[Metax] Fix Bug (#2472)

* Remove debug print statements in flash_attn_grad_kernel

Commented out debug print statements to clean up output.

* [Metax] Fix Bug (#224)

* [Metax] Fix eigen cmake bug

* [Metax] Fix allow_tf32_cublas

* fix patch

* [Metax] Fix error rnn

* [Metax] Fix error rnn

* [Metax] update paddle

* [Metax] Fix Bug

* [Metax] Fix Bug

---------

Co-authored-by: duqimeng <77875733+duqimeng@users.noreply.github.com>

Author: metax666

backends/metax_gpu/CMakeLists.txt

@@ -290,7 +290,7 @@ file(
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/unsqueeze_kernel.cc
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/squeeze_grad_kernel.cc
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/squeeze_kernel.cc
-  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/sign_kernel.cu.cc
+  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/sign_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/split_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/scatter_nd_add_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/scatter_nd_add_grad_kernel.cu

backends/metax_gpu/kernels/cuda_kernels/sign_kernel_register.cu

@@ -27,7 +27,7 @@ PD_CUSTOM_KERNEL_REGISTER(sign,
                           int64_t,
                           float,
                           double,
-                          phi::dtype::float16,
-                          phi::dtype::bfloat16,
-                          phi::dtype::complex<float>,
-                          phi::dtype::complex<double>) {}
+                          phi::float16,
+                          phi::bfloat16,
+                          phi::complex64,
+                          phi::complex128) {}

backends/metax_gpu/kernels/custom_kernel/flash_attn_grad_kernel.cu

@@ -48,7 +48,7 @@ void FlashAttnUnpaddedGradKernel_(
     DenseTensor* dk,
     DenseTensor* dv) {
 #ifdef PADDLE_WITH_FLASHATTN
-  printf("************ FlashAttnUnpaddedGradKernel_ ****************\n");
+  // printf("************ FlashAttnUnpaddedGradKernel_ ****************\n");
   ctx.template Alloc<T>(dq);
 
   DenseTensor dk_tmp;
@@ -214,7 +214,7 @@ void FlashAttnGradKernel(const Context& ctx,
   ctx.template Alloc<T>(dq);
   DenseTensor dk_tmp;
   if (dk) {
-    printf("dk input pointer is not nullptr!\n");
+    // printf("dk input pointer is not nullptr!\n");
     if (!is_mha) {
       dk_tmp.Resize({batch_size, seqlen_k, num_heads, head_size_og});
       ctx.template Alloc<T>(&dk_tmp);
@@ -224,7 +224,7 @@ void FlashAttnGradKernel(const Context& ctx,
     }
 
   } else {
-    printf("dk input pointer is nullptr!\n");
+    // printf("dk input pointer is nullptr!\n");
     dk_tmp = EmptyLike<T, Context>(ctx, k);
   }
 
@@ -242,9 +242,9 @@ void FlashAttnGradKernel(const Context& ctx,
   }
 
   if (dk_tmp.meta().is_contiguous()) {
-    printf("dk_tmp alloc memory is contiguous!\n");
+    // printf("dk_tmp alloc memory is contiguous!\n");
   } else {
-    printf("dk_tmp alloc memory is not contiguous!\n");
+    // printf("dk_tmp alloc memory is not contiguous!\n");
   }
 
   FlashAttnParamsBwd params = FlashAttnParamsBwd(ctx,
@@ -278,7 +278,7 @@ void FlashAttnGradKernel(const Context& ctx,
     VLOG(10) << "[FlashAttn Backward] attn_mask.shape=["
              << (attn_mask.get_ptr())->dims() << "]";
   }
-  // printf("params.dq dims[2]:%d, params.dk dims[2]:%d, params.dv
+  // // printf("params.dq dims[2]:%d, params.dk dims[2]:%d, params.dv
   // dims[2]:%d\n", params.dq->head_num, params.dk->head_num,
   // params.dv->head_num);
   print_tensor_info(params.dq);
@@ -326,7 +326,7 @@ void FlashAttnGradKernel(const Context& ctx,
         phi::SumKernel<T, Context>(ctx, dk_tmp, {3}, dk->type(), false, dk);
       } else {
         // kvReduceBatchedForGQA<T, Context>(ctx, dk_tmp, dk);
-        printf("[%s: %d] - Need to complete!\n", __func__, __LINE__);
+        // printf("[%s: %d] - Need to complete!\n", __func__, __LINE__);
       }
     }
 
@@ -340,7 +340,7 @@ void FlashAttnGradKernel(const Context& ctx,
         phi::SumKernel<T, Context>(ctx, dv_tmp, {3}, dv->type(), false, dv);
       } else {
         // kvReduceBatchedForGQA<T, Context>(ctx, dv_tmp, dv);
-        printf("[%s: %d] - Need to complete!\n", __func__, __LINE__);
+        // printf("[%s: %d] - Need to complete!\n", __func__, __LINE__);
       }
     }
   }