[BugFix] FA2 MLA Accuracy Issue (#18807)

Signed-off-by: LucasWilkinson <lwilkinson@neuralmagic.com>

csrc/attention/merge_attn_states.cu

@@ -143,6 +143,14 @@ void merge_attn_states_launcher(torch::Tensor& output,
   const uint pack_size = 16 / sizeof(scalar_t);
   TORCH_CHECK(head_size % pack_size == 0,
               "headsize must be multiple of pack_size:", pack_size);
+  TORCH_CHECK(output.stride(-2) == head_size && output.stride(-1) == 1,
+              "output heads must be contiguous in memory");
+  TORCH_CHECK(
+      prefix_output.stride(-2) == head_size && prefix_output.stride(-1) == 1,
+      "prefix_output heads must be contiguous in memory");
+  TORCH_CHECK(
+      suffix_output.stride(-2) == head_size && suffix_output.stride(-1) == 1,
+      "suffix_output heads must be contiguous in memory");
   float* output_lse_ptr = nullptr;
   if (output_lse.has_value()) {
     output_lse_ptr = output_lse.value().data_ptr<float>();

vllm/attention/backends/mla/common.py

@@ -1093,10 +1093,6 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
         if isinstance(attn_out, tuple):
             attn_out, *rest = attn_out
 
-        # unpad if necessary
-        if self._pad_v:
-            attn_out = attn_out[..., :v.shape[-1]]
-
         # Remain consistent with old `flash_attn_varlen_func` where there
         # is only one output tensor if `return_softmax_lse` is False.
         if return_softmax_lse:
@@ -1294,6 +1290,10 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
                 suffix_lse=suffix_lse,
             )
 
+        # unpad if necessary
+        if self._pad_v:
+            output = output[..., :v.shape[-1]]
+
         return output.flatten(start_dim=-2)
 
     @abstractmethod

vllm/v1/attention/backends/mla/common.py

@@ -653,10 +653,6 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
         if isinstance(attn_out, tuple):
             attn_out, lse = attn_out[0], attn_out[1]
 
-        # unpad if necessary
-        if self._pad_v:
-            attn_out = attn_out[..., :v.shape[-1]]
-
         # Remain consistent with old `flash_attn_varlen_func` where there
         # is only one output tensor if `return_softmax_lse` is False.
         if return_softmax_lse:
@@ -839,6 +835,10 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
                 suffix_lse=suffix_lse,
             )
 
+        # unpad if necessary
+        if self._pad_v:
+            output = output[..., :v.shape[-1]]
+
         return output.flatten(start_dim=-2)
 
     @abstractmethod