Fix sdpa flash attention op for et llama deployment (#4322)

Summary:
Pull Request resolved: #4322

We retropfitted flash attention cpu from aten. The retrofit we did was to
make it work to cacluate attention for a) batched prefill and b) decode with
different start_pos. For b, there was a bug when kv cache's seqlen dim is
split.
As a result attention calculation is not right. There is a detail in the code
to explain the issue.

bypass-github-export-checks
ghstack-source-id: 234634902

Reviewed By: larryliu0820

Differential Revision: D60011925

fbshipit-source-id: 50921846b329e449a4a767cf28c7a55d507217bd

examples/models/llama2/custom_ops/op_sdpa.cpp

@@ -239,6 +239,12 @@ void cpu_flash_attention(
   at::Tensor value = v.transpose(1, 2);
   */
 
+  // Without this we have out-of-bounds writes for
+  // causal masking
+  static_assert(
+      kv_split_size > q_split_size,
+      "KV_split_size must be greater than q_split_size");
+
   constexpr bool is_reduced_type =
       torch::executor::is_reduced_floating_point<scalar_t>::value;
 
@@ -417,7 +423,35 @@ void cpu_flash_attention(
       // Initialize max and sum
       fill_stub(
           qk_max_data, -std::numeric_limits<accum_t>::infinity(), qBlockSize);
-      int64_t num_keys = is_causal ? std::min(m + qBlockSize, kvSize) : kvSize;
+      // Original flash sdpa wasnt really meant to be used
+      // for decode the way we are using via start_pos here.
+      // Thus when num_keys is 1 during decode phase, we
+      // still need to iterate through all the kv_splits
+      // Take start_pos = 130 and k_split_size = 128
+      // Here we have to produce [1x130] of q @ k.T
+      // when seq_len = 1
+      // But if num_keys = 1 then we dont really loop over
+      // all kv_splits.
+      // When k_split_size > 130, this is not an issue because
+      // there is only one iteration of the following loop anyway.
+      // Outside of determining how many loop iterations are needed
+      // num_keys participates only in causal attention.
+      // Rest of the calculation of q @ k.T and @ v.T is same.
+      // We dont run into this bug when k_split_size < start_pos + seqlen
+      // since there is only one iteration and that applies
+      // causal attention correctly.
+      // Howeve when k_split_size > start_pos + seqlen, we have
+      // more than one iteration, however if we dont adjust num_keys
+      // we dont get more than one iteration
+      // This is unique to this deployment of flash attention since
+      // original implementation wasnt deployed on this way.
+
+      // Some of these bugs can be resolved by relying on attention mask
+      // but that requires storing attention mask in float as the current
+      // code doesnt support bool attention mask.
+      // However, lets just fix that as well.
+      int64_t num_keys =
+          is_causal ? std::min(m + start_pos + qBlockSize, kvSize) : kvSize;
       auto j_kv = j / num_reps;
       for (int64_t n = 0; n < num_keys; n += kvSplitSize) {
         int64_t kvBlockSize = std::min(kvSplitSize, kvSize - n);
@@ -452,6 +486,7 @@ void cpu_flash_attention(
         // entries need masked out. In our example n = 4
         // will qualify for that
         if (is_causal && num_keys - n <= kvSplitSize) {
+          // For this fn to work k_split_size > q_split_size
           for (int32_t row = 0; row < qBlockSize; ++row) {
             int64_t last_col = m + (row + start_pos) - n;
             accum_t* row_ptr = qk_data + row * kvBlockSize;

