DeformableDetrModel support fp16 (#29013)

* Update ms_deform_attn_cuda.cu

* Update ms_deform_attn_cuda.cuh

* Update modeling_deformable_detr.py

* Update src/transformers/models/deformable_detr/modeling_deformable_detr.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update modeling_deformable_detr.py

* python utils/check_copies.py --fix_and_overwrite

* Fix dtype missmatch error

* Update test_modeling_deformable_detr.py

* Update test_modeling_deformable_detr.py

* Update modeling_deformable_detr.py

* Update modeling_deformable_detr.py

---------

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

src/transformers/kernels/deformable_detr/cuda/ms_deform_attn_cuda.cu

@@ -64,7 +64,7 @@ at::Tensor ms_deform_attn_cuda_forward(
     for (int n = 0; n < batch/im2col_step_; ++n)
     {
         auto columns = output_n.select(0, n);
-        AT_DISPATCH_FLOATING_TYPES(value.type(), "ms_deform_attn_forward_cuda", ([&] {
+        AT_DISPATCH_FLOATING_TYPES_AND_HALF(value.type(), "ms_deform_attn_forward_cuda", ([&] {
             ms_deformable_im2col_cuda(at::cuda::getCurrentCUDAStream(),
                 value.data<scalar_t>() + n * im2col_step_ * per_value_size,
                 spatial_shapes.data<int64_t>(),
@@ -134,7 +134,7 @@ std::vector<at::Tensor> ms_deform_attn_cuda_backward(
     for (int n = 0; n < batch/im2col_step_; ++n)
     {
         auto grad_output_g = grad_output_n.select(0, n);
-        AT_DISPATCH_FLOATING_TYPES(value.type(), "ms_deform_attn_backward_cuda", ([&] {
+        AT_DISPATCH_FLOATING_TYPES_AND_HALF(value.type(), "ms_deform_attn_backward_cuda", ([&] {
             ms_deformable_col2im_cuda(at::cuda::getCurrentCUDAStream(),
                                     grad_output_g.data<scalar_t>(),
                                     value.data<scalar_t>() + n * im2col_step_ * per_value_size,

src/transformers/kernels/deformable_detr/cuda/ms_deform_attn_cuda.cuh

@@ -72,7 +72,7 @@ at::Tensor ms_deform_attn_cuda_forward(
     for (int n = 0; n < batch/im2col_step_; ++n)
     {
         auto columns = output_n.select(0, n);
-        AT_DISPATCH_FLOATING_TYPES(value.type(), "ms_deform_attn_forward_cuda", ([&] {
+        AT_DISPATCH_FLOATING_TYPES_AND_HALF(value.type(), "ms_deform_attn_forward_cuda", ([&] {
             ms_deformable_im2col_cuda(at::cuda::getCurrentCUDAStream(),
                 value.data<scalar_t>() + n * im2col_step_ * per_value_size,
                 spatial_shapes.data<int64_t>(),
@@ -142,7 +142,7 @@ std::vector<at::Tensor> ms_deform_attn_cuda_backward(
     for (int n = 0; n < batch/im2col_step_; ++n)
     {
         auto grad_output_g = grad_output_n.select(0, n);
-        AT_DISPATCH_FLOATING_TYPES(value.type(), "ms_deform_attn_backward_cuda", ([&] {
+        AT_DISPATCH_FLOATING_TYPES_AND_HALF(value.type(), "ms_deform_attn_backward_cuda", ([&] {
             ms_deformable_col2im_cuda(at::cuda::getCurrentCUDAStream(),
                                     grad_output_g.data<scalar_t>(),
                                     value.data<scalar_t>() + n * im2col_step_ * per_value_size,

src/transformers/models/deformable_detr/modeling_deformable_detr.py

@@ -617,7 +617,8 @@ class DeformableDetrMultiscaleDeformableAttention(nn.Module):
 
