Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.flake8
#	.github/workflows/bench.yml
#	.github/workflows/python-lint.yml
#	.pre-commit-config.yaml
#	Makefile
#	README.md
#	models/ggml-vocab-bert-bge.gguf.inp
#	models/ggml-vocab-bert-bge.gguf.out
#	models/ggml-vocab-deepseek-coder.gguf.inp
#	models/ggml-vocab-deepseek-coder.gguf.out
#	models/ggml-vocab-deepseek-llm.gguf.inp
#	models/ggml-vocab-deepseek-llm.gguf.out
#	models/ggml-vocab-falcon.gguf.inp
#	models/ggml-vocab-falcon.gguf.out
#	models/ggml-vocab-gpt-2.gguf.inp
#	models/ggml-vocab-gpt-2.gguf.out
#	models/ggml-vocab-llama-bpe.gguf.inp
#	models/ggml-vocab-llama-bpe.gguf.out
#	models/ggml-vocab-llama-spm.gguf.inp
#	models/ggml-vocab-llama-spm.gguf.out
#	models/ggml-vocab-mpt.gguf.inp
#	models/ggml-vocab-mpt.gguf.out
#	models/ggml-vocab-phi-3.gguf
#	models/ggml-vocab-phi-3.gguf.inp
#	models/ggml-vocab-phi-3.gguf.out
#	models/ggml-vocab-refact.gguf
#	models/ggml-vocab-starcoder.gguf.inp
#	models/ggml-vocab-starcoder.gguf.out
#	requirements/requirements-convert.txt
#	scripts/compare-llama-bench.py
#	scripts/run-with-preset.py
#	scripts/verify-checksum-models.py
#	tests/CMakeLists.txt
#	tests/test-tokenizer-0.cpp

examples/gguf-split/gguf-split.cpp

@@ -33,6 +33,7 @@ struct split_params {
     int n_split_tensors = 128;
     std::string input;
     std::string output;
+    bool no_tensor_first_split = false;
     bool dry_run = false;
 };
 
@@ -50,6 +51,7 @@ static void split_print_usage(const char * executable) {
     printf("  --merge                 merge multiple GGUF to a single GGUF\n");
     printf("  --split-max-tensors     max tensors in each split (default: %d)\n", default_params.n_split_tensors);
     printf("  --split-max-size N(M|G) max size per split\n");
+    printf("  --no-tensor-first-split do not add tensors to the first split (disabled by default)\n");
     printf("  --dry-run               only print out a split plan and exit, without writing any new files\n");
     printf("\n");
 }
@@ -101,6 +103,10 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
             arg_found = true;
             params.dry_run = true;
         }
+        if (arg == "--no-tensor-first-split") {
+            arg_found = true;
+            params.no_tensor_first_split = true;
+        }
 
         if (is_op_set) {
             throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
@@ -201,10 +207,10 @@ struct split_strategy {
         // because we need to know list of tensors for each file in advance, we will build all the ctx_out for all output splits
         int i_split = -1;
         struct gguf_context * ctx_out = NULL;
-        auto new_ctx_out = [&]() {
+        auto new_ctx_out = [&](bool allow_no_tensors) {
             i_split++;
             if (ctx_out != NULL) {
-                if (gguf_get_n_tensors(ctx_out) == 0) {
+                if (gguf_get_n_tensors(ctx_out) == 0 && !allow_no_tensors) {
                     fprintf(stderr, "error: one of splits have 0 tensors. Maybe size or tensors limit is too small\n");
                     exit(EXIT_FAILURE);
                 }
@@ -221,7 +227,12 @@ struct split_strategy {
         };
 
         // initialize ctx_out for the first split
-        new_ctx_out();
+        new_ctx_out(false);
+
+        // skip first split if no_tensor_first_split is set
+        if (params.no_tensor_first_split) {
+            new_ctx_out(true);
+        }
 
         // process tensors one by one
         size_t curr_tensors_size = 0; // current size by counting only tensors size (without metadata)
@@ -231,7 +242,7 @@ struct split_strategy {
             size_t n_bytes = GGML_PAD(ggml_nbytes(t), GGUF_DEFAULT_ALIGNMENT);
             size_t next_tensors_size = curr_tensors_size + n_bytes;
             if (should_split(i, next_tensors_size)) {
-                new_ctx_out();
+                new_ctx_out(false);
                 curr_tensors_size = n_bytes;
             } else {
                 curr_tensors_size = next_tensors_size;

examples/gguf-split/tests.sh

@@ -55,15 +55,15 @@ $MAIN --model $WORK_PATH/ggml-model-merge.gguf --random-prompt --n-predict 32
 echo PASS
 echo
 
-# 4. Split with no tensor in metadata
-#$SPLIT --split-max-tensors 32 --no-tensor-in-metadata $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-32-tensors
-#echo PASS
-#echo
+# 4. Split with no tensors in the first split
+$SPLIT --split-max-tensors 32 --no-tensor-first-split $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-32-tensors
+echo PASS
+echo
 
 # 4b. Test the sharded model is loading properly
-#$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00006.gguf --random-prompt --n-predict 32
-#echo PASS
-#echo
+$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00007.gguf --random-prompt --n-predict 32
+echo PASS
+echo
 
 # 5. Merge
 #$SPLIT --merge $WORK_PATH/ggml-model-split-32-tensors-00001-of-00006.gguf $WORK_PATH/ggml-model-merge-2.gguf