[rocm][V0] fix selection logic for custom PA in V0 (#16426)

Signed-off-by: Divakar Verma <[email protected]>

Author: divakar-amd

vllm/platforms/rocm.py

@@ -109,8 +109,11 @@ def use_rocm_custom_paged_attention(qtype: torch.dtype, head_size: int,
     ON_MI250_MI300 = any(arch in GPU_ARCH for arch in ["gfx90a", "gfx942"])
 
     # rocm custom page attention not support on navi (gfx1*)
+    # custom paged attn always supported on V0. On V1, requires sliding window
+    # disabled due to observed numerical discrepancy.
     return (ON_MI250_MI300 and not ON_NAVI
-            and (sliding_window == 0 or sliding_window == (-1, -1))
+            and (not envs.VLLM_USE_V1 or sliding_window == 0
+                 or sliding_window == (-1, -1))
             and (qtype == torch.half or qtype == torch.bfloat16)
             and (head_size == 64 or head_size == 128)
             and (block_size == 16 or block_size == 32)