server : fix draft model fit vs load inconsistency (#25056)

* fix: draft model fit vs load inconsistency

* refactor(server): unify draft/mtp parameter initialization, model, and context load
- moves speculative init to speculative.cpp
- changes server_context_impl model_dft and ctx_dft to use raw pointers

- fix: don't throttle progress callback when loading draft model
- refactor: rename draft model/ctx load method

* fix: valign

diff --git a/common/speculative.cpp b/common/speculative.cpp
index 5b26597f62c6ab38cfb48c15d951a175828ba792..1d2977e424f817f213926ce1da4c98ec0051915c 100644 (file)
--- a/common/speculative.cpp
+++ b/common/speculative.cpp
@@ -2221,6 +2221,113 @@ int32_t common_speculative_n_max(const common_params_speculative * spec) {
     return n_max;
 }
 
+common_params common_base_params_to_speculative(const common_params & params) {
+    const bool has_draft = params.speculative.has_dft();
+
+    const auto & params_spec = params.speculative.draft;
+    common_params result = params;
+
+    if (has_draft) {
+        result.devices               = params_spec.devices;
+        result.model                 = params_spec.mparams;
+        result.n_gpu_layers          = params_spec.n_gpu_layers;
+        result.tensor_buft_overrides = params_spec.tensor_buft_overrides;
+
+        if (params_spec.cpuparams.n_threads > 0) {
+            result.cpuparams.n_threads       = params_spec.cpuparams.n_threads;
+            result.cpuparams_batch.n_threads = params_spec.cpuparams_batch.n_threads;
+        }
+    }
+
+    result.cache_type_k  = params_spec.cache_type_k;
+    result.cache_type_v  = params_spec.cache_type_v;
+    result.n_outputs_max = params.n_parallel;
+
+    return result;
+}
+
+struct common_init_speculative_result::impl {
+    impl() = default;
+    ~impl() = default;
+
+    // note: the order in which model, context, etc. are declared matters because their destructors will be called bottom-to-top
+    llama_model_ptr   model;
+    llama_context_ptr context;
+};
+
+common_init_speculative_result::common_init_speculative_result(
+    common_params & params,
+      llama_model * model_tgt,
+    llama_context * ctx_tgt) :
+    pimpl(new impl{}) {
+    const bool has_draft = params.speculative.has_dft();
+    const bool spec_mtp = std::find(params.speculative.types.begin(),
+                                    params.speculative.types.end(),
+                                    COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
+    GGML_ASSERT(has_draft || spec_mtp);
+
+    auto mparams = common_model_params_to_llama(params);
+    auto cparams = common_context_params_to_llama(params);
+
+    if (spec_mtp) {
+        cparams.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
+    }
+
+    // note: for small models maybe we can set this to the maximum possible draft from all speculative types
+    //       the extra memory for small models is likely negligible?
+    cparams.n_rs_seq  = 0;
+    cparams.ctx_other = ctx_tgt;
+
+    std::string model_path;
+    if (has_draft) {
+        model_path = params.speculative.draft.mparams.path;
+        LOG_TRC("%s: loading draft model '%s'\n", __func__, model_path.c_str());
+
+        llama_model * model_dft = llama_model_load_from_file(params.model.path.c_str(), mparams);
+        if (model_dft == NULL) {
+            LOG_ERR("%s: failed to load draft model, '%s'\n", __func__, model_path.c_str());
+            return;
+        }
+
+        pimpl->model.reset(model_dft);
+
+        llama_context * ctx_dft = llama_init_from_model(model_dft, cparams);
+        if (ctx_dft == nullptr) {
+            LOG_ERR("%s: failed to create MTP context\n", __func__);
+            return;
+        }
+
+        pimpl->context.reset(ctx_dft);
+    } else if (spec_mtp) {
+        model_path = params.model.path;
+
+        LOG_TRC("%s: creating MTP draft context against the target model '%s'\n", 
+            __func__, model_path.c_str());
+
+        llama_context * ctx_dft = llama_init_from_model(model_tgt, cparams);
+        if (ctx_dft == nullptr) {
+            LOG_ERR("%s: failed to create MTP context\n", __func__);
+            return;
+        }
+
+        pimpl->context.reset(ctx_dft);
+    }
+}
+
+common_init_speculative_result::~common_init_speculative_result() = default;
+
+llama_model * common_init_speculative_result::model() {
+    return pimpl->model.get();
+}
+
+llama_context * common_init_speculative_result::context() {
+    return pimpl->context.get();
+}
+
+common_init_speculative_result_ptr common_init_speculative_from_params(common_params & params, llama_model * model_tgt, llama_context * ctx_tgt) {
+    return std::make_unique<common_init_speculative_result>(params, model_tgt, ctx_tgt);
+}
+
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq) {

diff --git a/common/speculative.h b/common/speculative.h
index c58fac3cc6d05a6ba379a625b80447355680cd79..2f6d225b6f4a34fe270d852a40cee0f6696e7bf1 100644 (file)
--- a/common/speculative.h
+++ b/common/speculative.h
@@ -23,6 +23,8 @@ std::string common_speculative_type_to_str(enum common_speculative_type type);
 // return the max number of draft tokens based on the speculative parameters
 int32_t common_speculative_n_max(const common_params_speculative * spec);
 
