From 95b6e5212f5e4e1419de1d833d7f8d788f9f2227 Mon Sep 17 00:00:00 2001
From: Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com>
Date: Mon, 28 Aug 2023 23:33:27 -0700
Subject: [PATCH 01/21] added `struct` to llama_dump_timing_info_yaml's
 `llama_context` (#2857)

fixes C compat.
---
 llama.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/llama.h b/llama.h
index b38d3be20..6e5e1df63 100644
--- a/llama.h
+++ b/llama.h
@@ -521,7 +521,7 @@ extern "C" {
     // If this is not called, or NULL is supplied, everything is output on stderr.
     LLAMA_API void llama_log_set(llama_log_callback log_callback, void * user_data);
 
-    LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx);
+    LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
 
 #ifdef __cplusplus
 }

From 611363ac791435497e66278dfe31ac8a4e11fa4f Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Tue, 29 Aug 2023 10:50:30 +0300
Subject: [PATCH 02/21] scripts : add pipefail

---
 scripts/qnt-all.sh      | 1 +
 scripts/run-all-perf.sh | 1 +
 scripts/run-all-ppl.sh  | 1 +
 3 files changed, 3 insertions(+)

diff --git a/scripts/qnt-all.sh b/scripts/qnt-all.sh
index 1376e4194..b4c2a159e 100755
--- a/scripts/qnt-all.sh
+++ b/scripts/qnt-all.sh
@@ -20,6 +20,7 @@ fi
 model="$1"
 out="../tmp/results-${model}"
 
+set -o pipefail
 set -e
 
 mkdir -p ${out}
diff --git a/scripts/run-all-perf.sh b/scripts/run-all-perf.sh
index 7391e3dd5..6384e364d 100755
--- a/scripts/run-all-perf.sh
+++ b/scripts/run-all-perf.sh
@@ -20,6 +20,7 @@ fi
 model="$1"
 out="../tmp/results-${model}"
 
+set -o pipefail
 set -e
 
 mkdir -p ${out}
diff --git a/scripts/run-all-ppl.sh b/scripts/run-all-ppl.sh
index f643ca3ae..e04d61d7f 100755
--- a/scripts/run-all-ppl.sh
+++ b/scripts/run-all-ppl.sh
@@ -17,6 +17,7 @@ if [ ! -z "$3" ]; then
     args="$3"
 fi
 
+set -o pipefail
 set -e
 
 model="$1"

From 3a007648f230ea37d6cca5e63850f04ebb12d2cf Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Tue, 29 Aug 2023 11:33:46 +0300
Subject: [PATCH 03/21] metal : add option to disable debug logs (close #2764)

---
 CMakeLists.txt |  2 +-
 Makefile       |  2 +-
 ggml-metal.m   | 71 +++++++++++++++++++++++---------------------------
 3 files changed, 35 insertions(+), 40 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index ba008bcc6..1eae2d670 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -301,7 +301,7 @@ if (LLAMA_METAL)
     set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
 
     add_compile_definitions(GGML_USE_METAL)
-    add_compile_definitions(GGML_METAL_NDEBUG)
+    #add_compile_definitions(GGML_METAL_NDEBUG)
 
     # get full path to the file
     #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
diff --git a/Makefile b/Makefile
index e60821dd5..a64374e7d 100644
--- a/Makefile
+++ b/Makefile
@@ -305,7 +305,7 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 endif # LLAMA_HIPBLAS
 
 ifdef LLAMA_METAL
-	CFLAGS   += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
+	CFLAGS   += -DGGML_USE_METAL #-DGGML_METAL_NDEBUG
 	CXXFLAGS += -DGGML_USE_METAL
 	LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
 	OBJS     += ggml-metal.o
diff --git a/ggml-metal.m b/ggml-metal.m
index ad2ee8cf5..e929c4b07 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -11,6 +11,7 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+// TODO: temporary - reuse llama.cpp logging
 #ifdef GGML_METAL_NDEBUG
 #define metal_printf(...)
 #else
@@ -113,7 +114,7 @@ static NSString * const msl_library_source = @"see metal.metal";
 @end
 
 struct ggml_metal_context * ggml_metal_init(int n_cb) {
-    fprintf(stderr, "%s: allocating\n", __func__);
+    metal_printf("%s: allocating\n", __func__);
 
     struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
 
@@ -132,7 +133,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 
         ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
         if (error) {
-            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
             return NULL;
         }
     }
@@ -146,11 +147,11 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
         NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
         NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
-        fprintf(stderr, "%s: loading '%s'\n", __func__, [path UTF8String]);
+        metal_printf("%s: loading '%s'\n", __func__, [path UTF8String]);
 
         NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
         if (error) {
-            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
             return NULL;
         }
 
@@ -162,7 +163,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
 #endif
         if (error) {
-            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
             return NULL;
         }
     }
@@ -174,11 +175,11 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #define GGML_METAL_ADD_KERNEL(name) \
         ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
         ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
-        fprintf(stderr, "%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
+        metal_printf("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
                 (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
                 (int) ctx->pipeline_##name.threadExecutionWidth); \
         if (error) { \
-            fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
+            metal_printf("%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
             return NULL; \
         }
 
@@ -230,19 +231,19 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #undef GGML_METAL_ADD_KERNEL
     }
 
-    fprintf(stderr, "%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-    fprintf(stderr, "%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    metal_printf("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+    metal_printf("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
     if (ctx->device.maxTransferRate != 0) {
-        fprintf(stderr, "%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+        metal_printf("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
     } else {
-        fprintf(stderr, "%s: maxTransferRate               = built-in GPU\n", __func__);
+        metal_printf("%s: maxTransferRate               = built-in GPU\n", __func__);
     }
 
     return ctx;
 }
 
 void ggml_metal_free(struct ggml_metal_context * ctx) {
-    fprintf(stderr, "%s: deallocating\n", __func__);
+    metal_printf("%s: deallocating\n", __func__);
 #define GGML_METAL_DEL_KERNEL(name) \
     [ctx->function_##name release]; \
     [ctx->pipeline_##name release];
@@ -311,7 +312,7 @@ void * ggml_metal_host_malloc(size_t n) {
     void * data = NULL;
     const int result = posix_memalign((void **) &data, getpagesize(), n);
     if (result != 0) {
-        fprintf(stderr, "%s: error: posix_memalign failed\n", __func__);
+        metal_printf("%s: error: posix_memalign failed\n", __func__);
         return NULL;
     }
 
@@ -339,7 +340,7 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
 // Metal buffer based on the host memory pointer
 //
 static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
-    //fprintf(stderr, "%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
+    //metal_printf("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
 
     const int64_t tsize = ggml_nbytes(t);
 
@@ -350,13 +351,13 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
         if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
             *offs = (size_t) ioffs;
 
-            //fprintf(stderr, "%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
+            //metal_printf("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
 
             return ctx->buffers[i].metal;
         }
     }
 
-    fprintf(stderr, "%s: error: buffer is nil\n", __func__);
+    metal_printf("%s: error: buffer is nil\n", __func__);
 
     return nil;
 }
@@ -368,7 +369,7 @@ bool ggml_metal_add_buffer(
                          size_t   size,
                          size_t   max_size) {
     if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
-        fprintf(stderr, "%s: too many buffers\n", __func__);
+        metal_printf("%s: too many buffers\n", __func__);
         return false;
     }
 
@@ -378,7 +379,7 @@ bool ggml_metal_add_buffer(
             const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
 
             if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
-                fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
+                metal_printf("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
                 return false;
             }
         }
@@ -399,11 +400,11 @@ bool ggml_metal_add_buffer(
             ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
             if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
+                metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
                 return false;
             }
 
-            fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
+            metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
 
             ++ctx->n_buffers;
         } else {
@@ -423,27 +424,27 @@ bool ggml_metal_add_buffer(
                 ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
                 if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                    fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
+                    metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
                     return false;
                 }
 
-                fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
+                metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
                 if (i + size_step < size) {
-                    fprintf(stderr, "\n");
+                    metal_printf("\n");
                 }
 
                 ++ctx->n_buffers;
             }
         }
 
-        fprintf(stderr, ", (%8.2f / %8.2f)",
+        metal_printf(", (%8.2f / %8.2f)",
                 ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
                 ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
 
         if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
-            fprintf(stderr, ", warning: current allocated size is greater than the recommended max working set size\n");
+            metal_printf(", warning: current allocated size is greater than the recommended max working set size\n");
         } else {
-            fprintf(stderr, "\n");
+            metal_printf("\n");
         }
     }
 
@@ -453,8 +454,6 @@ bool ggml_metal_add_buffer(
 void ggml_metal_set_tensor(
         struct ggml_metal_context * ctx,
         struct ggml_tensor * t) {
-    metal_printf("%s: set input for tensor '%s'\n", __func__, t->name);
-
     size_t offs;
     id<MTLBuffer> id_dst = ggml_metal_get_buffer(ctx, t, &offs);
 
@@ -464,8 +463,6 @@ void ggml_metal_set_tensor(
 void ggml_metal_get_tensor(
         struct ggml_metal_context * ctx,
         struct ggml_tensor * t) {
-    metal_printf("%s: extract results for tensor '%s'\n", __func__, t->name);
-
     size_t offs;
     id<MTLBuffer> id_src = ggml_metal_get_buffer(ctx, t, &offs);
 
@@ -560,15 +557,13 @@ void ggml_metal_graph_find_concurrency(
     }
 
     if (ctx->concur_list_len > GGML_MAX_CONCUR) {
-        fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
+        metal_printf("%s: too many elements for metal ctx->concur_list!\n", __func__);
     }
 }
 
 void ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
-    metal_printf("%s: evaluating graph\n", __func__);
-
     @autoreleasepool {
 
     // if there is ctx->concur_list, dispatch concurrently
@@ -616,7 +611,7 @@ void ggml_metal_graph_compute(
                     continue;
                 }
 
-                metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+                //metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
 
                 struct ggml_tensor * src0 = gf->nodes[i]->src[0];
                 struct ggml_tensor * src1 = gf->nodes[i]->src[1];
@@ -764,7 +759,7 @@ void ggml_metal_graph_compute(
                                 } break;
                             default:
                                 {
-                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                                    metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
                                     GGML_ASSERT(false);
                                 }
                         } break;
@@ -923,7 +918,7 @@ void ggml_metal_graph_compute(
                                         } break;
                                     default:
                                         {
-                                            fprintf(stderr, "Asserting on type %d\n",(int)src0t);
+                                            metal_printf("Asserting on type %d\n",(int)src0t);
                                             GGML_ASSERT(false && "not implemented");
                                         }
                                 };
@@ -1161,7 +1156,7 @@ void ggml_metal_graph_compute(
                         } break;
                     default:
                         {
-                            fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                            metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
                             GGML_ASSERT(false);
                         }
                 }
@@ -1186,7 +1181,7 @@ void ggml_metal_graph_compute(
 
         MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
         if (status != MTLCommandBufferStatusCompleted) {
-            fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status);
+            metal_printf("%s: command buffer %d failed with status %lu\n", __func__, i, status);
             GGML_ASSERT(false);
         }
     }

From d4b5e16c32ba9c5fa6bbd035e80a99c113050cde Mon Sep 17 00:00:00 2001
From: Cebtenzzre <cebtenzzre@gmail.com>
Date: Tue, 29 Aug 2023 04:42:41 -0400
Subject: [PATCH 04/21] make : fix clang tests build, add missing examples
 (#2859)

* make : do not pass headers to the compiler

This fixes building tests with clang.

* make : add missing examples

* make : fix build-info.h dependencies
---
 Makefile | 49 ++++++++++++++++++++++++++++++++-----------------
 1 file changed, 32 insertions(+), 17 deletions(-)

diff --git a/Makefile b/Makefile
index a64374e7d..02ba3e36d 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,5 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple server embd-input-test gguf llama-bench
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam_search
 
 # Binaries only useful for tests
 TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1
@@ -356,7 +356,7 @@ OBJS += ggml-alloc.o
 llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-common.o: common/common.cpp common/common.h
+common.o: common/common.cpp common/common.h build-info.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 console.o: common/console.cpp common/console.h
@@ -369,7 +369,7 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch convert-llama2c-to-ggml embd-input-test gguf llama-bench build-info.h $(TEST_TARGETS)
+	rm -vf *.o *.so *.dll benchmark-matmult build-info.h $(BUILD_TARGETS) $(TEST_TARGETS)
 
 #
 # Examples
@@ -409,18 +409,33 @@ $(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-in
 embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
 
-gguf: examples/gguf/gguf.cpp                                  build-info.h ggml.o llama.o $(OBJS)
+gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp    build-info.h ggml.o llama.o common.o $(OBJS)
+train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp    build-info.h ggml.o llama.o $(OBJS)
+convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+beam_search: examples/beam_search/beam_search.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
+BUILD_TARGETS += metal
+endif
+
+ifdef LLAMA_METAL
+metal: examples/metal/metal.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+endif
+
 build-info.h: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
@@ -443,34 +458,34 @@ vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
 tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o common.o grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-tokenizer-1: tests/test-tokenizer-1.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

From 74e0caeb82fc9db77fa2cc93070bb919a9a935dd Mon Sep 17 00:00:00 2001
From: Jhen-Jie Hong <iainst0409@gmail.com>
Date: Tue, 29 Aug 2023 17:30:10 +0800
Subject: [PATCH 05/21] readme : add react-native binding (#2869)

---
 README.md | 1 +
 1 file changed, 1 insertion(+)

diff --git a/README.md b/README.md
index bf3eb0b76..8d54a558d 100644
--- a/README.md
+++ b/README.md
@@ -113,6 +113,7 @@ as the main playground for developing new features for the [ggml](https://github
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
 - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
+- React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
 
 **UI:**
 

From bcce96ba4dd95482824700c4ce2455fe8c49055a Mon Sep 17 00:00:00 2001
From: jameswu2014 <545426914@qq.com>
Date: Tue, 29 Aug 2023 17:48:41 +0800
Subject: [PATCH 06/21] convert.py : fix baichuan7B support (#2870)

* [Fix]: convert.py support baichuan7B

* convert.py : fix trailing whitespaces

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---
 convert.py | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/convert.py b/convert.py
index a15e6ccd2..3f0a1c932 100755
--- a/convert.py
+++ b/convert.py
@@ -469,7 +469,7 @@ class UnquantizedTensor(Tensor):
 
     def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor':
         r = self.ndarray.shape[0] // 3
-        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head))
+        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head))
 
     def part(self, n_part: int) -> 'UnquantizedTensor':
         r = self.ndarray.shape[0] // 3
@@ -952,9 +952,10 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
            #tmp[f"model.layers.{i}.self_attn.v_proj.weight"] =              model[f"model.layers.{i}.self_attn.v_proj.weight"]
         elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
             print(f"Unpacking and permuting layer {i}")
-            tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head)
-            tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv)
+            tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head)
+            tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head)
             tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy        (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2)
+            del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
         else:
             break
 

From 53885d7256909ec3e2176cdc2477f3986c15ec69 Mon Sep 17 00:00:00 2001
From: maddes8cht <55592906+maddes8cht@users.noreply.github.com>
Date: Tue, 29 Aug 2023 15:51:02 +0200
Subject: [PATCH 07/21] py : fix "usage" messages (#2873)

convert-to-gguf python scripts
---
 convert-falcon-hf-to-gguf.py    | 2 +-
 convert-gptneox-hf-to-gguf.py   | 2 +-
 convert-llama-7b-pth-to-gguf.py | 2 +-
 convert-llama-hf-to-gguf.py     | 2 +-
 4 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py
index 411cbf682..168bcf17f 100755
--- a/convert-falcon-hf-to-gguf.py
+++ b/convert-falcon-hf-to-gguf.py
@@ -48,7 +48,7 @@ def count_model_parts(dir_model: str) -> int:
 
 
 if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
+    print(f"Usage: python {sys.argv[0]} dir-model ftype\n")
     print("  ftype == 0 -> float32")
     print("  ftype == 1 -> float16")
     sys.exit(1)
diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py
index 6eeff5bb1..d9c42d76b 100755
--- a/convert-gptneox-hf-to-gguf.py
+++ b/convert-gptneox-hf-to-gguf.py
@@ -50,7 +50,7 @@ def count_model_parts(dir_model: str) -> int:
 
 
 if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
+    print(f"Usage: python {sys.argv[0]} dir-model ftype\n")
     print("  ftype == 0 -> float32")
     print("  ftype == 1 -> float16")
     sys.exit(1)
diff --git a/convert-llama-7b-pth-to-gguf.py b/convert-llama-7b-pth-to-gguf.py
index f103f5f61..2ab082383 100755
--- a/convert-llama-7b-pth-to-gguf.py
+++ b/convert-llama-7b-pth-to-gguf.py
@@ -32,7 +32,7 @@ def count_model_parts(dir_model: str) -> int:
 
 
 if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
+    print(f"Usage: python {sys.argv[0]} dir-model ftype\n")
     print("  ftype == 0 -> float32")
     print("  ftype == 1 -> float16")
 
diff --git a/convert-llama-hf-to-gguf.py b/convert-llama-hf-to-gguf.py
index 08fde238b..b00810dbb 100755
--- a/convert-llama-hf-to-gguf.py
+++ b/convert-llama-hf-to-gguf.py
@@ -44,7 +44,7 @@ def count_model_parts(dir_model: str) -> int:
 
 
 if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
+    print(f"Usage: python {sys.argv[0]} dir-model ftype\n")
     print("  ftype == 0 -> float32")
     print("  ftype == 1 -> float16")
 

From e37e69dcc3d52f21222a63cafed2a71b3f6b53c6 Mon Sep 17 00:00:00 2001
From: Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Date: Tue, 29 Aug 2023 23:55:03 +0300
Subject: [PATCH 08/21] 10X faster BPE tokenizer (#2876)

