llama-batch: add n_keep_tail in split_equal for recurrent models (#25278)

diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp
index 6bf76939cddcbb7c06232ad7e88bf47e5bd57bc0..5436717c42711a4f16a32df6714dc82a6361eddc 100644 (file)
--- a/src/llama-batch.cpp
+++ b/src/llama-batch.cpp
@@ -505,7 +505,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
     return ubatch_add(idxs, idxs.size(), false);
 }
 
-llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
+llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential, uint32_t n_keep_tail) {
     if (sequential && has_cpl) {
         LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch (you may need to use the -kvu flag)\n", __func__);
 
@@ -548,7 +548,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential)
         }
     }
 
-    const uint32_t n_seqs = cur_seq_set.size();
+    uint32_t n_seqs = cur_seq_set.size();
 
     // we are done
     if (n_seqs == 0) {
@@ -569,7 +569,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential)
     std::vector<idx_vec_t> idxs_per_seq(n_seqs);
 
     while (true) {
-        // we can only add new n_seq_tokens tokens if all the sequence sets have at least one more unused token and
+        // we can only add new n_seq_tokens tokens if all the sequence sets have at least 1 more unused tokens and
         //   if we haven't reached n_ubatch
         bool can_expand = true;
 
@@ -600,6 +600,72 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential)
         }
     }
 
+    // if n_keep_tail > 0, keep only the seqs that either finish in this ubatch or have at least
+    //   n_keep_tail tokens remaining for a future ubatch, so that the trailing n_keep_tail tokens
+    //   of each seq are never split across ubatches
+    if (n_keep_tail > 0) {
+        GGML_ASSERT(n_ubatch > n_keep_tail);
+
+        auto n_remaining = [&](uint32_t s) {
+            return (uint32_t) (seq_set_map[cur_seq_set[s]].size() - cur_idx[s]);
+        };
+
+        // keep the longest prefix of seqs that satisfy the constraint, to preserve sequential seq ids
+        uint32_t n_keep = 0;
+        while (n_keep < n_seqs) {
+            const uint32_t remaining = n_remaining(n_keep);
+
+            if (remaining != 0 && remaining < n_keep_tail) {
+                break;
+            }
+
+            n_keep++;
+        }
+
+        // all seqs violate the constraint - resolve the first one directly and emit it alone
+        if (n_keep == 0) {
+            auto & idxs = idxs_per_seq[0];
+
+            const auto & seq_idxs = seq_set_map[cur_seq_set[0]];
+
+            if (idxs.size() + n_remaining(0) <= n_ubatch) {
+                // extend the seq to completion
+                while (n_remaining(0) > 0) {
+                    const int32_t idx = seq_idxs[cur_idx[0]];
+
+                    idxs.push_back(idx);
+
+                    used[idx] = true;
+                    ++n_used;
+
+                    ++cur_idx[0];
+                }
+            } else {
+                // truncate the seq so that at least n_keep_tail tokens remain
+                while (n_remaining(0) < n_keep_tail) {
+                    used[idxs.back()] = false;
+                    --n_used;
+
+                    idxs.pop_back();
+
+                    --cur_idx[0];
+                }
+            }
+
+            n_keep = 1;
+        }
+
+        // return the tokens of the deferred seqs back to the pool
+        for (uint32_t s = n_keep; s < n_seqs; ++s) {
+            for (const int32_t idx : idxs_per_seq[s]) {
+                used[idx] = false;
+                --n_used;
+            }
+        }
+
+        n_seqs = n_keep;
+    }
+
     // concat the per-sequence-set lists
     std::vector<int32_t> idxs;
 
@@ -814,7 +880,7 @@ void llama_batch_allocr::ubatch_print(const llama_ubatch & ubatch, int debug) {
         LLAMA_LOG_DEBUG("%s:   output     = %p\n", __func__, (void *) ubatch.output);
         LLAMA_LOG_DEBUG("%s:   n_outputs  = %d\n", __func__, n_outputs);
 
-        if (debug > 1) {
+        if (debug > 0) {
             int seq_id_max = 0;
             for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
                 for (int s = 0; s < ubatch.n_seq_id[i]; ++s) {

diff --git a/src/llama-batch.h b/src/llama-batch.h
index f77520e86c3ca45ffb5541597790005b76d5d4cd..a3d1889d4a041be695a91b72a2a7f3badcb46fca 100644 (file)
--- a/src/llama-batch.h
+++ b/src/llama-batch.h
@@ -104,7 +104,8 @@ public:
 
