Report a single-sequence cache's length on SequenceControl

backends/mlx/runtime/MLXSequenceCache.h

@@ -875,7 +875,7 @@ class MLXSequenceCache final : public cache::Cache,
   int capacity() const override {
     return capacity_;
   }
-  int length() const {
+  int length() const override {
     return length_;
   }
   // Includes persistent growth in a partially constructed forward.

extension/llm/cache/cache.h

@@ -90,6 +90,9 @@ class ET_EXPERIMENTAL SequenceControl : public CacheControl {
 
   // Room for n more tokens within capacity.
   virtual bool can_extend(int n = 1) const = 0;
+  // Where the sequence stands: one past its newest position, and so where it
+  // continues. Positions are dense from 0, so this is also how many it holds.
+  virtual int length() const = 0;
   // Keep [0, position), dropping the rest. False = a target it has not
   // reached, or one older than an evicting layer still retains.
   virtual bool rewind(int position) = 0;

extension/llm/cache/sequence_cache.h

@@ -189,9 +189,7 @@ class ET_EXPERIMENTAL SequenceCache : public Cache,
     length_ = 0;
     written_ = 0;
   }
-  // Positions are dense from 0, so this is both what the sequence holds and
-  // where its next token goes.
-  int length() const {
+  int length() const override {
     return length_;
   }
   // Whether rewind(position) would be accepted, without moving the length.

extension/llm/cache/test/cache_test.cpp

@@ -244,6 +244,21 @@ TEST_F(CacheTest, FaceRecoveryReturnsSameObject) {
   EXPECT_EQ(plan->read[0].len, 1);
 }
 
+TEST_F(CacheTest, ControlFaceReportsTheLength) {
+  SequenceCache cache(CacheGeometry{{flat_layer()}}, config(8));
+  SequenceControl* control = static_cast<Cache&>(cache).as<SequenceControl>();
+  ASSERT_NE(control, nullptr);
+  EXPECT_EQ(control->length(), 0);
+  auto plan = cache.plan(0, 0, 3);
+  ASSERT_TRUE(plan.has_value());
+  cache.commit(*plan);
+  EXPECT_EQ(control->length(), 3);
+  ASSERT_TRUE(control->rewind(1));
+  EXPECT_EQ(control->length(), 1);
+  control->clear();
+  EXPECT_EQ(control->length(), 0);
+}
+
 TEST_F(CacheTest, NullFaceIdsNeverMatch) {
   using executorch::extension::llm::cache::same_face;
   EXPECT_FALSE(same_face(nullptr, nullptr));