Add configurables for norm/rope/activation/scale/residual connection.

PiperOrigin-RevId: 648971168

backprop/backward-inl.h

@@ -355,7 +355,7 @@ void CrossEntropyLossBackwardPass(const Prompt& prompt,
   static constexpr size_t kLayers = TConfig::kLayers;
   const float kEmbScaling = EmbeddingScaling<TConfig>();
   static_assert(!TConfig::kAbsolutePE);
-  static_assert(!TConfig::kPostNormScale);
+  static_assert(TConfig::kPostNorm == PostNormType::None);
   static_assert(TConfig::kKVHeads == 1);
 
   HWY_DASSERT(prompt.context_size > 0);

backprop/backward_scalar_test.cc

@@ -388,7 +388,7 @@ struct TestConfig : ConfigCapNoSSM {
       FixedLayerConfig<2>(LayerAttentionType::kGemma);
   static constexpr int kLayers = kLayerConfig.size();
   static constexpr bool kAbsolutePE = false;
-  static constexpr bool kPostNormScale = false;
+  static constexpr PostNormType kPostNorm = PostNormType::None;
 
   static constexpr int kKVHeads = 1;
   static constexpr int kGemmaLayers = kLayers;

backprop/backward_test.cc

@@ -193,7 +193,7 @@ struct TestConfig : public ConfigCapNoSSM {
       FixedLayerConfig<2>(LayerAttentionType::kGemma);
   static constexpr int kLayers = kLayerConfig.size();
   static constexpr bool kAbsolutePE = false;
-  static constexpr bool kPostNormScale = false;
+  static constexpr PostNormType kPostNorm = PostNormType::None;
 
   static constexpr int kKVHeads = 1;
   static constexpr int kGemmaLayers = kLayers;

backprop/forward-inl.h

@@ -233,7 +233,7 @@ float CrossEntropyLossForwardPass(const std::vector<int>& prompt,
   static constexpr size_t kLayers = TConfig::kLayers;
   const float kEmbScaling = EmbeddingScaling<TConfig>();
   static_assert(!TConfig::kAbsolutePE);
-  static_assert(!TConfig::kPostNormScale);
+  static_assert(TConfig::kPostNorm == PostNormType::None);
   static_assert(TConfig::kKVHeads == 1);
 
   HWY_DASSERT(context_size > 0);

gemma/configs.h

@@ -52,6 +52,32 @@ enum class LayerAttentionType {
   kGriffinRecurrentBlock,
 };
 
+// Post attention and ffw normalization type.
+enum class PostNormType {
+  None,
+  Scale,
+};
+
+// Post qk projection operation type.
+enum class PostQKType {
+  Rope,
+};
+
+// FFW activation function.
+enum class ActivationType {
+  Gelu,
+};
+
+// Attention query scale.
+enum class QueryScaleType {
+  Sqrt,
+};
+
+// Residual connection type.
+enum class ResidualType {
+  Add,
+};
+
 template <size_t kNum>
 constexpr std::array<LayerAttentionType, kNum> FixedLayerConfig(
     LayerAttentionType type) {
@@ -107,6 +133,11 @@ struct ConfigNoSSM {
   static constexpr bool kUseLocalAttention = false;
   static constexpr bool kInterleaveQKV = true;
   static constexpr int kNumTensorScales = 0;
+
+  static constexpr PostQKType kPostQK = PostQKType::Rope;
+  static constexpr ActivationType kActivation = ActivationType::Gelu;
+  static constexpr QueryScaleType kQueryScale = QueryScaleType::Sqrt;
+  static constexpr ResidualType kResidual = ResidualType::Add;
 };
 
 struct ConfigNoCapNoSSM : ConfigNoSSM {
@@ -143,7 +174,7 @@ struct ConfigGemma27B : public ConfigCapNoSSM {
   static constexpr int kQKVDim = 128;  // query size == key size == value size
   static constexpr int kTopK = gcpp::kTopK;
   static constexpr bool kAbsolutePE = false;
-  static constexpr bool kPostNormScale = true;
+  static constexpr PostNormType kPostNorm = PostNormType::Scale;
 };
 
