Add `--no-op-offload` to improve `-ot` pp perf in MoE models like llama4 400B (#13386)

diff --git a/common/arg.cpp b/common/arg.cpp
index e0f1d998f6056e484935e9f980b0b4c71c004bda..a1fd4c9651b9cc78aa344fae4bd8b4fe3c89bdc0 100644 (file)
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -2437,6 +2437,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ));
+    add_opt(common_arg(
+        {"--no-op-offload"},
+        string_format("disable offloading host tensor operations to device (default: %s)", params.no_op_offload ? "true" : "false"),
+        [](common_params & params) {
+            params.no_op_offload = true;
+        }
+    ));
     add_opt(common_arg(
         {"--lora"}, "FNAME",
         "path to LoRA adapter (can be repeated to use multiple adapters)",

diff --git a/common/common.cpp b/common/common.cpp
index bd20af233695c2f76594bd05fb1e0916f5e83096..710bf1fe2a3c52ad30cfadffa59235a856064d4c 100644 (file)
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -1113,6 +1113,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.offload_kqv       = !params.no_kv_offload;
     cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
+    cparams.op_offload        = !params.no_op_offload;
 
     if (params.reranking) {
         cparams.embeddings    = true;

diff --git a/common/common.h b/common/common.h
index d051d4ec971c44d27b57874f35f5aa57997ba469..e15356b12a7c1799a3dedb3db30342391cac9541 100644 (file)
--- a/common/common.h
+++ b/common/common.h
@@ -332,6 +332,7 @@ struct common_params {
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
+    bool no_op_offload     = false; // globally disable offload host tensor operations to device
 
     bool single_turn       = false; // single turn chat conversation
 

diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
index ea2c1a402cca102f5c4b3efd9e8a9b0e7593b443..778927f68217ae2eaad89b080a3f03829b4321e0 100644 (file)
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -248,7 +248,7 @@ extern "C" {
         // preferrably to run on the same backend as the buffer
         ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false, true);
 
         // initialize buffers from a max size graph (optional)
         reserve_graph = build_graph(sched, max_batch_size);
@@ -289,7 +289,7 @@ extern "C" {
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
     // Initialize a backend scheduler, backends with low index are given priority over backends with high index
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
     // Initialize backend buffers from a measure graph

diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index c36b5abfb74224cfdfcd6d3ec9a1c4eb612aee78..6f69d895f170d039a1dcea8f89403c26e64ab8b5 100644 (file)
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -674,6 +674,8 @@ struct ggml_backend_sched {
     char * context_buffer;
     size_t context_buffer_size;
 
+    bool op_offload;
+
     int debug;
 };
 
@@ -766,7 +768,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
-            if (src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
+            if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
                 for (int b = 0; b < src_backend_id; b++) {
                     if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                         SET_CAUSE(tensor, "1.off");
@@ -1452,7 +1454,8 @@ ggml_backend_sched_t ggml_backend_sched_new(
         ggml_backend_buffer_type_t * bufts,
         int n_backends,
         size_t graph_size,
-        bool parallel) {
+        bool parallel,
+        bool op_offload) {
     GGML_ASSERT(n_backends > 0);
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
     GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
@@ -1497,6 +1500,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
     }
 
     sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
+    sched->op_offload = op_offload;
 
     ggml_backend_sched_reset(sched);
 

diff --git a/include/llama.h b/include/llama.h
index 7d5f9d559816d8975695d5d90fa1941ff931d2f8..6c6d377f85fb11b2eaf1aa3c2624c452fa0b3ab0 100644 (file)
--- a/include/llama.h
+++ b/include/llama.h
@@ -363,6 +363,7 @@ extern "C" {
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
         bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
         bool no_perf;     // whether to measure performance timings
+        bool op_offload;  // whether to offload host tensor operations to device
     };
 
