache = Tuple[torch.Tensor, torch.Tensor]
 
 
@@ -539,7 +541,7 @@ class JambaModel(nn.Module):
         return hidden_states
 
 
-class JambaForCausalLM(nn.Module, HasInnerState):
+class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA):
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",