GPU offload policy (#405)

* Adding GPU offload policy

* Minor

---------

Co-authored-by: Iwan Kawrakow <[email protected]>

Author: ikawrakow

common/common.cpp

@@ -1213,6 +1213,17 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         }
         return true;
     }
+    if (arg == "--offload-policy" || arg == "-op") {
+        CHECK_ARG
+        auto p = string_split_pairs<int,int>(argv[i], ',');
+        if (p.empty()) {
+            fprintf(stderr, "error: Invalid offload policy argument: %s\n", argv[i]);
+            invalid_param = true;
+        } else {
+            params.offload_policy.insert(params.offload_policy.end(), p.begin(), p.end());
+        }
+        return true;
+    }
     if (arg == "--host") {
         CHECK_ARG
         params.hostname = argv[i];
@@ -2222,6 +2233,10 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         return iparams;
     }
 
+    for (auto [op, on_off] : params.offload_policy) {
+        llama_set_offload_policy(lctx, op, on_off);
+    }
+
     if (!params.control_vectors.empty()) {
         if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
         if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_n_layer(model);
@@ -2418,6 +2433,8 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
     cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
 
+    if (!params.offload_policy.empty()) cparams.offload_policy = (void *)&params.offload_policy;
+
     return cparams;
 }
 

common/common.h

@@ -143,6 +143,7 @@ struct gpt_params {
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
     std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
+    std::vector<std::pair<int,int>> offload_policy;
 
     bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
     std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale

ggml/include/ggml-backend.h

@@ -208,6 +208,9 @@ extern "C" {
     // Set a callback to be called for each resulting node during graph compute
     GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
 
+    // enable or disable op offload for a given op
+    GGML_API void                 ggml_backend_sched_set_op_offload(ggml_backend_sched_t sched, enum ggml_op op, bool on_or_off);
+
     //
     // Utils
     //

ggml/src/ggml-backend.c

@@ -1104,9 +1104,34 @@ struct ggml_backend_sched {
     char * context_buffer;
     size_t context_buffer_size;
 
+    uint32_t op_offload[(GGML_OP_COUNT + 31)/32];
+
     bool debug;
 };
 
+void ggml_backend_sched_set_op_offload(ggml_backend_sched_t sched, enum ggml_op op, bool on_or_off) {
+    int int_op = (int)op;
+    if (!sched) return;
+    if (int_op < 0 || int_op >= (int)GGML_OP_COUNT) {
+        uint32_t mask = on_or_off ? 0xffffffff : 0;
+        for (int i = 0; i < (GGML_OP_COUNT + 31)/32; ++i) sched->op_offload[i] = mask;
+        return;
+    }
+    int i = int_op >> 5;
+    int j = int_op & 31;
+    if (on_or_off) {
+        sched->op_offload[i] |= (1u << j);
+    } else {
+        sched->op_offload[i] &= (~(1u << j));
+    }
+}
+
+static inline bool ggml_backend_sched_offload_enabled(ggml_backend_sched_t sched, enum ggml_op op) {
+    int int_op = (int)op;
+    if (!sched || op < 0 || op >= GGML_OP_COUNT) return false;
+    return sched->op_offload[int_op >> 5] & (1u << (int_op & 31));
+}
+
 #define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
 #define tensor_backend_id(tensor) sched->hv_tensor_backend_ids[hash_id(tensor)]
 #define tensor_id_copy(id, backend_id, copy_id) sched->hv_tensor_copies[(id) * sched->n_backends * sched->n_copies + (backend_id) * sched->n_copies + (copy_id)]
@@ -1181,6 +1206,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
     }
 
