llama : add high-throughput mode

common/arg.cpp

@@ -1464,6 +1464,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.swa_full = true;
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
+    add_opt(common_arg(
+        {"--attn-streams", "-as"},
+        string_format("use multiple streams when computing the attention (default: %s)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/14363)", params.attn_streams ? "true" : "false"),
+        [](common_params & params) {
+            params.attn_streams = true;
+        }
+    ).set_env("LLAMA_ARG_ATTN_STREAMS"));
     add_opt(common_arg(
         {"--no-context-shift"},
         string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),

common/common.cpp

@@ -1157,6 +1157,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.no_perf           = params.no_perf;
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;
+    cparams.attn_streams      = params.attn_streams;
 
     cparams.type_k = params.cache_type_k;
     cparams.type_v = params.cache_type_v;

common/common.h

@@ -330,6 +330,7 @@ struct common_params {
     bool no_perf           = false; // disable performance metrics
     bool ctx_shift         = true;  // context shift on inifinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+    bool attn_streams      = false; // multi-stream attention and KV cache buffers
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool use_mmap          = true;  // use mmap for faster loads

examples/parallel/parallel.cpp

@@ -224,6 +224,7 @@ int main(int argc, char ** argv) {
         auto & client = clients[i];
         client.id = i;
         client.smpl = common_sampler_init(model, params.sampling);
+        //params.sampling.seed++;
     }
 
     std::vector<llama_token> tokens_system;
@@ -345,7 +346,7 @@ int main(int argc, char ** argv) {
                     client.n_decoded = 0;
                     client.i_batch   = batch.n_tokens - 1;
 
-                    LOG_INF("\033[31mClient %3d, seq %4d, junk = %4d, started decoding ...\033[0m\n", client.id, client.seq_id, n_junk_cur);
+                    LOG_INF("\033[31mClient %3d, seq %4d, junk = %4d, prompt = %d, started decoding ...\033[0m\n", client.id, client.seq_id, n_junk_cur, client.n_prompt);
 
                     g_seq_id += 1;
 

include/llama.h

@@ -374,6 +374,10 @@ extern "C" {
         bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
                           // NOTE: setting to false when n_seq_max > 1 can cause bad performance in some cases
                           //       ref: https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573
+
+        bool attn_streams; // if enabled, use multiple streams during the attention (determined by n_seq_max)
+                           // NOTE: this requires support for the ggml_set_rows() operator
+                           //       this flag can improve the performance for parallel, multi-sequence use cases
     };
 
     // model quantization parameters

src/llama-context.cpp

@@ -101,7 +101,8 @@ llama_context::llama_context(
 
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
 
-    cparams.op_offload = params.op_offload;
+    cparams.op_offload   = params.op_offload;
+    cparams.attn_streams = params.attn_streams;
 
     const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
 
@@ -112,6 +113,7 @@ llama_context::llama_context(
     LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
     LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
     LLAMA_LOG_INFO("%s: flash_attn    = %d\n",   __func__, cparams.flash_attn);
+    LLAMA_LOG_INFO("%s: attn_streams  = %s\n",   __func__, cparams.attn_streams ? "true" : "false");
     LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
     LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
 
@@ -267,7 +269,7 @@ llama_context::llama_context(
 
     // reserve worst-case graph
     if (!hparams.vocab_only && memory) {
-        const uint32_t n_seqs = cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.attn_streams ? cparams.n_seq_max : 1;
         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);
@@ -300,7 +302,7 @@ 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");
             }
@@ -311,6 +313,10 @@ llama_context::llama_context(
 
         // reserve again with pp graph to avoid ggml-alloc reallocations during inference
         {
+            // TODO: not sure if the following graph would be worster case for multi-stream KV caches:
+            //
+            // auto * gf = graph_reserve(n_tokens, 1, n_tokens, mctx.get());
+            //
             auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
@@ -475,7 +481,7 @@ bool llama_context::kv_self_update(bool optimize) {
             throw std::runtime_error("failed to initialize memory context");
         }
 