+common_params common_base_params_to_speculative(const common_params & params);
+
 common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq);
 
 void common_speculative_free(common_speculative * spec);
@@ -80,3 +82,19 @@ struct common_speculative_deleter {
 };
 
 typedef std::unique_ptr<common_speculative, common_speculative_deleter> common_speculative_ptr;
+
+struct common_init_speculative_result {
+    common_init_speculative_result(common_params & params, llama_model * model_tgt, llama_context * ctx_tgt);
+    ~common_init_speculative_result();
+
+    llama_model   * model();
+    llama_context * context();
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+using common_init_speculative_result_ptr = std::unique_ptr<common_init_speculative_result>;
+
+common_init_speculative_result_ptr common_init_speculative_from_params(common_params & params, llama_model * model_tgt, llama_context * ctx_tgt);

diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index bb3b91ab5eda4b00fad41ff3d21470855a98fd8f..3890fe5f05c1734eb8d235225cc97d177ecd094d 100644 (file)
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -897,8 +897,10 @@ private:
 
     server_batch batch;
 
-    llama_model_ptr model_dft;
-    llama_context_ptr ctx_dft;
+    llama_model   * model_dft = nullptr;
+    llama_context * ctx_dft   = nullptr;
+
+    common_init_speculative_result_ptr spec_init;
 
     common_context_seq_rm_type ctx_tgt_seq_rm_type = COMMON_CONTEXT_SEQ_RM_TYPE_NO;
     common_context_seq_rm_type ctx_dft_seq_rm_type = COMMON_CONTEXT_SEQ_RM_TYPE_NO;
@@ -939,8 +941,10 @@ private:
 
     void destroy() {
         spec.reset();
-        ctx_dft.reset();
-        model_dft.reset();
+        spec_init.reset();
+
+        ctx_dft   = nullptr;
+        model_dft = nullptr;
 
         llama_init.reset();
 
@@ -1084,30 +1088,15 @@ private:
         // optionally reserve VRAM for the draft / MTP context before fitting the target model
         if (params_base.fit_params) {
             if (has_spec) {
-                common_params params_dft = params_base;
-                bool measure_model_bytes = true;
-
-                if (has_draft) {
-                    const auto & params_spec = params_base.speculative.draft;
-                    params_dft.devices               = params_spec.devices;
-                    params_dft.model                 = params_spec.mparams;
-                    params_dft.n_gpu_layers          = params_spec.n_gpu_layers;
-                    params_dft.cache_type_k          = params_spec.cache_type_k;
-                    params_dft.cache_type_v          = params_spec.cache_type_v;
-                    params_dft.tensor_buft_overrides = params_spec.tensor_buft_overrides;
-                } else {
-                    // MTP draft context lives on the target model, only context+compute are new
-                    measure_model_bytes = false;
-                }
+                // MTP draft context lives on the target model, only context+compute are new
+                bool measure_model_bytes = has_draft;
 
-                params_dft.n_outputs_max = params_base.n_parallel;
+                common_params params_dft = common_base_params_to_speculative(params_base);
 
                 auto mparams_dft = common_model_params_to_llama(params_dft);
                 auto cparams_dft = common_context_params_to_llama(params_dft);
                 if (spec_mtp) {
                     cparams_dft.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
-                    cparams_dft.type_k   = params_base.speculative.draft.cache_type_k;
-                    cparams_dft.type_v   = params_base.speculative.draft.cache_type_v;
                 }
                 cparams_dft.n_rs_seq = 0;
 
@@ -1175,82 +1164,35 @@ private:
 
         add_bos_token = llama_vocab_get_add_bos(vocab);
 
