Batch-aware torch.ops.llama.sdpa_with_kv_cache (#4822)

Summary:
Pull Request resolved: #4822

This changes makes torch.ops.llama.sdpa_with_kv_cache batch aware. This is needed for batched sdpa cases, for example llm beam search.

* Makes update_cache update across the batch dimension

As a performance optimization, update_cache implements the following operation
```
    k_cache[:, start_pos : start_pos + seq_len, :, :] = k
    v_cache[:, start_pos : start_pos + seq_len, :, :] = v
```
as part of the fused sdpa_with_kv_cache op. A naiive export of this code inserts expensive slice-scatter ops. sdpa_with_kv_cache fuses this update with the flash attention op for tensors that follow a predetermined format [batch, length, heads, dim]. This change removes the assumption that batch == 1.

* Makes sdpa_with_kv_cache apply cpu_flash_attention for all batch lines as well.

ExecuTorch-exported Llama models are implemented with a greedy search, so it has not been necessary for this op to be batch-aware. However when working with other models, or when doing LLM beam search, this is no longer true.

Reviewed By: kimishpatel

Differential Revision: D61605316

Author: meta-emilian

extension/llm/custom_ops/op_sdpa.cpp

@@ -700,10 +700,23 @@ void update_cache(
     const Tensor& cache,
     int64_t start_pos,
     int64_t seq_length) { // NOLINT: unused parameter 'seq_length'
+  // 1) Cache shape should be [bs, max_seq_len, num heads, head dim]
+  // 2) projected_value shape should be [bs, seq_len, num heads, head dim]
+  // 3) We're updating the cache with projected_value, at position start_pos
+
+  ET_CHECK_MSG(
+      projected_value.size(0) == cache.size(0),
+      "projected_value batch size should be equal to the cache batch size.");
+  ET_CHECK_MSG(
+      projected_value.size(2) == cache.size(2),
+      "projected_value number of heads should be equal to the cache number of heads.");
   ET_CHECK_MSG(
-      projected_value.size(0) == 1,
-      "projected_value must have batch size of 1");
-  ET_CHECK_MSG(cache.size(0) == 1, "cache must have batch size of 1");
+      projected_value.size(3) == cache.size(3),
+      "projected_value embedding dimension should be equal to the cache embedding dimension.");
+  ET_CHECK_MSG(
+      projected_value.element_size() == cache.element_size(),
+      "projected_value data type size should be equal to the cache data type size.");
+
   ET_CHECK_MSG(
       is_contiguous_dim_order(
           projected_value.dim_order().data(), projected_value.dim()),
@@ -714,16 +727,31 @@ void update_cache(
   ET_CHECK_MSG(projected_value_data != nullptr, "projected_value data is null");
   ET_CHECK_MSG(cache_data, "cache data is null");
 
-  auto strides = cache.strides();
-  exec_aten::StridesType seq_dim_stride = strides[1];
-  exec_aten::SizesType pos_offset = start_pos * seq_dim_stride;
-  exec_aten::SizesType pos_offset_bytes =
-      pos_offset * projected_value.element_size();
-  exec_aten::SizesType num_bytes =
-      projected_value.numel() * projected_value.element_size();
-  // NOLINTNEXTLINE
-  std::memcpy(
-      (uint8_t*)cache_data + pos_offset_bytes, projected_value_data, num_bytes);
+  auto cache_strides = cache.strides();
+  exec_aten::StridesType cache_batch_dim_stride = cache_strides[0];
+  exec_aten::StridesType cache_seq_dim_stride = cache_strides[1];
+
+  auto value_strides = projected_value.strides();
+  exec_aten::StridesType value_batch_dim_stride = value_strides[0];
+
+  exec_aten::SizesType num_bytes_to_copy =
+      (projected_value.numel() / projected_value.size(0)) *
+      projected_value.element_size();
+
+  for (int64_t batch_line = 0; batch_line < projected_value.size(0);
+       ++batch_line) {
+    exec_aten::SizesType cache_pos_offset =
+        (batch_line * cache_batch_dim_stride +
+         start_pos * cache_seq_dim_stride) *
+        cache.element_size();
+    exec_aten::SizesType value_pos_offset =
+        (batch_line * value_batch_dim_stride) * cache.element_size();
+
+    std::memcpy(
+        (uint8_t*)cache_data + cache_pos_offset,
+        (uint8_t*)projected_value_data + value_pos_offset,
+        num_bytes_to_copy);
+  }
 }
 