-        const uint32_t n_seqs   = cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.attn_streams ? cparams.n_seq_max : 1;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
@@ -741,6 +747,8 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     const uint32_t n_tokens = balloc->get_n_tokens();
 
+    // [TAG_NO_CACHE_PAD]
+    // TODO: add new split mode where we pad the input sequences so that ubatch.equal_seqs == true
     const llama_ubatch ubatch = balloc->split_simple(n_tokens);
 
     // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
@@ -2187,6 +2195,7 @@ llama_context_params llama_context_default_params() {
         /*.no_perf                     =*/ true,
         /*.op_offload                  =*/ true,
         /*.swa_full                    =*/ true,
+        /*.attn_streams                =*/ false,
     };
 
     return result;

src/llama-cparams.h

@@ -11,8 +11,8 @@ struct llama_cparams {
     uint32_t n_batch;
     uint32_t n_ubatch;
     uint32_t n_seq_max;
-    int      n_threads;       // number of threads to use for generation
-    int      n_threads_batch; // number of threads to use for batch processing
+    int32_t  n_threads;       // number of threads to use for generation
+    int32_t  n_threads_batch; // number of threads to use for batch processing
 
     float rope_freq_base;
     float rope_freq_scale;
@@ -33,6 +33,7 @@ struct llama_cparams {
     bool no_perf;
     bool warmup;
     bool op_offload;
+    bool attn_streams;
 
     enum llama_pooling_type pooling_type;
 

src/llama-graph.cpp

@@ -1000,12 +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_kv     = inp->mctx->get_attn()->get_n_kv();
+        const auto n_stream = cparams.attn_streams ? ubatch.n_seqs_unq : 1;
 
         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, GGML_KQ_MASK_PAD), 1, 1);
+        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;
@@ -1032,13 +1033,16 @@ ggml_tensor * llm_graph_context::build_attn_mha(
              float     kq_scale) const {
     const bool v_trans = v->nb[1] > v->nb[2];
 
+    // split the batch into streams if needed
+    const auto n_stream = k->ne[3];
+
+    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);
     v = ggml_permute(ctx0, v, 0, 2, 1, 3);
 
-    const auto n_tokens = q->ne[1];
-    const auto n_head   = q->ne[2];
-    const auto n_kv     = k->ne[1];
+    const auto n_kv = k->ne[1];
 
     ggml_tensor * cur;
 
@@ -1080,7 +1084,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 #endif
         }
 
-        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
     } else {
         ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
 
@@ -1125,7 +1129,8 @@ 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);
+        // recombine streams
+        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
 
         if (!cparams.offload_kqv) {
             // all nodes between the KV store and the attention output are run on the CPU
@@ -1172,6 +1177,10 @@ ggml_tensor * llm_graph_context::build_attn(
 
     const auto & kq_mask = inp->get_kq_mask();
 
+    // [TAG_NO_CACHE_PAD]
+    // TODO: if ubatch.equal_seqs == true, we can split the three tensors below into ubatch.n_seqs_unq streams
+    assert(ubatch.equal_seqs == false);
+
     ggml_tensor * q = q_cur;
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
@@ -1202,12 +1211,13 @@ 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_kv     = mctx_cur->get_n_kv();
+        const auto n_stream = cparams.attn_streams ? ubatch.n_seqs_unq : 1;
 
         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, GGML_KQ_MASK_PAD), 1, 1);
+        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;
@@ -1449,13 +1459,15 @@ 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_stream = cparams.attn_streams ? ubatch.n_seqs_unq : 1;
+
     {
         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, GGML_KQ_MASK_PAD), 1, 1);
+        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;
@@ -1469,7 +1481,7 @@ 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, GGML_KQ_MASK_PAD), 1, 1);
+        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

@@ -255,10 +255,10 @@ class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
     ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
 
     ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
-    ggml_tensor * self_v_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, 1, 1]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
+    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;
@@ -289,14 +289,14 @@ class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
     ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
 
     ggml_tensor * self_k_idxs     = nullptr; // I64 [n_batch]
-    ggml_tensor * self_v_idxs     = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs     = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
     ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
-    ggml_tensor * self_v_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, 1, 1]
-    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
+    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, 1, 1]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
+    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-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;