llama : add llama_kv_cache_compress (EXPERIMENTAL)

examples/passkey/passkey.cpp

@@ -146,9 +146,10 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_cache_seq_add(ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_cache_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_cache_seq_div(ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
-            llama_kv_cache_update (ctx);
+            llama_kv_cache_compress(ctx, 0);
+            llama_kv_cache_update  (ctx);
 
             n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
         }

llama.cpp

@@ -1733,6 +1733,12 @@ struct llama_kv_cache {
     // computed before each graph build
     uint32_t n = 0;
 
+    ggml_type type_k = GGML_TYPE_F16;
+    ggml_type type_v = GGML_TYPE_F16;
+
+    // if non-negative, compress data on next update
+    llama_pos compress_delta = -1;
+
     std::vector<llama_kv_cell> cells;
 
     std::vector<struct ggml_tensor *> k_l; // per layer
@@ -1968,8 +1974,8 @@ struct llama_context {
 static bool llama_kv_cache_init(
              struct llama_kv_cache & cache,
                  const llama_model & model,
-                         ggml_type   ktype,
+                         ggml_type   type_k,
-                         ggml_type   vtype,
+                         ggml_type   type_v,
                           uint32_t   n_ctx,
                               bool   offload) {
     const struct llama_hparams & hparams = model.hparams;
@@ -1984,6 +1990,9 @@ static bool llama_kv_cache_init(
     cache.size = n_ctx;
     cache.used = 0;
 
+    cache.type_k = type_k;
+    cache.type_v = type_v;
+
     cache.cells.clear();
     cache.cells.resize(n_ctx);
 
@@ -2024,8 +2033,8 @@ static bool llama_kv_cache_init(
 
     for (int i = 0; i < (int) n_layer; i++) {
         struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front();
-        ggml_tensor * k = ggml_new_tensor_1d(ctx, ktype, n_embd_k_gqa*n_ctx);
+        ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*n_ctx);
-        ggml_tensor * v = ggml_new_tensor_1d(ctx, vtype, n_embd_v_gqa*n_ctx);
+        ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*n_ctx);
         ggml_format_name(k, "cache_k_l%d", i);
         ggml_format_name(v, "cache_v_l%d", i);
         cache.k_l.push_back(k);
@@ -2265,6 +2274,10 @@ static llama_pos llama_kv_cache_seq_pos_max(struct llama_kv_cache & cache, llama
     return result;
 }
 