 } // anonymous namespace
@@ -859,6 +887,8 @@ Tensor& sdpa_with_kv_cache_out(
       sliced_key_dim_order.data(),
       util::kKVDim,
       sliced_key_strides.data());
+  // since the cache is sliced, the batch stride needs to stay the same.
+  sliced_key_strides[0] = key_cache.strides()[0];
   void* key_cache_data = key_cache.mutable_data_ptr();
   TensorImpl k_impl = TensorImpl(
       key_cache.scalar_type(),
@@ -883,6 +913,8 @@ Tensor& sdpa_with_kv_cache_out(
       sliced_value_dim_order.data(),
       util::kKVDim,
       sliced_value_strides.data());
+  // since the cache is sliced, the batch stride needs to stay the same.
+  sliced_value_strides[0] = value_cache.strides()[0];
   void* value_cache_data = value_cache.mutable_data_ptr();
   TensorImpl value_impl = TensorImpl(
       value_cache.scalar_type(),

extension/llm/custom_ops/test_sdpa_with_kv_cache.py

@@ -373,10 +373,10 @@ class SDPATestCommon(unittest.TestCase):
 
     def setup_caches(self):
         self.k_cache = torch.zeros(
-            (1, self.max_seq_len, self.n_heads_kv, self.head_dim)
+            (self.n_batch, self.max_seq_len, self.n_heads_kv, self.head_dim)
         )
         self.v_cache = torch.zeros(
-            (1, self.max_seq_len, self.n_heads_kv, self.head_dim)
+            (self.n_batch, self.max_seq_len, self.n_heads_kv, self.head_dim)
         )
         self.mask = torch.full(
             (self.max_seq_len, self.max_seq_len),
@@ -386,6 +386,7 @@ def setup_caches(self):
 
     def setUp(self):
         torch.manual_seed(42)
+        self.n_batch = 2
         self.n_heads_kv = 32
         self.n_heads_q = 32
         self.head_dim = 128
@@ -418,19 +419,19 @@ def _test_sdpa_common(
         self.max_seq_len = max_seq_len
         self.setup_caches()
         q = self._scale_tensor(
-            torch.rand((1, seq_len, self.n_heads_kv, self.head_dim)),
+            torch.rand((self.n_batch, seq_len, self.n_heads_kv, self.head_dim)),
             tensor_scale_max,
             tensor_scale_min,
             scale_tensors,
         )
         k = self._scale_tensor(
-            torch.rand((1, seq_len, self.n_heads_kv, self.head_dim)),
+            torch.rand((self.n_batch, seq_len, self.n_heads_kv, self.head_dim)),
             tensor_scale_max,
             tensor_scale_min,
             scale_tensors,
         )
         v = self._scale_tensor(
-            torch.rand((1, seq_len, self.n_heads_kv, self.head_dim)),
+            torch.rand((self.n_batch, seq_len, self.n_heads_kv, self.head_dim)),
             tensor_scale_max,
             tensor_scale_min,
             scale_tensors,
@@ -466,6 +467,15 @@ def _test_sdpa_common(
             scale_tensors,
         )
 
+        q = torch.rand(
+            (self.n_batch, next_iter_seq_len, self.n_heads_kv, self.head_dim)
+        )
+        k = torch.rand(
+            (self.n_batch, next_iter_seq_len, self.n_heads_kv, self.head_dim)
+        )
+        v = torch.rand(
+            (self.n_batch, next_iter_seq_len, self.n_heads_kv, self.head_dim)
+        )
         start_pos = seq_len
         seq_len = q.size(1)
         attn_mask = self.mask[start_pos : start_pos + seq_len, :]