* 10X faster BPE tokenizer

* Remove comment that no longer applies

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
---
 llama.cpp | 19 +++++++++----------
 1 file changed, 9 insertions(+), 10 deletions(-)

diff --git a/llama.cpp b/llama.cpp
index 7cb468538..fcd6f276a 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -3211,7 +3211,7 @@ private:
 
 struct llm_bigram_bpe {
     struct comparator {
-        bool operator()(llm_bigram_bpe & l, llm_bigram_bpe & r) {
+        bool operator()(const llm_bigram_bpe & l, const llm_bigram_bpe & r) const {
             return l.rank > r.rank || (l.rank == r.rank && l.left > r.left);
         }
     };
@@ -3359,23 +3359,22 @@ private:
     }
 
     // probably not 100% correct
-    // TODO: this is quite slow - how to make it more efficient?
-    static std::vector<std::string> bpe_gpt2_preprocess(std::string text) {
+    static std::vector<std::string> bpe_gpt2_preprocess(const std::string & text) {
         std::vector<std::string> words;
 
         // ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53
         const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
         const std::regex re(pattern);
-        std::smatch m;
 
-        while (std::regex_search(text, m, re)) {
-            for (auto x : m) {
-                words.push_back(x);
-            }
-            text = m.suffix();
+        auto words_begin = std::sregex_iterator(text.begin(), text.end(), re);
+        auto words_end = std::sregex_iterator();
+        auto n_words = std::distance(words_begin, words_end);
+        words.reserve(n_words);
+        for (auto it = words_begin; it != words_end; ++it) {
+            words.push_back(it->str());
         }
-
         return words;
+
     }
 
     const llama_vocab & vocab;

From fa3582f509a2715e80a473e79f88dcd1ebff44c2 Mon Sep 17 00:00:00 2001
From: Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Date: Tue, 29 Aug 2023 23:55:45 +0300
Subject: [PATCH 09/21] Tell users attmepting to run perplexity with too few
 tokens to use more (#2882)

Closes #2858

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
---
 examples/perplexity/perplexity.cpp | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
index aeb774c5f..7c02b6d40 100644
--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -142,6 +142,14 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params)
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    if (int(tokens.size()) < 2*params.n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*params.n_ctx,
+                params.n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return {std::move(tokens), 0., {}, {}};
+    }
+
     std::vector<float>       logit_history;
     std::vector<float>       prob_history;
 
@@ -274,6 +282,13 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
     auto tim2 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
 
+    if (int(tokens.size()) < 2*params.n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*params.n_ctx,
+                params.n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return {std::move(tokens), 0., {}, {}};
+    }
+
     std::vector<float> logit_history;
     logit_history.resize(tokens.size());
 

From c03a243abf9f30889f31fefdfa94fe9d7034820c Mon Sep 17 00:00:00 2001
From: slaren <slarengh@gmail.com>
Date: Tue, 29 Aug 2023 23:17:34 +0200
Subject: [PATCH 10/21] remove outdated references to -eps and -gqa from README
 (#2881)

---
 README.md | 2 --
 1 file changed, 2 deletions(-)

diff --git a/README.md b/README.md
index 8d54a558d..a880fd29f 100644
--- a/README.md
+++ b/README.md
@@ -729,8 +729,6 @@ python3 convert.py pygmalion-7b/ --outtype q4_1
   - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGML)
   - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML)
   - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGML)
-- Specify `-eps 1e-5` for best generation quality
-- Specify `-gqa 8` for 70B models to work
 
 ### Verifying the model files
 

From 06abf8eebabe086ca4003dee2754ab45032cd3fd Mon Sep 17 00:00:00 2001
From: slaren <slarengh@gmail.com>
Date: Tue, 29 Aug 2023 23:24:42 +0200
Subject: [PATCH 11/21] ggml : add view_src and view_offs to ggml_tensor for
 views (#2874)

* ggml : add view_src and view_offs

* update ggml-alloc to use view_src

* update ggml_diag_mask to work correctly with automatic inplace

* exclude other ops that set an inplace flag from automatic inplace
---
 ggml-alloc.c |  53 ++----------
 ggml.c       | 233 +++++++++++++++++++++++----------------------------
 ggml.h       |   5 +-
 3 files changed, 113 insertions(+), 178 deletions(-)

diff --git a/ggml-alloc.c b/ggml-alloc.c
index 63beb1d4e..f07a4a217 100644
--- a/ggml-alloc.c
+++ b/ggml-alloc.c
@@ -321,8 +321,7 @@ bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
 //////////// compute graph allocator
 
 static bool ggml_is_view(struct ggml_tensor * t) {
-    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
-           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
+    return t->view_src != NULL;
 }
 
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
@@ -340,28 +339,6 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml
     return true;
 }
 
-static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
-    switch (t->op) {
-        case GGML_OP_PERMUTE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_VIEW:
-            return t->src[0];
-        case GGML_OP_CPY:
-            return t->src[1];
-        default:
-            return NULL;
-    }
-}
-
-static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
-    struct ggml_tensor * parent = t;
-    do {
-        parent = get_view_parent(parent);
-    } while (ggml_is_view(parent));
-    return parent;
-}
-
 static bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {
         case GGML_OP_SCALE:
@@ -369,7 +346,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_ADD:
         case GGML_OP_ADD1:
-        case GGML_OP_ACC:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -379,7 +355,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_UNARY:
         case GGML_OP_ROPE:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_SET:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_CONT:
             return true;
@@ -393,24 +368,8 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     struct hash_node * ht = alloc->hash_table;
     if (node->data == NULL) {
         if (ggml_is_view(node)) {
-            size_t offset;
-            switch(node->op) {
-                case GGML_OP_VIEW:
-                    memcpy(&offset, node->op_params, sizeof(size_t));
-                    node->data = (char *) node->src[0]->data + offset;
-                    break;
-                case GGML_OP_PERMUTE:
-                case GGML_OP_RESHAPE:
-                case GGML_OP_TRANSPOSE:
-                    node->data = node->src[0]->data;
-                    break;
-                case GGML_OP_CPY:
-                    node->data = node->src[1]->data;
-                    break;
-                default:
-                    GGML_ASSERT(!"unknown view op");
-                    break;
-            }
+            assert(node->view_src->data != NULL);
+            node->data = (char *)node->view_src->data + node->view_offs;
         } else {
             // see if we can reuse a parent's buffer (inplace)
             if (ggml_op_can_inplace(node->op)) {
@@ -430,7 +389,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     struct hash_node * p_hn = hash_get(ht, parent);
                     if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
                         if (ggml_is_view(parent)) {
-                            struct ggml_tensor * view_src = get_view_source(parent);
+                            struct ggml_tensor * view_src = parent->view_src;
                             struct hash_node * view_src_hn = hash_get(ht, view_src);
                             if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
                                 // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
@@ -472,7 +431,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
             struct ggml_tensor * node = gf->nodes[i];
 
             if (ggml_is_view(node)) {
-                struct ggml_tensor * view_src = get_view_source(node);
+                struct ggml_tensor * view_src = node->view_src;
                 hash_get(ht, view_src)->n_views += 1;
             }
 
@@ -557,7 +516,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
 
                         if (p_hn->n_children == 0 && p_hn->n_views == 0) {
                             if (ggml_is_view(parent)) {
-                                struct ggml_tensor * view_src = get_view_source(parent);
+                                struct ggml_tensor * view_src = parent->view_src;
                                 struct hash_node * view_src_hn = hash_get(ht, view_src);
                                 view_src_hn->n_views -= 1;
                                 AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);
diff --git a/ggml.c b/ggml.c
index 9a787863d..46ce4a581 100644
--- a/ggml.c
+++ b/ggml.c
@@ -4104,16 +4104,11 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) {
 }
 
 size_t ggml_nbytes(const struct ggml_tensor * tensor) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
-
-    // this should handle cases where the tensor is not contiguous in memory
-    // probaby just:
-    //
-    //     return tensor->ne[3]*tensor->nb[3]
-    //
-    // is enough, but just in case, adding the second part
-
-    return MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*ggml_type_size(tensor->type))/ggml_blck_size(tensor->type));
+    size_t nbytes = tensor->ne[0]*tensor->nb[0]/ggml_blck_size(tensor->type);
+    for (int i = 1; i < GGML_MAX_DIMS; ++i) {
+        nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
+    }
+    return nbytes;
 }
 
 size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
@@ -4567,36 +4562,51 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         enum   ggml_type      type,
         int                   n_dims,
         const int64_t       * ne,
-        void                * data) {
+        struct ggml_tensor  * view_src,
+        size_t                view_offs) {
 
     assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
 
-    size_t data_size = 0;
+    // find the base tensor and absolute offset
+    if (view_src != NULL && view_src->view_src != NULL) {
+        view_offs += view_src->view_offs;
+        view_src   = view_src->view_src;
+    }
 
-    if (data == NULL && !ctx->no_alloc) {
-        data_size += ggml_type_size(type)*(ne[0]/ggml_blck_size(type));
-        for (int i = 1; i < n_dims; i++) {
-            data_size *= ne[i];
+    size_t data_size = ggml_type_size(type)*(ne[0]/ggml_blck_size(type));
+    for (int i = 1; i < n_dims; i++) {
+        data_size *= ne[i];
+    }
+
+    GGML_ASSERT(view_src == NULL || data_size + view_offs <= ggml_nbytes(view_src));
+
+    void * data = view_src != NULL ? view_src->data : NULL;
+    if (data != NULL) {
+        data = (char *) data + view_offs;
+    }
+
+    size_t obj_alloc_size = 0;
+
+    if (view_src == NULL && ctx->no_alloc == false) {
+        if (ctx->scratch.data != NULL) {
+            // allocate tensor data in the scratch buffer
+            if (ctx->scratch.offs + data_size > ctx->scratch.size) {
+                GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n",
+                        __func__, ctx->scratch.offs + data_size, ctx->scratch.size);
+                assert(false);
+                return NULL;
+            }
+
+            data = (char * const) ctx->scratch.data + ctx->scratch.offs;
+
+            ctx->scratch.offs += data_size;
+        } else {
+            // allocate tensor data in the context's memory pool
+            obj_alloc_size = data_size;
         }
     }
 
-    if (ctx->scratch.data != NULL && data == NULL) {
-        // allocate tensor data in the scratch buffer
-        if (ctx->scratch.offs + data_size > ctx->scratch.size) {
-            GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n",
-                    __func__, ctx->scratch.offs + data_size, ctx->scratch.size);
-            assert(false);
-            return NULL;
-        }
-
-        data = (char * const) ctx->scratch.data + ctx->scratch.offs;
-
-        ctx->scratch.offs += data_size;
-
-        data_size = 0;
-    }
-
-    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + data_size);
+    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size);
 
     // TODO: for recoverable errors, we would need to free the data allocated from the scratch buffer here
 
@@ -4616,7 +4626,9 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.perf_runs    =*/ 0,
         /*.perf_cycles  =*/ 0,
         /*.perf_time_us =*/ 0,
-        /*.data         =*/ (data == NULL && !ctx->no_alloc) ? (void *)(result + 1) : data,
+        /*.view_src     =*/ view_src,
+        /*.view_offs    =*/ view_offs,
+        /*.data         =*/ obj_alloc_size > 0 ? (void *)(result + 1) : data,
         /*.name         =*/ { 0 },
         /*.extra        =*/ NULL,
         /*.padding      =*/ { 0 },
@@ -4640,28 +4652,12 @@ static struct ggml_tensor * ggml_new_tensor_impl(
     return result;
 }
 
-static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
-    GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
-    assert(params_size <= GGML_MAX_OP_PARAMS);
-    memcpy(tensor->op_params, params, params_size);
-}
-
-static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_t i) {
-    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
-    return ((const int32_t *)(tensor->op_params))[i];
-}
-
-static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
-    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
-    ((int32_t *)(tensor->op_params))[i] = value;
-}
-
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type      type,
         int                   n_dims,
         const int64_t       * ne) {
-    return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL);
+    return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL, 0);
 }
 
 struct ggml_tensor * ggml_new_tensor_1d(
@@ -4726,7 +4722,23 @@ struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value) {
 }
 
 struct ggml_tensor * ggml_dup_tensor(struct ggml_context * ctx, const struct ggml_tensor * src) {
-    return ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, NULL);
+    return ggml_new_tensor(ctx, src->type, src->n_dims, src->ne);
+}
+
+static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
+    GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
+    assert(params_size <= GGML_MAX_OP_PARAMS);
+    memcpy(tensor->op_params, params, params_size);
+}
+
+static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_t i) {
+    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
+    return ((const int32_t *)(tensor->op_params))[i];
+}
+
+static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
+    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
+    ((int32_t *)(tensor->op_params))[i] = value;
 }
 
 struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) {
@@ -5012,14 +5024,13 @@ struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char *
 
 struct ggml_tensor * ggml_view_tensor(
         struct ggml_context * ctx,
-        const struct ggml_tensor * src) {
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src->data);
+        struct ggml_tensor  * src) {
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src, 0);
     ggml_format_name(result, "%s (view)", src->name);
 
-    result->nb[0] = src->nb[0];
-    result->nb[1] = src->nb[1];
-    result->nb[2] = src->nb[2];
-    result->nb[3] = src->nb[3];
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        result->nb[i] = src->nb[i];
+    }
 
     return result;
 }
@@ -5592,7 +5603,7 @@ struct ggml_tensor * ggml_repeat_back(
 
 // ggml_concat
 
-struct ggml_tensor* ggml_concat(
+struct ggml_tensor * ggml_concat(
     struct ggml_context* ctx,
     struct ggml_tensor* a,
     struct ggml_tensor* b) {
@@ -6201,7 +6212,7 @@ struct ggml_tensor * ggml_reshape(
         //GGML_ASSERT(false);
     }
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6225,7 +6236,7 @@ struct ggml_tensor * ggml_reshape_1d(
     }
 
     const int64_t ne[1] = { ne0 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6250,7 +6261,7 @@ struct ggml_tensor * ggml_reshape_2d(
     }
 
     const int64_t ne[2] = { ne0, ne1 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6276,7 +6287,7 @@ struct ggml_tensor * ggml_reshape_3d(
     }
 
     const int64_t ne[3] = { ne0, ne1, ne2 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6286,7 +6297,6 @@ struct ggml_tensor * ggml_reshape_3d(
     return result;
 }
 
-
 struct ggml_tensor * ggml_reshape_4d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -6304,7 +6314,7 @@ struct ggml_tensor * ggml_reshape_4d(
     }
 
     const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6314,34 +6324,12 @@ struct ggml_tensor * ggml_reshape_4d(
     return result;
 }
 
-// ggml_view_1d
-
-static struct ggml_tensor * ggml_view_tensor_offset(
+static struct ggml_tensor * ggml_view_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         int                   n_dims,
         const int64_t       * ne,
         size_t                offset) {
-    // don't calculate an offset from an unallocated tensor
-    void * data = NULL;
-    if (a->data != NULL) {
-        data = (char *) a->data + offset;
-    }
-
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, data);
-
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
-
-    return result;
-}
-
-struct ggml_tensor * ggml_view_1d(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        int64_t               ne0,
-        size_t                offset) {
 
     bool is_node = false;
 
@@ -6349,7 +6337,10 @@ struct ggml_tensor * ggml_view_1d(
         is_node = true;
     }
 
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 1, &ne0, offset);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, a, offset);
+    ggml_format_name(result, "%s (view)", a->name);
+
+    ggml_set_op_params(result, &offset, sizeof(offset));
 
     result->op   = GGML_OP_VIEW;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6358,6 +6349,19 @@ struct ggml_tensor * ggml_view_1d(
     return result;
 }
 
+// ggml_view_1d
+
+struct ggml_tensor * ggml_view_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        size_t                offset) {
+
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 1, &ne0, offset);
+
+    return result;
+}
+
 // ggml_view_2d
 
 struct ggml_tensor * ggml_view_2d(
@@ -6368,24 +6372,14 @@ struct ggml_tensor * ggml_view_2d(
         size_t                nb1,
         size_t                offset) {
 
-    bool is_node = false;
+    const int64_t ne[2] = { ne0, ne1 };
 
-    if (a->grad) {
-        is_node = true;
-    }
-
-    const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 };
-
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 2, ne, offset);
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 2, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = result->nb[1]*ne1;
     result->nb[3] = result->nb[2];
 
-    result->op   = GGML_OP_VIEW;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-
     return result;
 }
 
@@ -6401,24 +6395,14 @@ struct ggml_tensor * ggml_view_3d(
         size_t                nb2,
         size_t                offset) {
 
-    bool is_node = false;
+    const int64_t ne[3] = { ne0, ne1, ne2 };
 
-    if (a->grad) {
-        is_node = true;
-    }
-
-    const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, 1 };
-
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 3, ne, offset);
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 3, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = nb2;
     result->nb[3] = result->nb[2]*ne2;
 
-    result->op   = GGML_OP_VIEW;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-
     return result;
 }
 
@@ -6436,24 +6420,14 @@ struct ggml_tensor * ggml_view_4d(
         size_t                nb3,
         size_t                offset) {
 
-    bool is_node = false;
+    const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
 
-    if (a->grad) {
-        is_node = true;
-    }
-
-    const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, ne3 };
-
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 4, ne, offset);
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 4, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = nb2;
     result->nb[3] = nb3;
 
-    result->op   = GGML_OP_VIEW;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-
     return result;
 }
 
@@ -6640,7 +6614,7 @@ static struct ggml_tensor * ggml_diag_mask_inf_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    int32_t params[] = { n_past, inplace ? 1 : 0 };
+    int32_t params[] = { n_past };
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_DIAG_MASK_INF;
@@ -6657,7 +6631,6 @@ struct ggml_tensor * ggml_diag_mask_inf(
     return ggml_diag_mask_inf_impl(ctx, a, n_past, false);
 }
 
-
 struct ggml_tensor * ggml_diag_mask_inf_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -6680,7 +6653,7 @@ static struct ggml_tensor * ggml_diag_mask_zero_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    int32_t params[] = { n_past, inplace ? 1 : 0 };
+    int32_t params[] = { n_past };
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_DIAG_MASK_ZERO;
@@ -11935,8 +11908,8 @@ static void ggml_compute_forward_diag_mask_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int  n_past  =       ((int32_t *) dst->op_params)[0];
-    const bool inplace = (bool)((int32_t *) dst->op_params)[1];
+    const int  n_past  = ((int32_t *) dst->op_params)[0];
+    const bool inplace = src0->data == dst->data;
 
