diff --git a/vllm_omni/deploy/voxcpm2.yaml b/vllm_omni/deploy/voxcpm2.yaml
index 4e3f7f3f6..5d366521b 100644
--- a/vllm_omni/deploy/voxcpm2.yaml
+++ b/vllm_omni/deploy/voxcpm2.yaml
@@ -1,5 +1,6 @@
 # VoxCPM2 deploy: single-stage AR pipeline with per-request state batching.
-# Verified on 1x H20 141GB.
+# Runs on large (H20 141GB) and small (24GB L4) cards alike: the KV cache is
+# pinned to a right-sized budget instead of ballooning to fill the card.
 #
 # VoxCPM2-local runtime knobs are passed through hf_overrides.
 async_chunk: false
@@ -9,6 +10,16 @@ stages:
   - stage_id: 0
     max_num_seqs: 32
     gpu_memory_utilization: 0.9
+    # Right-size the KV cache instead of letting it fill gpu_memory_utilization.
+    # max_num_seqs(32) * max_model_len(4096) ~= 131k tokens; 6 GiB holds ~175k
+    # (42x concurrency, above the 32-seq scheduler cap, so no throughput loss).
+    # The remaining VRAM stays free for the VoxCPM2 diffusion side-path
+    # (LocDiT / AudioVAE / code2wav + batch CUDA-graph pools), which allocates
+    # outside vLLM's KV accounting during inference. Keeps peak ~13 GiB on any
+    # card and avoids per-rebase OOM tuning on small CI GPUs.
+    # Assumes prefix caching off and max_num_seqs*max_model_len <= ~175k tokens;
+    # raise this cap if either grows or enable_prefix_caching is turned on.
+    kv_cache_memory_bytes: 6442450944  # 6 GiB
     enforce_eager: true
     trust_remote_code: true
     enable_prefix_caching: false