llama : use "stream" vs "virtual sequence"

ggml-ci

Author: ggerganov

examples/parallel/parallel.cpp

@@ -236,7 +236,7 @@ int main(int argc, char ** argv) {
 
     // the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
     // users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
-    llama_batch batch = llama_batch_init(n_ctx*n_clients, 0, 1);
+    llama_batch batch = llama_batch_init(n_ctx, 0, 1);
 
     int32_t n_total_prompt = 0;
     int32_t n_total_gen    = 0;
@@ -290,7 +290,6 @@ int main(int argc, char ** argv) {
             // all sequences have ended - clear the entire KV cache
             for (int i = 1; i <= n_clients; ++i) {
                 llama_memory_seq_rm(mem, i, -1, -1);
-
                 // but keep the system prompt
                 llama_memory_seq_cp(mem, 0, i, -1, -1);
             }

src/llama-context.cpp

@@ -34,7 +34,7 @@ llama_context::llama_context(
     }
 
     const char * LLAMA_HT = getenv("LLAMA_HT");
-    cparams.n_seq_virt = LLAMA_HT ? cparams.n_seq_max : 1;
+    cparams.kv_unified = (LLAMA_HT && atoi(LLAMA_HT) > 0) ? false : true;
 
     cparams.n_threads        = params.n_threads;
     cparams.n_threads_batch  = params.n_threads_batch;
@@ -270,7 +270,7 @@ llama_context::llama_context(
 
     // reserve worst-case graph
     if (!hparams.vocab_only && memory) {
-        const uint32_t n_seqs = 1; // reserve worst-case graph for single-sequence batches
+        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
@@ -303,7 +303,7 @@ llama_context::llama_context(
 
         // reserve with tg graph to get the number of splits and nodes
         {
-            auto * gf = graph_reserve(1, 1, 1, mctx.get());
+            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }

src/llama-cparams.h

@@ -11,7 +11,6 @@ struct llama_cparams {
     uint32_t n_batch;
     uint32_t n_ubatch;
     uint32_t n_seq_max;
-    uint32_t n_seq_virt;
     int32_t  n_threads;       // number of threads to use for generation
     int32_t  n_threads_batch; // number of threads to use for batch processing
 
@@ -34,6 +33,7 @@ struct llama_cparams {
     bool no_perf;
     bool warmup;
     bool op_offload;
+    bool kv_unified;
 
     enum llama_pooling_type pooling_type;
 

src/llama-graph.cpp

@@ -1000,13 +1000,13 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
     {
         GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Hybrid recurrent is not supported with SWA attention layers");
 
-        const auto n_kv   = inp->mctx->get_attn()->get_n_kv();
-        const auto n_seqs = cparams.n_seq_virt > 1 ? ubatch.n_seqs_unq : 1;
+        const auto n_kv     = inp->mctx->get_attn()->get_n_kv();
+        const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
 
         inp->self_k_idxs = mctx_cur->get_attn()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->get_attn()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_seqs, GGML_KQ_MASK_PAD), 1, n_seqs);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1033,9 +1033,10 @@ ggml_tensor * llm_graph_context::build_attn_mha(
              float     kq_scale) const {
     const bool v_trans = v->nb[1] > v->nb[2];
 
-    const auto n_seqs = cparams.n_seq_virt > 1 ? ubatch.n_seqs_unq : 1;
+    // split the batch into streams if needed
+    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
 
-    q = ggml_reshape_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_seqs, n_seqs);
+    q = ggml_reshape_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream);
 
     q = ggml_permute(ctx0, q, 0, 2, 1, 3);
     k = ggml_permute(ctx0, k, 0, 2, 1, 3);
@@ -1085,7 +1086,8 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 #endif
         }
 
-        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens*n_seqs);
+        // recombine streams
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens*n_stream);
     } else {
         ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
 
@@ -1130,7 +1132,8 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
 
-        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens*n_seqs);
+        // recombine streams
+        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens*n_stream);
 
         if (!cparams.offload_kqv) {
             // all nodes between the KV store and the attention output are run on the CPU
@@ -1207,13 +1210,13 @@ llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified()
     {
         GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA");
 
-        const auto n_kv   = mctx_cur->get_n_kv();
-        const auto n_seqs = cparams.n_seq_virt > 1 ? ubatch.n_seqs_unq : 1;
+        const auto n_kv     = mctx_cur->get_n_kv();
+        const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
 
         inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_seqs, GGML_KQ_MASK_PAD), 1, n_seqs);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1455,15 +1458,15 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
 
     auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, mctx_cur);
 
