kv-cache : prepare K/V buffers for separation

ggml-ci

src/llama-hparams.cpp

@@ -65,6 +65,46 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     return n_embd_head_v * n_head_kv;
 }
 
+bool llama_hparams::is_n_embd_k_gqa_variable() const {
+    const uint32_t val = n_embd_k_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (val != n_embd_k_gqa(il)) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+bool llama_hparams::is_n_embd_v_gqa_variable() const {
+    const uint32_t val = n_embd_v_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (val != n_embd_v_gqa(il)) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+uint32_t llama_hparams::n_embd_k_gqa_max() const {
+    uint32_t val = n_embd_k_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        val = std::max(val, n_embd_k_gqa(il));
+    }
+
+    return val;
+}
+
+uint32_t llama_hparams::n_embd_v_gqa_max() const {
+    uint32_t val = n_embd_v_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        val = std::max(val, n_embd_v_gqa(il));
+    }
+
+    return val;
+}
+
 uint32_t llama_hparams::n_embd_r() const {
     if (wkv_head_size != 0) {
         // for RWKV models

src/llama-hparams.h

@@ -189,6 +189,14 @@ struct llama_hparams {
     // dimension of value embeddings across all k-v heads
     uint32_t n_embd_v_gqa(uint32_t il = 0) const;
 
+    // true if any layer has a different n_embd_k_gqa/n_embd_v_gqa
+    bool is_n_embd_k_gqa_variable() const;
+    bool is_n_embd_v_gqa_variable() const;
+
+    // return the maximum n_embd_k_gqa/n_embd_v_gqa across all layers
+    uint32_t n_embd_k_gqa_max() const;
+    uint32_t n_embd_v_gqa_max() const;
+
     // dimension of the rolling state embeddings
     // corresponds to Mamba's conv_states size or RWKV's token_shift states size
     uint32_t n_embd_r() const;

src/llama-kv-cache-unified.cpp

@@ -68,14 +68,21 @@ llama_kv_cache_unified::llama_kv_cache_unified(
 
     cells.resize(kv_size);
 
+    // [TAG_V_CACHE_VARIABLE]
+    if (v_trans && hparams.is_n_embd_v_gqa_variable()) {
+        LLAMA_LOG_WARN("%s: the V embeddings have different sizes across layers and FA is not enabled - padding V cache to %d\n",
+                __func__, hparams.n_embd_v_gqa_max());
+    }
+
     for (uint32_t il = 0; il < n_layer_cache; il++) {
         if (filter && !filter(il)) {
             LLAMA_LOG_DEBUG("%s: layer %3d: skipped\n", __func__, il);
             continue;
         }
 
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
+        // [TAG_V_CACHE_VARIABLE]
-        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
+        const uint32_t n_embd_k_gqa =            hparams.n_embd_k_gqa(il);
+        const uint32_t n_embd_v_gqa = !v_trans ? hparams.n_embd_v_gqa(il) : hparams.n_embd_v_gqa_max();
 
         const char * dev_name = "CPU";
 
@@ -98,8 +105,8 @@ llama_kv_cache_unified::llama_kv_cache_unified(
         ggml_tensor * k;
         ggml_tensor * v;
 
-        k = ggml_new_tensor_2d(ctx, type_k, n_embd_k_gqa, kv_size);
+        k = ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, 1);
-        v = ggml_new_tensor_2d(ctx, type_v, n_embd_v_gqa, kv_size);
+        v = ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, 1);
 
         ggml_format_name(k, "cache_k_l%d", il);
         ggml_format_name(v, "cache_v_l%d", il);
@@ -785,11 +792,17 @@ ggml_tensor * llama_kv_cache_unified::get_k(ggml_context * ctx, int32_t il, uint
 
     auto * k = layers[ikv].k;
 
