[Feature] ISS-60: Implement Self Extend

gemma/configs.h

@@ -127,6 +127,10 @@ struct LayerConfig {
   size_t conv1d_width = 0;
   bool ff_biases = false;
   bool softmax_attn_output_biases = false;
+  bool self_extend = false;
+  size_t ngb_size = 0;
+  size_t grp_size = 1;
+
   PostNormType post_norm = PostNormType::None;
   LayerAttentionType type = LayerAttentionType::kGemma;
   ActivationType activation = ActivationType::Gelu;

gemma/gemma-inl.h

@@ -312,21 +312,29 @@ class GemmaAttention {
                 const size_t interleaved_idx = task / layer_config_.kv_heads;
                 const size_t query_idx = interleaved_idx % num_queries_;
                 const size_t batch_idx = interleaved_idx / num_queries_;
-                const size_t pos = queries_pos_[query_idx] + batch_idx;
+                size_t pos = queries_pos_[query_idx] + batch_idx;
                 const size_t cache_pos = div_seq_len_.Remainder(pos);
                 const size_t kv_offset = cache_pos * cache_pos_size_ +
                                          layer_ * cache_layer_size_ +
                                          head * layer_config_.qkv_dim * 2;
                 KVCache& kv_cache = kv_caches_[query_idx];
+
+                const size_t grp_size = layer_config_.grp_size;
+                const size_t ngb_size = layer_config_.ngb_size;
+                const bool self_extend = layer_config_.self_extend;
+
                 float* HWY_RESTRICT kv = kv_cache.kv_cache.get() + kv_offset;
                 const float* HWY_RESTRICT mha_kv =
                     activations_.q.Batch(interleaved_idx) + head * q_stride_ +
                     layer_config_.qkv_dim;
 
+                // When embedding position, we will use grouped key position
+                if (self_extend && pos > ngb_size) {
+                  pos /= grp_size;
+                }
                 // Copy from `q` if MHA, or apply in-place.
                 PositionalEncodingQK(is_mha_ ? mha_kv : kv, pos, layer_, 1.0f,
                                      kv);
-
                 // If MHA, also copy V into KVCache.
                 if (is_mha_) {
                   hwy::CopyBytes(mha_kv + layer_config_.qkv_dim,
@@ -411,12 +419,22 @@ class GemmaAttention {
                 const size_t batch_idx = interleaved_idx / num_queries_;
                 const size_t head_offset =
                     (head / kHeadGroups) * layer_config_.qkv_dim * 2;
+
+                const size_t grp_size = layer_config_.grp_size;
+                const size_t ngb_size = layer_config_.ngb_size;
+                const bool self_extend = layer_config_.self_extend;
                 KVCache& kv_cache = kv_caches_[query_idx];
                 float* HWY_RESTRICT q =
                     activations_.q.Batch(interleaved_idx) + head * q_stride_;
 
                 // Apply rope and scaling to Q.
-                const size_t pos = queries_pos_[query_idx] + batch_idx;
+                size_t pos = queries_pos_[query_idx] + batch_idx;
+                if (self_extend && pos > ngb_size) {
+                  const size_t grp_pos = pos / grp_size;
+                  const size_t shift = ngb_size - ngb_size / grp_size;
+                  const size_t shifted_grouped_pos = grp_pos + shift;
+                  pos = shifted_grouped_pos;
+                }
                 PositionalEncodingQK(q, pos, layer_, query_scale, q);
 
                 const size_t start_pos = StartPos(pos, layer_);