kv-cache : refactor + add llama_memory_state_i (#13746)

* kv-cache : simplify the "struct llama_kv_cache" interface

ggml-ci

* kv-cache : revert the (n_swa + n_ubatch) change (for next PR)

ggml-ci

* kv-cache : some comments

ggml-ci

* context : fix graph reserve for multiple sequences

ggml-ci

* kv-cache : fix typo [no ci]

* kv-cache : fix find_slot() logic for free slots

ggml-ci

* llama : add TODO for deprecating the defrag API in the future

* kv-cache : improve find_slot() using min/max seq pos info

ggml-ci

* llama : handle aborts and compute errors

ggml-ci

* memory : extract state into llama_memory_state

ggml-ci

* kv-cache : add comments

ggml-ci

* server : update batching logic to reset n_batch on successful decode

* server : upon full re-processing, remove the sequence from the cache

* kv-cache : add TODO for doing split_equal when split_simple fails

ggml-ci

Author: ggerganov

examples/parallel/parallel.cpp

@@ -362,15 +362,17 @@ int main(int argc, char ** argv) {
         // process in chunks of params.n_batch
         int32_t n_batch = params.n_batch;
 
-        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
+        int32_t i_next = 0;
+
+        for (int32_t i = 0; i < batch.n_tokens; i = i_next) {
             // experiment: process in powers of 2
             //if (i + n_batch > (int32_t) batch.n_tokens && n_batch > 32) {
             //    n_batch /= 2;
             //    i -= n_batch;
             //    continue;
             //}
 
-            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+            const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i);
 
             llama_batch batch_view = {
                 n_tokens,
@@ -396,13 +398,18 @@ int main(int argc, char ** argv) {
 
                 // retry with half the batch size to try to find a free slot in the KV cache
                 n_batch /= 2;
-                i -= n_batch;
 
                 continue;
             }
 
             LOG_DBG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
 
+            // move the head of the batch forward with the number of tokens we just processed
+            i_next = i + n_tokens;
+
+            // on successful decode, restore the original batch size
+            n_batch = params.n_batch;
+
             for (auto & client : clients) {
                 if (client.i_batch < (int) i || client.i_batch >= (int) (i + n_tokens)) {
                     continue;

include/llama.h

@@ -259,9 +259,9 @@ extern "C" {
         llama_token  *  token;
         float        *  embd;
         llama_pos    *  pos;
-        int32_t      *  n_seq_id;
-        llama_seq_id ** seq_id;
-        int8_t       *  logits; // TODO: rename this to "output"
+        int32_t      *  n_seq_id; // TODO: remove, should belong to only 1 sequence
+        llama_seq_id ** seq_id;   // TODO: become llama_seq_id * seq_id;
+        int8_t       *  logits;   // TODO: rename this to "output"
     } llama_batch;
 
     enum llama_model_kv_override_type {
@@ -677,12 +677,14 @@ extern "C" {
 
     // Returns the smallest position present in the KV cache for the specified sequence
     // This is typically non-zero only for SWA caches
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
     // Return -1 if the sequence is empty
     LLAMA_API llama_pos llama_kv_self_seq_pos_min(
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
     // Returns the largest position present in the KV cache for the specified sequence
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
     // Return -1 if the sequence is empty
     LLAMA_API llama_pos llama_kv_self_seq_pos_max(
             struct llama_context * ctx,
@@ -692,12 +694,14 @@ extern "C" {
     // This will be applied:
     //   - lazily on next llama_decode()
     //   - explicitly with llama_kv_self_update()
+    // TODO: deprecate and always update the cache lazily [TAG: API_KV_NO_DEFRAG]
     LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx);
 
     // Check if the context supports KV cache shifting
     LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx);
 
     // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
+    // TODO: deprecate and always update the cache lazily [TAG: API_KV_NO_DEFRAG]
     LLAMA_API void llama_kv_self_update(struct llama_context * ctx);
 