 template <typename TWeight>
@@ -169,7 +200,7 @@ struct ConfigGemma9B : public ConfigCapNoSSM {
   static constexpr int kQKVDim = 256;  // query size == key size == value size
   static constexpr int kTopK = gcpp::kTopK;
   static constexpr bool kAbsolutePE = false;
-  static constexpr bool kPostNormScale = true;
+  static constexpr PostNormType kPostNorm = PostNormType::Scale;
 };
 
 template <typename TWeight>
@@ -191,7 +222,7 @@ struct ConfigGemma7B : public ConfigNoCapNoSSM {
   static constexpr int kQKVDim = 256;  // query size == key size == value size
   static constexpr int kTopK = gcpp::kTopK;
   static constexpr bool kAbsolutePE = false;
-  static constexpr bool kPostNormScale = false;
+  static constexpr PostNormType kPostNorm = PostNormType::None;
 };
 
 template <typename TWeight>
@@ -213,7 +244,7 @@ struct ConfigGemma2B : public ConfigNoCapNoSSM {
   static constexpr int kQKVDim = 256;  // query size == key size == value size
   static constexpr int kTopK = gcpp::kTopK;
   static constexpr bool kAbsolutePE = false;
-  static constexpr bool kPostNormScale = false;
+  static constexpr PostNormType kPostNorm = PostNormType::None;
 };
 
 template <typename TWeight>
@@ -235,7 +266,7 @@ struct ConfigGemmaTiny : public ConfigNoSSM {
   static constexpr int kQKVDim = 16;  // query size == key size == value size
   static constexpr int kTopK = gcpp::kTopK;
   static constexpr bool kAbsolutePE = false;
-  static constexpr bool kPostNormScale = false;
+  static constexpr PostNormType kPostNorm = PostNormType::None;
 
   static constexpr float kAttCap = 0.0f;
   // This is required for optimize_test to pass.
@@ -294,7 +325,7 @@ struct ConfigGriffin2B {
   static constexpr int kQKVDim = 256;  // query size == key size == value size
   static constexpr int kTopK = gcpp::kTopK;
   static constexpr bool kAbsolutePE = false;
-  static constexpr bool kPostNormScale = false;
+  static constexpr PostNormType kPostNorm = PostNormType::None;
 
   // No SoftCap.
   static constexpr float kAttCap = 0.0f;
@@ -308,6 +339,9 @@ struct ConfigGriffin2B {
   static constexpr bool kUseLocalAttention = true;
   static constexpr bool kInterleaveQKV = false;
   static constexpr int kNumTensorScales = 140;
+  static constexpr PostQKType kPostQK = PostQKType::Rope;
+  static constexpr QueryScaleType kQueryScale = QueryScaleType::Sqrt;
+  static constexpr ResidualType kResidual = ResidualType::Add;
 };
 
 }  // namespace gcpp

gemma/gemma.cc

@@ -295,8 +295,9 @@ HWY_NOINLINE void Attention(
   constexpr size_t kHeads = TConfig::kHeads;
   constexpr size_t kKVHeads = TConfig::kKVHeads;
   constexpr size_t kSeqLen = TConfig::kSeqLen;
-  GEMMA_CONSTEXPR_SQRT const float kQueryScale =
+  GEMMA_CONSTEXPR_SQRT float kQueryScale =
       1.0f / Sqrt(static_cast<float>(kQKVDim));
+
   constexpr bool kIsMHA = TActivations::kIsMHA;  // Multi-Head Attention
   const size_t batch_start = batch_and_query_start / num_queries;
   const size_t num_tokens_and_queries = num_tokens * num_queries;
@@ -350,7 +351,9 @@ HWY_NOINLINE void Attention(
           // Skip past the Q part of `q`, and copy KV to `kv`.
           memcpy(kv, q + kQKVDim, 2 * kQKVDim * sizeof(float));
         }
+        if (TConfig::kPostQK == PostQKType::Rope) {
           Rope(kv, TConfig::kUseHalfRope ? kQKVDim / 2 : kQKVDim, pos);
+        }
       });
 
   static_assert((kHeads % kKVHeads) == 0,
@@ -373,7 +376,10 @@ HWY_NOINLINE void Attention(
         activations.att.data() + head * kSeqLen
         + batch_and_query_idx * kHeads * kSeqLen;
 