     GGML_ASSERT(n_past >= 0);
 
diff --git a/ggml.h b/ggml.h
index 8b410cc85..c936823d6 100644
--- a/ggml.h
+++ b/ggml.h
@@ -479,6 +479,9 @@ extern "C" {
         int64_t perf_cycles;
         int64_t perf_time_us;
 
+        struct ggml_tensor * view_src;
+        size_t               view_offs;
+
         void * data;
 
         char name[GGML_MAX_NAME];
@@ -661,7 +664,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
 
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
-    GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src);
+    GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
 
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 

From 849408957c687cde4ab32c147107f643fc55130b Mon Sep 17 00:00:00 2001
From: Cebtenzzre <cebtenzzre@gmail.com>
Date: Wed, 30 Aug 2023 02:20:26 -0400
Subject: [PATCH 12/21] tests : add a C compliance test (#2848)

* tests : add a C compliance test

* make : build C compliance test by default

* make : fix clean and make sure C test fails on clang

* make : move -Werror=implicit-int to CFLAGS
---
 CMakeLists.txt       | 1 +
 Makefile             | 9 ++++++---
 tests/CMakeLists.txt | 5 +++++
 tests/test-c.c       | 3 +++
 4 files changed, 15 insertions(+), 3 deletions(-)
 create mode 100644 tests/test-c.c

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 1eae2d670..d6c1b3b33 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -402,6 +402,7 @@ if (LLAMA_ALL_WARNINGS)
             -Wstrict-prototypes
             -Wpointer-arith
             -Wmissing-prototypes
+            -Werror=implicit-int
         )
         set(cxx_flags
             -Wall
diff --git a/Makefile b/Makefile
index 02ba3e36d..44e68b7fc 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,5 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam_search
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam_search tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1
@@ -64,7 +64,7 @@ endif
 
 # warnings
 CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \
-			-Wmissing-prototypes
+			-Wmissing-prototypes -Werror=implicit-int
 CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
 
 # OS specific
@@ -369,7 +369,7 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o *.so *.dll benchmark-matmult build-info.h $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h $(BUILD_TARGETS) $(TEST_TARGETS)
 
 #
 # Examples
@@ -489,3 +489,6 @@ tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml
 
 tests/test-tokenizer-1: tests/test-tokenizer-1.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-c.o: tests/test-c.c llama.h
+	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index ca1f39d31..483210d7b 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -37,3 +37,8 @@ llama_build_and_test_executable(test-grammar-parser.cpp)
 llama_build_and_test_executable(test-llama-grammar.cpp)
 llama_build_and_test_executable(test-grad0.cpp) # SLOW
 # llama_build_and_test_executable(test-opt.cpp) # SLOW
+
+# dummy executable - not installed
+get_filename_component(TEST_TARGET test-c.c NAME_WE)
+add_executable(${TEST_TARGET} test-c.c)
+target_link_libraries(${TEST_TARGET} PRIVATE llama)
diff --git a/tests/test-c.c b/tests/test-c.c
new file mode 100644
index 000000000..a05071080
--- /dev/null
+++ b/tests/test-c.c
@@ -0,0 +1,3 @@
+#include "llama.h"
+
+int main(void) {}

From 8341a25957b319a03d4a811176cd5ad7f2b0fbd4 Mon Sep 17 00:00:00 2001
From: staviq <staviq@gmail.com>
Date: Wed, 30 Aug 2023 08:29:32 +0200
Subject: [PATCH 13/21] main : log file (#2748)

* initial, base LOG macro

* add *.log to .gitignore

* added basic log file handler

* reverted log auto endline to better mimic printf

* remove atomics and add dynamic log target

* log_enable/disable, LOG_TEE, basic usage doc

* update .gitignore

* mv include to common, params, help msg

* log tostring helpers, token vectors pretty prints

* main: replaced fprintf/LOG_TEE, some trace logging

* LOG_DISABLE_LOGS compile flag, wrapped f in macros

* fix LOG_TEELN and configchecker

* stub LOG_DUMP_CMDLINE for WIN32 for now

* fix msvc

* cleanup main.cpp:273

* fix stray whitespace after master sync

* log : fix compile warnings

- do not use C++20 stuff
- use PRIu64 to print uint64_t
- avoid string copies by using const ref
- fix ", ##__VA_ARGS__" warnings
- compare strings with == and !=

* log : do not append to existing log + disable file line func by default

* log : try to fix Windows build

* main : wip logs

* main : add trace log

* review: macro f lowercase, str append to sstream

* review: simplify ifs and str comparisons

* fix MSVC, formatting, FMT/VAL placeholders

* review: if/else cleanup

* review: if/else cleanup (2)

* replace _ prefix with _impl suffix

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---
 .gitignore             |   1 +
 Makefile               |   7 +-
 common/common.cpp      |  22 ++
 common/common.h        |   3 +
 common/log.h           | 643 +++++++++++++++++++++++++++++++++++++++++
 examples/chat.sh       |   2 +-
 examples/main/main.cpp | 288 +++++++++++-------
 7 files changed, 859 insertions(+), 107 deletions(-)
 create mode 100644 common/log.h

diff --git a/.gitignore b/.gitignore
index 7a3f3fff4..54ea2b522 100644
--- a/.gitignore
+++ b/.gitignore
@@ -5,6 +5,7 @@
 *.bin
 *.exe
 *.dll
+*.log
 .DS_Store
 .build/
 .cache/
diff --git a/Makefile b/Makefile
index 44e68b7fc..c8b8a92d7 100644
--- a/Makefile
+++ b/Makefile
@@ -326,6 +326,11 @@ k_quants.o: k_quants.c k_quants.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_NO_K_QUANTS
 
+ifdef LLAMA_DISABLE_LOGS
+	CFLAGS   += -DLOG_DISABLE_LOGS
+	CXXFLAGS += -DLOG_DISABLE_LOGS
+endif # LLAMA_DISABLE_LOGS
+
 #
 # Print build information
 #
@@ -356,7 +361,7 @@ OBJS += ggml-alloc.o
 llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-common.o: common/common.cpp common/common.h build-info.h
+common.o: common/common.cpp common/common.h build-info.h common/log.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 console.o: common/console.cpp common/console.h
diff --git a/common/common.cpp b/common/common.cpp
index 90fe2e84e..ed09fc27d 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -480,6 +480,9 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             }
         } else if (arg == "-h" || arg == "--help") {
             gpt_print_usage(argc, argv, default_params);
+#ifndef LOG_DISABLE_LOGS
+            log_print_usage();
+#endif // LOG_DISABLE_LOGS
             exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
@@ -519,6 +522,25 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 std::istreambuf_iterator<char>(),
                 std::back_inserter(params.grammar)
             );
+#ifndef LOG_DISABLE_LOGS
+        // Parse args for logging parameters
+        } else if ( log_param_single_parse( argv[i] ) ) {
+            // Do nothing, log_param_single_parse automatically does it's thing
+            //  and returns if a match was found and parsed.
+        } else if ( log_param_pair_parse( /*check_but_dont_parse*/ true, argv[i] ) ) {
+            // We have a matching known parameter requiring an argument,
+            //  now we need to check if there is anything after this argv
+            //  and flag invalid_param or parse it.
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            if( !log_param_pair_parse( /*check_but_dont_parse*/ false, argv[i-1], argv[i]) ) {
+                invalid_param = true;
+                break;
+            }
+        // End of Parse args for logging parameters
+#endif // LOG_DISABLE_LOGS
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             gpt_print_usage(argc, argv, default_params);
diff --git a/common/common.h b/common/common.h
index c15373144..5a379688e 100644
--- a/common/common.h
+++ b/common/common.h
@@ -4,6 +4,9 @@
 
 #include "llama.h"
 
+#define LOG_NO_FILE_LINE_FUNCTION
+#include "log.h"
+
 #include <string>
 #include <vector>
 #include <random>
diff --git a/common/log.h b/common/log.h
new file mode 100644
index 000000000..c1364187d
--- /dev/null
+++ b/common/log.h
@@ -0,0 +1,643 @@
+#pragma once
+
+#include <chrono>
+#include <cstring>
+#include <sstream>
+#include <iostream>
+#include <thread>
+#include <vector>
+#include <algorithm>
+#include <cinttypes>
+
+// --------------------------------
+//
+// Basic usage:
+//
+// --------
+//
+//  The LOG() and LOG_TEE() macros are ready to go by default
+//   they do not require any initialization.
+//
+//  LOGLN() and LOG_TEELN() are variants which automatically
+//   include \n character at the end of the log string.
+//
+//  LOG() behaves exactly like printf, by default writing to a logfile.
+//  LOG_TEE() additionally, prints to the screen too ( mimics Unix tee command ).
+//
+//  Default logfile is named
+//   "llama.<threadID>.log"
+//  Default LOG_TEE() secondary output target is
+//   stderr
+//
+//  Logs can be dynamically disabled or enabled using functions:
+//   log_disable()
+//  and
+//   log_enable()
+//
+//  A log target can be changed with:
+//   log_set_target( string )
+//    creating and opening, or re-opening a file by string filename
+//  or
+//   log_set_target( FILE* )
+//    allowing to point at stderr, stdout, or any valid FILE* file handler.
+//
+// --------
+//
+// End of Basic usage.
+//
+// --------------------------------
+
+// Specifies a log target.
+//  default uses log_handler() with "llama.log" log file
+//  this can be changed, by defining LOG_TARGET
+//  like so:
+//
+//  #define LOG_TARGET (a valid FILE*)
+//  #include "log.h"
+//
+//  or it can be simply redirected to stdout or stderr
+//  like so:
+//
+//  #define LOG_TARGET stderr
+//  #include "log.h"
+//
+//  The log target can also be redirected to a diffrent function
+//  like so:
+//
+//  #define LOG_TARGET log_handler_diffrent()
+//  #include "log.h"
+//
+//  FILE* log_handler_diffrent()
+//  {
+//      return stderr;
+//  }
+//
+//  or:
+//
+//  #define LOG_TARGET log_handler_another_one("somelog.log")
+//  #include "log.h"
+//
+//  FILE* log_handler_another_one(char*filename)
+//  {
+//      static FILE* logfile = nullptr;
+//      (...)
+//      if( !logfile )
+//      {
+//          fopen(...)
+//      }
+//      (...)
+//      return logfile
+//  }
+//
+#ifndef LOG_TARGET
+    #define LOG_TARGET log_handler()
+#endif
+
+#ifndef LOG_TEE_TARGET
+    #define LOG_TEE_TARGET stderr
+#endif
+
+// Utility to obtain "pid" like unique process id and use it when creating log files.
+inline std::string log_get_pid()
+{
+    static std::string pid;
+    if (pid.empty())
+    {
+        // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+        //  it's not the same as "pid" but is unique enough to solve multiple instances
+        //  trying to write to the same log.
+        std::stringstream ss;
+        ss << std::this_thread::get_id();
+        pid = ss.str();
+    }
+
+    return pid;
+}
+
+// Utility function for generating log file names with unique id based on thread id.
+//  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
+//  where the number is a runtime id of the current thread.
+
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension)
+
+// INTERNAL, DO NOT USE
+inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension)
+{
+    std::stringstream buf;
+
+    buf << log_file_basename;
+    buf << ".";
+    buf << log_get_pid();
+    buf << ".";
+    buf << log_file_extension;
+
+    return buf.str();
+}
+
+#ifndef LOG_DEFAULT_FILE_NAME
+    #define LOG_DEFAULT_FILE_NAME log_filename_generator("llama", "log")
+#endif
+
+// Utility for turning #define values into string literals
+//  so we can have a define for stderr and
+//  we can print "stderr" instead of literal stderr, etc.
+#define LOG_STRINGIZE1(s) #s
+#define LOG_STRINGIZE(s) LOG_STRINGIZE1(s)
+
+#define LOG_TEE_TARGET_STRING LOG_STRINGIZE(LOG_TEE_TARGET)
+
+// Allows disabling timestamps.
+//  in order to disable, define LOG_NO_TIMESTAMPS
+//  like so:
+//
+//  #define LOG_NO_TIMESTAMPS
+//  #include "log.h"
+//
+#ifndef LOG_NO_TIMESTAMPS
+    #ifndef _WIN32
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TIMESTAMP_FMT "%s"
+    #define LOG_TIMESTAMP_VAL ,""
+#endif
+
+#ifdef LOG_TEE_TIMESTAMPS
+    #ifndef _WIN32
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TEE_TIMESTAMP_FMT "%s"
+    #define LOG_TEE_TIMESTAMP_VAL ,""
+#endif
+
+// Allows disabling file/line/function prefix
+//  in order to disable, define LOG_NO_FILE_LINE_FUNCTION
+//  like so:
+//
+//  #define LOG_NO_FILE_LINE_FUNCTION
+//  #include "log.h"
+//
+#ifndef LOG_NO_FILE_LINE_FUNCTION
+    #ifndef _WIN32
+        #define LOG_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_FLF_FMT "%s"
+    #define LOG_FLF_VAL ,""
+#endif
+
+#ifdef LOG_TEE_FILE_LINE_FUNCTION
+    #ifndef _WIN32
+        #define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_TEE_FLF_FMT "%s"
+    #define LOG_TEE_FLF_VAL ,""
+#endif
+
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
+
+// INTERNAL, DO NOT USE
+//  USE LOG() INSTEAD
+//
+#ifndef _WIN32
+    #define LOG_IMPL(str, ...)                                                                                          \
+    {                                                                                                               \
+        if (LOG_TARGET != nullptr)                                                                                  \
+        {                                                                                                           \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                     \
+        }                                                                                                           \
+    }
+#else
+    #define LOG_IMPL(str, ...)                                                                                               \
+    {                                                                                                                    \
+        if (LOG_TARGET != nullptr)                                                                                       \
+        {                                                                                                                \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                          \
+        }                                                                                                                \
+    }
+#endif
+
+// INTERNAL, DO NOT USE
+//  USE LOG_TEE() INSTEAD
+//
+#ifndef _WIN32
+    #define LOG_TEE_IMPL(str, ...)                                                                                                          \
+    {                                                                                                                                   \
+        if (LOG_TARGET != nullptr)                                                                                                      \
+        {                                                                                                                               \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                         \
+        }                                                                                                                               \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                         \
+        {                                                                                                                               \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                     \
+        }                                                                                                                               \
+    }
+#else
+    #define LOG_TEE_IMPL(str, ...)                                                                                                               \
+    {                                                                                                                                        \
+        if (LOG_TARGET != nullptr)                                                                                                           \
+        {                                                                                                                                    \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                              \
+        }                                                                                                                                    \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                              \
+        {                                                                                                                                    \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                          \
+        }                                                                                                                                    \
+    }
+#endif
+
+// The '\0' as a last argument, is a trick to bypass the silly
+//  "warning: ISO C++11 requires at least one argument for the "..." in a variadic macro"
+//  so we can have a single macro which can be called just like printf.
+
+// Main LOG macro.
+//  behaves like printf, and supports arguments the exact same way.
+//
+#ifndef _WIN32
+    #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "")
+#endif
+
+// Main TEE macro.
+//  does the same as LOG
+//  and
+//  simultaneously writes stderr.
+//
+// Secondary target can be changed just like LOG_TARGET
+//  by defining LOG_TEE_TARGET
+//
+#ifndef _WIN32
+    #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "")
+#endif
+
+// LOG macro variants with auto endline.
+#ifndef _WIN32
+    #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
+    #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
+#else
+    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "\n")
+    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "\n")
+#endif
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+{
+    static bool _initialized{false};
+    static bool _disabled{(filename.empty() && target == nullptr)};
+    static std::string log_current_filename{filename};
+    static FILE *log_current_target{target};
+    static FILE *logfile = nullptr;
+
+    if (change)
+    {
+        if (disable == LogTriStateTrue)
+        {
+            // Disable primary target
+            _disabled = true;
+        }
+        // If previously disabled, only enable, and keep previous target
+        else if (disable == LogTriStateFalse)
+        {
+            _disabled = false;
+        }
+        // Otherwise, process the arguments
+        else if (log_current_filename != filename || log_current_target != target)
+        {
+            _initialized = false;
+        }
+    }
+
+    if (_initialized)
+    {
+        if (_disabled)
+        {
+            // Log is disabled
+            return nullptr;
+        }
+
+        // with fallback in case something went wrong
+        return logfile ? logfile : stderr;
+    }
+
+    // do the (re)initialization
+    if (target != nullptr)
+    {
+        if (logfile != nullptr && logfile != stdout && logfile != stderr)
+        {
+            fclose(logfile);
+        }
+
+        log_current_filename = LOG_DEFAULT_FILE_NAME;
+        log_current_target = target;
+
+        logfile = target;
+    }
+    else
+    {
+        if (log_current_filename != filename)
+        {
+            if (logfile != nullptr && logfile != stdout && logfile != stderr)
+            {
+                fclose(logfile);
+            }
+        }
+
+        logfile = fopen(filename.c_str(), "w");
+    }
+
+    if (!logfile)
+    {
+        //  Verify whether the file was opened, otherwise fallback to stderr
+        logfile = stderr;
+
+        fprintf(stderr, "Failed to open logfile '%s' with error '%s'\n", filename.c_str(), std::strerror(errno));
+        fflush(stderr);
+
+        // At this point we let the init flag be to true below, and let the target fallback to stderr
+        //  otherwise we would repeatedly fopen() which was already unsuccessful
+    }
+
+    _initialized = true;
+
+    return logfile ? logfile : stderr;
+}
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+{
+    return log_handler1_impl(change, disable, filename, target);
+}
+
+// Disables logs entirely at runtime.
+//  Makes LOG() and LOG_TEE() produce no output,
+//  untill enabled back.
+#define log_disable() log_disable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_disable_impl()
+{
+    return log_handler1_impl(true, LogTriStateTrue);
+}
+
+// Enables logs at runtime.
+#define log_enable() log_enable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_enable_impl()
+{
+    return log_handler1_impl(true, LogTriStateFalse);
+}
+
+// Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
+#define log_set_target(target) log_set_target_impl(target)
+
+// INTERNAL, DO NOT USE
+inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); }
+inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); }
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler() { return log_handler1_impl(); }
+
+inline void log_test()
+{
+    log_disable();
+    LOG("01 Hello World to nobody, because logs are disabled!\n")
+    log_enable();
+    LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET))
+    LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n")
+    log_set_target(stderr);
+    LOG("04 Hello World to stderr!\n")
+    LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n")
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("06 Hello World to default log file!\n")
+    log_set_target(stdout);
+    LOG("07 Hello World to stdout!\n")
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("08 Hello World to default log file again!\n")
+    log_disable();
+    LOG("09 Hello World _1_ into the void!\n")
+    log_enable();
+    LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n")
+    log_disable();
+    log_set_target("llama.anotherlog.log");
+    LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n")
+    log_enable();
+    LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n")
+    log_set_target("llama.yetanotherlog.log");
+    LOG("13 Hello World this time in yet new file?\n")
+    log_set_target(log_filename_generator("llama_autonamed", "log"));
+    LOG("14 Hello World in log with generated filename!\n")
+#ifdef _WIN32
+    LOG_TEE("15 Hello msvc TEE without arguments\n")
+    LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test")
+    LOG_TEELN("17 Hello msvc TEELN without arguments\n")
+    LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test")
+    LOG("19 Hello msvc LOG without arguments\n")
+    LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test")
+    LOGLN("21 Hello msvc LOGLN without arguments\n")
+    LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test")
+#endif
+}
+
+inline bool log_param_single_parse(const std::string & param)
+{
+    if ( param == "--log-test")
+    {
+        log_test();
+        return true;
+    }
+
+    if ( param == "--log-disable")
+    {
+        log_disable();
+        return true;
+    }
+
+    if ( param == "--log-enable")
+    {
+        log_enable();
+        return true;
+    }
+
+    return false;
+}
+
+inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string & param, const std::string & next = std::string())
+{
+    if ( param == "--log-file")
+    {
+        if (!check_but_dont_parse)
+        {
+            log_set_target(log_filename_generator(next.empty() ? "unnamed" : next, "log"));
+        }
+
+        return true;
+    }
+
+    return false;
+}
+
+inline void log_print_usage()
+{
+    fprintf(stdout, "log options:\n");
+    /* format
+    fprintf(stdout, "  -h, --help            show this help message and exit\n");*/
+    /* spacing
+    fprintf(stdout, "__-param----------------Description\n");*/
+    fprintf(stdout, "  --log-test            Run simple logging test\n");
+    fprintf(stdout, "  --log-disable         Disable trace logs\n");
+    fprintf(stdout, "  --log-enable          Enable trace logs\n");
+    fprintf(stdout, "  --log-file            Specify a log filename (without extension)\n");
+    fprintf(stdout, "                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
+}
+
+#define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
+
+// INTERNAL, DO NOT USE
+inline void log_dump_cmdline_impl(int argc, char **argv)
+{
+    std::stringstream buf;
+    for (int i = 0; i < argc; ++i)
+    {
+        if (std::string(argv[i]).find(' ') != std::string::npos)
+        {
+            buf << " \"" << argv[i] <<"\"";
+        }
+        else
+        {
+            buf << " " << argv[i];
+        }
+    }
+    LOGLN("Cmd:%s", buf.str().c_str())
+}
+
+#define log_tostr(var) log_var_to_string_impl(var).c_str()
+
+inline std::string log_var_to_string_impl(bool var)
+{
+    return var ? "true" : "false";
+}
+
+inline std::string log_var_to_string_impl(std::string var)
+{
+    return var;
+}
+
+inline std::string log_var_to_string_impl(const std::vector<int> & var)
+{
+    std::stringstream buf;
+    buf << "[ ";
+    bool first = true;
+    for (auto e : var)
+    {
+        if (first)
+        {
+            first = false;
+        }
+        else
+        {
+            buf << ", ";
+        }
+        buf << std::to_string(e);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+#define LOG_TOKENS_TOSTR_PRETTY(ctx, tokens)                                 \
+    [&tokens, &ctx]()                                                        \
+    {                                                                        \
+        std::stringstream buf;                                               \
+        buf << "[ ";                                                         \
+                                                                             \
+        bool first = true;                                                   \
+        for (const auto &token : tokens)                                     \
+        {                                                                    \
+            if (!first)                                                      \
+                buf << ", ";                                                 \
+            else                                                             \
+                first = false;                                               \
+                                                                             \
+            auto detokenized = llama_token_to_piece(ctx, token);             \
+                                                                             \
+            detokenized.erase(                                               \
+                std::remove_if(                                              \
+                    detokenized.begin(),                                     \
+                    detokenized.end(),                                       \
+                    [](const unsigned char c) { return !std::isprint(c); }), \
+                detokenized.end());                                          \
+                                                                             \
+            buf                                                              \
+                << "'" << detokenized << "'"                                 \
+                << ":" << std::to_string(token);                             \
+        }                                                                    \
+        buf << " ]";                                                         \
+                                                                             \
+        return buf.str();                                                    \
+    }()                                                                      \
+        .c_str()
+
+#ifdef LOG_DISABLE_LOGS
+
+#undef LOG
+#define LOG(...) // dummy stub
+#undef LOGLN
+#define LOGLN(...) // dummy stub
+
+#undef LOG_TEE
+#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__); // convert to normal fprintf
+
+#undef LOG_TEELN
+#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__); // convert to normal fprintf
+
+#undef LOG_DISABLE
+#define LOG_DISABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_SET_TARGET
+#define LOG_SET_TARGET(...) // dummy stub
+
+#undef LOG_DUMP_CMDLINE
+#define LOG_DUMP_CMDLINE(...) // dummy stub
+
+#endif // LOG_DISABLE_LOGS
diff --git a/examples/chat.sh b/examples/chat.sh
index 9a928ef05..d567acecd 100755
--- a/examples/chat.sh
+++ b/examples/chat.sh
@@ -11,6 +11,6 @@ cd ..
 #
 #   "--keep 48" is based on the contents of prompts/chat-with-bob.txt
 #
-./main -m ./models/7B/ggml-model-q4_0.bin -c 512 -b 1024 -n 256 --keep 48 \
+./main -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
     --repeat_penalty 1.0 --color -i \
     -r "User:" -f prompts/chat-with-bob.txt
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index 89cc4f602..7117db4b0 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -4,6 +4,7 @@
 #endif
 