     // make ubatches of equal-length sequences sets
     // if sequential == true, the tokens in the ubatch will have increasing sequential sequence ids
-    llama_ubatch split_equal(uint32_t n_ubatch, bool sequential);
+    // n_keep_tail = minimum trailing tokens of a seq that must land in the same ubatch
+    llama_ubatch split_equal(uint32_t n_ubatch, bool sequential, uint32_t n_keep_tail);
 
     // sequence-set-wise split - each ubatch contains a single sequence-set
     llama_ubatch split_seq(uint32_t n_ubatch);

diff --git a/src/llama-kv-cache-dsa.cpp b/src/llama-kv-cache-dsa.cpp
index 916ab65375635d9cca6203f942b7da7bb0d93f1f..241c50365a1371f6641525348d72169bf874d169 100644 (file)
--- a/src/llama-kv-cache-dsa.cpp
+++ b/src/llama-kv-cache-dsa.cpp
@@ -113,7 +113,7 @@ llama_memory_context_ptr llama_kv_cache_dsa::init_batch(
 
         std::vector<llama_ubatch> ubatches;
         while (true) {
-            auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true);
+            auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true, 0);
 
             if (ubatch.n_tokens == 0) {
                 break;

diff --git a/src/llama-kv-cache-dsv4.cpp b/src/llama-kv-cache-dsv4.cpp
index 3a698d719ed7e514585c107fc919fef1ca236174..9fccf347ed6676bd89636225949402b03d5ade24 100644 (file)
--- a/src/llama-kv-cache-dsv4.cpp
+++ b/src/llama-kv-cache-dsv4.cpp
@@ -1110,7 +1110,7 @@ llama_memory_context_ptr llama_kv_cache_dsv4::init_batch(
             if (has_coupled) {
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                ubatch = balloc.split_equal(n_ubatch, raw_per_seq || comp_per_seq);
+                ubatch = balloc.split_equal(n_ubatch, raw_per_seq || comp_per_seq, 0);
             }
 
             if (ubatch.n_tokens == 0) {

diff --git a/src/llama-kv-cache-iswa.cpp b/src/llama-kv-cache-iswa.cpp
index 2fcf238d9173e11700d66fc1ac86931738180c99..e91866469adf31aa4fa2b1190cc8432a38b61b24 100644 (file)
--- a/src/llama-kv-cache-iswa.cpp
+++ b/src/llama-kv-cache-iswa.cpp
@@ -206,7 +206,7 @@ llama_memory_context_ptr llama_kv_cache_iswa::init_batch(llama_batch_allocr & ba
 
         std::vector<llama_ubatch> ubatches;
         while (true) {
-            auto ubatch = balloc.split_equal(n_ubatch, !unified);
+            auto ubatch = balloc.split_equal(n_ubatch, !unified, 0);
 
             if (ubatch.n_tokens == 0) {
                 break;

diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
index 680de5144a86048947b07a2c63d3fe3bba6eeeb1..e70583e64152ec6c57feeaf5f6e23fe58796ac89 100644 (file)
--- a/src/llama-kv-cache.cpp
+++ b/src/llama-kv-cache.cpp
@@ -706,7 +706,7 @@ llama_memory_context_ptr llama_kv_cache::init_batch(
 
         std::vector<llama_ubatch> ubatches;
         while (true) {
-            auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true);
+            auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true, 0);
 
             if (ubatch.n_tokens == 0) {
                 break;

diff --git a/src/llama-memory-hybrid-iswa.cpp b/src/llama-memory-hybrid-iswa.cpp
index c7d4bcd413e8f6d633ea53bd44d4dde5a0457b11..06f7fd5428c4e92c445b5c0170e92e557b2f62df 100644 (file)
--- a/src/llama-memory-hybrid-iswa.cpp
+++ b/src/llama-memory-hybrid-iswa.cpp
@@ -77,15 +77,15 @@ llama_memory_context_ptr llama_memory_hybrid_iswa::init_batch(llama_batch_allocr
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                if (mem_recr->n_rs_seq > 0) {
-                    // [TAG_RECURRENT_ROLLBACK_SPLITS]
-                    // TODO: recurrent state rollback does not support equal splits
-                    ubatch = balloc.split_seq(n_ubatch);
-                } else {
-                    // Use non-sequential split when KV cache is unified (needed for hellaswag/winogrande/multiple-choice)
-                    const bool unified = (mem_attn->get_base()->get_n_stream() == 1);
-                    ubatch = balloc.split_equal(n_ubatch, !unified);
-                }
+                // Use non-sequential split when KV cache is unified (needed for hellaswag/winogrande/multiple-choice)
+                const bool unified = (mem_attn->get_base()->get_n_stream() == 1);
+
+                // [TAG_RECURRENT_ROLLBACK_SPLITS]
+                // the trailing (1 + n_rs_seq) tokens of each seq must stay in the same ubatch
+                //   so that the rollback snapshots remain valid
+                const uint32_t n_rs_seq = mem_recr->n_rs_seq;
+
+                ubatch = balloc.split_equal(n_ubatch, !unified, n_rs_seq > 0 ? n_rs_seq + 1 : 0);
             }
 