+    if (TConfig::kPostQK == PostQKType::Rope) {
       Rope(q, TConfig::kUseHalfRope ? kQKVDim / 2 : kQKVDim, pos);
+    }
+
     MulByConst(kQueryScale, q, kQKVDim);
 
     // Compute Q dot K scores
@@ -465,10 +471,12 @@ HWY_NOINLINE void FFW(Activations<TConfig, kBatchSize>& activations,
     namespace hn = hwy::HWY_NAMESPACE;
     using DF = hn::ScalableTag<float>;
     using VF = hn::Vec<DF>;
+    if (TConfig::kActivation == ActivationType::Gelu) {
       hn::Transform1(DF(), activations.C1.data(), kFFHiddenDim * num_tokens,
                      activations.C2.data(), [](DF df, VF v, VF mul) HWY_ATTR {
                        return hn::Mul(mul, Gelu(df, v));
                      });
+    }
 
     MatMul_4x4_Batch<kFFHiddenDim, kModelDim>(num_tokens, activations.C1.data(),
                                               layer_weights->linear_w.data(),
@@ -560,29 +568,34 @@ HWY_NOINLINE void TransformerLayer(
                       layer_weights, kv_caches, pool);
     }
   }
-  if (TConfig::kPostNormScale) {
+
+  if (TConfig::kPostNorm == PostNormType::Scale) {
     RMSNormInplaceBatched<kBatchSize * kQueryBatchSize>(
         num_tokens_and_queries,
         layer_weights->post_attention_norm_scale.data(),
         activations.att_post2.data(), kModelDim);
   }
-  AddFromBatched<kBatchSize * kQueryBatchSize>(num_tokens_and_queries,
+  if (TConfig::kResidual == ResidualType::Add) {
-                                               activations.att_post2.data(),
+    AddFromBatched<kBatchSize * kQueryBatchSize>(
+        num_tokens_and_queries, activations.att_post2.data(),
         activations.x.data(), kModelDim);
+  }
   RMSNormBatched<kBatchSize * kQueryBatchSize>(
       num_tokens_and_queries, activations.x.data(),
       layer_weights->pre_ffw_norm_scale.data(),
       activations.bf_pre_ffw_rms_out.data(), kModelDim);
   FFW<TConfig, kBatchSize * kQueryBatchSize>(
       activations, num_tokens_and_queries, layer_weights, pool);
-  if (TConfig::kPostNormScale) {
+  if (TConfig::kPostNorm == PostNormType::Scale) {
     RMSNormInplaceBatched<kBatchSize * kQueryBatchSize>(
         num_tokens_and_queries, layer_weights->post_ffw_norm_scale.data(),
         activations.ffw_out.data(), kModelDim);
   }
+  if (TConfig::kResidual == ResidualType::Add) {
     AddFromBatched<kBatchSize * kQueryBatchSize>(
         num_tokens_and_queries, activations.ffw_out.data(),
         activations.x.data(), kModelDim);
+  }
 }
 
 template <class TConfig, size_t kBatchSize, size_t kQueryBatchSize>

gemma/weights.h

@@ -50,7 +50,7 @@ struct CompressedLayer {
   static constexpr size_t kGatingEinsumWSize = 2 * kFFHiddenDim * kModelDim;
   static constexpr size_t kConv1dWidth = TConfig::kConv1dWidth;
   static constexpr bool kFFBiases = TConfig::kFFBiases;
-  static constexpr bool kPostNormScale = TConfig::kPostNormScale;
+  static constexpr PostNormType kPostNorm = TConfig::kPostNorm;
   static constexpr size_t kAOBiasDim =
       TConfig::kSoftmaxAttnOutputBiases ? kModelDim : 0;
   static constexpr size_t kGriffinDim =
@@ -86,9 +86,10 @@ struct CompressedLayer {
   // We don't yet have an RMSNorm that accepts all Weight.
   ArrayT<WeightF32OrBF16, kModelDim> pre_attention_norm_scale;
   ArrayT<WeightF32OrBF16, kModelDim> pre_ffw_norm_scale;
-  ArrayT<WeightF32OrBF16, kPostNormScale ? kModelDim : 0>
+  ArrayT<WeightF32OrBF16, kPostNorm == PostNormType::Scale ? kModelDim : 0>
       post_attention_norm_scale;
-  ArrayT<WeightF32OrBF16, kPostNormScale ? kModelDim : 0> post_ffw_norm_scale;
+  ArrayT<WeightF32OrBF16, kPostNorm == PostNormType::Scale ? kModelDim : 0>
+      post_ffw_norm_scale;
 