-        if (has_draft) {
-            // TODO speculative: move to common/speculative.cpp?
-            const auto & params_spec = params_base.speculative.draft;
-
-            SRV_TRC("loading draft model '%s'\n", params_spec.mparams.path.c_str());
-
-            auto params_dft = params_base;
-
-            params_dft.devices      = params_spec.devices;
-            params_dft.model        = params_spec.mparams;
-            params_dft.n_gpu_layers = params_spec.n_gpu_layers;
-            params_dft.cache_type_k = params_spec.cache_type_k;
-            params_dft.cache_type_v = params_spec.cache_type_v;
-
-            if (params_spec.cpuparams.n_threads > 0) {
-                params_dft.cpuparams.n_threads       = params_spec.cpuparams.n_threads;
-                params_dft.cpuparams_batch.n_threads = params_spec.cpuparams_batch.n_threads;
-            }
-
-            params_dft.tensor_buft_overrides = params_spec.tensor_buft_overrides;
-
-            auto mparams_dft = common_model_params_to_llama(params_dft);
-
-            // progress callback
-            mparams_dft.progress_callback           = load_progress_callback;
-            mparams_dft.progress_callback_user_data = &load_progress_spec;
-
-            model_dft.reset(llama_model_load_from_file(params_dft.model.path.c_str(), mparams_dft));
-            if (model_dft == nullptr) {
-                SRV_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str());
-                return false;
-            }
-
-            auto cparams = common_context_params_to_llama(params_dft);
-
-            if (spec_mtp) {
-                cparams.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
-            }
+        if (has_spec) {
+            // spec_mtp doesn't use load a model internally, so we report 0.0 and 1.0 manually
+            load_progress_callback(0.0f, &load_progress_spec);
+            load_progress_spec.t_last_load_progress_ms = 0;  // reset so internal cbs aren't delayed
 
-            // note: for small models maybe we can set this to the maximum possible draft from all speculative types
-            //       the extra memory for small models is likely negligible?
-            cparams.n_rs_seq  = 0;
-            cparams.ctx_other = ctx_tgt;
+            {
+                common_params params_dft = common_base_params_to_speculative(params_base);
+                // progress callback
+                params_dft.load_progress_callback           = load_progress_callback;
+                params_dft.load_progress_callback_user_data = &load_progress_spec;
 
-            ctx_dft.reset(llama_init_from_model(model_dft.get(), cparams));
-            if (ctx_dft == nullptr) {
-                SRV_ERR("%s", "failed to create draft context\n");
-                return false;
-            }
+                spec_init = common_init_speculative_from_params(params_dft, model_tgt, ctx_tgt);
+                model_dft = spec_init->model();
+                ctx_dft   = spec_init->context();
 
-            params_base.speculative.draft.ctx_tgt = ctx_tgt;
-            params_base.speculative.draft.ctx_dft = ctx_dft.get();
-        } else if (spec_mtp) {
-            // no new model load, so we simply report 0.0 and 1.0 progress
-            load_progress_callback(0.0f, &load_progress_spec);
-
-            SRV_TRC("creating MTP draft context against the target model '%s'\n",
-                    params_base.model.path.c_str());
+                if (has_draft && model_dft == nullptr) {
+                    SRV_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str());
+                    return false;
+                }
 
-            auto cparams_mtp = common_context_params_to_llama(params_base);
-            cparams_mtp.ctx_type      = LLAMA_CONTEXT_TYPE_MTP;
-            cparams_mtp.type_k        = params_base.speculative.draft.cache_type_k;
-            cparams_mtp.type_v        = params_base.speculative.draft.cache_type_v;
-            cparams_mtp.n_rs_seq      = 0;
-            cparams_mtp.n_outputs_max = params_base.n_parallel;
-            cparams_mtp.ctx_other     = ctx_tgt;
+                if (ctx_dft == nullptr) {
+                    SRV_ERR("%s", "failed to create MTP context\n");
+                    return false;
+                }
 
-            ctx_dft.reset(llama_init_from_model(model_tgt, cparams_mtp));
-            if (ctx_dft == nullptr) {
-                SRV_ERR("%s", "failed to create MTP context\n");
-                return false;
+                params_base.speculative.draft.ctx_tgt = ctx_tgt;
+                params_base.speculative.draft.ctx_dft = ctx_dft;
             }
 
-            params_base.speculative.draft.ctx_tgt = ctx_tgt;
-            params_base.speculative.draft.ctx_dft = ctx_dft.get();
-
             load_progress_callback(1.0f, &load_progress_spec);
         }
 
@@ -1343,13 +1285,15 @@ private:
         }
 
         if (ctx_dft) {
-            ctx_dft_seq_rm_type = common_context_can_seq_rm(ctx_dft.get());
+            ctx_dft_seq_rm_type = common_context_can_seq_rm(ctx_dft);
         }
 
         if (spec) {
             SRV_TRC("%s", "speculative decoding context initialized\n");
         } else {
-            ctx_dft.reset();
+            spec_init.reset();
+            ctx_dft   = nullptr;
+            model_dft = nullptr;
         }
 
         for (int i = 0; i < params_base.n_parallel; i++) {
@@ -1357,7 +1301,7 @@ private:
 
             slot.id      = i;
             slot.ctx_tgt = ctx_tgt;
-            slot.ctx_dft = ctx_dft.get();
+            slot.ctx_dft = ctx_dft;
             slot.spec    = spec.get();
             slot.n_ctx   = n_ctx_slot;
 