examples/models/llama2/custom_ops/test_sdpa_with_kv_cache.py

@@ -365,3 +365,135 @@ def test_sdpa_with_cache_mqa_3(self):
             q, k, v, self.k_cache, self.v_cache, 1, 1, None, 0, False
         )
         self.assertTrue(torch.allclose(ref_output, op_output))
+
+
+class SDPATestForLargeSeqLength(unittest.TestCase):
+
+    def setup_caches(self):
+        self.k_cache = torch.zeros(
+            (1, 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.mask = torch.full(
+            (self.max_seq_len, self.max_seq_len),
+            float("-inf"),
+        )
+        self.mask = torch.triu(self.mask, diagonal=1)
+
+    def setUp(self):
+        torch.manual_seed(42)
+        self.n_heads_kv = 32
+        self.n_heads_q = 32
+        self.head_dim = 128
+        self.max_seq_len = 2048
+        self.setup_caches()
+
+    def test_sdpa_with_cache_seq_len_130(self):
+        self.n_heads_kv = 32
+        self.n_heads_q = 32
+        self.head_dim = 128
+        self.max_seq_len = 2048
+        self.setup_caches()
+        seq_len = 130
+        q = torch.rand((1, seq_len, self.n_heads_kv, self.head_dim))
+        k = torch.rand((1, seq_len, self.n_heads_kv, self.head_dim))
+        v = torch.rand((1, seq_len, self.n_heads_kv, self.head_dim))
+        start_pos = 0
+        attn_mask = self.mask[start_pos : start_pos + seq_len, :]
+        attn_mask = attn_mask[:, : start_pos + seq_len]
+        ref_output = _sdpa_with_kv_cache_ref(
+            q, k, v, self.k_cache, self.v_cache, attn_mask, start_pos, seq_len
+        )
+        op_output = torch.ops.llama.sdpa_with_kv_cache(
+            q, k, v, self.k_cache, self.v_cache, start_pos, seq_len, None, 0, True
+        )
+        self.assertTrue(torch.allclose(ref_output, op_output))
+
+        q = torch.rand((1, 1, self.n_heads_kv, self.head_dim))
+        k = torch.rand((1, 1, self.n_heads_kv, self.head_dim))
+        v = torch.rand((1, 1, 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, :]
+        attn_mask = attn_mask[:, : start_pos + seq_len]
+        ref_output = _sdpa_with_kv_cache_ref(
+            q, k, v, self.k_cache, self.v_cache, attn_mask, start_pos, seq_len
+        )
+        op_output = torch.ops.llama.sdpa_with_kv_cache(
+            q, k, v, self.k_cache, self.v_cache, start_pos, seq_len, None, 0, True
+        )
+        self.assertTrue(torch.allclose(ref_output, op_output))
+
+    def test_sdpa_with_cache_seq_len_small(self):
+        self.n_heads_kv = 4
+        self.n_heads_q = 4
+        self.head_dim = 4
+        self.max_seq_len = 8
+        self.setup_caches()
+        q = torch.rand((1, 4, self.n_heads_q, 4))
+        k = torch.rand((1, 4, self.n_heads_q, 4))
+        v = torch.rand((1, 4, self.n_heads_q, 4))
+        start_pos = 0
+        seq_len = q.size(1)
+        attn_mask = self.mask[start_pos : start_pos + seq_len, :]
+        attn_mask = attn_mask[:, : start_pos + seq_len]
+        ref_output = _sdpa_with_kv_cache_ref(
+            q, k, v, self.k_cache, self.v_cache, attn_mask, start_pos, seq_len
+        )
+        op_output = torch.ops.llama.sdpa_with_kv_cache(
+            q, k, v, self.k_cache, self.v_cache, start_pos, seq_len, None, 0, True
+        )
+        self.assertTrue(torch.allclose(ref_output, op_output))
+
+        q = torch.rand((1, 1, self.n_heads_q, 4))
+        k = torch.rand((1, 1, self.n_heads_q, 4))
+        v = torch.rand((1, 1, self.n_heads_q, 4))
+        start_pos = 4
+        seq_len = q.size(1)
+        attn_mask = self.mask[start_pos : start_pos + seq_len, :]
+        attn_mask = attn_mask[:, : start_pos + seq_len]
+        ref_output = _sdpa_with_kv_cache_ref(
+            q, k, v, self.k_cache, self.v_cache, attn_mask, start_pos, seq_len
+        )
+        op_output = torch.ops.llama.sdpa_with_kv_cache(
+            q, k, v, self.k_cache, self.v_cache, start_pos, seq_len, None, 0, True
+        )
+        self.assertTrue(torch.allclose(ref_output, op_output))
+
+    def test_sdpa_with_cache_seq_len_llava_example(self):
+        self.n_heads_kv = 32
+        self.n_heads_q = 32
+        self.head_dim = 128
+        self.max_seq_len = 2048
+        self.setup_caches()
+        seq_len = 634
+        q = torch.rand((1, seq_len, self.n_heads_kv, self.head_dim))
+        k = torch.rand((1, seq_len, self.n_heads_kv, self.head_dim))
+        v = torch.rand((1, seq_len, self.n_heads_kv, self.head_dim))
+        start_pos = 0
+        attn_mask = self.mask[start_pos : start_pos + seq_len, :]
+        attn_mask = attn_mask[:, : start_pos + seq_len]
+        ref_output = _sdpa_with_kv_cache_ref(
+            q, k, v, self.k_cache, self.v_cache, attn_mask, start_pos, seq_len
+        )
+        op_output = torch.ops.llama.sdpa_with_kv_cache(
+            q, k, v, self.k_cache, self.v_cache, start_pos, seq_len, None, 0, True
+        )
+        self.assertTrue(torch.allclose(ref_output, op_output))
+
+        q = torch.rand((1, 1, self.n_heads_kv, self.head_dim))
+        k = torch.rand((1, 1, self.n_heads_kv, self.head_dim))
+        v = torch.rand((1, 1, 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, :]
+        attn_mask = attn_mask[:, : start_pos + seq_len]
+        ref_output = _sdpa_with_kv_cache_ref(
+            q, k, v, self.k_cache, self.v_cache, attn_mask, start_pos, seq_len
+        )
+        op_output = torch.ops.llama.sdpa_with_kv_cache(
+            q, k, v, self.k_cache, self.v_cache, start_pos, seq_len, None, 0, True
+        )
+        self.assertTrue(torch.allclose(ref_output, op_output))