+static void llama_kv_cache_compress(struct llama_kv_cache & cache, llama_pos delta) {
+    cache.compress_delta = delta;
+}
+
 //
 // model loading and saving
 //
@@ -8034,6 +8047,191 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) {
             }
         }
     }
+
+    // compress the KV cache data if needed:
+    //
+    //   - determine which KV cell pairs (i0, i1) to merge:
+    //
+    //     abs(cell[i0].pos - cell[i1].pos) <= compress_delta
+    //
+    //   - move the KV cache to the Host memory for easier maniiplation
+    //   - processing is done layer-by-layer
+    //   - convert the KV data to F32
+    //   - merge the KV data (different ways to merge)
+    //   - convert the KV data back to the original type
+    //   - move the KV cache back to the device memory
+    //   - update the KV cache metadata
+    //
+    // as a side effect, the new KV cache is defragmented
+    //
+    if (lctx.kv_self.compress_delta >= 0) {
+        auto & kv_self = lctx.kv_self;
+
+        const auto & hparams = lctx.model.hparams;
+
+        const uint32_t n_layer      = hparams.n_layer;
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa();
+        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa();
+        const uint32_t kv_size      = kv_self.size;
+
+        std::vector<uint8_t> buf_q;
+
+        std::vector<float> buf_src_f32;
+        std::vector<float> buf_dst_f32;
+
+        const int64_t t_start = ggml_time_us();
+
+        struct c_pair { uint32_t i0, i1; };
+        struct c_info { bool merged; uint32_t id, cnt;};
+
+        std::vector<c_info> infos(kv_size, { false, 0, 0 });
+
+        // the destination cell in the new KV cache
+        uint32_t id = 0;
+
+        // number of pairs merged
+        uint32_t n_merges = 0;
+
+        // determine which KV cells to merge
+        for (uint32_t i0 = 0; i0 < kv_size; ++i0) {
+            const auto & cell0 = kv_self.cells[i0];
+
+            if (!cell0.is_empty() && !infos[i0].merged) {
+                infos[i0] = { true, id, 0 };
+                infos[id].cnt = 1;
+
+                const llama_pos p0 = cell0.pos;
+
+                for (uint32_t i1 = i0 + 1; i1 < kv_size; ++i1) {
+                    const auto & cell1 = kv_self.cells[i1];
+
+                    if (i0 != i1 && cell0.is_same_seq(cell1)) {
+                        const llama_pos p1 = cell1.pos;
+
+                        if (std::abs(p0 - p1) <= kv_self.compress_delta) {
+                            infos[i1] = { true, id, 0 };
+                            infos[id].cnt++;
+                            n_merges++;
+                        }
+                    }
+                }
+
+                if (i0 != id) {
+                    kv_self.cells[id] = cell0;
+                }
+
+                id++;
+            }
+        }
+
+        kv_self.head = id;
+        kv_self.used = id;
+
+        for (uint32_t i = id; i < kv_size; ++i) {
+            kv_self.cells[i] = llama_kv_cell();
+        }
+
+        LLAMA_LOG_INFO("(tmp log) KV compress pairs: %u\n", n_merges);
+
+        ggml_type_traits_t tt_k;
+        ggml_type_traits_t tt_v;
+
+        tt_k = ggml_internal_get_type_traits(kv_self.type_k);
+        tt_v = ggml_internal_get_type_traits(kv_self.type_v);
+
+        for (uint32_t il = 0; il < n_layer; ++il) {
+            // update keys
+            {
+                const int64_t ne = n_embd_k_gqa*kv_size;
+
+                const size_t k_size = ggml_row_size(kv_self.k_l[il]->type, ne);
+
+                buf_q.resize(k_size);
+
+                buf_src_f32.resize(ne);
+                buf_dst_f32.resize(ne);
+
+                ggml_backend_tensor_get(kv_self.k_l[il], buf_q.data(), 0, buf_q.size());
+
+                tt_k.to_float(buf_q.data(), buf_src_f32.data(), ne);
+
+                std::fill(buf_dst_f32.begin(), buf_dst_f32.end(), 0);
+
+                for (uint32_t i = 0; i < kv_size; ++i) {
+                    if (!infos[i].merged) {
+                        continue;
+                    }
+
+                    const uint32_t id = infos[i].id;
+
+                    // merge using averaging
+                    {
+                        const float scale = 1.0f/float(infos[id].cnt);
+
+                        const int64_t os =  i*n_embd_k_gqa;
+                        const int64_t od = id*n_embd_k_gqa;
+
+                        for (uint32_t j = 0; j < n_embd_k_gqa; ++j) {
+                            buf_dst_f32[od + j] += buf_src_f32[os + j]*scale;
+                        }
+                    }
+                }
+
+                tt_k.from_float(buf_dst_f32.data(), buf_q.data(), ne);
+
+                ggml_backend_tensor_set(kv_self.k_l[il], buf_q.data(), 0, buf_q.size());
+            }
+
+            // update values (note: they are transposed)
+            {
+                const int64_t ne = n_embd_v_gqa*kv_size;
+
+                const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, ne);
+
+                buf_q.resize(v_size);
+
+                buf_src_f32.resize(ne);
+                buf_dst_f32.resize(ne);
+
+                ggml_backend_tensor_get(kv_self.v_l[il], buf_q.data(), 0, buf_q.size());
+
+                tt_v.to_float(buf_q.data(), buf_src_f32.data(), ne);
+
+                std::fill(buf_dst_f32.begin(), buf_dst_f32.end(), 0);
+
+                for (uint32_t i = 0; i < kv_size; ++i) {
+                    if (!infos[i].merged) {
+                        continue;
+                    }
+
+                    const uint32_t id = infos[i].id;
+
+                    // merge using averaging
+                    {
+                        const float scale = 1.0f/float(infos[id].cnt);
+                        //printf("i: %d -> id: %d, scale: %f\n", i, id, scale);
+
+                        const int64_t os =  i;
+                        const int64_t od = id;
+
+                        for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                            buf_dst_f32[od + j*kv_size] += buf_src_f32[os + j*kv_size]*scale;
+                        }
+                    }
+                }
+
+                tt_v.from_float(buf_dst_f32.data(), buf_q.data(), ne);
+
+                ggml_backend_tensor_set(kv_self.v_l[il], buf_q.data(), 0, buf_q.size());
+            }
+        }
+
+        const int64_t t_end = ggml_time_us();
+
+        LLAMA_LOG_INFO("(tmp log) KV compress time: %.3f ms\n", (t_end - t_start)/1000.0);
+
+        kv_self.compress_delta = -1;
+    }
 }
 
 //
@@ -12083,6 +12281,10 @@ llama_pos llama_kv_cache_seq_pos_max(struct llama_context * ctx, llama_seq_id se
     return llama_kv_cache_seq_pos_max(ctx->kv_self, seq_id);
 }
 
+void llama_kv_cache_compress(struct llama_context * ctx, llama_pos delta) {
+    llama_kv_cache_compress(ctx->kv_self, delta);
+}
+
 void llama_kv_cache_update(struct llama_context * ctx) {
     llama_kv_cache_update_internal(*ctx);
 }

llama.h

@@ -552,6 +552,11 @@ extern "C" {
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
+    // [EXPERIMENTAL] Compress the data in the KV cache
+    LLAMA_API void llama_kv_cache_compress(
+            struct llama_context * ctx,
+                       llama_pos   delta);
+
     // Apply the KV cache updates (such as K-shifts) to the KV data
     LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);