MPT : clone wte to output at load time

convert-hf-to-gguf.py

@@ -462,11 +462,6 @@ def write_tensors(self):
 
             self.gguf_writer.add_tensor(new_name, data)
 
-            # note: MPT output is tied to (same as) wte in original model;
-            # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
-            if new_name == "token_embd.weight":
-                self.gguf_writer.add_tensor("output.weight", data)
-
 
 class BaichuanModel(Model):
     def set_vocab(self):

ggml.c

@@ -18278,10 +18278,11 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
         // the ggml_tensor structs to the appropriate locations in the binary blob
 
         // compute the exact size needed for the new ggml_context
+        int n_tensors = ctx->header.n_tensors + params.extra_tensors;
         const size_t mem_size =
             params.no_alloc ?
-            (ctx->header.n_tensors    )*ggml_tensor_overhead() :
-            (ctx->header.n_tensors + 1)*ggml_tensor_overhead() + ctx->size;
+            (n_tensors    )*ggml_tensor_overhead() :
+            (n_tensors + 1)*ggml_tensor_overhead() + ctx->size;
 
         struct ggml_init_params pdata = {
             .mem_size   = mem_size,
@@ -18590,6 +18591,10 @@ size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i) {
     return ctx->infos[i].offset;
 }
 
+void gguf_set_tensor_offset(const struct gguf_context * ctx, int i, size_t offset) {
+    ctx->infos[i].offset = offset;
+}
+
 char * gguf_get_tensor_name(const struct gguf_context * ctx, int i) {
     return ctx->infos[i].name.data;
 }

ggml.h

@@ -2010,6 +2010,7 @@ extern "C" {
 
         // if not NULL, create a ggml_context and allocate the tensor data in it
         struct ggml_context ** ctx;
+        int extra_tensors;
     };
 
     GGML_API struct gguf_context * gguf_init_empty(void);
@@ -2053,6 +2054,7 @@ extern "C" {
     GGML_API int    gguf_get_n_tensors    (const struct gguf_context * ctx);
     GGML_API int    gguf_find_tensor      (const struct gguf_context * ctx, const char * name);
     GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
+    GGML_API void   gguf_set_tensor_offset(const struct gguf_context * ctx, int i, size_t offset);
     GGML_API char * gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
 
     // overrides existing values or adds a new one

llama.cpp

@@ -1792,8 +1792,9 @@ struct llama_model_loader {
 
     llama_model_loader(const std::string & fname, bool use_mmap) : file(fname.c_str(), "rb") {
         struct gguf_init_params params = {
-            /*.no_alloc = */ true,
-            /*.ctx      = */ &ctx_meta,
+            /*.no_alloc      = */ true,
+            /*.ctx           = */ &ctx_meta,
+            /*.extra_tensors = */ 1,
         };
 
         ctx_gguf = gguf_init_from_file(fname.c_str(), params);
@@ -2100,6 +2101,25 @@ struct llama_model_loader {
             done_size += ggml_nbytes(cur);
         }
     }
+
+    // must be called before calc_sizes
+    void clone_tensor(const char * src_name, const char * dst_name) {
+        int src_idx = gguf_find_tensor(ctx_gguf, src_name);
+        GGML_ASSERT(src_idx >= 0);
+
+        struct ggml_tensor * src = ggml_get_tensor(ctx_meta, src_name);
+        size_t src_offset = gguf_get_tensor_offset(ctx_gguf, src_idx);
+
+        struct ggml_tensor * cur = ggml_new_tensor(ctx_meta, src->type, src->n_dims, src->ne);
+        GGML_ASSERT(cur);
+
+        ggml_set_name(cur, dst_name);
+        gguf_add_tensor(ctx_gguf, cur);
+        gguf_set_tensor_offset(ctx_gguf, n_tensors, src_offset);
+        n_tensors++;
+        n_elements += ggml_nelements(cur);
+        n_bytes += ggml_nbytes(cur);
+    }
 };
 
 //
@@ -2666,6 +2686,11 @@ static void llm_load_tensors(
 
     model.n_gpu_layers = n_gpu_layers;
 
+    // MPT output is tied to (same as) wte in original model
+    if (model.arch == LLM_ARCH_MPT) {
+        ml.clone_tensor("token_embd.weight", "output.weight");
+    }
+
     size_t ctx_size;
     size_t mmapped_size;