[Bugfix][Hardware][CPU] Enable Gemma2 with SDPA on CPU backend (vllm-…

…project#11169)

vllm/attention/backends/torch_sdpa.py

@@ -13,7 +13,7 @@
 from vllm.attention.backends.utils import CommonAttentionState
 from vllm.attention.ops.ipex_attn import PagedAttention
 from vllm.attention.ops.paged_attn import PagedAttentionMetadata
-from vllm.utils import make_tensor_with_pad
+from vllm.utils import make_tensor_with_pad, print_warning_once
 from vllm.worker.cpu_model_runner import ModelInputForCPUBuilder
 
 
@@ -395,7 +395,8 @@ def __init__(
             raise ValueError(
                 "Torch SPDA does not support block-sparse attention.")
         if logits_soft_cap is not None:
-            raise ValueError("Torch SPDA does not support logits soft cap.")
+            print_warning_once("Torch SPDA does not support logits soft cap. "
+                               "Outputs may be slightly off.")
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)
@@ -619,7 +620,7 @@ def _run_sdpa_forward(
                 value[None, :, start_kv:end_kv, :],
                 attn_mask=mask,
                 dropout_p=0.0,
-                is_causal=causal_attn and not self.need_mask,
+                is_causal=causal_attn and mask is None,
                 scale=self.scale).squeeze(0).movedim(query.dim() - 2, 0)
             output[start_q:end_q, :, :] = sub_out
             start_q, start_kv = end_q, end_kv