Upgrade to vllm 0.17.0 corex v4.1 overlay

vllm/model_executor/layers/mamba/ops/causal_conv1d.py

@@ -1159,7 +1159,7 @@ def causal_conv1d_update(
                 f"ERROR: conv_state_indices should have shape ({batch},*) but got {conv_state_indices.shape}"
             )
 
-        # assert num_cache_lines >= batch
+        assert num_cache_lines >= batch
         assert weight.stride(1) == 1  # Need this
 
     # adopt the strategy in vLLM that overwrite on 'x' directly, rather than creating a new tensor 'o'

vllm/model_executor/layers/mamba/ops/mamba_ssm.py

@@ -497,6 +497,8 @@ def selective_scan_fn(
     block_idx_first_scheduled_token=None,
     block_idx_last_scheduled_token=None,
     initial_state_idx=None,
+    cu_chunk_seqlen=None,
+    last_chunk_indices=None,
 ) -> torch.Tensor:
     """
     u: (dim, total_length) for varlen or (batch, dim, seqlen)
@@ -588,6 +590,8 @@ def selective_scan_fn(
         block_idx_first_scheduled_token,
         block_idx_last_scheduled_token,
         initial_state_idx,
+        cu_chunk_seqlen,
+        last_chunk_indices,
     )
 
     if z is None: