Symbolic shape tracing on jagged op (#1758)

Summary:
Pull Request resolved: #1758

fbgemm's jagged op doesn't have symint support in the meta function, resulting in all the shapes specialized (materialized int baked in the model instead of symbolic shape). Fixing them.

Reviewed By: q10

Differential Revision:
D44736488

Privacy Context Container: L1156430

fbshipit-source-id: af969188c914aae32651beccc491f6c76bd23536

fbgemm_gpu/include/fbgemm_gpu/sparse_ops.h

@@ -431,7 +431,7 @@ std::tuple<at::Tensor, std::vector<at::Tensor>> jagged_dense_elementwise_mul(
 std::tuple<at::Tensor, std::vector<at::Tensor>> dense_to_jagged(
     const at::Tensor& dense,
     const std::vector<at::Tensor>& offsets,
-    const c10::optional<int64_t>& total_L);
+    const c10::optional<at::SymInt>& total_L);
 
 std::tuple<at::Tensor, std::vector<at::Tensor>>
 jagged_dense_elementwise_add_jagged_output(

fbgemm_gpu/src/jagged_tensor_ops/dense_to_jagged_forward.cu

@@ -15,18 +15,18 @@ namespace fbgemm_gpu {
 Tensor dense_to_jagged_forward(
     const Tensor& dense,
     const std::vector<Tensor>& offsets,
-    const c10::optional<int64_t>& total_L) {
+    const c10::optional<at::SymInt>& total_L) {
   // D is the embedding dimension
   auto D = dense.size(-1);
 
   // If total_L is not given then compute it
-  int64_t total_L_computed;
+  at::SymInt total_L_computed;
   if (total_L.has_value()) {
     total_L_computed = total_L.value();
   } else {
     total_L_computed = (int64_t)offsets.back().max().item<int64_t>();
   }
-  auto values = at::empty({total_L_computed, D}, dense.options());
+  auto values = at::empty_symint({total_L_computed, D}, dense.options());
   auto output = at::empty_like(values);
 
   at::cuda::OptionalCUDAGuard device_guard;

fbgemm_gpu/src/jagged_tensor_ops/jagged_to_padded_dense_backward.cu

@@ -15,7 +15,7 @@ namespace fbgemm_gpu {
 at::Tensor jagged_to_padded_dense_backward(
     const Tensor& grad_output,
     const std::vector<Tensor>& offsets,
-    const int64_t total_L) {
+    const at::SymInt& total_L) {
   auto grad_padded_values = grad_output;
   at::cuda::OptionalCUDAGuard device_guard;
   device_guard.set_index(grad_padded_values.get_device());
@@ -29,7 +29,8 @@ at::Tensor jagged_to_padded_dense_backward(
 
   // Initialize with zeros so output will be zero for the portion truncated
   // in forward.
-  auto grad_values = at::zeros({total_L, D}, grad_padded_values.options());
+  auto grad_values =
+      at::zeros_symint({total_L, D}, grad_padded_values.options());
 
   AT_DISPATCH_FLOATING_TYPES_AND2(
       at::ScalarType::Half,

fbgemm_gpu/src/jagged_tensor_ops/jagged_to_padded_dense_forward.cu

@@ -21,7 +21,7 @@ namespace fbgemm_gpu {
 at::Tensor jagged_to_padded_dense_forward(
     const Tensor& values,
     const std::vector<Tensor>& offsets,
-    const std::vector<int64_t>& max_lengths,
+    const at::ArrayRef<at::SymInt>& max_lengths,
     const double padding_value) {
   const size_t num_jagged_dim = offsets.size();
   TORCH_CHECK(
@@ -40,7 +40,7 @@ at::Tensor jagged_to_padded_dense_forward(
            values.sizes().end(),
            1,
            std::multiplies<size_t>())});
-  at::DimVector padded_values_shape({offsets[0].size(0) - 1});
+  at::SymDimVector padded_values_shape({at::SymInt(offsets[0].size(0) - 1)});
   padded_values_shape.insert(
       padded_values_shape.end(), max_lengths.begin(), max_lengths.end());
 
@@ -50,7 +50,8 @@ at::Tensor jagged_to_padded_dense_forward(
   if (!D_folded) {
     padded_values_shape.push_back(values.size(-1));
   }
-  Tensor padded_values = at::empty(padded_values_shape, values.options());
+  Tensor padded_values =
+      at::empty_symint(padded_values_shape, values.options());
   Tensor padded_values_view =
       D_folded ? padded_values.unsqueeze(-1) : padded_values;
 
@@ -121,7 +122,7 @@ std::vector<Tensor> stacked_jagged_1d_to_dense_gpu(
     padded_values_per_key.push_back(jagged_to_padded_dense_forward(
         values.slice(0, offset_per_key[t], offset_per_key[t + 1]),
         {offsets},
-        {max_L},
+        at::ArrayRef<at::SymInt>({max_L}),
         padding_value));
   }
   return padded_values_per_key;
@@ -179,7 +180,7 @@ stacked_jagged_2d_to_dense_forward_cuda(
     padded_values_per_key.push_back(jagged_to_padded_dense_forward(
         values.slice(0, offset_per_key[t], offset_per_key[t + 1]),
         {offsets},
-        {max_L},
+        at::ArrayRef<at::SymInt>({max_L}),
         padding_value));
   }
 
@@ -301,7 +302,10 @@ Tensor jagged_2d_to_dense_gpu_forward(
     Tensor offsets,
     int64_t max_sequence_length) {
   return jagged_to_padded_dense_forward(
-      values, {offsets}, {max_sequence_length}, /*padding_value=*/0);
+      values,
+      {offsets},
+      c10::ArrayRef<c10::SymInt>({max_sequence_length}),
+      /*padding_value=*/0);
 }
 
 namespace {
@@ -369,7 +373,8 @@ class JaggedDenseAddJaggedOutputGPUOp
     Tensor dense_values_grad = jagged_to_padded_dense_forward(
         grad_outputs[0],
         offsets,
-        std::vector<int64_t>(dense_shape.begin() + 1, dense_shape.end() - 1),
+        c10::fromIntArrayRefKnownNonNegative(std::vector<int64_t>(
+            dense_shape.begin() + 1, dense_shape.end() - 1)),
         /*padding_value=*/0);
     TORCH_CHECK(dense_values_grad.sizes() == dense_shape);
 

fbgemm_gpu/src/jagged_tensor_ops_autograd.cpp

@@ -11,6 +11,7 @@
 #include <ATen/core/dispatch/Dispatcher.h>
 #include <torch/csrc/autograd/custom_function.h>
 #include <torch/library.h>
+#include <torch/torch.h>
 
 #include "ATen/TensorUtils.h"
 #include "fbgemm_gpu/sparse_ops.h"
@@ -35,17 +36,18 @@ class JaggedToPaddedDenseOp
       const std::vector<int64_t>& max_lengths,
       const double padding_value) {
     ctx->save_for_backward(offsets);
-    ctx->saved_data["total_L"] = values.size(0);
+    ctx->saved_data["total_L"] = values.sym_size(0);
 
     static auto op =
         c10::Dispatcher::singleton()
             .findSchemaOrThrow("fbgemm::jagged_to_padded_dense_forward", "")
             .typed<at::Tensor(
                 const Tensor& values,
                 const std::vector<Tensor>& offsets,
-                const std::vector<int64_t>& max_lengths,
+                const at::ArrayRef<at::SymInt>& max_lengths,
                 const double padding_value)>();
-    Tensor padded_values = op.call(values, offsets, max_lengths, padding_value);
+    Tensor padded_values = op.call(
+        values, offsets, c10::fromIntArrayRefSlow(max_lengths), padding_value);
 
     return {padded_values};
   }
@@ -54,7 +56,7 @@ class JaggedToPaddedDenseOp
       torch::autograd::AutogradContext* ctx,
       torch::autograd::variable_list grad_outputs) {
     auto offsets = ctx->get_saved_variables();
-    int32_t total_L = ctx->saved_data["total_L"].toInt();
+    at::SymInt total_L = ctx->saved_data["total_L"].toSymInt();
     TORCH_CHECK(grad_outputs.size() == 1);
 
     TORCH_CHECK(total_L >= 0);
@@ -64,7 +66,7 @@ class JaggedToPaddedDenseOp
             .typed<at::Tensor(
                 const Tensor& grad_output,
                 const std::vector<Tensor>& offsets,
-                const int64_t total_L)>();
+                const at::SymInt& total_L)>();
     auto grad_values = op.call(grad_outputs[0], {offsets}, total_L);
 
     return {
@@ -86,7 +88,15 @@ class JaggedDenseDenseAddJaggedOutputOp
       const Tensor& dense_0,
       const Tensor& dense_1) {
     ctx->save_for_backward(offsets);
+#if TORCH_VERSION_MAJOR > 2 || \
+    (TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR >= 1)
+    // toSymIntVector support is from a recent PR
+    // https://github.com/pytorch/pytorch/pull/101056,
+    // so protect it under a version guard for compatibility
+    ctx->saved_data["dense_shape"] = dense_0.sym_sizes();
+#else
     ctx->saved_data["dense_shape"] = dense_0.sizes();
+#endif
 
     static auto op =
         c10::Dispatcher::singleton()
@@ -107,7 +117,12 @@ class JaggedDenseDenseAddJaggedOutputOp
       torch::autograd::AutogradContext* ctx,
       torch::autograd::variable_list grad_outputs) {
     auto offsets = ctx->get_saved_variables();
+#if TORCH_VERSION_MAJOR > 2 || \
+    (TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR >= 1)
+    auto dense_shape = ctx->saved_data["dense_shape"].toSymIntVector();
+#else
     auto dense_shape = ctx->saved_data["dense_shape"].toIntVector();
+#endif
     TORCH_CHECK(grad_outputs.size() == 1);
 
     static auto op =
@@ -116,12 +131,12 @@ class JaggedDenseDenseAddJaggedOutputOp
             .typed<at::Tensor(
                 const Tensor& values,
                 const std::vector<Tensor>& offsets,
-                const std::vector<int64_t>& max_lengths,
+                const at::ArrayRef<at::SymInt>& max_lengths,
                 const double padding_value)>();
     Tensor dense_values_grad_0 = op.call(
         grad_outputs[0],
         offsets,
-        std::vector<int64_t>(dense_shape.begin() + 1, dense_shape.end() - 1),
+        std::vector<at::SymInt>(dense_shape.begin() + 1, dense_shape.end() - 1),
         /*padding_value=*/0);
     Tensor dense_values_grad_1 = dense_values_grad_0;
 
@@ -249,19 +264,27 @@ class DenseToJaggedOp : public torch::autograd::Function<DenseToJaggedOp> {
       torch::autograd::AutogradContext* ctx,
       const Tensor& dense,
       const std::vector<Tensor>& offsets,
-      const c10::optional<int64_t>& total_L) {
+      const c10::optional<at::SymInt>& total_L) {
     ctx->save_for_backward(offsets);
 
     // dims of dense tensor: <batch, [maxlen0, maxlen1, ...], embedding_dim>
+#if TORCH_VERSION_MAJOR > 2 || \
+    (TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR >= 1)
+    // toSymIntVector support is from a recent PR
+    // https://github.com/pytorch/pytorch/pull/101056,
+    // so protect it under a version guard for compatibility
+    ctx->saved_data["dense_shape"] = dense.sym_sizes();
+#else
     ctx->saved_data["dense_shape"] = dense.sizes();
+#endif
 
     static auto op =
         c10::Dispatcher::singleton()
             .findSchemaOrThrow("fbgemm::dense_to_jagged_forward", "")
             .typed<Tensor(
                 const Tensor& dense,
                 const std::vector<Tensor>& offsets,
-                const c10::optional<int64_t>& total_L)>();
+                const c10::optional<at::SymInt>& total_L)>();
     auto output = op.call(dense, offsets, total_L);
 
     return {output};
@@ -271,7 +294,12 @@ class DenseToJaggedOp : public torch::autograd::Function<DenseToJaggedOp> {
       torch::autograd::AutogradContext* ctx,
       torch::autograd::variable_list grad_outputs) {
     auto offsets = ctx->get_saved_variables();
+#if TORCH_VERSION_MAJOR > 2 || \
+    (TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR >= 1)
+    auto dense_shape = ctx->saved_data["dense_shape"].toSymIntVector();
+#else
     auto dense_shape = ctx->saved_data["dense_shape"].toIntVector();
+#endif
     TORCH_CHECK(grad_outputs.size() == 1);
 
     static auto op =
@@ -280,15 +308,20 @@ class DenseToJaggedOp : public torch::autograd::Function<DenseToJaggedOp> {
             .typed<Tensor(
                 const Tensor& values,
                 const std::vector<Tensor>& offsets,
-                const std::vector<int64_t>& max_lengths,
+                const at::ArrayRef<at::SymInt>& max_lengths,
                 const double padding_value)>();
     auto dense_values_grad = op.call(
         grad_outputs[0],
         offsets,
-        std::vector<int64_t>(dense_shape.begin() + 1, dense_shape.end() - 1),
+        std::vector<at::SymInt>(dense_shape.begin() + 1, dense_shape.end() - 1),
         /*padding_value=*/0);
 
+#if TORCH_VERSION_MAJOR > 2 || \
+    (TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR >= 1)
+    TORCH_CHECK(dense_values_grad.sym_sizes() == dense_shape);
+#else
     TORCH_CHECK(dense_values_grad.sizes() == dense_shape);
+#endif
 
     return {
         dense_values_grad,
@@ -730,7 +763,7 @@ Tensor batched_dense_vec_jagged_2d_mul(
 std::tuple<Tensor, std::vector<Tensor>> dense_to_jagged(
     const Tensor& dense,
     const std::vector<Tensor>& offsets,
-    const c10::optional<int64_t>& total_L) {
+    const c10::optional<at::SymInt>& total_L) {
   return {DenseToJaggedOp::apply(dense, offsets, total_L)[0], offsets};
 }