llama: avoid copying logits during prompt decode in MTP (#23198)

* llama: avoid copying logits during prompt decode in MTP

* review: update comment

* llama-graph: call set_output for t_h_pre_norm

diff --git a/common/speculative.cpp b/common/speculative.cpp
index 3488b9393c5a4ebb29d0d9f7cc3773f75a420ac4..e591bab875dbabe6ceb6f7cda3331fcdf46a07f1 100644 (file)
--- a/common/speculative.cpp
+++ b/common/speculative.cpp
@@ -146,8 +146,11 @@ struct common_speculative_impl {
 
     virtual void accept(llama_seq_id seq_id, uint16_t n_accepted) = 0;
 
-    // true if this implementation requires the target context to extract embeddings
+    // true if this implementation requires the target context to extract post-norm embeddings
     virtual bool need_embd() const = 0;
+
+    // true if this implementation requires the target context to extract pre-norm embeddings
+    virtual bool need_embd_pre_norm() const { return false; }
 };
 
 struct common_speculative_impl_draft_simple : public common_speculative_impl {
@@ -429,8 +432,8 @@ struct common_speculative_state_draft_mtp : public common_speculative_impl {
             s.reset(common_sampler_init(llama_get_model(ctx_dft), sparams));
         }
 
-        llama_set_embeddings_pre_norm(ctx_tgt, true);
-        llama_set_embeddings_pre_norm(ctx_dft, true);
+        llama_set_embeddings_pre_norm(ctx_tgt, true, /*masked*/ false);
+        llama_set_embeddings_pre_norm(ctx_dft, true, /*masked*/ true);
 
         pending_h.assign(n_seq, std::vector<float>(n_embd, 0.0f));
 
@@ -691,6 +694,10 @@ struct common_speculative_state_draft_mtp : public common_speculative_impl {
     }
 
     bool need_embd() const override {
+        return false;
+    }
+
+    bool need_embd_pre_norm() const override {
         return true;
     }
 };
@@ -1408,6 +1415,20 @@ bool common_speculative_need_embd(common_speculative * spec) {
     return false;
 }
 
+bool common_speculative_need_embd_pre_norm(common_speculative * spec) {
+    if (spec == nullptr) {
+        return false;
+    }
+
+    for (auto & impl : spec->impls) {
+        if (impl->need_embd_pre_norm()) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
 void common_speculative_draft(common_speculative * spec) {
     if (spec == nullptr) {
         return;

diff --git a/common/speculative.h b/common/speculative.h
index 614db9b1b5093e961ec4dab5682ffc1a19e7d458..f24bac79edb70df3183add4348842dad87033e2b 100644 (file)
--- a/common/speculative.h
+++ b/common/speculative.h
@@ -53,9 +53,12 @@ void common_speculative_begin(common_speculative * spec, llama_seq_id seq_id, co
 // process the batch and update the internal state of the speculative context
 bool common_speculative_process(common_speculative * spec, const llama_batch & batch);
 
-// true if any implementation requires target embeddings to be extracted
+// true if any implementation requires target post-norm embeddings to be extracted
 bool common_speculative_need_embd(common_speculative * spec);
 
+// true if any implementation requires target pre-norm embeddings to be extracted
+bool common_speculative_need_embd_pre_norm(common_speculative * spec);
+
 // generate drafts for the sequences specified with `common_speculative_get_draft_params`
 void common_speculative_draft(common_speculative * spec);
 

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index d62abc4009b85622a8f40dba1a978185880fc448..b1b12d017c0a9d0a2599c871db979db2bc5e3715 100644 (file)
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -895,8 +895,17 @@ float * llama_context::get_embeddings_pre_norm_ith(int32_t i) {
             throw std::runtime_error("no pre-norm embeddings");
         }
 
-        const int64_t j = output_resolve_row(i);
         const uint32_t n_embd = model.hparams.n_embd;
+
+        if (!cparams.embeddings_pre_norm_masked) {
+            // unmasked: pre-norm rows are stored densely, indexed by raw token position.
+            if (i < 0 || (size_t)(i + 1) * n_embd > embd_pre_norm.size) {
+                throw std::runtime_error(format("out of range [0, %zu)", embd_pre_norm.size / n_embd));
+            }
+            return embd_pre_norm.data + (size_t) i * n_embd;
+        }
+
+        const int64_t j = output_resolve_row(i);
         return embd_pre_norm.data + j*n_embd;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: invalid pre-norm embeddings id %d, reason: %s\n", __func__, i, err.what());
@@ -1088,10 +1097,11 @@ void llama_context::set_embeddings(bool value) {
     //sched_need_reserve = true;
 }
 
-void llama_context::set_embeddings_pre_norm(bool value) {
-    LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value);
+void llama_context::set_embeddings_pre_norm(bool value, bool masked) {
+    LLAMA_LOG_DEBUG("%s: value = %d, masked = %d\n", __func__, value, masked);
 
