llama: update vendored code to commit 46e3556 (#8308)

llama/patches/0003-embeddings.patch

@@ -4,14 +4,15 @@ Date: Mon, 16 Sep 2024 15:53:14 -0700
 Subject: [PATCH] embeddings
 
 ---
- src/llama.cpp | 9 ++++++---
- 1 file changed, 6 insertions(+), 3 deletions(-)
+ src/llama-context.cpp | 2 +-
+ src/llama.cpp         | 6 ++++--
+ 2 files changed, 5 insertions(+), 3 deletions(-)
 
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 626c3e3f..9e292c4f 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -17419,7 +17419,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
+diff --git a/src/llama-context.cpp b/src/llama-context.cpp
+index 38a55fb2..b9c4a5bf 100644
+--- a/src/llama-context.cpp
++++ b/src/llama-context.cpp
+@@ -475,7 +475,7 @@ size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs) {
      const auto n_embd  = hparams.n_embd;
  
      // TODO: use a per-batch flag for logits presence instead
@@ -20,7 +21,11 @@ index 626c3e3f..9e292c4f 100644
      const bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
  
      const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
-@@ -17714,7 +17714,6 @@ static int llama_decode_internal(
+diff --git a/src/llama.cpp b/src/llama.cpp
+index ea78ea48..4eb3f6b9 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -10876,7 +10876,6 @@ static int llama_decode_internal(
              res  = nullptr;
              embd = nullptr;
          } else if (cparams.embeddings) {
@@ -28,7 +33,7 @@ index 626c3e3f..9e292c4f 100644
              embd = nullptr;
              for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) {
                  if (strcmp(ggml_graph_node(gf, i)->name, "result_embd_pooled") == 0) {
-@@ -17722,11 +17721,15 @@ static int llama_decode_internal(
+@@ -10884,12 +10883,15 @@ static int llama_decode_internal(
                      break;
                  }
              }
@@ -37,7 +42,7 @@ index 626c3e3f..9e292c4f 100644
              embd = nullptr; // do not extract embeddings when not needed
              GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
          }
-+
+ 
 +        if (!cparams.causal_attn) {
 +            res = nullptr; // do not extract logits when not needed
 +        }