     // model quantization parameters

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index fd64622b8e02d2e2a789e109e331ecc35906f069..a12849f0e0df459bab806862b7a17ef07e73e880 100644 (file)
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -93,6 +93,7 @@ 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;
 
     const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
 
@@ -243,7 +244,7 @@ llama_context::llama_context(
             }
         }
 
-        sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel));
+        sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel, cparams.op_offload));
 
         if (pipeline_parallel) {
             LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
@@ -1871,6 +1872,7 @@ llama_context_params llama_context_default_params() {
         /*.offload_kqv                 =*/ true,
         /*.flash_attn                  =*/ false,
         /*.no_perf                     =*/ true,
+        /*.op_offload                  =*/ true,
     };
 
     return result;

diff --git a/src/llama-cparams.h b/src/llama-cparams.h
index 30e550f023a9e323fae37fd364d9fcb8a3745e41..246fa5777deea1f6d4b94581d9b07b258b9434a2 100644 (file)
--- a/src/llama-cparams.h
+++ b/src/llama-cparams.h
@@ -30,6 +30,7 @@ struct llama_cparams {
     bool flash_attn;
     bool no_perf;
     bool warmup;
+    bool op_offload;
 
     enum llama_pooling_type pooling_type;
 

diff --git a/tests/test-opt.cpp b/tests/test-opt.cpp
index f90c92b4b8ecfc8eff87c8dbbb9b368c55b8b1ce..1bc160511357186c827287e6a7a9c9f59cc47764 100644 (file)
--- a/tests/test-opt.cpp
+++ b/tests/test-opt.cpp
@@ -853,7 +853,7 @@ int main(void) {
         backends_modded.insert(backends_modded.end(), backends.begin(), backends.end());
 
         ggml_backend_sched_t backend_sched = ggml_backend_sched_new(
-            backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false);
+            backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false, true);
 
         printf("Backend %zu/%zu: %s\n", i + 1, dev_count, ggml_backend_dev_name(devs[i]));
         printf("  Device description: %s\n", ggml_backend_dev_description(devs[i]));

diff --git a/tools/llama-bench/llama-bench.cpp b/tools/llama-bench/llama-bench.cpp
index 0786594296e94cb04e65618fcfaa4b8ec9de056b..5d26b506bd935e73fb1776ce0b5277d17084de22 100644 (file)
--- a/tools/llama-bench/llama-bench.cpp
+++ b/tools/llama-bench/llama-bench.cpp
@@ -219,6 +219,7 @@ struct cmd_params {
     std::vector<std::vector<llama_model_tensor_buft_override>> tensor_buft_overrides;
     std::vector<bool>                use_mmap;
     std::vector<bool>                embeddings;
+    std::vector<bool>                no_op_offload;
     ggml_numa_strategy               numa;
     int                              reps;
     ggml_sched_priority              prio;
@@ -253,6 +254,7 @@ static const cmd_params cmd_params_defaults = {
     /* tensor_buft_overrides*/ { std::vector<llama_model_tensor_buft_override>{{nullptr,nullptr}} },
     /* use_mmap             */ { true },
     /* embeddings           */ { false },
+    /* no_op_offload        */ { false },
     /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
     /* reps                 */ 5,
     /* prio                 */ GGML_SCHED_PRIO_NORMAL,
@@ -311,6 +313,7 @@ static void print_usage(int /* argc */, char ** argv) {
            join(cmd_params_defaults.embeddings, ",").c_str());
     printf("  -ts, --tensor-split <ts0/ts1/..>          (default: 0)\n");
     printf("  -ot --override-tensors <tensor name pattern>=<buffer type>;... (default: disabled)\n");
+    printf("  -nopo, --no-op-offload <i>                (default: 0)\n");
     printf("  -r, --repetitions <n>                     (default: %d)\n", cmd_params_defaults.reps);
     printf("  --prio <0|1|2|3>                          (default: %d)\n", cmd_params_defaults.prio);
     printf("  --delay <0...N> (seconds)                 (default: %d)\n", cmd_params_defaults.delay);
@@ -588,6 +591,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = string_split<bool>(argv[i], split_delim);
             params.embeddings.insert(params.embeddings.end(), p.begin(), p.end());
+        } else if (arg == "-nopo" || arg == "--no-op-offload") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<bool>(argv[i], split_delim);
+            params.no_op_offload.insert(params.no_op_offload.end(), p.begin(), p.end());
         } else if (arg == "-ts" || arg == "--tensor-split") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -794,6 +804,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.embeddings.empty()) {
         params.embeddings = cmd_params_defaults.embeddings;
     }
+    if (params.no_op_offload.empty()) {
+        params.no_op_offload = cmd_params_defaults.no_op_offload;
+    }
     if (params.n_threads.empty()) {
         params.n_threads = cmd_params_defaults.n_threads;
     }
@@ -833,6 +846,7 @@ struct cmd_params_instance {
     std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
     bool               use_mmap;
     bool               embeddings;
+    bool               no_op_offload;
 