-    return ggml_view_3d(ctx, k,
+    const uint64_t kv_size      = get_size();
-            hparams.n_embd_head_k, hparams.n_head_kv(il), n_kv,
+    const uint64_t n_embd_k_gqa = k->ne[0];
+
+    assert(n_embd_k_gqa == hparams.n_embd_k_gqa(il));
+
+    return ggml_view_4d(ctx, k,
+            hparams.n_embd_head_k, hparams.n_head_kv(il), n_kv, 1,
             ggml_row_size(k->type, hparams.n_embd_head_k),
-            ggml_row_size(k->type, hparams.n_embd_k_gqa(il)),
+            ggml_row_size(k->type, n_embd_k_gqa),
-            0);
+            ggml_row_size(k->type, n_embd_k_gqa*kv_size),
+            ggml_row_size(k->type, n_embd_k_gqa*kv_size)*0);
 }
 
 ggml_tensor * llama_kv_cache_unified::get_v(ggml_context * ctx, int32_t il, uint32_t n_kv) const {
@@ -797,21 +810,29 @@ ggml_tensor * llama_kv_cache_unified::get_v(ggml_context * ctx, int32_t il, uint
 
     auto * v = layers[ikv].v;
 
+    const uint64_t kv_size      = get_size();
+    const uint64_t n_embd_v_gqa = v->ne[0];
+
+    // [TAG_V_CACHE_VARIABLE]
+    assert(n_embd_v_gqa >= hparams.n_embd_v_gqa(il));
+
     if (!v_trans) {
         // note: v->nb[1] <= v->nb[2]
-        return ggml_view_3d(ctx, v,
+        return ggml_view_4d(ctx, v,
-                hparams.n_embd_head_v, hparams.n_head_kv(il), n_kv,
+                hparams.n_embd_head_v, hparams.n_head_kv(il), n_kv, 1,
-                ggml_row_size(v->type, hparams.n_embd_head_v),    // v->nb[1]
+                ggml_row_size(v->type, hparams.n_embd_head_v), // v->nb[1]
-                ggml_row_size(v->type, hparams.n_embd_v_gqa(il)), // v->nb[2]
+                ggml_row_size(v->type, n_embd_v_gqa),          // v->nb[2]
-                0);
+                ggml_row_size(v->type, n_embd_v_gqa*kv_size),  // v->nb[3]
+                ggml_row_size(v->type, n_embd_v_gqa*kv_size)*0);
     }
 
     // note: v->nb[1] > v->nb[2]
-    return ggml_view_3d(ctx, v,
+    return ggml_view_4d(ctx, v,
-            n_kv, hparams.n_head_kv(il), hparams.n_embd_head_v,
+            n_kv, hparams.n_head_kv(il), hparams.n_embd_head_v, 1,
-            ggml_row_size(v->type, v->ne[1]*hparams.n_embd_head_v), // v->nb[1]
+            ggml_row_size(v->type, kv_size*hparams.n_embd_head_v), // v->nb[1]
-            ggml_row_size(v->type, v->ne[1]),                       // v->nb[2]
+            ggml_row_size(v->type, kv_size),                       // v->nb[2]
-            0);
+            ggml_row_size(v->type, kv_size*n_embd_v_gqa),          // v->nb[3]
+            ggml_row_size(v->type, kv_size*n_embd_v_gqa)*0);
 }
 
 ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const {
@@ -825,6 +846,10 @@ ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_
     k_cur = ggml_reshape_2d(ctx, k_cur, k->ne[0], n_tokens);
 
     if (k_idxs && supports_set_rows) {
+        if (k->ne[2] > 1) {
+            k = ggml_reshape_2d(ctx, k, k->ne[0], k->ne[1]*k->ne[2]);
+        }
+
         return ggml_set_rows(ctx, k, k_cur, k_idxs);
     }
 
@@ -843,31 +868,30 @@ ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_
 
     auto * v = layers[ikv].v;
 