             if (ubatch.n_tokens == 0) {

diff --git a/src/llama-memory-hybrid.cpp b/src/llama-memory-hybrid.cpp
index f2d49cbce545429bb56e375da6c101c0719e92e5..42c7381a9e6fa507c925ba66a35a7e433c4447b8 100644 (file)
--- a/src/llama-memory-hybrid.cpp
+++ b/src/llama-memory-hybrid.cpp
@@ -78,15 +78,15 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                if (mem_recr->n_rs_seq > 0) {
-                    // [TAG_RECURRENT_ROLLBACK_SPLITS]
-                    // TODO: recurrent state rollback does not support equal splits
-                    ubatch = balloc.split_seq(n_ubatch);
-                } else {
-                    // Use non-sequential split when KV cache is unified (needed for hellaswag/winogrande/multiple-choice)
-                    const bool unified = (mem_attn->get_n_stream() == 1);
-                    ubatch = balloc.split_equal(n_ubatch, !unified);
-                }
+                // Use non-sequential split when KV cache is unified (needed for hellaswag/winogrande/multiple-choice)
+                const bool unified = (mem_attn->get_n_stream() == 1);
+
+                // [TAG_RECURRENT_ROLLBACK_SPLITS]
+                // the trailing (1 + n_rs_seq) tokens of each seq must stay in the same ubatch
+                //   so that the rollback snapshots remain valid
+                const uint32_t n_rs_seq = mem_recr->n_rs_seq;
+
+                ubatch = balloc.split_equal(n_ubatch, !unified, n_rs_seq > 0 ? n_rs_seq + 1 : 0);
             }
 
             if (ubatch.n_tokens == 0) {

diff --git a/src/llama-memory-recurrent.cpp b/src/llama-memory-recurrent.cpp
index 6a4892fb471d7393dd8a2bd2e6d52f91bedf53f1..3d6c6db876b4217d404d4a9999a8d1d31c22d83c 100644 (file)
--- a/src/llama-memory-recurrent.cpp
+++ b/src/llama-memory-recurrent.cpp
@@ -416,15 +416,12 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr &
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                if (n_rs_seq > 0) {
-                    // [TAG_RECURRENT_ROLLBACK_SPLITS]
-                    // TODO: recurrent state rollback does not support equal splits
-                    ubatch = balloc.split_seq(n_ubatch);
-                } else {
-                    // TODO: non-sequential equal split can be done if using unified KV cache
-                    //       for simplicity, we always use sequential equal split for now
-                    ubatch = balloc.split_equal(n_ubatch, true);
-                }
+                // TODO: non-sequential equal split can be done if using unified KV cache
+                //       for simplicity, we always use sequential equal split for now
+                // [TAG_RECURRENT_ROLLBACK_SPLITS]
+                // the trailing (1 + n_rs_seq) tokens of each seq must stay in the same ubatch
+                //   so that the rollback snapshots remain valid
+                ubatch = balloc.split_equal(n_ubatch, true, n_rs_seq > 0 ? n_rs_seq + 1 : 0);
             }
 
             if (ubatch.n_tokens == 0) {

diff --git a/src/models/delta-net-base.cpp b/src/models/delta-net-base.cpp
index ad9ce77140840c2c3d631f1c27685b95bc200e94..cf5e380950191dc654dd37285abd258389b47c6a 100644 (file)
--- a/src/models/delta-net-base.cpp
+++ b/src/models/delta-net-base.cpp
@@ -496,8 +496,8 @@ ggml_tensor * llm_build_delta_net_base::build_conv_state(
         ggml_build_forward_expand(gf, ggml_cpy(ctx0, conv_state_last, conv_state_update));
     } else {
         // [TAG_RECURRENT_ROLLBACK_SPLITS]
-        // TODO: this logic incorrectly assumes that the last (n_rs_seq + 1) tokens of a sequence in a batch are
-        //       inside the same ubatch. currently with `split_equal()` this is not correct
+        // this logic assumes that the last (n_rs_seq + 1) tokens of a sequence in a batch are inside
+        //   the same ubatch, which `split_equal()` guarantees via its n_keep_tail argument
 
         const int64_t K = (int64_t) cparams.n_rs_seq + 1;