-    cparams.embeddings_pre_norm = value;
+    cparams.embeddings_pre_norm        = value;
+    cparams.embeddings_pre_norm_masked = masked;
 }
 
 void llama_context::set_causal_attn(bool value) {
@@ -1737,6 +1747,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     };
 
     int64_t n_outputs_prev = 0;
+    int64_t n_tokens_prev  = 0;
 
     do {
         const auto & ubatch = mctx->get_ubatch();
@@ -1882,16 +1893,21 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
         // extract pre-norm embeddings (hidden state before the final output norm)
         // only meaningful in LLAMA_POOLING_TYPE_NONE (per-token); other pooling modes are ignored.
-        if (embd_pre_norm.data && t_h_pre_norm && n_outputs > 0 && cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
-            ggml_backend_t backend_h = ggml_backend_sched_get_tensor_backend(sched.get(), t_h_pre_norm);
-            GGML_ASSERT(backend_h != nullptr);
+        {
+            const bool masked    = cparams.embeddings_pre_norm_masked;
+            const int64_t n_rows = masked ? n_outputs       : (int64_t) ubatch.n_tokens;
+            const int64_t offset = masked ? n_outputs_prev  : n_tokens_prev;
+
+            if (embd_pre_norm.data && t_h_pre_norm && n_rows > 0 && cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
+                ggml_backend_t backend_h = ggml_backend_sched_get_tensor_backend(sched.get(), t_h_pre_norm);
+                GGML_ASSERT(backend_h != nullptr);
 
-            const uint32_t n_embd = hparams.n_embd;
-            float * embd_pre_norm_out = embd_pre_norm.data + n_outputs_prev*n_embd;
+                const uint32_t n_embd = hparams.n_embd;
+                float * embd_pre_norm_out = embd_pre_norm.data + offset*n_embd;
 
-            GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all);
-            GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd <= (int64_t) embd_pre_norm.size);
-            ggml_backend_tensor_get_async(backend_h, t_h_pre_norm, embd_pre_norm_out, 0, n_outputs*n_embd*sizeof(float));
+                GGML_ASSERT((offset + n_rows)*n_embd <= (int64_t) embd_pre_norm.size);
+                ggml_backend_tensor_get_async(backend_h, t_h_pre_norm, embd_pre_norm_out, 0, n_rows*n_embd*sizeof(float));
+            }
         }
 
         // Copy backend sampling output if this ubatch produced any sampling tensors.
@@ -1908,6 +1924,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
         }
 
         n_outputs_prev += n_outputs;
+        n_tokens_prev  += ubatch.n_tokens;
     } while (mctx->next());
 
     // set to total number of outputs in the batch, for use in llama_get_logits_ith
@@ -1999,6 +2016,12 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     embd.size          = has_embd          ? n_embd_out*n_outputs_max  : 0;
     embd_pre_norm.size = has_embd_pre_norm ? n_embd*n_outputs_max      : 0;
 
+    if (has_embd_pre_norm && !cparams.embeddings_pre_norm_masked) {
+        // unmasked: pre-norm row exists for every token in the batch, not just
+        // those flagged via batch.logits[i] -> size by token count instead.
+        embd_pre_norm.size = (size_t) n_embd * n_batch;
+    }
+
     // Allocate backend sampling output buffers if there are backend samplers configured.
     const bool has_sampling = !sampling.samplers.empty();
     if (has_sampling) {
@@ -3547,8 +3570,8 @@ float * llama_get_embeddings_seq(llama_context * ctx, llama_seq_id seq_id) {
     return ctx->get_embeddings_seq(seq_id);
 }
 
-void llama_set_embeddings_pre_norm(llama_context * ctx, bool value) {
-    ctx->set_embeddings_pre_norm(value);
+void llama_set_embeddings_pre_norm(llama_context * ctx, bool value, bool masked) {
+    ctx->set_embeddings_pre_norm(value, masked);
 }
 
 float * llama_get_embeddings_pre_norm(llama_context * ctx) {

diff --git a/src/llama-context.h b/src/llama-context.h
index e16ac4c618baea4429cd4570123baaf1a42962db..d03f681d4a136fe5616f49f48977610129e06422 100644 (file)
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -110,7 +110,7 @@ struct llama_context {
     void set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data);
 
     void set_embeddings (bool value);
-    void set_embeddings_pre_norm(bool value);
+    void set_embeddings_pre_norm(bool value, bool masked);
     void set_causal_attn(bool value);
     void set_warmup(bool value);
 

diff --git a/src/llama-cparams.h b/src/llama-cparams.h
index 5898a1c38d511ef6135e6e25f52b338b9a252a71..20ec59fe3357ee3232c67f0f0181c9d83646b3da 100644 (file)
--- a/src/llama-cparams.h
+++ b/src/llama-cparams.h
@@ -28,7 +28,8 @@ struct llama_cparams {
     float yarn_beta_slow;
 