-    const int64_t n_embd_v_gqa = v->ne[0];
+    const int64_t n_embd_v_gqa = v_cur->ne[0]*v_cur->ne[1];
-    const int64_t n_tokens = v_cur->ne[2];
+    const int64_t n_tokens     = v_cur->ne[2];
 
     v_cur = ggml_reshape_2d(ctx, v_cur, n_embd_v_gqa, n_tokens);
 
     if (v_idxs && supports_set_rows) {
         if (!v_trans) {
+            if (v->ne[2] > 1) {
+                v = ggml_reshape_2d(ctx, v, v->ne[0], v->ne[1]*v->ne[2]);
+            }
+
             return ggml_set_rows(ctx, v, v_cur, v_idxs);
         }
 
+        // [TAG_V_CACHE_VARIABLE]
+        if (n_embd_v_gqa < v->ne[0]) {
+            v_cur = ggml_pad(ctx, v_cur, v->ne[0] - n_embd_v_gqa, 0, 0, 0);
+        }
+
         // the row becomes a single element
-        ggml_tensor * v_view = ggml_reshape_3d(ctx, v, 1, v->ne[1], v->ne[0]);
+        ggml_tensor * v_view = ggml_reshape_2d(ctx, v, 1, v->ne[0]*v->ne[1]*v->ne[2]);
 
-        // note: the V cache is transposed when not using flash attention
+        v_cur = ggml_reshape_2d(ctx, v_cur, 1, v_cur->ne[0]*v_cur->ne[1]);
-        v_cur = ggml_permute(ctx, ggml_reshape_3d(ctx, v_cur, v_cur->ne[0], 1, v_cur->ne[1]), 2, 0, 1, 3);
 
-        // note: we can be more explicit here at the cost of extra cont
-        //       however, above we take advantage that a row of single element is always continuous regardless of the row stride
-        //v_cur = ggml_transpose(ctx, v_cur);
-        //v_cur = ggml_cont_3d(ctx, v_cur, 1, v_cur->ne[0], v_cur->ne[1]);
-
-        // we broadcast the KV indices n_embd_v_gqa times
-        // v      [1,        n_kv,     n_embd_v_gqa]
-        // v_cur  [1,        n_tokens, n_embd_v_gqa]
-        // v_idxs [n_tokens, 1,        1]
         return ggml_set_rows(ctx, v_view, v_cur, v_idxs);
     }
 
@@ -904,7 +928,13 @@ ggml_tensor * llama_kv_cache_unified::build_input_k_idxs(ggml_context * ctx, con
 ggml_tensor * llama_kv_cache_unified::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
     const uint32_t n_tokens = ubatch.n_tokens;
 
-    ggml_tensor * v_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens);
+    ggml_tensor * v_idxs;
+
+    if (!v_trans) {
+        v_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens);
+    } else {
+        v_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens*hparams.n_embd_v_gqa_max());
+    }
 
     ggml_set_input(v_idxs);
 
@@ -921,7 +951,7 @@ void llama_kv_cache_unified::set_input_k_idxs(ggml_tensor * dst, const llama_uba
     GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
     int64_t * data = (int64_t *) dst->data;
 
-    for (int64_t i = 0; i < n_tokens; ++i) {
+    for (uint32_t i = 0; i < n_tokens; ++i) {
         data[i] = sinfo.idxs.at(i);
     }
 }
@@ -936,8 +966,21 @@ void llama_kv_cache_unified::set_input_v_idxs(ggml_tensor * dst, const llama_uba
     GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
     int64_t * data = (int64_t *) dst->data;
 
-    for (int64_t i = 0; i < n_tokens; ++i) {
+    if (!v_trans) {
-        data[i] = sinfo.idxs.at(i);
+        for (uint32_t i = 0; i < n_tokens; ++i) {
+            data[i] = sinfo.idxs.at(i);
+        }
+    } else {
+        // note: the V cache is transposed when not using flash attention
+        const int64_t kv_size = get_size();
+
+        const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa_max();
+
+        for (uint32_t i = 0; i < n_tokens; ++i) {
+            for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                data[i*n_embd_v_gqa + j] = j*kv_size + sinfo.idxs.at(i);
+            }
+        }
     }
 }