feat:flash_mla，q去掉pad

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

@@ -924,14 +924,14 @@ class FlashMLASparseImpl(MLACommonBaseImpl[FlashMLASparseMetadata]):
         padded_num_heads = self.fp8_decode_padded_heads
 
         # Pad query if needed (kernel only supports h_q = 64 or 128)
-        if actual_num_heads < padded_num_heads:
-            logger.warning_once(
-                f"Padding num_heads from {actual_num_heads} to "
-                f"{padded_num_heads} for FP8 sparse decode kernel"
-            )
-            q_padded = q.new_zeros((q.size(0), q.size(1), padded_num_heads, q.size(3)))
-            q_padded[:, :, :actual_num_heads, :] = q
-            q = q_padded
+        #if actual_num_heads < padded_num_heads:
+        #    logger.warning_once(
+        #        f"Padding num_heads from {actual_num_heads} to "
+        #        f"{padded_num_heads} for FP8 sparse decode kernel"
+        #    )
+        #    q_padded = q.new_zeros((q.size(0), q.size(1), padded_num_heads, q.size(3)))
+        #    q_padded[:, :, :actual_num_heads, :] = q
+        #    q = q_padded
 
         out, lse = flash_mla_with_kvcache(
             q=q,
@@ -946,8 +946,8 @@ class FlashMLASparseImpl(MLACommonBaseImpl[FlashMLASparseMetadata]):
         )
 
         # Slice output back to actual head count if we padded
-        if actual_num_heads < padded_num_heads:
-            out = out[:, :, :actual_num_heads, :]
+        #if actual_num_heads < padded_num_heads:
+        #    out = out[:, :, :actual_num_heads, :]
 
         return out, lse