     //

src/llama-batch.cpp

@@ -15,24 +15,31 @@ llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) {
             break;
         }
     }
-    ubatch_token.resize(!has_embd ? n_ubatch : 0);
-    ubatch_embd.resize(has_embd ? n_embd * n_ubatch : 0);
-    ubatch_pos.resize(n_ubatch);
-    ubatch_n_seq_id.resize(n_ubatch);
-    ubatch_seq_id.resize(n_ubatch);
-    ubatch_output.resize(n_ubatch);
+
+    udatas.push_back({});
+
+    auto & udata = udatas.back();
+
+    udata.token.resize(!has_embd ? n_ubatch : 0);
+    udata.embd.resize(has_embd ? n_embd * n_ubatch : 0);
+    udata.pos.resize(n_ubatch);
+    udata.n_seq_id.resize(n_ubatch);
+    udata.seq_id.resize(n_ubatch);
+    udata.output.resize(n_ubatch);
+
     llama_ubatch ubatch = {
         /*equal_seqs   =*/ true,
         /*n_tokens     =*/ 0,
         /*n_seq_tokens =*/ 0,
         /*n_seqs       =*/ 0,
-        /*token        =*/ !has_embd ? ubatch_token.data() : nullptr,
-        /*embd         =*/ has_embd  ? ubatch_embd.data()  : nullptr,
-        /*pos          =*/ ubatch_pos.data(),
-        /*n_seq_id     =*/ ubatch_n_seq_id.data(),
-        /*seq_id       =*/ ubatch_seq_id.data(),
-        /*output       =*/ ubatch_output.data(),
+        /*token        =*/ !has_embd ? udata.token.data() : nullptr,
+        /*embd         =*/ has_embd  ? udata.embd.data()  : nullptr,
+        /*pos          =*/ udata.pos.data(),
+        /*n_seq_id     =*/ udata.n_seq_id.data(),
+        /*seq_id       =*/ udata.seq_id.data(),
+        /*output       =*/ udata.output.data(),
     };
+
     return ubatch;
 }
 

src/llama-batch.h

@@ -11,15 +11,15 @@ struct llama_ubatch {
     bool equal_seqs;
     // TODO: whole_seqs for embeddings?
 
-    uint32_t n_tokens; // total tokens (n_seq_tokens * n_seqs)
+    uint32_t n_tokens;     // total tokens (n_seq_tokens * n_seqs)
     uint32_t n_seq_tokens; // tokens per sequence
     uint32_t n_seqs;
 
     llama_token  *  token;    // [n_tokens]
     float        *  embd;     // [n_embd, n_tokens]
     llama_pos    *  pos;      // [n_tokens]
-    int32_t      *  n_seq_id; // [n_seqs]
-    llama_seq_id ** seq_id;   // [n_seqs]
+    int32_t      *  n_seq_id; // [n_seqs] // TODO: remove, should belong to only 1 sequence
+    llama_seq_id ** seq_id;   // [n_seqs] // TODO: become llama_seq_id * seq_id;
     int8_t       *  output;   // [n_tokens]
 };
 
@@ -49,13 +49,18 @@ struct llama_sbatch {
 
     const llama_batch * batch = nullptr;
 
-    // buffers for the ubatch
-    std::vector<llama_token>    ubatch_token;
-    std::vector<float>          ubatch_embd;
-    std::vector<llama_pos>      ubatch_pos;
-    std::vector<int32_t>        ubatch_n_seq_id;
-    std::vector<llama_seq_id *> ubatch_seq_id;
-    std::vector<int8_t>         ubatch_output;
+    // buffers for the ubatches
+    // TODO: very hacky, this needs a complete rework
+    struct ubatch_data {
+        std::vector<llama_token>    token;
+        std::vector<float>          embd;
+        std::vector<llama_pos>      pos;
+        std::vector<int32_t>        n_seq_id;
+        std::vector<llama_seq_id *> seq_id;
+        std::vector<int8_t>         output;
+    };
+
+    std::vector<ubatch_data> udatas;
 
     llama_ubatch reserve_ubatch(size_t n_ubatch, bool has_embd = false);