     bool embeddings;
-    bool embeddings_pre_norm; // also extract the hidden state before the final output norm
+    bool embeddings_pre_norm;        // also extract the hidden state before the final output norm
+    bool embeddings_pre_norm_masked; // extract for only rows where batch.logits != 0
     bool causal_attn;
     bool offload_kqv;
     bool flash_attn;

diff --git a/src/llama-ext.h b/src/llama-ext.h
index 11f1986676a571d9744846c3d07025a5edd0c252..edfa71c207c59c870405b6a5fb8943a40f8959d5 100644 (file)
--- a/src/llama-ext.h
+++ b/src/llama-ext.h
@@ -93,14 +93,14 @@ LLAMA_API llama_memory_breakdown llama_get_memory_breakdown(const struct llama_c
 // pre-norm embeddings (hidden state before the final output norm)
 //
 
-// mirrors:
-// LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings);
-LLAMA_API void llama_set_embeddings_pre_norm(struct llama_context * ctx, bool value);
+// Set whether the context outputs pre-norm embeddings or not
+// If masked == true,  output the embeddings only for the tokens with batch.logits != 0
+// If masked == false, output the embeddings for all tokens in the batch regardless of batch.logits
+LLAMA_API void llama_set_embeddings_pre_norm(struct llama_context * ctx, bool value, bool masked);
 
 // mirrors:
 // LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
-LLAMA_API float * llama_get_embeddings_pre_norm(struct llama_context * ctx);
+LLAMA_API float * llama_get_embeddings_pre_norm    (struct llama_context * ctx);
 
-// mirrors:
 // LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
 LLAMA_API float * llama_get_embeddings_pre_norm_ith(struct llama_context * ctx, int32_t i);

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 858c297dd76288ce1f4207241109cf2eb24bc3ac..31cf41a1c2d236582c2ea19911363ebeb028c4e0 100644 (file)
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -848,6 +848,9 @@ void llm_graph_result::set_outputs() {
     if (t_embd_pooled != nullptr) {
         ggml_set_output(t_embd_pooled);
     }
+    if (t_h_pre_norm != nullptr) {
+        ggml_set_output(t_h_pre_norm);
+    }
     for (auto & [seq_id, t] : t_sampled) {
         if (t != nullptr) {
             ggml_set_output(t);

diff --git a/src/models/qwen35.cpp b/src/models/qwen35.cpp
index 2b4d5b14cd42cd595d92a4b01af851d731abd062..361d7538a0385f8ae57ecfe69b5de296b4ac3517 100644 (file)
--- a/src/models/qwen35.cpp
+++ b/src/models/qwen35.cpp
@@ -176,7 +176,7 @@ llama_model_qwen35::graph::graph(const llama_model & model, const llm_graph_para
             cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
         }
 
-        if (il == n_transformer_layers - 1 && inp_out_ids) {
+        if (il == n_transformer_layers - 1 && inp_out_ids && cparams.embeddings_pre_norm_masked) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }
@@ -211,6 +211,10 @@ llama_model_qwen35::graph::graph(const llama_model & model, const llm_graph_para
     cb(cur, "h_pre_norm", -1);
     res->t_h_pre_norm = cur;
 
+    if (!cparams.embeddings_pre_norm_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
     // Final norm
     cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
 

diff --git a/src/models/qwen35moe.cpp b/src/models/qwen35moe.cpp
index 22e3e1107655386af47c074f9f53f636559d4a0e..4f63c410d668b326b77bd7cc980ab0ca8f5ca0f6 100644 (file)
--- a/src/models/qwen35moe.cpp
+++ b/src/models/qwen35moe.cpp
@@ -199,7 +199,7 @@ llama_model_qwen35moe::graph::graph(const llama_model & model, const llm_graph_p
             cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
         }
 
-        if (il == n_transformer_layers - 1 && inp_out_ids) {
+        if (il == n_transformer_layers - 1 && inp_out_ids && cparams.embeddings_pre_norm_masked) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }
@@ -234,6 +234,10 @@ llama_model_qwen35moe::graph::graph(const llama_model & model, const llm_graph_p
     cb(cur, "h_pre_norm", -1);
     res->t_h_pre_norm = cur;
 
+    if (!cparams.embeddings_pre_norm_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
     // Final norm
     cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
 

diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index 1ce7f0958279ccaa67844b83d479d188a97b242c..0f3fb9efa3cacca52c4a0970b3626edac8c77d49 100644 (file)
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -243,6 +243,11 @@ struct server_slot {
         return task->need_embd() || (spec && common_speculative_need_embd(spec));
     }
 
+    bool need_embd_pre_norm() const {
+        GGML_ASSERT(task);
+        return spec && common_speculative_need_embd_pre_norm(spec);
+    }
+
     // if the context does not have a memory module then all embeddings have to be computed within a single ubatch
     // also we cannot split if the pooling would require any past tokens
     // (MTP supports splitting — uses task->need_embd() not need_embd())