[Bugfix][Attention] Make the query contiguous for TRTLLM ragged MLA prefill (#61634)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

vllm/v1/attention/backends/mla/prefill/trtllm_ragged.py

@@ -156,7 +156,8 @@ class TrtllmRaggedPrefillBackend(MLAPrefillBackend):
             return out
 
         ret = trtllm_ragged_attention_deepseek(
-            query=q,
+            # The kernel reads the query as packed, ignoring its strides
+            query=q.contiguous(),
             key=k,
             value=v,
             workspace_buffer=self._workspace_buffer,
@@ -207,7 +208,8 @@ class TrtllmRaggedPrefillBackend(MLAPrefillBackend):
             )
 
         attn_out, lse = trtllm_ragged_attention_deepseek(
-            query=q,
+            # The kernel reads the query as packed, ignoring its strides
+            query=q.contiguous(),
             key=k,
             value=v,
             workspace_buffer=self._workspace_buffer,