 #include "common.h"
+
 #include "console.h"
 #include "llama.h"
 #include "build-info.h"
@@ -112,6 +113,15 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("main", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc,argv);
+#endif // LOG_DISABLE_LOGS
+
+    // TODO: Dump params ?
+    //LOG("Params perplexity: %s\n", LOG_TOSTR(params.perplexity));
+
     // save choice to use color for later
     // (note for later: this is a slightly awkward choice)
     console::init(params.simple_io, params.use_color);
@@ -134,34 +144,35 @@ int main(int argc, char ** argv) {
     }
 
     if (params.rope_freq_base != 10000.0) {
-        fprintf(stderr, "%s: warning: changing RoPE frequency base to %g (default 10000.0)\n", __func__, params.rope_freq_base);
+        LOG_TEE("%s: warning: changing RoPE frequency base to %g (default 10000.0)\n", __func__, params.rope_freq_base);
     }
 
     if (params.rope_freq_scale != 1.0) {
-        fprintf(stderr, "%s: warning: scaling RoPE frequency by %g (default 1.0)\n", __func__, params.rope_freq_scale);
+        LOG_TEE("%s: warning: scaling RoPE frequency by %g (default 1.0)\n", __func__, params.rope_freq_scale);
     }
 
     if (params.n_ctx > 2048) {
         // TODO: determine the actual max context of the model (e.g. 4096 for LLaMA v2) and use that instead of 2048
-        fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified)\n", __func__, params.n_ctx);
+        LOG_TEE("%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified)\n", __func__, params.n_ctx);
     } else if (params.n_ctx < 8) {
-        fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
         params.n_ctx = 8;
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
 
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
+    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
         params.prompt = gpt_random_prompt(rng);
     }
 
+    LOG("%s: llama backend init\n", __func__);
     llama_backend_init(params.numa);
 
     llama_model * model;
@@ -171,6 +182,7 @@ int main(int argc, char ** argv) {
     g_ctx = &ctx;
 
     // load the model and apply lora adapter, if any
+    LOG("%s: load the model and apply lora adapter, if any\n", __func__);
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (params.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
@@ -178,14 +190,14 @@ int main(int argc, char ** argv) {
     }
 
     if (model == NULL) {
-        fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        LOG_TEE("%s: error: unable to load model\n", __func__);
         return 1;
     }
 
     // print system information
     {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
+        LOG_TEE("\n");
+        LOG_TEE("system_info: n_threads = %d / %d | %s\n",
                 params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
     }
 
@@ -193,7 +205,7 @@ int main(int argc, char ** argv) {
     // uncomment the "used_mem" line in llama.cpp to see the results
     if (params.mem_test) {
         {
-            fprintf(stderr, "%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx);
+            LOG_TEE("%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx);
 
             const std::vector<llama_token> tmp(params.n_batch, llama_token_bos(ctx));
             llama_eval(ctx, tmp.data(), tmp.size(), params.n_ctx, params.n_threads);
@@ -219,7 +231,7 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> session_tokens;
 
     if (!path_session.empty()) {
-        fprintf(stderr, "%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
+        LOG_TEE("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
 
         // fopen to check for existing session
         FILE * fp = std::fopen(path_session.c_str(), "rb");
@@ -229,33 +241,38 @@ int main(int argc, char ** argv) {
             session_tokens.resize(params.n_ctx);
             size_t n_token_count_out = 0;
             if (!llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
-                fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
+                LOG_TEE("%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
                 return 1;
             }
             session_tokens.resize(n_token_count_out);
             llama_set_rng_seed(ctx, params.seed);
 
-            fprintf(stderr, "%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
+            LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
         } else {
-            fprintf(stderr, "%s: session file does not exist, will create\n", __func__);
+            LOG_TEE("%s: session file does not exist, will create\n", __func__);
         }
     }
 
-    // Add BOS if SPM tokenizer
     const bool add_bos = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM;
+    LOG("add_bos: %d\n", add_bos);
 
-    // tokenize the prompt
     std::vector<llama_token> embd_inp;
 
     if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
+        LOG("tokenize the prompt\n");
         embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
     } else {
+        LOG("use session tokens\n");
         embd_inp = session_tokens;
     }
 
+    LOG("prompt: \"%s\"\n", log_tostr(params.prompt));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+
     // Should not run without any tokens
     if (embd_inp.empty()) {
         embd_inp.push_back(llama_token_bos(ctx));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
     }
 
     // Tokenize negative prompt
@@ -263,23 +280,31 @@ int main(int argc, char ** argv) {
     int guidance_offset = 0;
     int original_prompt_len = 0;
     if (ctx_guidance) {
+        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(params.cfg_negative_prompt));
+
         guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
 
         std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
+
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
+        LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
+        LOG("guidance_offset:     %s", log_tostr(guidance_offset));
     }
 
     const int n_ctx = llama_n_ctx(ctx);
+    LOG("n_ctx: %d\n", n_ctx);
 
     if ((int) embd_inp.size() > n_ctx - 4) {
-        fprintf(stderr, "%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
         return 1;
     }
 
     // debug message about similarity of saved session, if applicable
     size_t n_matching_session_tokens = 0;
-    if (session_tokens.size()) {
+    if (session_tokens.size() > 0) {
         for (llama_token id : session_tokens) {
             if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) {
                 break;
@@ -287,22 +312,27 @@ int main(int argc, char ** argv) {
             n_matching_session_tokens++;
         }
         if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
-            fprintf(stderr, "%s: using full prompt from session file\n", __func__);
+            LOG_TEE("%s: using full prompt from session file\n", __func__);
         } else if (n_matching_session_tokens >= embd_inp.size()) {
-            fprintf(stderr, "%s: session file has exact match for prompt!\n", __func__);
+            LOG_TEE("%s: session file has exact match for prompt!\n", __func__);
         } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
-            fprintf(stderr, "%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
+            LOG_TEE("%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
                 __func__, n_matching_session_tokens, embd_inp.size());
         } else {
-            fprintf(stderr, "%s: session file matches %zu / %zu tokens of prompt\n",
+            LOG_TEE("%s: session file matches %zu / %zu tokens of prompt\n",
                 __func__, n_matching_session_tokens, embd_inp.size());
         }
     }
 
+    LOGLN(
+            "recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu, embd_inp.size() %zu",
+            log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size(), embd_inp.size());
+
     // if we will use the cache for the full prompt without reaching the end of the cache, force
     // reevaluation of the last token token to recalculate the cached logits
-    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() &&
-            session_tokens.size() > embd_inp.size()) {
+    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {
+        LOGLN("recalculate the cached logits (do): session_tokens.resize( %zu )", embd_inp.size() - 1);
+
         session_tokens.resize(embd_inp.size() - 1);
     }
 
@@ -315,6 +345,9 @@ int main(int argc, char ** argv) {
     const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
     const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false);
 
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
+
     // in instruct mode, we inject a prefix and a suffix to each input by the user
     if (params.instruct) {
         params.interactive_first = true;
@@ -327,30 +360,30 @@ int main(int argc, char ** argv) {
     }
 
     if (params.verbose_prompt) {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        LOG_TEE("\n");
+        LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
         for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+            LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
         if (ctx_guidance) {
-            fprintf(stderr, "\n");
-            fprintf(stderr, "%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
-            fprintf(stderr, "%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
+            LOG_TEE("\n");
+            LOG_TEE("%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
+            LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
             for (int i = 0; i < (int) guidance_inp.size(); i++) {
-                fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
+                LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
             }
         }
 
         if (params.n_keep > 0) {
-        fprintf(stderr, "%s: static prompt based on n_keep: '", __func__);
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
-                fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
+                LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
             }
-            fprintf(stderr, "'\n");
+            LOG_TEE("'\n");
         }
-        fprintf(stderr, "\n");
+        LOG_TEE("\n");
     }
 
     if (params.interactive) {
@@ -367,30 +400,30 @@ int main(int argc, char ** argv) {
         SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-        fprintf(stderr, "%s: interactive mode on.\n", __func__);
+        LOG_TEE("%s: interactive mode on.\n", __func__);
 
         if (params.antiprompt.size()) {
-            for (auto antiprompt : params.antiprompt) {
-                fprintf(stderr, "Reverse prompt: '%s'\n", antiprompt.c_str());
+            for (const auto & antiprompt : params.antiprompt) {
+                LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
             }
         }
 
         if (params.input_prefix_bos) {
-            fprintf(stderr, "Input prefix with BOS\n");
+            LOG_TEE("Input prefix with BOS\n");
         }
 
         if (!params.input_prefix.empty()) {
-            fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str());
+            LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
         }
 
         if (!params.input_suffix.empty()) {
-            fprintf(stderr, "Input suffix: '%s'\n", params.input_suffix.c_str());
+            LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    fprintf(stderr, "sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
             params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
-    fprintf(stderr, "generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
-    fprintf(stderr, "\n\n");
+    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+    LOG_TEE("\n\n");
 
     grammar_parser::parse_state parsed_grammar;
     llama_grammar *             grammar = NULL;
@@ -400,14 +433,14 @@ int main(int argc, char ** argv) {
         if (parsed_grammar.rules.empty()) {
             return 1;
         }
-        fprintf(stderr, "%s: grammar:\n", __func__);
+        LOG_TEE("%s: grammar:\n", __func__);
         grammar_parser::print_grammar(stderr, parsed_grammar);
-        fprintf(stderr, "\n");
+        LOG_TEE("\n");
 
         {
             auto it = params.logit_bias.find(llama_token_eos(ctx));
             if (it != params.logit_bias.end() && it->second == -INFINITY) {
-                fprintf(stderr, "%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
             }
         }
 
@@ -430,11 +463,11 @@ int main(int argc, char ** argv) {
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }
-        fprintf(stderr, "== Running in interactive mode. ==\n"
+        LOG_TEE("== Running in interactive mode. ==\n");
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-               " - Press Ctrl+C to interject at any time.\n"
+        LOG_TEE(       " - Press Ctrl+C to interject at any time.\n");
 #endif
-               "%s\n", control_message);
+        LOG_TEE(       "%s\n", control_message);
 
         is_interacting = params.interactive_first;
     }
@@ -459,8 +492,9 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    // do one empty run to warm up the model
     {
+        LOG("warming up the model with an empty run\n");
+
         const std::vector<llama_token> tmp = { llama_token_bos(ctx), };
         llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
         llama_reset_timings(ctx);
@@ -471,15 +505,17 @@ int main(int argc, char ** argv) {
         if (embd.size() > 0) {
             // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
             // --prompt or --file which uses the same value.
-            auto max_embd_size = n_ctx - 4;
+            int max_embd_size = n_ctx - 4;
+
             // Ensure the input doesn't exceed the context size by truncating embd if necessary.
-            if ((int)embd.size() > max_embd_size) {
-                auto skipped_tokens = embd.size() - max_embd_size;
+            if ((int) embd.size() > max_embd_size) {
+                const int skipped_tokens = (int) embd.size() - max_embd_size;
+                embd.resize(max_embd_size);
+
                 console::set_display(console::error);
-                printf("<<input too long: skipped %zu token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
                 console::set_display(console::reset);
                 fflush(stdout);
-                embd.resize(max_embd_size);
             }
 
             // infinite text generation via context swapping
@@ -488,28 +524,26 @@ int main(int argc, char ** argv) {
             // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
             if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
                 if (params.n_predict == -2) {
-                    fprintf(stderr, "\n\n%s: context full, stopping generation\n", __func__);
+                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                     break;
                 }
 
                 const int n_left = n_past - params.n_keep;
+                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d\n", n_past, n_left, n_ctx, params.n_keep);
+
                 // always keep the first token - BOS
-                n_past = std::max(1, params.n_keep);
+                n_past          = std::max(1, params.n_keep);
                 n_past_guidance = std::max(1, params.n_keep + guidance_offset);
 