     def _reset_parameters(self):
         nn.init.constant_(self.sampling_offsets.weight.data, 0.0)
-        thetas = torch.arange(self.n_heads, dtype=torch.int64).float() * (2.0 * math.pi / self.n_heads)
+        default_dtype = torch.get_default_dtype()
+        thetas = torch.arange(self.n_heads, dtype=torch.int64).to(default_dtype) * (2.0 * math.pi / self.n_heads)
         grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
         grid_init = (
             (grid_init / grid_init.abs().max(-1, keepdim=True)[0])
@@ -1171,8 +1172,8 @@ class DeformableDetrEncoder(DeformableDetrPreTrainedModel):
         reference_points_list = []
         for level, (height, width) in enumerate(spatial_shapes):
             ref_y, ref_x = meshgrid(
-                torch.linspace(0.5, height - 0.5, height, dtype=torch.float32, device=device),
+                torch.linspace(0.5, height - 0.5, height, dtype=valid_ratios.dtype, device=device),
-                torch.linspace(0.5, width - 0.5, width, dtype=torch.float32, device=device),
+                torch.linspace(0.5, width - 0.5, width, dtype=valid_ratios.dtype, device=device),
                 indexing="ij",
             )
             # TODO: valid_ratios could be useless here. check https://github.com/fundamentalvision/Deformable-DETR/issues/36
@@ -1540,15 +1541,15 @@ class DeformableDetrModel(DeformableDetrPreTrainedModel):
         for name, param in self.backbone.conv_encoder.model.named_parameters():
             param.requires_grad_(True)
 
-    def get_valid_ratio(self, mask):
+    def get_valid_ratio(self, mask, dtype=torch.float32):
         """Get the valid ratio of all feature maps."""
 
         _, height, width = mask.shape
         valid_height = torch.sum(mask[:, :, 0], 1)
         valid_width = torch.sum(mask[:, 0, :], 1)
-        valid_ratio_heigth = valid_height.float() / height
+        valid_ratio_height = valid_height.to(dtype) / height
-        valid_ratio_width = valid_width.float() / width
+        valid_ratio_width = valid_width.to(dtype) / width
-        valid_ratio = torch.stack([valid_ratio_width, valid_ratio_heigth], -1)
+        valid_ratio = torch.stack([valid_ratio_width, valid_ratio_height], -1)
         return valid_ratio
 
     def get_proposal_pos_embed(self, proposals):
@@ -1721,7 +1722,7 @@ class DeformableDetrModel(DeformableDetrPreTrainedModel):
         lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)
         spatial_shapes = torch.as_tensor(spatial_shapes, dtype=torch.long, device=source_flatten.device)
         level_start_index = torch.cat((spatial_shapes.new_zeros((1,)), spatial_shapes.prod(1).cumsum(0)[:-1]))
-        valid_ratios = torch.stack([self.get_valid_ratio(m) for m in masks], 1)
+        valid_ratios = torch.stack([self.get_valid_ratio(m, dtype=source_flatten.dtype) for m in masks], 1)
         valid_ratios = valid_ratios.float()
 
         # Fourth, sent source_flatten + mask_flatten + lvl_pos_embed_flatten (backbone + proj layer output) through encoder

src/transformers/models/deta/modeling_deta.py

@@ -1549,15 +1549,15 @@ class DetaModel(DetaPreTrainedModel):
             param.requires_grad_(True)
 
     # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrModel.get_valid_ratio
-    def get_valid_ratio(self, mask):
+    def get_valid_ratio(self, mask, dtype=torch.float32):
         """Get the valid ratio of all feature maps."""
 
         _, height, width = mask.shape
         valid_height = torch.sum(mask[:, :, 0], 1)
         valid_width = torch.sum(mask[:, 0, :], 1)
-        valid_ratio_heigth = valid_height.float() / height
+        valid_ratio_height = valid_height.to(dtype) / height
-        valid_ratio_width = valid_width.float() / width
+        valid_ratio_width = valid_width.to(dtype) / width
-        valid_ratio = torch.stack([valid_ratio_width, valid_ratio_heigth], -1)
+        valid_ratio = torch.stack([valid_ratio_width, valid_ratio_height], -1)
         return valid_ratio
 
     # Copied from transformers.models.deformable_detr.modeling_deformable_detr.DeformableDetrModel.get_proposal_pos_embed

tests/models/deformable_detr/test_modeling_deformable_detr.py

@@ -583,6 +583,18 @@ class DeformableDetrModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineT
         loss = model(**inputs).loss
         loss.backward()
 
+    def create_and_check_model_fp16_forward(self):
+        model_class = DeformableDetrForObjectDetection
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        model = model_class(config)
+        model.to(torch_device)
+        model.half()
+        model.eval()
+        inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+        output = model(**inputs)["last_hidden_state"]
+        self.parent.assertFalse(torch.isnan(output).any().item())
+
 
 TOLERANCE = 1e-4