Fix Griffin model:

- use HalfRope position encodings
- zero-initialize the caches for each Generate at position 0

The lack of the latter made the tests in gemma_test dependent on each other. The new entropy values are the same now independent of which tests are executed before.

PiperOrigin-RevId: 694230456

evals/gemma_test.cc

@@ -246,7 +246,7 @@ TEST_F(GemmaTest, CrossEntropySmall) {
       EXPECT_NEAR(entropy, 2.8f, 0.2f);
       break;
     case gcpp::Model::GRIFFIN_2B:
-      EXPECT_NEAR(entropy, 1.57f, 0.02f);
+      EXPECT_NEAR(entropy, 2.61f, 0.02f);
       break;
     case gcpp::Model::GEMMA2_2B:
       EXPECT_NEAR(entropy, 1.14f, 0.02f);
@@ -277,7 +277,7 @@ TEST_F(GemmaTest, CrossEntropyJingleBells) {
       EXPECT_NEAR(entropy, 1.07f, 0.05f);
       break;
     case gcpp::Model::GRIFFIN_2B:
-      EXPECT_NEAR(entropy, 2.09f, 0.02f);
+      EXPECT_NEAR(entropy, 1.62f, 0.02f);
       break;
     case gcpp::Model::GEMMA2_2B:
       EXPECT_NEAR(entropy, 0.49f, 0.02f);
@@ -308,7 +308,7 @@ TEST_F(GemmaTest, CrossEntropyGettysburg) {
       EXPECT_NEAR(entropy, 0.75f, 0.1f);
       break;
     case gcpp::Model::GRIFFIN_2B:
-      EXPECT_NEAR(entropy, 0.86f, 0.02f);
+      EXPECT_NEAR(entropy, 0.71f, 0.02f);
       break;
     case gcpp::Model::GEMMA2_2B:
       EXPECT_NEAR(entropy, 0.20f, 0.02f);

gemma/configs.cc

@@ -183,7 +183,7 @@ static ModelConfig ConfigGriffin2B() {
       .softmax_attn_output_biases = true,
       .type = LayerAttentionType::kGriffinRecurrentBlock,
       .activation = ActivationType::Gelu,
-      .post_qk = PostQKType::Rope,
+      .post_qk = PostQKType::HalfRope,
   };
   config.layer_configs = {26, layer_config};
   for (size_t i = 2; i < config.layer_configs.size(); i += 3) {

gemma/configs_test.cc

@@ -397,7 +397,11 @@ void AssertMatch(const ModelConfig& config) {
     ASSERT_EQ(TConfig::kPostNorm, config.layer_configs[i].post_norm);
     ASSERT_EQ(TConfig::kLayerConfig[i], config.layer_configs[i].type);
     ASSERT_EQ(TConfig::kActivation, config.layer_configs[i].activation);
-    ASSERT_EQ(TConfig::kPostQK, config.layer_configs[i].post_qk);
+    PostQKType post_qk = TConfig::kPostQK;
+    if (TConfig::kUseHalfRope) {
+      post_qk = PostQKType::HalfRope;
+    }
+    ASSERT_EQ(post_qk, config.layer_configs[i].post_qk);
   }
 
   ASSERT_EQ(TConfig::kAttentionWindowSizes.size(),

gemma/gemma-inl.h

@@ -1240,8 +1240,12 @@ void GenerateT(const ModelWeightsStorage& model, Activations& activations,
                const QueriesPos& queries_prefix_end,
                const size_t query_idx_start, const KVCaches& kv_caches,
                TimingInfo& timing_info) {
-  const size_t vocab_size = model.Config().vocab_size;
-  const ModelWeightsPtrs<T>& weights = *model.GetWeightsOfType<T>();
+  // Griffin assumes that the recurrent block cache is zero-initialized.
+  for (int i = 0; i < kv_caches.size(); ++i) {
+    if (queries_pos_in[i] == 0) {
+      kv_caches[i].ZeroGriffinCache();  // No-op for non-Griffin models.
+    }
+  }
 