@@ -2363,7 +2307,7 @@ private:
         cur.update_pos(slot.prompt.n_tokens() - n_tokens_cur, pos_min, pos_max);
 
         cur.update_tgt(ctx_tgt,       slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
-        cur.update_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+        cur.update_dft(ctx_dft,       slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
         // stash the draft's speculative state with the checkpoint
         common_speculative_get_state(spec.get(), slot.id, cur.data_spec);
 
@@ -2899,8 +2843,8 @@ private:
                 common_context_seq_add(ctx_tgt, slot.id, n_keep + n_discard, slot.prompt.n_tokens(), -n_discard);
 
                 if (ctx_dft) {
-                    common_context_seq_rm (ctx_dft.get(), slot.id, n_keep            , n_keep + n_discard);
-                    common_context_seq_add(ctx_dft.get(), slot.id, n_keep + n_discard, slot.prompt.tokens.pos_next(), -n_discard);
+                    common_context_seq_rm (ctx_dft, slot.id, n_keep            , n_keep + n_discard);
+                    common_context_seq_add(ctx_dft, slot.id, n_keep + n_discard, slot.prompt.tokens.pos_next(), -n_discard);
                 }
 
                 // add generated tokens to cache
@@ -2972,7 +2916,7 @@ private:
                                 llama_memory_seq_pos_max(llama_get_memory(ctx_tgt), slot.id));
 
                         if (use_ckpt_dft) {
-                            slot.spec_ckpt.update_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                            slot.spec_ckpt.update_dft(ctx_dft, slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
                         }
 
                         slot.spec_prompt = slot.prompt.tokens.get_text_tokens();
@@ -3009,10 +2953,10 @@ private:
 
             if (ctx_dft) {
                 if (use_ckpt_dft) {
-                    ckpt.load_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                    ckpt.load_dft(ctx_dft, slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
                 }
 
-                common_context_seq_rm(ctx_dft.get(), slot.id, ckpt.pos_max + 1, -1);
+                common_context_seq_rm(ctx_dft, slot.id, ckpt.pos_max + 1, -1);
             }
 
             if (!draft.empty()) {
@@ -3021,7 +2965,7 @@ private:
                    (ctx_tgt_seq_rm_type == COMMON_CONTEXT_SEQ_RM_TYPE_RS && draft.size() > llama_n_rs_seq(ctx_tgt));
 
                 const bool use_ckpt_dft =
-                   (ctx_dft_seq_rm_type == COMMON_CONTEXT_SEQ_RM_TYPE_RS && draft.size() > llama_n_rs_seq(ctx_dft.get()));
+                   (ctx_dft_seq_rm_type == COMMON_CONTEXT_SEQ_RM_TYPE_RS && draft.size() > llama_n_rs_seq(ctx_dft));
 
                 if (use_ckpt_tgt) {
                     //const int64_t t_start = ggml_time_us();
@@ -3038,7 +2982,7 @@ private:
                 }
 
                 if (use_ckpt_dft) {
-                    ckpt.update_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                    ckpt.update_dft(ctx_dft, slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
                 }
             }
         });
@@ -3219,8 +3163,8 @@ private:
                                             common_context_seq_add(ctx_tgt, slot.id, head_c, head_c + n_match, kv_shift);
 
                                             if (ctx_dft) {
-                                                common_context_seq_rm (ctx_dft.get(), slot.id, head_p, head_c);
-                                                common_context_seq_add(ctx_dft.get(), slot.id, head_c, head_c + n_match, kv_shift);
+                                                common_context_seq_rm (ctx_dft, slot.id, head_p, head_c);
+                                                common_context_seq_add(ctx_dft, slot.id, head_c, head_c + n_match, kv_shift);
                                             }
 
                                             for (size_t i = 0; i < n_match; i++) {
@@ -3321,7 +3265,7 @@ private:
                                     if (!do_reset) {
                                         // restore the context checkpoint
                                         it->load_tgt(ctx_tgt,       slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
-                                        it->load_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                                        it->load_dft(ctx_dft,       slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
                                         // restore the draft's speculative state
                                         common_speculative_set_state(spec.get(), slot.id, it->data_spec);
 
@@ -3395,7 +3339,7 @@ private:
 
                     common_context_seq_rm(ctx_tgt, slot.id, p0, -1);
                     if (ctx_dft) {
-                        common_context_seq_rm(ctx_dft.get(), slot.id, p0, -1);
+                        common_context_seq_rm(ctx_dft, slot.id, p0, -1);
                     }
 
                     // If using an alora, there may be uncached tokens that come