+                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+
                 // insert n_left/2 tokens at the start of embd from last_n_tokens
                 embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size());
 
-                // stop saving session if we run out of context
-                path_session.clear();
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
 
-                //printf("\n---\n");
-                //printf("resetting: '");
-                //for (int i = 0; i < (int) embd.size(); i++) {
-                //    printf("%s", llama_token_to_piece(ctx, embd[i]));
-                //}
-                //printf("'\n");
-                //printf("\n---\n");
+                LOG("clear session path\n");
+                path_session.clear();
             }
 
             // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
@@ -539,7 +573,7 @@ int main(int argc, char ** argv) {
 
             if (ctx_guidance) {
                 int input_size = 0;
-                llama_token* input_buf = NULL;
+                llama_token * input_buf = NULL;
 
                 if (n_past_guidance < (int) guidance_inp.size()) {
                     // Guidance context should have the same data with these modifications:
@@ -555,22 +589,19 @@ int main(int argc, char ** argv) {
                         );
                     }
 
-                    input_buf = embd_guidance.data();
+                    input_buf  = embd_guidance.data();
                     input_size = embd_guidance.size();
-                    //fprintf(stderr, "\n---------------------\n");
-                    //for (int i = 0; i < (int) embd_guidance.size(); i++) {
-                        //fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_guidance[i]));
-                    //}
-                    //fprintf(stderr, "\n---------------------\n");
+
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
                 } else {
-                    input_buf = embd.data();
+                    input_buf  = embd.data();
                     input_size = embd.size();
                 }
 
                 for (int i = 0; i < input_size; i += params.n_batch) {
                     int n_eval = std::min(input_size - i, params.n_batch);
                     if (llama_eval(ctx_guidance, input_buf + i, n_eval, n_past_guidance, params.n_threads)) {
-                        fprintf(stderr, "%s : failed to eval\n", __func__);
+                        LOG_TEE("%s : failed to eval\n", __func__);
                         return 1;
                     }
 
@@ -583,11 +614,17 @@ int main(int argc, char ** argv) {
                 if (n_eval > params.n_batch) {
                     n_eval = params.n_batch;
                 }
+
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
                 if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) {
-                    fprintf(stderr, "%s : failed to eval\n", __func__);
+                    LOG_TEE("%s : failed to eval\n", __func__);
                     return 1;
                 }
+
                 n_past += n_eval;
+
+                LOG("n_past = %d\n", n_past);
             }
 
             if (embd.size() > 0 && !path_session.empty()) {
@@ -600,7 +637,6 @@ int main(int argc, char ** argv) {
         embd_guidance.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
-            // out of user input, sample next token
             const float   temp            = params.temp;
             const int32_t top_k           = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
             const float   top_p           = params.top_p;
@@ -619,6 +655,8 @@ int main(int argc, char ** argv) {
             if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {
                 need_to_save_session = false;
                 llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
+
+                LOG("saved session to %s\n", path_session.c_str());
             }
 
             llama_token id = 0;
@@ -638,55 +676,68 @@ int main(int argc, char ** argv) {
                     candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
                 }
 
-                llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+                llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
 
                 if (ctx_guidance) {
-                    llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale);
+                    llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale);
                 }
 
                 // Apply penalties
                 float nl_logit = logits[llama_token_nl(ctx)];
                 auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
-                llama_sample_repetition_penalty(ctx, &candidates_p,
+                llama_sample_repetition_penalty(ctx, &cur_p,
                     last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
                     last_n_repeat, repeat_penalty);
-                llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
+                llama_sample_frequency_and_presence_penalties(ctx, &cur_p,
                     last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
                     last_n_repeat, alpha_frequency, alpha_presence);
                 if (!penalize_nl) {
-                    for (size_t idx = 0; idx < candidates_p.size; idx++) {
-                        if (candidates_p.data[idx].id == llama_token_nl(ctx)) {
-                            candidates_p.data[idx].logit = nl_logit;
+                    for (size_t idx = 0; idx < cur_p.size; idx++) {
+                        if (cur_p.data[idx].id == llama_token_nl(ctx)) {
+                            cur_p.data[idx].logit = nl_logit;
                             break;
                         }
                     }
                 }
 
                 if (grammar != NULL) {
-                    llama_sample_grammar(ctx, &candidates_p, grammar);
+                    llama_sample_grammar(ctx, &cur_p, grammar);
                 }
 
                 if (temp <= 0) {
                     // Greedy sampling
-                    id = llama_sample_token_greedy(ctx, &candidates_p);
+                    id = llama_sample_token_greedy(ctx, &cur_p);
                 } else {
                     if (mirostat == 1) {
                         static float mirostat_mu = 2.0f * mirostat_tau;
                         const int mirostat_m = 100;
-                        llama_sample_temperature(ctx, &candidates_p, temp);
-                        id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
+                        llama_sample_temperature(ctx, &cur_p, temp);
+                        id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
                     } else if (mirostat == 2) {
                         static float mirostat_mu = 2.0f * mirostat_tau;
-                        llama_sample_temperature(ctx, &candidates_p, temp);
-                        id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+                        llama_sample_temperature(ctx, &cur_p, temp);
+                        id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu);
                     } else {
                         // Temperature sampling
-                        llama_sample_top_k(ctx, &candidates_p, top_k, 1);
-                        llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
-                        llama_sample_typical(ctx, &candidates_p, typical_p, 1);
-                        llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-                        llama_sample_temperature(ctx, &candidates_p, temp);
-                        id = llama_sample_token(ctx, &candidates_p);
+                        llama_sample_top_k      (ctx, &cur_p, top_k, 1);
+                        llama_sample_tail_free  (ctx, &cur_p, tfs_z, 1);
+                        llama_sample_typical    (ctx, &cur_p, typical_p, 1);
+                        llama_sample_top_p      (ctx, &cur_p, top_p, 1);
+                        llama_sample_temperature(ctx, &cur_p, temp);
+
+                        {
+                            const int n_top = 10;
+                            LOG("top %d candidates:\n", n_top);
+
+                            for (int i = 0; i < n_top; i++) {
+                                const llama_token id = cur_p.data[i].id;
+                                LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p);
+                            }
+                        }
+
+                        id = llama_sample_token(ctx, &cur_p);
+
+                        LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str());
                     }
                 }
                 // printf("`%d`", candidates_p.size);
@@ -697,9 +748,10 @@ int main(int argc, char ** argv) {
 
                 last_n_tokens.erase(last_n_tokens.begin());
                 last_n_tokens.push_back(id);
+
+                LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_n_tokens));
             }
 
-            // add it to the context
             embd.push_back(id);
 
             // echo this to console
@@ -707,8 +759,11 @@ int main(int argc, char ** argv) {
 
             // decrement remaining sampling budget
             --n_remain;
+
+            LOG("n_remain: %d\n", n_remain);
         } else {
             // some user input remains from prompt or interaction, forward it to processing
+            LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
                 last_n_tokens.erase(last_n_tokens.begin());
@@ -736,13 +791,12 @@ int main(int argc, char ** argv) {
             fflush(stdout);
         }
         // reset color to default if we there is no pending user input
-        if (input_echo && (int)embd_inp.size() == n_consumed) {
+        if (input_echo && (int) embd_inp.size() == n_consumed) {
             console::set_display(console::reset);
         }
 
         // if not currently processing queued inputs;
         if ((int) embd_inp.size() <= n_consumed) {
-
             // check for reverse prompt
             if (params.antiprompt.size()) {
                 std::string last_output;
@@ -760,7 +814,7 @@ int main(int argc, char ** argv) {
                         ? last_output.length() - static_cast<size_t>(antiprompt.length() + extra_padding)
                         : 0;
 
-                    if (last_output.find(antiprompt.c_str(), search_start_pos) != std::string::npos) {
+                    if (last_output.find(antiprompt, search_start_pos) != std::string::npos) {
                         if (params.interactive) {
                             is_interacting = true;
                             console::set_display(console::user_input);
@@ -770,10 +824,16 @@ int main(int argc, char ** argv) {
                         break;
                     }
                 }
+
+                if (is_antiprompt) {
+                    LOG("found antiprompt: %s\n", last_output.c_str());
+                }
             }
 
             // deal with end of text token in interactive mode
             if (last_n_tokens.back() == llama_token_eos(ctx)) {
+                LOG("found EOS token\n");
+
                 if (params.interactive) {
                     if (params.antiprompt.size() != 0) {
                         // tokenize and inject first reverse prompt
@@ -792,16 +852,20 @@ int main(int argc, char ** argv) {
             }
 
             if (n_past > 0 && is_interacting) {
+                LOG("waiting for user input\n");
+
                 if (params.instruct) {
                     printf("\n> ");
                 }
 
                 if (params.input_prefix_bos) {
+                    LOG("adding input prefix BOS token\n");
                     embd_inp.push_back(llama_token_bos(ctx));
                 }
 
                 std::string buffer;
                 if (!params.input_prefix.empty()) {
+                    LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
                     buffer += params.input_prefix;
                     printf("%s", buffer.c_str());
                 }
@@ -821,23 +885,30 @@ int main(int argc, char ** argv) {
                 if (buffer.length() > 1) {
                     // append input suffix if any
                     if (!params.input_suffix.empty()) {
+                        LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
                         buffer += params.input_suffix;
                         printf("%s", params.input_suffix.c_str());
                     }
 
+                    LOG("buffer: '%s'\n", buffer.c_str());
+
                     const size_t original_size = embd_inp.size();
 
                     // instruct mode: insert instruction prefix
                     if (params.instruct && !is_antiprompt) {
+                        LOG("inserting instruction prefix\n");
                         n_consumed = embd_inp.size();
                         embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
                     }
 
-                    auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
+
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
 
                     // instruct mode: insert response suffix
                     if (params.instruct) {
+                        LOG("inserting instruction suffix\n");
                         embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
                     }
 
@@ -848,6 +919,9 @@ int main(int argc, char ** argv) {
                     }
 
                     n_remain -= line_inp.size();
+                    LOG("n_remain: %d\n", n_remain);
+                } else {
+                    LOG("empty line, passing control back\n");
                 }
 
                 input_echo = false; // do not echo this again
@@ -871,7 +945,7 @@ int main(int argc, char ** argv) {
 
         // end of text token
         if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || params.interactive)) {
-            fprintf(stderr, " [end of text]\n");
+            LOG_TEE(" [end of text]\n");
             break;
         }
 
@@ -884,7 +958,7 @@ int main(int argc, char ** argv) {
     }
 
     if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
-        fprintf(stderr, "\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
+        LOG_TEE("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
         llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
     }
 
@@ -900,5 +974,9 @@ int main(int argc, char ** argv) {
     }
     llama_backend_free();
 
+#ifndef LOG_DISABLE_LOGS
+    LOG_TEE("Log end\n")
+#endif // LOG_DISABLE_LOGS
+
     return 0;
 }

From ad9ddcff6ef322db5cf13785bd7c856b610d242e Mon Sep 17 00:00:00 2001
From: chaihahaha <chai836275709@gmail.com>
Date: Wed, 30 Aug 2023 14:50:55 +0800
Subject: [PATCH 14/21] llm.vim : stop generation at multiple linebreaks, bind
 to <F2> (#2879)

---
 examples/llm.vim | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/examples/llm.vim b/examples/llm.vim
index 594a28549..d580a3d00 100644
--- a/examples/llm.vim
+++ b/examples/llm.vim
@@ -8,7 +8,7 @@ function! Llm()
   let buffer_content = join(getline(1, '$'), "\n")
 
   " Create the JSON payload
-  let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":10,"stream": v:false}
+  let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":256,"stop": ["\n\n\n"],"stream": v:false}
   let json_payload.prompt = buffer_content
 
   " Define the curl command
@@ -25,3 +25,4 @@ function! Llm()
 endfunction
 
 command! Llm call Llm()
+noremap <F2> :Llm<CR>

From dc07dc492ef9640bbb82904d7c7679f7bdcf6d76 Mon Sep 17 00:00:00 2001
From: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
Date: Wed, 30 Aug 2023 02:25:50 -0600
Subject: [PATCH 15/21] convert : various script cleanups/fixes + merges and
 special token handling (#2842)

* convert: Fix permute calls and method/func definitions

* Cleanups for gguf-py

* Minor types cleanups.

* Initial implementation of handling merges and special tokens

* convert: Handle special tokens and merges in vocab only mode

convert: Vocab only mode no longer requires loading model tensors

* gguf: Refactor tensor name mapping

* convert: Fix type hint for special_token_types in SpecialVocab

* Use common special vocab handling in various conversion scripts

* First pass at implementing suggested changes

* Second pass

* gguf: SpecialVocab: Fix issue with special token content not in a dict

gguf: SpecialVocab: Allow skipping handling of merges

* convert-falcon-hf-to-gguf: Support --vocab-only option, bail out if no tokenizer.json

* convert-gptneox-hf-to-gguf and convert: Only handle merges for BPE tokenizer

* gguf: SpecialVocab: Actually set load_merges in object

* Uniform args parsing and vocab only mode for convert examples

* convert.py: Set gpt2 as tokenizer model when using BPE

* Squish last type warning in gguf.py - yay!
---
 convert-falcon-hf-to-gguf.py    | 170 +++++-----
 convert-gptneox-hf-to-gguf.py   | 175 ++++------
 convert-llama-7b-pth-to-gguf.py | 200 +++++-------
 convert-llama-ggmlv3-to-gguf.py |  28 +-
 convert-llama-hf-to-gguf.py     | 203 +++++-------
 convert-lora-to-ggml.py         |   6 +-
 convert.py                      | 142 ++++----
 gguf-py/gguf/gguf.py            | 551 +++++++++++++++++++-------------
 gguf-py/gguf/py.typed           |   0
 gguf-py/pyproject.toml          |   1 +
 10 files changed, 728 insertions(+), 748 deletions(-)
 create mode 100644 gguf-py/gguf/py.typed

diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py
index 168bcf17f..0fdea70e1 100755
--- a/convert-falcon-hf-to-gguf.py
+++ b/convert-falcon-hf-to-gguf.py
@@ -8,6 +8,7 @@ import struct
 import json
 import numpy as np
 import torch
+import argparse
 
 from typing import Any, List
 from pathlib import Path
@@ -32,11 +33,10 @@ def bytes_to_unicode():
             bs.append(b)
             cs.append(2**8+n)
             n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
+    return dict(zip(bs, (chr(n) for n in cs)))
 
 
-def count_model_parts(dir_model: str) -> int:
+def count_model_parts(dir_model: Path) -> int:
     num_parts = 0
     for filename in os.listdir(dir_model):
         if filename.startswith("pytorch_model-"):
@@ -47,17 +47,22 @@ def count_model_parts(dir_model: str) -> int:
     return num_parts
 
 
-if len(sys.argv) < 3:
-    print(f"Usage: python {sys.argv[0]} dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file")
+    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
+    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
+    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.bin)")
+    parser.add_argument("ftype",     type=int, choices=[0, 1],   help="output format - use 0 for float32, 1 for float16", default = 1)
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
     sys.exit(1)
 
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
 # possible tensor data types
 #   ftype == 0 -> float32
 #   ftype == 1 -> float16
@@ -65,25 +70,21 @@ last_dir = os.path.basename(os.path.normpath(dir_model))
 # map from ftype to string
 ftype_str = ["f32", "f16"]
 
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
 
-        sys.exit(1)
+print("gguf: loading model "+dir_model.name)
 
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 
 if hparams["architectures"][0] != "RWForCausalLM":
     print("Model architecture not supported: " + hparams["architectures"][0])
 
-    sys.exit()
+    sys.exit(1)
 
 # get number of model parts
 num_parts = count_model_parts(dir_model)
@@ -113,77 +114,58 @@ gguf_writer.add_file_type(ftype)
 
 print("gguf: get tokenizer metadata")
 
-tokens: List[str] = []
+tokens: List[bytearray] = []
 scores: List[float] = []
 toktypes: List[int] = []
-merges: List[str] = []
 
+tokenizer_json_file = dir_model / 'tokenizer.json'
+if not tokenizer_json_file.is_file():
+    print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr)
+    sys.exit(1)
 
-if Path(dir_model + "/tokenizer.json").is_file():
-    # gpt2 tokenizer
-    gguf_writer.add_tokenizer_model("gpt2")
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
 
-    print("gguf: get gpt2 tokenizer merges")
+with open(tokenizer_json_file, "r", encoding="utf-8") as f:
+    tokenizer_json = json.load(f)
 
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer_json = json.load(f)
-    merges = tokenizer_json["model"]["merges"]
+print("gguf: get gpt2 tokenizer vocab")
 
-    gguf_writer.add_token_merges(merges)
+vocab_size = len(tokenizer_json["model"]["vocab"])
 
-    print("gguf: get gpt2 tokenizer vocab")
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
 
-    vocab_size = len(tokenizer_json["model"]["vocab"])
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v: k for k, v in byte_encoder.items()}
 
-    # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-    tokenizer = AutoTokenizer.from_pretrained(dir_model)
+for i in range(vocab_size):
+    if i in reverse_vocab:
+        try:
+            text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+        except KeyError:
+            text = bytearray()
+            for c in reverse_vocab[i]:
+                if ord(c) < 256:  # single byte character
+                    text.append(byte_decoder[ord(c)])
+                else:  # multibyte special token character
+                    text.extend(c.encode('utf-8'))
+    else:
+        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
+        pad_token = f"[PAD{i}]".encode("utf8")
+        text = bytearray(pad_token)
 
-    reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-    byte_encoder = bytes_to_unicode()
-    byte_decoder = {v: k for k, v in byte_encoder.items()}
+    tokens.append(text)
+    scores.append(0.0)                      # dymmy
+    toktypes.append(gguf.TokenType.NORMAL)  # dummy
 