   // Copy so we can increment without requiring users to pass in a mutable span.
   std::vector<size_t> queries_pos_copy(queries_pos_in.cbegin(),
@@ -1268,7 +1272,7 @@ void GenerateT(const ModelWeightsStorage& model, Activations& activations,
   HWY_ASSERT(queries_pos_in.size() == num_queries);
   HWY_ASSERT(kv_caches.size() == num_queries);
   const hwy::Divisor div_seq_len(static_cast<uint32_t>(kv_caches[0].seq_len));
-
+  const ModelWeightsPtrs<T>& weights = *model.GetWeightsOfType<T>();
   size_t max_prompt_size = MaxQueryLength(queries_prompt);
   size_t max_generated_tokens = runtime_config.max_generated_tokens;
   RangeChecks(weights.weights_config, max_generated_tokens, max_prompt_size);
@@ -1314,6 +1318,7 @@ void GenerateT(const ModelWeightsStorage& model, Activations& activations,
                          0.0f);
   }
 
+  const size_t vocab_size = model.Config().vocab_size;
   const double gen_start = hwy::platform::Now();
   for (size_t gen = 0; gen < max_generated_tokens; ++gen) {
     // Decode generates one token per query and increments queries_mutable_pos.

gemma/kv_cache.cc

@@ -23,6 +23,17 @@
 
 namespace gcpp {
 
+void KVCache::ZeroGriffinCache() {
+  if (conv1d_cache_size != 0) {
+    hwy::ZeroBytes(conv1d_cache.get(),
+                   conv1d_cache_size * sizeof(conv1d_cache[0]));
+  }
+  if (rglru_cache_size != 0) {
+    hwy::ZeroBytes(rglru_cache.get(),
+                   rglru_cache_size * sizeof(rglru_cache[0]));
+  }
+}
+
 // prefill_tbatch_size is the maximum number of tokens from one query to
 // prefill at a time.
 KVCache KVCache::Create(const ModelConfig& weights_config,
@@ -37,9 +48,9 @@ KVCache KVCache::Create(const ModelConfig& weights_config,
     kv_cache.kv_cache =
         hwy::AllocateAligned<float>(kv_cache.seq_len * size_cache_pos);
   }
-  size_t num_griffin_layers = weights_config.NumLayersOfType(
-      LayerAttentionType::kGriffinRecurrentBlock);
 
+  const size_t num_griffin_layers = weights_config.NumLayersOfType(
+      LayerAttentionType::kGriffinRecurrentBlock);
   // TODO(patrickms): Add query batching support for Griffin.
   if (num_griffin_layers > 0) {
     size_t conv1d_width = 0;
@@ -49,20 +60,18 @@ KVCache KVCache::Create(const ModelConfig& weights_config,
     const size_t conv1d_cache_size =
         num_griffin_layers * (conv1d_width == 0 ? 0 : conv1d_width - 1) *
         weights_config.model_dim;
+    kv_cache.conv1d_cache_size = conv1d_cache_size;
     if (conv1d_cache_size != 0) {
       kv_cache.conv1d_cache = hwy::AllocateAligned<float>(conv1d_cache_size);
-      hwy::ZeroBytes(kv_cache.conv1d_cache.get(),
-                     conv1d_cache_size * sizeof(kv_cache.conv1d_cache[0]));
     }
 
     const size_t rglru_cache_size =
         num_griffin_layers * weights_config.model_dim;
+    kv_cache.rglru_cache_size = rglru_cache_size;
     if (rglru_cache_size != 0) {
       kv_cache.rglru_cache = hwy::AllocateAligned<float>(rglru_cache_size);
-      hwy::ZeroBytes(kv_cache.rglru_cache.get(),
-                     rglru_cache_size * sizeof(kv_cache.rglru_cache[0]));
     }
-  }  // kGriffinLayers
+  }  // num_griffin_layers
 
   return kv_cache;
 }

gemma/kv_cache.h

@@ -31,9 +31,15 @@ struct KVCache {
 
   // (kConv1dWidth - 1) * kModelDim * kGriffinLayers
   hwy::AlignedFreeUniquePtr<float[]> conv1d_cache;
+  size_t conv1d_cache_size = 0;
 
   // kModelDim * kGriffinLayers
   hwy::AlignedFreeUniquePtr<float[]> rglru_cache;
+  size_t rglru_cache_size = 0;
+
+  // Zero-initialize the Griffin recurrent block cache, i.e. the conv1d_cache
+  // and rglru_cache.
+  void ZeroGriffinCache();
 
   static KVCache Create(const ModelConfig& weights_config,
                         size_t prefill_tbatch_size);