Implement get_freqs for RopeWithAttentionSink

This PR implements the `get_freqs` function for `RopeWithAttentionSink`. It returns the `freqs_cos` and `freqs_sin` for given `input_pos` and `seq_len` after shifting tokens in the pre-computed `freqs_cos` and `freq_sin`.

Differential Revision: [D66525306](https://our.internmc.facebook.com/intern/diff/D66525306/)

ghstack-source-id: 255582545
Pull Request resolved: #7100

Co-authored-by: Lunwen He <[email protected]>

Author: 2 people

examples/models/llama/source_transformation/attention_sink.py

@@ -7,6 +7,8 @@
 # Components for supporting Attention Sink. See
 # https://arxiv.org/abs/2309.17453 for more details about Attention Sink.
 
+from typing import Optional
+
 import torch
 
 from executorch.examples.models.llama.llama_transformer import ModelArgs, Rope
@@ -23,12 +25,37 @@ class RopeWithAttentionSink(Rope):
     in KVCache instead of positions in the actual text.
     """
 
-    def __init__(self, params: ModelArgs):
+    def __init__(
+        self,
+        params: ModelArgs,
+        window_size: int,
+        sink_size: int,
+        eviction_batch_size: int,
+    ):
         super().__init__(params)
         if self.params.use_hf_rope:
             self.apply_rotary_emb_to_k = hf_apply_rotary_emb_to_k
         else:
             self.apply_rotary_emb_to_k = apply_rotary_emb_to_k
+        self.max_seq_length = window_size + sink_size
+        assert self.max_seq_length == self.params.max_seq_len
+        self.eviction_batch_size = eviction_batch_size
+        self.position_shift = 0
+
+    def get_freqs(self, input_pos: Optional[torch.Tensor], seq_len: int):
+        assert input_pos is not None
+
+        input_pos_item = input_pos.item()
+        torch._check_is_size(input_pos_item)
+        if input_pos_item + self.position_shift + seq_len > self.max_seq_length:
+            # There are not enough spaces in the cache to store the new tokens.
+            # We need to evict some old tokens and shift some recent tokens.
+            num_to_evict = max(
+                input_pos_item + self.position_shift - self.max_seq_length + seq_len,
+                self.eviction_batch_size,
+            )
+            self.position_shift -= num_to_evict  # pyre-ignore [8]
+        return super().get_freqs(input_pos + self.position_shift, seq_len)
 
     def rerotate_k(
         self,

examples/models/llama/source_transformation/test_attention_sink.py

@@ -17,10 +17,57 @@
 
 class RopeWithAttentionSinkTest(unittest.TestCase):
 
+    def _init_rope(self, params: ModelArgs, eviction_batch_size: int):
+        return RopeWithAttentionSink(
+            params=params,
+            window_size=252,
+            sink_size=4,
+            eviction_batch_size=eviction_batch_size,
+        )
+
     def setUp(self):
         torch.manual_seed(42)
-        self.params = ModelArgs(use_kv_cache=True, enable_dynamic_shape=True)
-        self.rope_with_attention_sink = RopeWithAttentionSink(params=self.params)
+        self.params = ModelArgs(
+            use_kv_cache=True, enable_dynamic_shape=True, max_seq_len=256
+        )
+        self.rope_with_attention_sink = self._init_rope(
+            params=self.params, eviction_batch_size=1
+        )
+
+    @parameterized.expand(
+        [
+            [0, 10, 1, 0],  # No shift
+            [250, 10, 1, 246],  # Some shift
+            [256, 10, 1, 246],  # All shift
+            [0, 10, 30, 0],  # No shift with batch eviction
+            [250, 10, 30, 220],  # Some shift with batch eviction
+            [256, 10, 30, 226],  # All shift with batch eviction
+        ]
+    )
+    def test_get_freqs(
+        self, input_pos, seq_len, eviction_batch_size, expected_result_pos
+    ):
+        self.rope_with_attention_sink = self._init_rope(
+            params=self.params, eviction_batch_size=eviction_batch_size
+        )
+
+        freqs_cos, freqs_sin = self.rope_with_attention_sink.get_freqs(
+            input_pos=torch.tensor([input_pos], dtype=torch.int32),
+            seq_len=seq_len,
+        )
+
+        torch.testing.assert_close(
+            freqs_cos,
+            self.rope_with_attention_sink.freqs_cos.narrow(
+                0, expected_result_pos, seq_len
+            ),
+        )
+        torch.testing.assert_close(
+            freqs_sin,
+            self.rope_with_attention_sink.freqs_sin.narrow(
+                0, expected_result_pos, seq_len
+            ),
+        )
 
     @parameterized.expand(
         [