     // operations with weights are preferably run on the same backend as the weights
+    bool offload_enabled = ggml_backend_sched_offload_enabled(sched, tensor->op);
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         const struct ggml_tensor * src = tensor->src[i];
         if (src == NULL) {
@@ -1189,7 +1215,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         if (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) {
+            if (offload_enabled && src_backend_id == sched->n_backends - 1) {
                 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");
@@ -1888,6 +1914,8 @@ ggml_backend_sched_t ggml_backend_sched_new(
 
     struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
 
+    for (int i = 0; i < (GGML_OP_COUNT + 31)/32; ++i) sched->op_offload[i] = 0xffffffff;
+
     sched->debug = getenv("GGML_SCHED_DEBUG") != NULL;
     sched->n_backends = n_backends;
     sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;

ggml/src/ggml-cuda.cu

@@ -3391,6 +3391,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                 struct ggml_tensor * a = op->src[0];
                 struct ggml_tensor * b = op->op == GGML_OP_MOE_FUSED_UP_GATE ? op->src[2] : op->src[1];
                 if (op->op == GGML_OP_MOE_FUSED_UP_GATE && a->type != op->src[1]->type) {
+                    printf("%s: returning false for GGML_OP_MOE_FUSED_UP_GATE because src0->type != src1->type\n", __func__);
                     return false;
                 }
                 //==================================================================
@@ -3399,6 +3400,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                 //}
                 //==================================================================
                 if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16 && !ggml_is_quantized(a->type)) {
+                    printf("%s: returning false for op %d because (case 1)\n", __func__, (int)op->op);
                     return false;
                 }
                 if (op->op == GGML_OP_MUL_MAT && a->ne[3] != b->ne[3]) {
@@ -3621,7 +3623,7 @@ GGML_CALL static bool ggml_backend_cuda_offload_op(ggml_backend_t backend, const
     const int min_batch_size = 32;
 
     return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
-           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
+           (op->ne[2] >= min_batch_size && (op->op == GGML_OP_MUL_MAT_ID || op->op == GGML_OP_MOE_FUSED_UP_GATE));
 
     GGML_UNUSED(backend);
 }

include/llama.h

@@ -408,6 +408,7 @@ extern "C" {
         // currently works only with CPU execution
         ggml_abort_callback abort_callback;
         void *              abort_callback_data;
+        void *              offload_policy;
     };
 
     // model quantization parameters
@@ -523,6 +524,8 @@ extern "C" {
                      struct llama_model * model,
             struct llama_context_params   params);
 
+    LLAMA_API void llama_set_offload_policy(struct llama_context * lctx, int op, bool on_or_off);
+
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 

src/llama.cpp

@@ -19980,6 +19980,7 @@ struct llama_context_params llama_context_default_params() {
         /*.thtesh_experts              =*/ 0.0f,
         /*.abort_callback              =*/ nullptr,
         /*.abort_callback_data         =*/ nullptr,
+        /*.offload_policy              =*/ nullptr,
     };
 
     return result;
@@ -20574,6 +20575,19 @@ struct llama_context * llama_new_context_with_model(
         }
     }
 
+    if (params.offload_policy) {
+        const std::vector<std::pair<int, int>>& policy = *(const std::vector<std::pair<int, int>>*)params.offload_policy;
+        for (auto [op, on_off] : policy) {
+            if (op < 0 || op >= int(GGML_OP_COUNT)) {
+                LLAMA_LOG_INFO("XXXXXXXXXXXXXXXXXXXXX Setting offload policy for all ops to %s\n", on_off ? "ON" : "OFF");
+            } else {
+                LLAMA_LOG_INFO("XXXXXXXXXXXXXXXXXXXXX Setting offload policy for op %s to %s\n",
+                        ggml_op_name(ggml_op(op)), on_off ? "ON" : "OFF");
+            }
+            ggml_backend_sched_set_op_offload(ctx->sched, ggml_op(op), on_off);
+        }
+    }
+
     return ctx;
 }
 
@@ -23222,3 +23236,10 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
     fputs(text, stderr);
     fflush(stderr);
 }
+
+void llama_set_offload_policy(struct llama_context * lctx, int op, bool on_or_off) {
+    if (!lctx || !lctx->sched) return;
+    const char * op_name = op < 0 || op >= int(GGML_OP_COUNT) ? "all ops" : ggml_op_name(ggml_op(op));
+    printf("XXXXXXXXXXXXXXXXXXXXXXXXXXXX offload(%s) = %d\n", op_name, on_or_off);
+    ggml_backend_sched_set_op_offload(lctx->sched, ggml_op(op), on_or_off);
+}