-    for i in range(vocab_size):
-        if i in reverse_vocab:
-            try:
-                text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
-            except KeyError:
-                text = bytearray()
-                for c in reverse_vocab[i]:
-                    if ord(c) < 256:  # single byte character
-                        text.append(byte_decoder[ord(c)])
-                    else:  # multibyte special token character
-                        text.extend(c.encode('utf-8'))
-        else:
-            print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
-            pad_token = f"[PAD{i}]".encode("utf8")
-            text = bytearray(pad_token)
-
-        tokens.append(text)
-        scores.append(0.0)                      # dymmy
-        toktypes.append(gguf.TokenType.NORMAL)  # dummy
-
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
-
-print("gguf: get special token ids")
-# Look for special tokens in config.json
-
-if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-    gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-    gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-    gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-    gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-    gguf_writer.add_pad_token_id(hparams["pad_token_id"])
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
 
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS
 
@@ -199,15 +181,17 @@ head_dim = hparams["hidden_size"] // n_head
 print("gguf: get tensor metadata")
 
 if num_parts == 0:
-    part_names = ("pytorch_model.bin",)
+    part_names = iter(("pytorch_model.bin",))
 else:
     part_names = (
         f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
     )
 
 for part_name in part_names:
+    if args.vocab_only:
+        break
     print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
 
     for name in model_part.keys():
         data = model_part[name]
@@ -238,11 +222,8 @@ for part_name in part_names:
         data = data.squeeze().numpy()
 
         # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
             print("Can not map tensor '" + name + "'")
             sys.exit()
 
@@ -261,19 +242,20 @@ for part_name in part_names:
         if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
             data = data.astype(np.float16)
 
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
 
-        gguf_writer.add_tensor(name, data)
+        gguf_writer.add_tensor(new_name, data)
 
 
 print("gguf: write header")
 gguf_writer.write_header_to_file()
 print("gguf: write metadata")
 gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
 
 gguf_writer.close()
 
-print("gguf: model successfully exported to '" + fname_out + "'")
+print(f"gguf: model successfully exported to '{fname_out}'")
 print("")
diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py
index d9c42d76b..38e71e03b 100755
--- a/convert-gptneox-hf-to-gguf.py
+++ b/convert-gptneox-hf-to-gguf.py
@@ -8,6 +8,7 @@ import struct
 import json
 import numpy as np
 import torch
+import argparse
 
 from typing import Any, List
 from pathlib import Path
@@ -34,11 +35,10 @@ def bytes_to_unicode():
             bs.append(b)
             cs.append(2**8+n)
             n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
+    return dict(zip(bs, (chr(n) for n in cs)))
 
 
-def count_model_parts(dir_model: str) -> int:
+def count_model_parts(dir_model: Path) -> int:
     num_parts = 0
     for filename in os.listdir(dir_model):
         if filename.startswith("pytorch_model-"):
@@ -49,17 +49,22 @@ def count_model_parts(dir_model: str) -> int:
     return num_parts
 
 
-if len(sys.argv) < 3:
-    print(f"Usage: python {sys.argv[0]} dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file")
+    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
+    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
+    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.bin)")
+    parser.add_argument("ftype",     type=int, choices=[0, 1],   help="output format - use 0 for float32, 1 for float16", default = 1)
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
     sys.exit(1)
 
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
 # possible tensor data types
 #   ftype == 0 -> float32
 #   ftype == 1 -> float16
@@ -67,19 +72,15 @@ last_dir = os.path.basename(os.path.normpath(dir_model))
 # map from ftype to string
 ftype_str = ["f32", "f16"]
 
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
 
-        sys.exit(1)
+print("gguf: loading model "+dir_model.name)
 
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 
 if hparams["architectures"][0] != "GPTNeoXForCausalLM":
@@ -97,7 +98,7 @@ print("gguf: get model metadata")
 
 block_count = hparams["num_hidden_layers"]
 
-gguf_writer.add_name(last_dir)
+gguf_writer.add_name(dir_model.name)
 gguf_writer.add_context_length(hparams["max_position_embeddings"])
 gguf_writer.add_embedding_length(hparams["hidden_size"])
 gguf_writer.add_block_count(block_count)
@@ -111,86 +112,52 @@ gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"])
 
 print("gguf: get tokenizer metadata")
 
-tokens: List[str] = []
-merges: List[str] = []
+tokens: List[bytearray] = []
 
+tokenizer_json_file = dir_model / 'tokenizer.json'
+if not tokenizer_json_file.is_file():
+    print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr)
+    sys.exit(1)
 
-if Path(dir_model + "/tokenizer.json").is_file():
-    # gpt2 tokenizer
-    gguf_writer.add_tokenizer_model("gpt2")
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
 
-    print("gguf: get gpt2 tokenizer merges")
+with open(tokenizer_json_file, "r", encoding="utf-8") as f:
+    tokenizer_json = json.load(f)
 
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer_json = json.load(f)
-    merges = tokenizer_json["model"]["merges"]
+print("gguf: get gpt2 tokenizer vocab")
 
-    gguf_writer.add_token_merges(merges)
+vocab_size = len(tokenizer_json["model"]["vocab"])
 
-    print("gguf: get gpt2 tokenizer vocab")
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
 
-    vocab_size = len(tokenizer_json["model"]["vocab"])
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v: k for k, v in byte_encoder.items()}
 
-    # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-    tokenizer = AutoTokenizer.from_pretrained(dir_model)
+for i in range(vocab_size):
+    if i in reverse_vocab:
+        try:
+            text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+        except KeyError:
+            text = bytearray()
+            for c in reverse_vocab[i]:
+                if ord(c) < 256:  # single byte character
+                    text.append(byte_decoder[ord(c)])
+                else:  # multibyte special token character
+                    text.extend(c.encode('utf-8'))
+    else:
+        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
+        pad_token = f"[PAD{i}]".encode("utf8")
+        text = bytearray(pad_token)
 
-    reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-    byte_encoder = bytes_to_unicode()
-    byte_decoder = {v: k for k, v in byte_encoder.items()}
+    tokens.append(text)
 
-    for i in range(vocab_size):
-        if i in reverse_vocab:
-            try:
-                text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
-            except KeyError:
-                text = bytearray()
-                for c in reverse_vocab[i]:
-                    if ord(c) < 256:  # single byte character
-                        text.append(byte_decoder[ord(c)])
-                    else:  # multibyte special token character
-                        text.extend(c.encode('utf-8'))
-        else:
-            print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
-            pad_token = f"[PAD{i}]".encode("utf8")
-            text = bytearray(pad_token)
-
-        tokens.append(text)
-
-    gguf_writer.add_token_list(tokens)
-
-    if "added_tokens" in tokenizer_json and Path(dir_model + "/tokenizer_config.json").is_file():
-        print("gguf: get special token ids")
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        # find special token ids
-
-        if "bos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]:
-                    gguf_writer.add_pad_token_id(key["id"])
+gguf_writer.add_token_list(tokens)
 
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS
 
@@ -200,13 +167,15 @@ tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
 print("gguf: get tensor metadata")
 
 if num_parts == 0:
-    part_names = ("pytorch_model.bin",)
+    part_names = iter(("pytorch_model.bin",))
 else:
     part_names = (
         f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
     )
 
 for part_name in part_names:
+    if args.vocab_only:
+        break
     print("gguf: loading model part '" + part_name + "'")
     model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
 
@@ -226,11 +195,8 @@ for part_name in part_names:
         data = data.squeeze().numpy()
 
         # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
             print("Can not map tensor '" + name + "'")
             sys.exit()
 
@@ -249,19 +215,20 @@ for part_name in part_names:
         if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
             data = data.astype(np.float16)
 
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
 
-        gguf_writer.add_tensor(name, data)
+        gguf_writer.add_tensor(new_name, data)
 
 
 print("gguf: write header")
 gguf_writer.write_header_to_file()
 print("gguf: write metadata")
 gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
 
 gguf_writer.close()
 
-print("gguf: model successfully exported to '" + fname_out + "'")
+print(f"gguf: model successfully exported to '{fname_out}'")
 print("")
diff --git a/convert-llama-7b-pth-to-gguf.py b/convert-llama-7b-pth-to-gguf.py
index 2ab082383..6e973a116 100755
--- a/convert-llama-7b-pth-to-gguf.py
+++ b/convert-llama-7b-pth-to-gguf.py
@@ -10,8 +10,9 @@ import struct
 import json
 import numpy as np
 import torch
+import argparse
 
-from typing import Any, List
+from typing import Any, List, TypeAlias
 from pathlib import Path
 from sentencepiece import SentencePieceProcessor
 
@@ -20,7 +21,7 @@ from sentencepiece import SentencePieceProcessor
 NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
 
 
-def count_model_parts(dir_model: str) -> int:
+def count_model_parts(dir_model: Path) -> int:
     num_parts = 0
     for filename in os.listdir(dir_model):
         if filename.startswith("consolidated."):
@@ -31,19 +32,22 @@ def count_model_parts(dir_model: str) -> int:
     return num_parts
 
 
-if len(sys.argv) < 3:
-    print(f"Usage: python {sys.argv[0]} dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a PyTorch 7B LLaMA model to a GGML compatible file")
+    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
+    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
+    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.bin)")
+    parser.add_argument("ftype",     type=int, choices=[0, 1],   help="output format - use 0 for float32, 1 for float16", default = 1)
+    return parser.parse_args()
 
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
     sys.exit(1)
 
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
-
 # possible tensor data types
 #   ftype == 0 -> float32
 #   ftype == 1 -> float16
@@ -51,19 +55,15 @@ last_dir = os.path.basename(os.path.normpath(dir_model))
 # map from ftype to string
 ftype_str = ["f32", "f16"]
 
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
 
-        sys.exit(1)
+print("gguf: loading model "+dir_model.name)
 
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 
 if hparams["architectures"][0] != "LlamaForCausalLM":
@@ -107,7 +107,7 @@ else:
     sys.exit()
 
 
-gguf_writer.add_name(last_dir)
+gguf_writer.add_name(dir_model.name)
 gguf_writer.add_source_hf_repo(hf_repo)
 gguf_writer.add_tensor_data_layout("Meta AI original pth")
 gguf_writer.add_context_length(ctx_length)
@@ -133,109 +133,60 @@ tokens: List[bytes] = []
 scores: List[float] = []
 toktypes: List[int] = []
 
-if Path(dir_model + "/tokenizer.model").is_file():
-    # vocab type sentencepiece
-    print("gguf: get sentencepiece tokenizer vocab and scores")
+tokenizer_model_file = dir_model / 'tokenizer.model'
+if not tokenizer_model_file.is_file():
+    print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr)
+    sys.exit(1)
 
-    tokenizer = SentencePieceProcessor(dir_model + "/tokenizer.model")
+# vocab type sentencepiece
+print("gguf: get sentencepiece tokenizer vocab and scores")
 
-    for i in range(tokenizer.vocab_size()):
-        text: bytes
-        score: float
+tokenizer = SentencePieceProcessor(str(tokenizer_model_file))
 
-        piece = tokenizer.id_to_piece(i)
-        text = piece.encode("utf-8")
-        score = tokenizer.get_score(i)
+for i in range(tokenizer.vocab_size()):
+    text: bytes
+    score: float
 
-        toktype = 1  # defualt to normal token type
-        if tokenizer.is_unknown(i):
-            toktype = 2
-        if tokenizer.is_control(i):
-            toktype = 3
+    piece = tokenizer.id_to_piece(i)
+    text = piece.encode("utf-8")
+    score = tokenizer.get_score(i)
 
-        # toktype = 4 is user-defined = tokens from added_tokens.json
+    toktype = 1  # defualt to normal token type
+    if tokenizer.is_unknown(i):
+        toktype = 2
+    if tokenizer.is_control(i):
+        toktype = 3
 
-        if tokenizer.is_unused(i):
-            toktype = 5
-        if tokenizer.is_byte(i):
-            toktype = 6
+    # toktype = 4 is user-defined = tokens from added_tokens.json
 
-        tokens.append(text)
-        scores.append(score)
-        toktypes.append(toktype)
+    if tokenizer.is_unused(i):
+        toktype = 5
+    if tokenizer.is_byte(i):
+        toktype = 6
 
-    if Path(dir_model + "/added_tokens.json").is_file():
-        with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
-            addtokens_json = json.load(f)
+    tokens.append(text)
+    scores.append(score)
+    toktypes.append(toktype)
 
-            print("gguf: get added tokens")
+added_tokens_file = dir_model / 'added_tokens.json'
+if added_tokens_file.is_file():
+    with open(added_tokens_file, "r", encoding="utf-8") as f:
+        addtokens_json = json.load(f)
 
-            for key in addtokens_json:
-                tokens.append( key.encode("utf-8") )
-                scores.append(-1000.0)
-                toktypes.append(4) # user-defined token type
+        print("gguf: get added tokens")
 
-    gguf_writer.add_tokenizer_model("llama")
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
+        for key in addtokens_json:
+            tokens.append( key.encode("utf-8") )
+            scores.append(-1000.0)
+            toktypes.append(4) # user-defined token type
 
+gguf_writer.add_tokenizer_model("llama")
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
 
-print("gguf: get special token ids")
-
-if Path(dir_model + "/tokenizer.json").is_file():
-    # Look for special tokens in tokenizer.json if it exists
-
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer = json.load(f)
-
-    if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file():
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]["content"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]["content"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]["content"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]["content"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]["content"]:
-                    gguf_writer.add_pad_token_id(key["id"])
-else:
-    # If no tokenizer.json: Look for special tokens in config.json
-
-    if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-        gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-    if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-        gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-    if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-        gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-    if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-        gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-    if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-        gguf_writer.add_pad_token_id(hparams["pad_token_id"])
-
+special_vocab = gguf.SpecialVocab(dir_model)
+special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS
 
@@ -247,6 +198,8 @@ print("gguf: get tensor metadata")
 part_names = (f"consolidated.{n:02}.pth" for n in range(0, num_parts))
 
 for part_name in part_names:
+    if args.vocab_only:
+        break
     print("gguf: loading model part '" + part_name + "'")
     model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
 
@@ -266,11 +219,8 @@ for part_name in part_names:
         data = data.squeeze().numpy()
 
         # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
             print("Can not map tensor '" + name + "'")
             sys.exit()
 
@@ -289,20 +239,20 @@ for part_name in part_names:
         if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
             data = data.astype(np.float16)
 
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
 
-        gguf_writer.add_tensor(name, data)
+        gguf_writer.add_tensor(new_name, data)
 
 
 print("gguf: write header")
 gguf_writer.write_header_to_file()
 print("gguf: write metadata")
 gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
 
 gguf_writer.close()
 
-
-print("gguf: model successfully exported to '" + fname_out + "'")
+print(f"gguf: model successfully exported to '{fname_out}'")
 print("")
diff --git a/convert-llama-ggmlv3-to-gguf.py b/convert-llama-ggmlv3-to-gguf.py
index 3bf93627d..c8e7f1761 100755
--- a/convert-llama-ggmlv3-to-gguf.py
+++ b/convert-llama-ggmlv3-to-gguf.py
@@ -75,7 +75,7 @@ class Tensor:
         self.dims = ()
         self.dtype = None
         self.start_offset = 0
-        self.len_bytes = 0
+        self.len_bytes = np.int64(0)
 
     def load(self, data, offset):
         orig_offset = offset
@@ -134,13 +134,14 @@ class GGMLV3Model:
         return offset
 
 class GGMLToGGUF:
-    def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None):
+    def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None):
         hp = ggml_model.hyperparameters
         self.model = ggml_model
         self.data = data
         self.cfg = cfg
         self.params_override = params_override
         self.vocab_override = vocab_override
+        self.special_vocab = special_vocab
         if params_override is not None:
             n_kv_head = params_override.n_head_kv
         else:
@@ -162,6 +163,8 @@ class GGMLToGGUF:
         gguf_writer = gguf.GGUFWriter(self.cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
         self.add_params(gguf_writer)
         self.add_vocab(gguf_writer)
+        if self.special_vocab is not None:
+            self.special_vocab.add_to_gguf(gguf_writer)
         self.add_tensors(gguf_writer)
         print("    gguf: write header")
         gguf_writer.write_header_to_file()
@@ -259,20 +262,13 @@ class GGMLToGGUF:
         gguf_writer.add_eos_token_id(2)
 
     def add_tensors(self, gguf_writer):
-        nm = self.name_map
+        tensor_map = self.name_map
         data = self.data
         print(f'* Adding {len(self.model.tensors)} tensor(s)')
         for tensor in self.model.tensors:
             name = str(tensor.name, 'UTF-8')
-            if name.endswith('.weight'):
-                name = name[:-7]
-                suffix = '.weight'
-            elif name.endswith('.bias'):
-                name = name[:-5]
-                suffix = '.bias'
-            mapped_name = nm.get(name)
+            mapped_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
             assert mapped_name is not None, f'Bad name {name}'
-            mapped_name += suffix
             tempdims = list(tensor.dims[:])
             if len(tempdims) > 1:
                 temp = tempdims[1]
@@ -302,8 +298,10 @@ def handle_metadata(cfg, hp):
     else:
         raise ValueError('Unable to load metadata')
     vocab = convert.load_vocab(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir, cfg.vocabtype)
+    # FIXME: Respect cfg.vocab_dir?
+    svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
     convert.check_vocab_size(params, vocab)
-    return (params, vocab)
+    return (params, vocab, svocab)
 
 def handle_args():
     parser = argparse.ArgumentParser(description = 'Convert GGMLv3 models to GGUF')
@@ -330,14 +328,16 @@ def main():
     print(f'* GGML model hyperparameters: {model.hyperparameters}')
     vocab_override = None
     params_override = None
+    special_vocab = None
     if cfg.model_metadata_dir is not None:
-        (params_override, vocab_override) = handle_metadata(cfg, model.hyperparameters)
+        (params_override, vocab_override, special_vocab) = handle_metadata(cfg, model.hyperparameters)
         print('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.')
         print(f'* Overriding params: {params_override}')
         print(f'* Overriding vocab: {vocab_override}')
+        print(f'* Special vocab: {special_vocab}')
     else:
         print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
-    converter = GGMLToGGUF(model, data, cfg, params_override = params_override, vocab_override = vocab_override)
+    converter = GGMLToGGUF(model, data, cfg, params_override = params_override, vocab_override = vocab_override, special_vocab = special_vocab)
     converter.save()
     print(f'* Successful completion. Output saved to: {cfg.output}')
 
diff --git a/convert-llama-hf-to-gguf.py b/convert-llama-hf-to-gguf.py
index b00810dbb..ab94b5eab 100755
--- a/convert-llama-hf-to-gguf.py
+++ b/convert-llama-hf-to-gguf.py
@@ -8,8 +8,9 @@ import struct
 import json
 import numpy as np
 import torch
+import argparse
 