     llama_model_params to_llama_mparams() const {
         llama_model_params mparams = llama_model_default_params();
@@ -902,6 +916,7 @@ struct cmd_params_instance {
         cparams.offload_kqv = !no_kv_offload;
         cparams.flash_attn  = flash_attn;
         cparams.embeddings  = embeddings;
+        cparams.op_offload  = !no_op_offload;
 
         return cparams;
     }
@@ -921,6 +936,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & ot : params.tensor_buft_overrides)
     for (const auto & mmp : params.use_mmap)
     for (const auto & embd : params.embeddings)
+    for (const auto & nopo : params.no_op_offload)
     for (const auto & nb : params.n_batch)
     for (const auto & nub : params.n_ubatch)
     for (const auto & tk : params.type_k)
@@ -959,6 +975,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
+                /* .no_op_offload= */ nopo,
             };
             instances.push_back(instance);
         }
@@ -990,6 +1007,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
+                /* .no_op_offload= */ nopo,
             };
             instances.push_back(instance);
         }
@@ -1021,6 +1039,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
+                /* .no_op_offload= */ nopo,
             };
             instances.push_back(instance);
         }
@@ -1056,6 +1075,7 @@ struct test {
     std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
     bool                     use_mmap;
     bool                     embeddings;
+    bool                     no_op_offload;
     int                      n_prompt;
     int                      n_gen;
     int                      n_depth;
@@ -1089,6 +1109,7 @@ struct test {
         tensor_buft_overrides = inst.tensor_buft_overrides;
         use_mmap       = inst.use_mmap;
         embeddings     = inst.embeddings;
+        no_op_offload  = inst.no_op_offload;
         n_prompt       = inst.n_prompt;
         n_gen          = inst.n_gen;
         n_depth        = inst.n_depth;
@@ -1134,7 +1155,7 @@ struct test {
             "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
             "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
             "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
-            "use_mmap",     "embeddings",   "n_prompt",       "n_gen",      "n_depth",      "test_time",
+            "use_mmap",     "embeddings",   "no_op_offload",   "n_prompt",       "n_gen",      "n_depth",      "test_time",
             "avg_ns",       "stddev_ns",    "avg_ts",         "stddev_ts",
         };
         return fields;
@@ -1146,7 +1167,7 @@ struct test {
         if (field == "build_number" || field == "n_batch" || field == "n_ubatch" || field == "n_threads" ||
             field == "poll" || field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" ||
             field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "n_depth" ||
-            field == "avg_ns" || field == "stddev_ns") {
+            field == "avg_ns" || field == "stddev_ns" || field == "no_op_offload") {
             return INT;
         }
         if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" || field == "flash_attn" ||
@@ -1222,6 +1243,7 @@ struct test {
                                             tensor_buft_overrides_str,
                                             std::to_string(use_mmap),
                                             std::to_string(embeddings),
+                                            std::to_string(no_op_offload),
                                             std::to_string(n_prompt),
                                             std::to_string(n_gen),
                                             std::to_string(n_depth),
@@ -1404,6 +1426,9 @@ struct markdown_printer : public printer {
         if (field == "test") {
             return 15;
         }
+        if (field == "no_op_offload") {
+            return 4;
+        }
 
         int width = std::max((int) field.length(), 10);
 
@@ -1435,6 +1460,9 @@ struct markdown_printer : public printer {
         if (field == "embeddings") {
             return "embd";
         }
+        if (field == "no_op_offload") {
+            return "nopo";
+        }
         if (field == "tensor_split") {
             return "ts";
         }
@@ -1503,6 +1531,9 @@ struct markdown_printer : public printer {
         if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) {
             fields.emplace_back("embeddings");
         }
+        if (params.no_op_offload.size() > 1 || params.no_op_offload != cmd_params_defaults.no_op_offload) {
+            fields.emplace_back("no_op_offload");
+        }
         fields.emplace_back("test");
         fields.emplace_back("t/s");
 

diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index 735dfe7f7802947c9fc125aea26dbcc5453cd02d..3f11c301a7212d40fd827d660a09a55f875175d1 100644 (file)
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -383,7 +383,7 @@ struct clip_ctx {
         backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu));
 
         sched.reset(
-            ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false)
+            ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false, true)
         );
     }