diff --git a/modal-deploy/README.md b/modal-deploy/README.md
index 1213d7dd..767e613b 100644
--- a/modal-deploy/README.md
+++ b/modal-deploy/README.md
@@ -34,8 +34,9 @@ modal setup
 
 6. test the deployed model with the client script, for example:
 ```
+cd Sunflower32b-Ultravox
 python client.py \
     --app-name Sunflower32b-Ultravox  \
     --prompt "Translate to English: " \
-    --audio_file "../sunflower-ultravox-vllm/audios/context_eng_1.wav"
+    --audio_file "../../sunflower-ultravox-vllm/audios/context_eng_1.wav"
 ```
\ No newline at end of file
diff --git a/modal-deploy/spark-tts-salt/vllm_inference.py b/modal-deploy/spark-tts-salt/vllm_inference.py
index 17b28d17..c6fe806a 100644
--- a/modal-deploy/spark-tts-salt/vllm_inference.py
+++ b/modal-deploy/spark-tts-salt/vllm_inference.py
@@ -60,8 +60,11 @@
 
 # cache model weights with Modal Volumes
 hf_cache_vol = modal.Volume.from_name("huggingface-cache", create_if_missing=True)
-HF_CACHE_DIR = "/root/.cache/huggingface"
+# cache some of vLLM's compilation artifacts in a Modal Volume.
+vllm_cache_vol = modal.Volume.from_name("vllm-cache", create_if_missing=True)
 
+HF_CACHE_DIR = "/root/.cache/huggingface"
+VLLM_CACHE_DIR = "/root/.cache/vllm"
 
 # ## The TTS model class
 
@@ -78,6 +81,7 @@
     secrets=[modal.Secret.from_name("huggingface-read")],
     volumes={
         HF_CACHE_DIR: hf_cache_vol,
+        VLLM_CACHE_DIR: vllm_cache_vol,
     }
 )
 @modal.concurrent(max_inputs=10)