-from typing import Any, List, Optional
+from typing import Any, List, Optional, TypeAlias
 from pathlib import Path
 from sentencepiece import SentencePieceProcessor
 
@@ -43,40 +44,38 @@ def count_model_parts(dir_model: str) -> int:
     return num_parts
 
 
-if len(sys.argv) < 3:
-    print(f"Usage: python {sys.argv[0]} dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a HuggingFace LLaMA model to a GGML compatible file")
+    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
+    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
+    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.bin)")
+    parser.add_argument("ftype",     type=int, choices=[0, 1],   help="output format - use 0 for float32, 1 for float16", default = 1)
+    return parser.parse_args()
 
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
     sys.exit(1)
 
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
-
 # possible tensor data types
 #   ftype == 0 -> float32
 #   ftype == 1 -> float16
 
-
 # map from ftype to string
 ftype_str = ["f32", "f16"]
 
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
 
-        sys.exit(1)
+print("gguf: loading model "+dir_model.name)
 
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 
 if hparams["architectures"][0] != "LlamaForCausalLM":
@@ -115,7 +114,7 @@ else:
     sys.exit()
 
 
-gguf_writer.add_name(last_dir)
+gguf_writer.add_name(dir_model.name)
 gguf_writer.add_source_hf_repo(hf_repo)
 gguf_writer.add_tensor_data_layout("Meta AI original pth")
 gguf_writer.add_context_length(ctx_length)
@@ -141,110 +140,61 @@ tokens: List[bytes] = []
 scores: List[float] = []
 toktypes: List[int] = []
 
-if Path(dir_model + "/tokenizer.model").is_file():
-    # vocab type sentencepiece
-    print("gguf: get sentencepiece tokenizer vocab, scores and token types")
+tokenizer_model_file = dir_model / 'tokenizer.model'
+if not tokenizer_model_file.is_file():
+    print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr)
+    sys.exit(1)
 
-    tokenizer = SentencePieceProcessor(dir_model + "/tokenizer.model")
+# vocab type sentencepiece
+print("gguf: get sentencepiece tokenizer vocab, scores and token types")
 
-    for i in range(tokenizer.vocab_size()):
-        text: bytes
-        score: float
+tokenizer = SentencePieceProcessor(str(tokenizer_model_file))
 
-        piece = tokenizer.id_to_piece(i)
-        text = piece.encode("utf-8")
-        score = tokenizer.get_score(i)
+for i in range(tokenizer.vocab_size()):
+    text: bytes
+    score: float
 
-        toktype = 1  # defualt to normal token type
-        if tokenizer.is_unknown(i):
-            toktype = 2
-        if tokenizer.is_control(i):
-            toktype = 3
+    piece = tokenizer.id_to_piece(i)
+    text = piece.encode("utf-8")
+    score = tokenizer.get_score(i)
 
-        # toktype = 4 is user-defined = tokens from added_tokens.json
+    toktype = 1  # defualt to normal token type
+    if tokenizer.is_unknown(i):
+        toktype = 2
+    if tokenizer.is_control(i):
+        toktype = 3
 
-        if tokenizer.is_unused(i):
-            toktype = 5
-        if tokenizer.is_byte(i):
-            toktype = 6
+    # toktype = 4 is user-defined = tokens from added_tokens.json
 
-        tokens.append(text)
-        scores.append(score)
-        toktypes.append(toktype)
+    if tokenizer.is_unused(i):
+        toktype = 5
+    if tokenizer.is_byte(i):
+        toktype = 6
 
-    if Path(dir_model + "/added_tokens.json").is_file():
-        with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
-            addtokens_json = json.load(f)
+    tokens.append(text)
+    scores.append(score)
+    toktypes.append(toktype)
 
-            print("gguf: get added tokens")
+added_tokens_file = dir_model / 'added_tokens.json'
+if added_tokens_file.is_file():
+    with open(added_tokens_file, "r", encoding="utf-8") as f:
+        addtokens_json = json.load(f)
 
-            for key in addtokens_json:
-                tokens.append( key.encode("utf-8") )
-                scores.append(-1000.0)
-                toktypes.append(4) # user-defined token type
+        print("gguf: get added tokens")
+
+        for key in addtokens_json:
+            tokens.append( key.encode("utf-8") )
+            scores.append(-1000.0)
+            toktypes.append(4) # user-defined token type
 
 
-    gguf_writer.add_tokenizer_model("llama")
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
-
-
-print("gguf: get special token ids")
-
-if Path(dir_model + "/tokenizer.json").is_file():
-    # Look for special tokens in tokenizer.json if it exists
-
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer = json.load(f)
-
-    if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file():
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]["content"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]["content"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]["content"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]["content"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]["content"]:
-                    gguf_writer.add_pad_token_id(key["id"])
-else:
-    # If no tokenizer.json: Look for special tokens in config.json
-
-    if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-        gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-    if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-        gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-    if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-        gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-    if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-        gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-    if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-        gguf_writer.add_pad_token_id(hparams["pad_token_id"])
+gguf_writer.add_tokenizer_model("llama")
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
 
+special_vocab = gguf.SpecialVocab(dir_model)
+special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS
 
@@ -254,13 +204,15 @@ tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
 print("gguf: get tensor metadata")
 
 if num_parts == 0:
-    part_names = ("pytorch_model.bin",)
+    part_names = iter(("pytorch_model.bin",))
 else:
     part_names = (
         f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
     )
 
 for part_name in part_names:
+    if args.vocab_only:
+        break
     print("gguf: loading model part '" + part_name + "'")
     model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
 
@@ -286,11 +238,8 @@ for part_name in part_names:
             data = reverse_hf_permute(data, head_count, head_count_kv)
 
         # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
             print("Can not map tensor '" + name + "'")
             sys.exit()
 
@@ -309,20 +258,20 @@ for part_name in part_names:
         if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
             data = data.astype(np.float16)
 
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
 
-        gguf_writer.add_tensor(name, data)
+        gguf_writer.add_tensor(new_name, data)
 
 
 print("gguf: write header")
 gguf_writer.write_header_to_file()
 print("gguf: write metadata")
 gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
 
 gguf_writer.close()
 
-
-print("gguf: model successfully exported to '" + fname_out + "'")
+print(f"gguf: model successfully exported to '{fname_out}'")
 print("")
diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py
index a94a7d0af..a00339b47 100755
--- a/convert-lora-to-ggml.py
+++ b/convert-lora-to-ggml.py
@@ -4,7 +4,7 @@ import os
 import re
 import struct
 import sys
-from typing import Any, Dict, Sequence, TextIO
+from typing import Any, Dict, Sequence, BinaryIO
 
 import numpy as np
 import torch
@@ -46,7 +46,7 @@ def translate_tensor_name(t: str) -> str:
         sys.exit(1)
 
 
-def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
+def write_file_header(fout: BinaryIO, params: Dict[str, Any]) -> None:
     fout.write(b"ggla"[::-1])  # magic (ggml lora)
     fout.write(struct.pack("i", 1))  # file version
     fout.write(struct.pack("i", params["r"]))
@@ -60,7 +60,7 @@ def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
 
 
 def write_tensor_header(
-    self, name: str, shape: Sequence[int], data_type: np.dtype
+    self, name: str, shape: Sequence[int], data_type: np.dtype[Any]
 ) -> None:
     sname = name.encode("utf-8")
     fout.write(
diff --git a/convert.py b/convert.py
index 3f0a1c932..448b6f0f3 100755
--- a/convert.py
+++ b/convert.py
@@ -25,7 +25,7 @@ import numpy as np
 from abc import ABCMeta, abstractmethod
 from dataclasses import dataclass
 from pathlib import Path
-from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Generator, Iterable, List, Literal, Optional, Sequence, Set, Tuple, TypeVar, Union)
+from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Generator, Iterable, List, Literal, Optional, Sequence, Set, Tuple, Type, TypeVar, Union)
 from sentencepiece import SentencePieceProcessor  # type: ignore
 
 if TYPE_CHECKING:
@@ -299,8 +299,10 @@ class Params:
             params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
         elif orig_config_path.exists():
             params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
-        else:
+        elif model_plus.format != 'none':
             params = Params.guessed(model_plus.model)
+        else:
+            raise ValueError('Cannot guess params when model format is none')
 
         params.path_model = model_plus.paths[0].parent
 
@@ -353,7 +355,7 @@ class BpeVocab:
         yield from self.added_tokens()
 
     def __repr__(self) -> str:
-        return f"BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+        return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
 class SentencePieceVocab:
@@ -416,7 +418,6 @@ class SentencePieceVocab:
 
 Vocab = Union[BpeVocab, SentencePieceVocab]
 
-
 #
 # data loading
 # TODO: reuse (probably move to gguf.py?)
@@ -439,14 +440,14 @@ class Tensor(metaclass=ABCMeta):
     @abstractmethod
     def permute(self, n_head: int, n_head_kv: int) -> 'Tensor': ...
     @abstractmethod
-    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': ...
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> 'UnquantizedTensor': ...
     @abstractmethod
     def part(self, n_part: int) -> 'UnquantizedTensor': ...
     @abstractmethod
     def to_ggml(self) -> 'GGMLCompatibleTensor': ...
 
 
-def bf16_to_fp32(bf16_arr: np.ndarray) -> np.ndarray:
+def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray:
     assert bf16_arr.dtype == np.uint16, f"Input array should be of dtype uint16, but got {bf16_arr.dtype}"
     fp32_arr = bf16_arr.astype(np.uint32) << 16
     return fp32_arr.view(np.float32)
@@ -467,9 +468,9 @@ class UnquantizedTensor(Tensor):
     def to_ggml(self) -> 'UnquantizedTensor':
         return self
 
-    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor':
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> 'UnquantizedTensor':
         r = self.ndarray.shape[0] // 3
-        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head))
+        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
 
     def part(self, n_part: int) -> 'UnquantizedTensor':
         r = self.ndarray.shape[0] // 3
@@ -531,7 +532,7 @@ LazyModel = Dict[str, LazyTensor]
 class ModelPlus:
     model: LazyModel
     paths: List[Path]  # Where this was read from.
-    format: Literal['ggml', 'torch', 'safetensors']
+    format: Literal['ggml', 'torch', 'safetensors', 'none']
     vocab: Optional[Vocab]  # For GGML models (which have vocab built in), the vocab.
 
 
@@ -597,12 +598,12 @@ def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTe
         return lazy_tensor.load().permute(n_head, n_head_kv)
     return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
-def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int) -> LazyTensor:
+def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_kv: int) -> LazyTensor:
     def load() -> Tensor:
-        return lazy_tensor.load().permute_part(n_part, n_head)
+        return lazy_tensor.load().permute_part(n_part, n_head, n_head_kv)
     s = lazy_tensor.shape.copy()
     s[0] = s[0] // 3
-    return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description)
+    return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
 def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
     def load() -> Tensor:
@@ -657,7 +658,7 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
         return LazyStorage(load=load, kind=pid[1], description=description)
 
-    # @staticmethod
+    @staticmethod
     def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,
                                # pyright: ignore[reportSelfClsParameterName]
                                requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
@@ -669,13 +670,15 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'pickled storage_offset={storage_offset} in {storage.description}'
         return LazyTensor(load, list(size), storage.kind.data_type, description)
 
-    # @staticmethod
+    @staticmethod
     def rebuild_from_type_v2(func, new_type, args, state):
         return func(*args)
 
-    CLASSES: Dict[Any, Any] = {
-        ('torch._tensor', '_rebuild_from_type_v2'): rebuild_from_type_v2,
-        ('torch._utils', '_rebuild_tensor_v2'): lazy_rebuild_tensor_v2,
+    CLASSES: Dict[Tuple[str, str], Any] = {
+        # getattr used here as a workaround for mypy not being smart enough to detrmine
+        # the staticmethods have a __func__ attribute.
+        ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
+        ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
         ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16),
         ('torch', 'HalfStorage'): LazyStorageKind(DT_F16),
         ('torch', 'FloatStorage'): LazyStorageKind(DT_F32),
@@ -751,7 +754,7 @@ def lazy_load_file(path: Path) -> ModelPlus:
 In = TypeVar('In')
 Out = TypeVar('Out')
 
-def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: Optional[int] = None, factory: Callable = ThreadPoolExecutor) -> Iterable[Out]:
+def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: Optional[int] = None, use_processpool_executor: bool = False) -> Iterable[Out]:
     '''Parallel map, but with backpressure.  If the caller doesn't call `next`
     fast enough, this will stop calling `func` at some point rather than
     letting results pile up in memory.  Specifically, there is a max of one
@@ -760,7 +763,12 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
         yield from map(func, iterable)
         # Not reached.
     iterable = iter(iterable)
-    with factory(max_workers = max_workers) as executor:
+    executor_class: Union[Type[ThreadPoolExecutor], Type[ProcessPoolExecutor]]
+    if use_processpool_executor:
+        executor_class = ProcessPoolExecutor
+    else:
+        executor_class = ThreadPoolExecutor
+    with executor_class(max_workers = max_workers) as executor:
         futures: List[concurrent.futures.Future[Out]] = []
         done = False
         for _ in range(concurrency):
@@ -838,11 +846,19 @@ class OutputFile:
             scores.append(score)
             toktypes.append(toktype)
 
-        self.gguf.add_tokenizer_model("llama")
+        if isinstance(vocab, SentencePieceVocab):
+            self.gguf.add_tokenizer_model("llama")
+        elif isinstance(vocab, BpeVocab):
+            self.gguf.add_tokenizer_model("gpt2")
+        else:
+            raise ValueError(f'Unknown vocab type: Not BpeVocab or SentencePieceVocab')
         self.gguf.add_token_list(tokens)
         self.gguf.add_token_scores(scores)
         self.gguf.add_token_types(toktypes)
 
+    def add_meta_special_vocab(self, svocab: gguf.SpecialVocab) -> None:
+        svocab.add_to_gguf(self.gguf)
+
     def add_tensor_info(self, name: str, tensor: LazyTensor) -> None:
         n_elements = int(np.prod(tensor.shape))
         raw_dtype = getattr(tensor.data_type, 'ggml_type', None)
@@ -861,7 +877,7 @@ class OutputFile:
         self.gguf.close()
 
     @staticmethod
-    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab) -> None:
+    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab) -> None:
         check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out)
@@ -869,6 +885,8 @@ class OutputFile:
         # meta data
         of.add_meta_arch(params)
         of.add_meta_vocab(vocab)
+        of.add_meta_special_vocab(svocab)
+
         of.write_meta()
 
         of.close()
@@ -887,7 +905,7 @@ class OutputFile:
         return dt.quantize(arr)
 
     @staticmethod
-    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
+    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
         check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out)
@@ -895,6 +913,7 @@ class OutputFile:
         # meta data
         of.add_meta_arch(params)
         of.add_meta_vocab(vocab)
+        of.add_meta_special_vocab(svocab)
 
         # tensor info
         for name, lazy_tensor in model.items():
@@ -906,7 +925,7 @@ class OutputFile:
         # tensor data
         ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
         if ftype == GGMLFileType.MostlyQ8_0:
-            ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, factory = ProcessPoolExecutor)
+            ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, use_processpool_executor = True)
         else:
             ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
 
@@ -939,7 +958,8 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
             for (name, tensor) in model.items()}
 
 def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
-    tmap = gguf.get_tensor_name_map(ARCH, params.n_layer)
+    tmap = gguf.TensorNameMap(ARCH, params.n_layer)
+    should_skip: Set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
     tmp = model
 
@@ -952,8 +972,8 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
            #tmp[f"model.layers.{i}.self_attn.v_proj.weight"] =              model[f"model.layers.{i}.self_attn.v_proj.weight"]
         elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
             print(f"Unpacking and permuting layer {i}")
-            tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head)
-            tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head)
+            tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head)
+            tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv)
             tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy        (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2)
             del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
         else:
@@ -961,23 +981,16 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
 
     out: LazyModel = {}
     for name, lazy_tensor in model.items():
-        name_new = name
-
-        if name in tmap:
-            name_new = tmap[name]
-        elif name.endswith(".weight") and name[:-7] in tmap:
-            name_new = tmap[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tmap:
-            name_new = tmap[name[:-5]] + ".bias"
-        else:
+        tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
+        if name_new is None:
             raise Exception(f"Unexpected tensor name: {name}")
 
-        if gguf.should_skip_tensor_TMP(ARCH, params.n_layer, name_new):
+        if tensor_type in should_skip:
             print(f"skipping tensor {name_new}")
             continue
-        else:
-            print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
-            out[name_new] = lazy_tensor
+
+        print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
+        out[name_new] = lazy_tensor
 
     return out
 
@@ -1117,8 +1130,16 @@ def main(args_in: Optional[List[str]] = None) -> None:
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
         do_dump_model(model_plus)
+        return
 
-    model_plus = load_some_model(args.model)
+    if not args.vocab_only:
+        model_plus = load_some_model(args.model)
+    else:
+        model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
+
+    if args.dump:
+        do_dump_model(model_plus)
+        return
 
     params = Params.load(model_plus)
     if params.n_ctx == -1:
@@ -1140,33 +1161,34 @@ def main(args_in: Optional[List[str]] = None) -> None:
 
     vocab: Vocab
     if args.vocab_only:
-        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
         assert args.outfile, "need --outfile if using --vocab-only"
+        # FIXME: Try to respect vocab_dir somehow?
+        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
+        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
         outfile = args.outfile
-        OutputFile.write_vocab_only(outfile, params, vocab)
+        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
         print(f"Wrote {outfile}")
+        return
+
+    if model_plus.vocab is not None and args.vocab_dir is None:
+        vocab = model_plus.vocab
     else:
-        if args.dump:
-            do_dump_model(model_plus)
-            return
+        vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
+        vocab = load_vocab(vocab_dir, args.vocabtype)
+    # FIXME: Try to respect vocab_dir somehow?
+    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
 