-    const auto n_seqs = cparams.n_seq_virt > 1 ? ubatch.n_seqs_unq : 1;
+    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
 
     {
         const auto n_kv = mctx_cur->get_base()->get_n_kv();
 
         inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_seqs, GGML_KQ_MASK_PAD), 1, n_seqs);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1477,7 +1480,7 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
         inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_seqs, GGML_KQ_MASK_PAD), 1, n_seqs);
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;

src/llama-graph.h

@@ -257,8 +257,8 @@ class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
     ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
     ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
 
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_seq, 1, n_seq]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_seq, 1, n_seq]
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -293,10 +293,10 @@ class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
     ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
     ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
 
-    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch/n_seq, 1, n_seq]
-    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch/n_seq, 1, n_seq]
-    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch/n_seq, 1, n_seq]
-    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch/n_seq, 1, n_seq]
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -343,8 +343,8 @@ class llm_graph_input_mem_hybrid : public llm_graph_input_i {
     ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
     ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
 
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_seq, 1, n_seq]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_seq, 1, n_seq]
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;

src/llama-kv-cache-unified-iswa.cpp

@@ -18,17 +18,17 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
                      bool   v_trans,
                      bool   offload,
                      bool   swa_full,
+                     bool   unified,
                  uint32_t   kv_size,
                  uint32_t   n_seq_max,
-                 uint32_t   n_seq_virt,
                  uint32_t   n_ubatch,
-                 uint32_t   n_pad) : hparams(model.hparams), n_seq_virt(n_seq_virt) {
+                 uint32_t   n_pad) : hparams(model.hparams), unified(unified) {
     llama_kv_cache_unified::layer_filter_cb filter_base = [&](int32_t il) { return !model.hparams.is_swa(il); };
     llama_kv_cache_unified::layer_filter_cb filter_swa  = [&](int32_t il) { return  model.hparams.is_swa(il); };
 
     const uint32_t size_base = kv_size;
 
-    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(n_seq_max/n_seq_virt) + n_ubatch, n_pad));
+    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch, n_pad));
 
     // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
     if (swa_full) {
@@ -42,14 +42,14 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
 
     kv_base = std::make_unique<llama_kv_cache_unified>(
             model, std::move(filter_base), type_k, type_v,
-            v_trans, offload, size_base, n_seq_max, n_seq_virt, n_pad,
+            v_trans, offload, unified, size_base, n_seq_max, n_pad,
             0, LLAMA_SWA_TYPE_NONE);
 
     LLAMA_LOG_INFO("%s: creating     SWA KV cache, size = %u cells\n", __func__, size_swa);
 
     kv_swa = std::make_unique<llama_kv_cache_unified>(
             model, std::move(filter_swa), type_k, type_v,
-            v_trans, offload, size_swa, n_seq_max, n_seq_virt, n_pad,
+            v_trans, offload, unified, size_swa, n_seq_max, n_pad,
             hparams.n_swa, hparams.swa_type);
 }
 
@@ -101,7 +101,7 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
 
     // first try simple split
     do {
-        if (n_seq_virt > 1) {
+        if (!unified) {
             // requires equal splits, so we skip the simple split
             break;
         }
@@ -146,7 +146,7 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
 
         std::vector<llama_ubatch> ubatches;
         while (true) {
-            auto ubatch = balloc.split_equal(n_ubatch, n_seq_virt > 1);
+            auto ubatch = balloc.split_equal(n_ubatch, !unified);
 
             if (ubatch.n_tokens == 0) {
                 break;

src/llama-kv-cache-unified-iswa.h

@@ -20,9 +20,9 @@ class llama_kv_cache_unified_iswa : public llama_memory_i {
                          bool   v_trans,
                          bool   offload,
                          bool   swa_full,
+                         bool   unified,
                      uint32_t   kv_size,
                      uint32_t   n_seq_max,
-                     uint32_t   n_seq_virt,
                      uint32_t   n_ubatch,
                      uint32_t   n_pad);
 
@@ -69,7 +69,7 @@ class llama_kv_cache_unified_iswa : public llama_memory_i {
 private:
     const llama_hparams & hparams;
 
-    const uint32_t n_seq_virt = 1;
+    const bool unified;
 
     std::unique_ptr<llama_kv_cache_unified> kv_base;
     std::unique_ptr<llama_kv_cache_unified> kv_swa;