   ArrayT<float, kFFBiases ? 2 * kFFHiddenDim : 0> ffw_gating_biases;
   ArrayT<float, kFFBiases ? kModelDim : 0> ffw_output_biases;
@@ -267,7 +268,7 @@ void ForEachTensor(RawWeightsPtr raw_weights,
       GEMMA_CALL_FUNC("gr_a", griffin.a);
     }
     GEMMA_CALL_FUNC("pre_att_ns", pre_attention_norm_scale);
-    if (TConfig::kPostNormScale) {
+    if (TConfig::kPostNorm == PostNormType::Scale) {
       GEMMA_CALL_FUNC("post_att_ns", post_attention_norm_scale);
       GEMMA_CALL_FUNC("post_ff_ns", post_ffw_norm_scale);
     }
@@ -331,7 +332,7 @@ void ForEachTensor(RawWeightsPtr raw_weights,
   GEMMA_CALL_LAYER_FUNC ## N("gating_ein", gating_einsum_w);                  \
   GEMMA_CALL_LAYER_FUNC ## N("linear_w", linear_w);                           \
   GEMMA_CALL_LAYER_FUNC ## N("pre_att_ns", pre_attention_norm_scale);         \
-  if (TConfig::kPostNormScale) {                                              \
+  if (TConfig::kPostNorm == PostNormType::Scale) {                            \
     GEMMA_CALL_LAYER_FUNC ## N("post_att_ns", post_attention_norm_scale);     \
     GEMMA_CALL_LAYER_FUNC ## N("post_ff_ns", post_ffw_norm_scale);            \
   }                                                                           \

gemma/weights_raw.h

@@ -25,6 +25,7 @@
 #include <random>
 
 #include "gemma/common.h"
+#include "gemma/configs.h"
 #include "hwy/aligned_allocator.h"
 #include "hwy/base.h"
 #include "hwy/contrib/thread_pool/thread_pool.h"
@@ -46,7 +47,7 @@ struct Layer {
   static constexpr size_t kGatingEinsumWSize = 2 * kFFHiddenDim * kModelDim;
   static constexpr size_t kConv1dWidth = TConfig::kConv1dWidth;
   static constexpr bool kFFBiases = TConfig::kFFBiases;
-  static constexpr bool kPostNormScale = TConfig::kPostNormScale;
+  static constexpr PostNormType kPostNorm = TConfig::kPostNorm;
   static constexpr size_t kAOBiasDim =
       TConfig::kSoftmaxAttnOutputBiases ? kModelDim : 0;
   static constexpr size_t kGriffinDim =
@@ -78,8 +79,10 @@ struct Layer {
   std::array<T, kModelDim * kFFHiddenDim> linear_w;
   std::array<T, kModelDim> pre_attention_norm_scale;
   std::array<T, kModelDim> pre_ffw_norm_scale;
-  std::array<T, kPostNormScale ? kModelDim : 0> post_attention_norm_scale;
+  std::array<T, kPostNorm == PostNormType::Scale ? kModelDim : 0>
-  std::array<T, kPostNormScale ? kModelDim : 0> post_ffw_norm_scale;
+      post_attention_norm_scale;
+  std::array<T, kPostNorm == PostNormType::Scale ? kModelDim : 0>
+      post_ffw_norm_scale;
 
   std::array<T, kFFBiases ? 2 * kFFHiddenDim : 0> ffw_gating_biases;
   std::array<T, kFFBiases ? kModelDim : 0> ffw_output_biases;

util/compress_weights.cc

@@ -159,7 +159,7 @@ struct LoadRawWeightsT {
       SCALE_WEIGHTS(linear_w);
       READ_WEIGHTS(pre_attention_norm_scale);
       READ_WEIGHTS(pre_ffw_norm_scale);
-      if (TConfig::kPostNormScale) {
+      if (TConfig::kPostNorm == PostNormType::Scale) {
         READ_WEIGHTS(post_attention_norm_scale);
         READ_WEIGHTS(post_ffw_norm_scale);
       }