cont : more consistent indexing in recurrent cache

ggml-ci

Author: ggerganov

src/llama-kv-cache-recurrent.cpp

@@ -422,9 +422,8 @@ bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatche
 }
 
 bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
-    const uint32_t n_seqs = ubatch.n_seqs;
-
     const uint32_t n_seq_tokens = ubatch.n_seq_tokens;
+    const uint32_t n_seqs       = ubatch.n_seqs;
 
     // if we have enough unused cells before the current head ->
     //   better to start searching from the beginning of the cache, hoping to fill it
@@ -444,9 +443,11 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 
     // everything should fit if all seq_ids are smaller than the max
     for (uint32_t s = 0; s < n_seqs; ++s) {
-        const uint32_t n_seq_id = ubatch.n_seq_id[s*n_seq_tokens];
+        const uint32_t i = s*n_seq_tokens; // first token of sequence set s
+        const uint32_t n_seq_id = ubatch.n_seq_id[i];
+
         for (uint32_t j = 0; j < n_seq_id; ++j) {
-            const llama_seq_id seq_id = ubatch.seq_id[s*n_seq_tokens][j];
+            const llama_seq_id seq_id = ubatch.seq_id[i][j];
 
             if (seq_id < 0 || (uint32_t) seq_id >= size) {
                 // too big seq_id
@@ -505,7 +506,9 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 
     // find usable cell range
     for (uint32_t s = 0; s < n_seqs; ++s) {
-        const llama_seq_id seq_id = ubatch.seq_id[s*n_seq_tokens][0];
+        const uint32_t i = s*n_seq_tokens;
+        const llama_seq_id seq_id = ubatch.seq_id[i][0];
+
         kv_cell & seq_meta = cells[seq_id];
         bool has_cell = false;
         if (seq_meta.tail >= 0) {
@@ -529,7 +532,7 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
             seq_meta.tail = next_empty_cell;
             // find next empty cell
             if (s + 1 < n_seqs) {
-                for (uint32_t i = 0; i < size; ++i) {
+                for (uint32_t j = 0; j < size; ++j) {
                     next_empty_cell += 1;
                     if (next_empty_cell >= size) { next_empty_cell -= size; }
                     kv_cell & cell = cells[next_empty_cell];
@@ -543,8 +546,9 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 
     // gather and re-order
     for (uint32_t s = 0; s < n_seqs; ++s) {
+        const uint32_t i = s*n_seq_tokens;
         const int32_t dst_id = s + min;
-        const int32_t src_id = cells[ubatch.seq_id[s*n_seq_tokens][0]].tail;
+        const int32_t src_id = cells[ubatch.seq_id[i][0]].tail;
         if (dst_id != src_id) {
             kv_cell & dst_cell = cells[dst_id];
             kv_cell & src_cell = cells[src_id];
@@ -554,8 +558,8 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
             std::swap(dst_cell.seq_id, src_cell.seq_id);
 
             // swap tails
-            for (uint32_t i = 0; i < size; ++i) {
-                int32_t & tail = cells[i].tail;
+            for (uint32_t j = 0; j < size; ++j) {
+                int32_t & tail = cells[j].tail;
                 if (tail == src_id) {
                     tail = dst_id;
                 } else if (tail == dst_id) {
@@ -567,20 +571,21 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 
     // update the pos of the used seqs
     for (uint32_t s = 0; s < n_seqs; ++s) {
-        const llama_pos last_pos = ubatch.pos[s*n_seq_tokens + n_seq_tokens - 1];
+        const uint32_t i = s*n_seq_tokens;
+        const llama_pos last_pos = ubatch.pos[i + n_seq_tokens - 1];
         const int32_t cell_id = s + min;
         kv_cell & cell = cells[cell_id];
 
         if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) {
             // What should happen when the pos backtracks or skips a value?
             // Clearing the state mid-batch would require special-casing which isn't done.
             LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n",
-                __func__, last_pos, cell.pos, ubatch.seq_id[s*n_seq_tokens][0], n_seq_tokens);
+                __func__, last_pos, cell.pos, ubatch.seq_id[i][0], n_seq_tokens);
         }
         cell.pos = last_pos;
         cell.seq_id.clear();
-        for (int32_t j = 0; j < ubatch.n_seq_id[s*n_seq_tokens]; ++j) {
-            const llama_seq_id seq_id = ubatch.seq_id[s*n_seq_tokens][j];
+        for (int32_t j = 0; j < ubatch.n_seq_id[i]; ++j) {
+            const llama_seq_id seq_id = ubatch.seq_id[i][j];
             cell.seq_id.insert(seq_id);
             cells[seq_id].tail = cell_id;
         }