-        if model_plus.vocab is not None and args.vocab_dir is None:
-            vocab = model_plus.vocab
-        else:
-            vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
-            vocab = load_vocab(vocab_dir, args.vocabtype)
+    model   = model_plus.model
+    model   = convert_model_names(model, params)
+    ftype   = pick_output_type(model, args.outtype)
+    model   = convert_to_output_type(model, ftype)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype)
 
-        model   = model_plus.model
-        model   = convert_model_names(model, params)
-        ftype   = pick_output_type(model, args.outtype)
-        model   = convert_to_output_type(model, ftype)
-        outfile = args.outfile or default_outfile(model_plus.paths, ftype)
+    params.ftype = ftype
+    print(f"Writing {outfile}, format {ftype}")
 
-        params.ftype = ftype
-        print(f"Writing {outfile}, format {ftype}")
-
-        OutputFile.write_all(outfile, ftype, params, model, vocab, concurrency = args.concurrency)
-        print(f"Wrote {outfile}")
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency)
+    print(f"Wrote {outfile}")
 
 
 if __name__ == '__main__':
diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py
index 838a2c0f8..de3edbc99 100644
--- a/gguf-py/gguf/gguf.py
+++ b/gguf-py/gguf/gguf.py
@@ -4,9 +4,13 @@ import sys
 import struct
 import tempfile
 import numpy as np
+import json
+import os
+from pathlib import Path
 
 from enum import IntEnum, auto
-from typing import Any, IO, List, Optional
+from io import BufferedWriter
+from typing import Any, BinaryIO, Callable, IO, Dict, List, Optional, Sequence, Tuple, Union
 
 #
 # constants
@@ -71,35 +75,35 @@ KEY_TOKENIZER_RWKV       = "tokenizer.rwkv.world"
 
 
 class MODEL_ARCH(IntEnum):
-    LLAMA   = auto()
-    FALCON  = auto()
-    GPT2    = auto()
-    GPTJ    = auto()
-    GPTNEOX = auto()
-    MPT     = auto()
+    LLAMA  : int = auto()
+    FALCON : int = auto()
+    GPT2   : int = auto()
+    GPTJ   : int = auto()
+    GPTNEOX: int = auto()
+    MPT    : int = auto()
 
 
 class MODEL_TENSOR(IntEnum):
-    TOKEN_EMBD    = auto()
-    POS_EMBD      = auto()
-    OUTPUT        = auto()
-    OUTPUT_NORM   = auto()
-    ROPE_FREQS    = auto()
-    ATTN_Q        = auto()
-    ATTN_K        = auto()
-    ATTN_V        = auto()
-    ATTN_QKV      = auto()
-    ATTN_OUT      = auto()
-    ATTN_NORM     = auto()
-    ATTN_NORM_2   = auto()
-    ATTN_ROT_EMBD = auto()
-    FFN_GATE      = auto()
-    FFN_DOWN      = auto()
-    FFN_UP        = auto()
-    FFN_NORM      = auto()
+    TOKEN_EMBD   : int = auto()
+    POS_EMBD     : int = auto()
+    OUTPUT       : int = auto()
+    OUTPUT_NORM  : int = auto()
+    ROPE_FREQS   : int = auto()
+    ATTN_Q       : int = auto()
+    ATTN_K       : int = auto()
+    ATTN_V       : int = auto()
+    ATTN_QKV     : int = auto()
+    ATTN_OUT     : int = auto()
+    ATTN_NORM    : int = auto()
+    ATTN_NORM_2  : int = auto()
+    ATTN_ROT_EMBD: int = auto()
+    FFN_GATE     : int = auto()
+    FFN_DOWN     : int = auto()
+    FFN_UP       : int = auto()
+    FFN_NORM     : int = auto()
 
 
-MODEL_ARCH_NAMES = {
+MODEL_ARCH_NAMES: Dict[MODEL_ARCH, str] = {
     MODEL_ARCH.LLAMA:   "llama",
     MODEL_ARCH.FALCON:  "falcon",
     MODEL_ARCH.GPT2:    "gpt2",
@@ -108,7 +112,7 @@ MODEL_ARCH_NAMES = {
     MODEL_ARCH.MPT:     "mpt",
 }
 
-MODEL_TENSOR_NAMES = {
+MODEL_TENSOR_NAMES: Dict[MODEL_ARCH, Dict[MODEL_TENSOR, str]] = {
     MODEL_ARCH.LLAMA: {
         MODEL_TENSOR.TOKEN_EMBD:    "token_embd",
         MODEL_TENSOR.OUTPUT_NORM:   "output_norm",
@@ -154,7 +158,7 @@ MODEL_TENSOR_NAMES = {
 }
 
 # tensors that will not be serialized
-MODEL_TENSOR_SKIP = {
+MODEL_TENSOR_SKIP: Dict[MODEL_ARCH, List[MODEL_TENSOR]] = {
     MODEL_ARCH.LLAMA: [
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_ROT_EMBD,
@@ -162,167 +166,198 @@ MODEL_TENSOR_SKIP = {
 }
 
 
-# TODO: the following helper functions should be removed
-#       instead, get_tensor_name_map should return tuples of (name, MODEL_TENSOR)
-#       however, my Python is very bad, and I couldn't figure out how to do this, hence these functions
-# REMOVE
-def should_skip_tensor_TMP(arch: MODEL_ARCH, n_blocks: int, name: str) -> bool:
-    for skip in MODEL_TENSOR_SKIP.get(arch, []):
-        for i in range(n_blocks):
-            if name == MODEL_TENSOR_NAMES[arch][skip].format(bid=i):
-                return True
+class TensorNameMap:
+    mappings_cfg: Dict[MODEL_TENSOR, Tuple[str, ...]] = {
+        # Token embeddings
+        MODEL_TENSOR.TOKEN_EMBD: (
+            "gpt_neox.embed_in",           # gptneox
+            "transformer.wte",             # gpt2 mpt
+            "transformer.word_embeddings", # falcon
+            "model.embed_tokens",          # llama-hf
+            "tok_embeddings",              # llama-pth
+        ),
 
-    return False
+        # Position embeddings
+        MODEL_TENSOR.POS_EMBD: (
+            "transformer.wpe", # gpt2
+        ),
 
+        # Output
+        MODEL_TENSOR.OUTPUT: (
+            "embed_out", # gptneox
+            "lm_head",   # gpt2 mpt falcon llama-hf
+            "output",    # llama-pth
+        ),
 
-def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> dict:
-    tensor_map = {}
+        # Output norm
+        MODEL_TENSOR.OUTPUT_NORM: (
+            "gpt_neox.final_layer_norm", # gptneox
+            "transformer.ln_f",          # gpt2 falcon
+            "model.norm",                # llama-hf
+            "norm",                      # llama-pth
+        ),
 
-    # Token embeddings
-    mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.TOKEN_EMBD, None)
+        # Rope frequencies
+        MODEL_TENSOR.ROPE_FREQS: (
+            "rope.freqs", # llama-pth
+        ),
+    }
 
-    tensor_map["gpt_neox.embed_in"]           = mapped_to  # gptneox
-    tensor_map["transformer.wte"]             = mapped_to  # gpt2 mpt
-    tensor_map["transformer.word_embeddings"] = mapped_to  # falcon
-    tensor_map["model.embed_tokens"]          = mapped_to  # llama-hf
-    tensor_map["tok_embeddings"]              = mapped_to  # llama-pth
-
-    # Position embeddings
-    mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.POS_EMBD, None)
-
-    tensor_map["transformer.wpe"] = mapped_to  # gpt2
-
-    # Output
-    mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.OUTPUT, None)
-
-    tensor_map["embed_out"] = mapped_to  # gptneox
-    tensor_map["lm_head"]   = mapped_to  # gpt2 mpt falcon llama-hf
-    tensor_map["output"]    = mapped_to  # llama-pth
-
-    # Output norm
-    mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.OUTPUT_NORM, None)
-
-    tensor_map["gpt_neox.final_layer_norm"] = mapped_to  # gptneox
-    tensor_map["transformer.ln_f"]          = mapped_to  # gpt2 falcon
-    tensor_map["transformer.norm_f"]        = mapped_to  # mpt
-    tensor_map["model.norm"]                = mapped_to  # llama-hf
-    tensor_map["norm"]                      = mapped_to  # llama-pth
-
-    # Rope frequencies
-    mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ROPE_FREQS, None)
-
-    tensor_map["rope.freqs"] = mapped_to  # llama-pth
-
-    # Attention and feed-forward blocks
-    for i in range(0, n_blocks):
+    block_mappings_cfg: Dict[MODEL_TENSOR, Tuple[str, ...]] = {
         # Attention norm
-        # TODO: is there are simpler way to write these 2 lines in Python?
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_NORM, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".input_layernorm"] = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".ln_1"]              = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".norm_1"]       = mapped_to  # mpt
-        tensor_map["transformer.h."+str(i)+".input_layernorm"]   = mapped_to  # falcon7b
-        tensor_map["transformer.h."+str(i)+".ln_mlp"]            = mapped_to  # falcon40b
-        tensor_map["model.layers."+str(i)+".input_layernorm"]    = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention_norm"]           = mapped_to  # llama-pth
+        MODEL_TENSOR.ATTN_NORM: (
+            "gpt_neox.layers.{bid}.input_layernorm", # gptneox
+            "transformer.h.{bid}.ln_1",              # gpt2
+            "transformer.blocks.{bid}.norm_1",       # mpt
+            "transformer.h.{bid}.input_layernorm",   # falcon7b
+            "transformer.h.{bid}.ln_mlp",            # falcon40b
+            "model.layers.{bid}.input_layernorm",    # llama-hf
+            "layers.{bid}.attention_norm",           # llama-pth
+        ),
 
         # Attention norm 2
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_NORM_2, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["transformer.h."+str(i)+".ln_attn"] = mapped_to  # falcon40b
+        MODEL_TENSOR.ATTN_NORM_2: (
+            "transformer.h.{bid}.ln_attn", # falcon40b
+        ),
 
         # Attention query-key-value
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_QKV, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".attention.query_key_value"]    = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".attn.c_attn"]                    = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".attn.Wqkv"]                 = mapped_to  # mpt
-        tensor_map["transformer.h."+str(i)+".self_attention.query_key_value"] = mapped_to  # falcon
+        MODEL_TENSOR.ATTN_QKV: (
+            "gpt_neox.layers.{bid}.attention.query_key_value",    # gptneox
+            "transformer.h.{bid}.attn.c_attn",                    # gpt2
+            "transformer.blocks.{bid}.attn.Wqkv",                 # mpt
+            "transformer.h.{bid}.self_attention.query_key_value", # falcon
+        ),
 
         # Attention query
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_Q, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["model.layers."+str(i)+".self_attn.q_proj"] = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention.wq"]           = mapped_to  # llama-pth
+        MODEL_TENSOR.ATTN_Q: (
+            "model.layers.{bid}.self_attn.q_proj", # llama-hf
+            "layers.{bid}.attention.wq",           # llama-pth
+        ),
 
         # Attention key
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_K, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["model.layers."+str(i)+".self_attn.k_proj"] = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention.wk"]           = mapped_to  # llama-pth
+        MODEL_TENSOR.ATTN_K: (
+            "model.layers.{bid}.self_attn.k_proj", # llama-hf
+            "layers.{bid}.attention.wk",           # llama-pth
+        ),
 
         # Attention value
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_V, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["model.layers."+str(i)+".self_attn.v_proj"] = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention.wv"]           = mapped_to  # llama-pth
+        MODEL_TENSOR.ATTN_V: (
+            "model.layers.{bid}.self_attn.v_proj", # llama-hf
+            "layers.{bid}.attention.wv",           # llama-pth
+        ),
 
         # Attention output
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_OUT, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".attention.dense"]    = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".attn.c_proj"]          = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".attn.out_proj"]   = mapped_to  # mpt
-        tensor_map["transformer.h."+str(i)+".self_attention.dense"] = mapped_to  # falcon
-        tensor_map["model.layers."+str(i)+".self_attn.o_proj"]      = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention.wo"]                = mapped_to  # llama-pth
+        MODEL_TENSOR.ATTN_OUT: (
+            "gpt_neox.layers.{bid}.attention.dense",    # gptneox
+            "transformer.h.{bid}.attn.c_proj",          # gpt2
+            "transformer.blocks.{bid}.attn.out_proj",   # mpt
+            "transformer.h.{bid}.self_attention.dense", # falcon
+            "model.layers.{bid}.self_attn.o_proj",      # llama-hf
+            "layers.{bid}.attention.wo",                # llama-pth
+        ),
 
         # Rotary embeddings
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_ROT_EMBD, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["model.layers."+str(i)+".self_attn.rotary_emb.inv_freq"]  = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention.inner_attention.rope.freqs"] = mapped_to  # llama-pth
+        MODEL_TENSOR.ATTN_ROT_EMBD: (
+            "model.layers.{bid}.self_attn.rotary_emb.inv_freq",  # llama-hf
+            "layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth
+        ),
 
         # Feed-forward norm
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_NORM, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".post_attention_layernorm"] = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".ln_2"]                       = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".norm_2"]                = mapped_to  # mpt
-        tensor_map["model.layers."+str(i)+".post_attention_layernorm"]    = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".ffn_norm"]                          = mapped_to  # llama-pth
+        MODEL_TENSOR.FFN_NORM: (
+            "gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox
+            "transformer.h.{bid}.ln_2",                       # gpt2
+            "transformer.blocks.{bid}.norm_2",                # mpt
+            "model.layers.{bid}.post_attention_layernorm",    # llama-hf
+            "layers.{bid}.ffn_norm",                          # llama-pth
+        ),
 
         # Feed-forward up
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_UP, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".mlp.dense_h_to_4h"] = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".mlp.c_fc"]            = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".ffn.up_proj"]    = mapped_to  # mpt
-        tensor_map["transformer.h."+str(i)+".mlp.dense_h_to_4h"]   = mapped_to  # falcon
-        tensor_map["model.layers."+str(i)+".mlp.up_proj"]          = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".feed_forward.w3"]            = mapped_to  # llama-pth
+        MODEL_TENSOR.FFN_UP: (
+            "gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox
+            "transformer.h.{bid}.mlp.c_fc",            # gpt2
+            "transformer.blocks.{bid}.ffn.up_proj",    # mpt
+            "transformer.h.{bid}.mlp.dense_h_to_4h",   # falcon
+            "model.layers.{bid}.mlp.up_proj",          # llama-hf
+            "layers.{bid}.feed_forward.w3",            # llama-pth
+        ),
 
         # Feed-forward gate
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_GATE, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["model.layers."+str(i)+".mlp.gate_proj"] = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".feed_forward.w1"]     = mapped_to  # llama-pth
+        MODEL_TENSOR.FFN_GATE: (
+            "model.layers.{bid}.mlp.gate_proj", # llama-hf
+            "layers.{bid}.feed_forward.w1",     # llama-pth
+        ),
 
         # Feed-forward down
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_DOWN, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
+        MODEL_TENSOR.FFN_DOWN: (
+            "gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox
+            "transformer.h.{bid}.mlp.c_proj",          # gpt2
+            "transformer.blocks.{bid}.ffn.down_proj",  # mpt
+            "transformer.h.{bid}.mlp.dense_4h_to_h",   # falcon
+            "model.layers.{bid}.mlp.down_proj",        # llama-hf
+            "layers.{bid}.feed_forward.w2",            # llama-pth
+        ),
+    }
 
-        tensor_map["gpt_neox.layers."+str(i)+".mlp.dense_4h_to_h"] = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".mlp.c_proj"]          = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".ffn.down_proj"]  = mapped_to  # mpt
-        tensor_map["transformer.h."+str(i)+".mlp.dense_4h_to_h"]   = mapped_to  # falcon
-        tensor_map["model.layers."+str(i)+".mlp.down_proj"]        = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".feed_forward.w2"]            = mapped_to  # llama-pth
+    mapping: Dict[str, Tuple[MODEL_TENSOR, str]]
 
-    return tensor_map
+    tensor_names: Dict[MODEL_TENSOR, str]
 
+    def __init__(self, arch: MODEL_ARCH, n_blocks: int):
+        mapping = self.mapping = {}
+        tensor_names = self.tensor_names = MODEL_TENSOR_NAMES[arch]
+        for tensor, keys in self.mappings_cfg.items():
+            tensor_name = tensor_names.get(tensor)
+            if tensor_name is None:
+                continue
+            for key in keys:
+                mapping[key] = (tensor, tensor_name)
+        for bid in range(n_blocks):
+            for tensor, keys in self.block_mappings_cfg.items():
+                tensor_name = tensor_names.get(tensor)
+                if tensor_name is None:
+                    continue
+                tensor_name = tensor_name.format(bid = bid)
+                for key in keys:
+                    key = key.format(bid = bid)
+                    mapping[key] = (tensor, tensor_name)
+
+    def get_type_and_name(self, key: str, try_suffixes: Sequence[str]) -> Optional[Tuple[MODEL_TENSOR, str]]:
+        result = self.mapping.get(key)
+        if result is not None:
+            return result
+        for suffix in try_suffixes:
+            if key.endswith(suffix):
+                result = self.mapping.get(key[:-len(suffix)])
+                if result is not None:
+                    return (result[0], result[1] + suffix)
+        return None
+
+    def get_name(self, key: str, try_suffixes: Sequence[str]) -> Optional[str]:
+        result = self.get_type_and_name(key, try_suffixes = try_suffixes)
+        if result is None:
+            return None
+        return result[1]
+
+    def get_type(self, key: str, try_suffixes: Sequence[str]) -> Optional[MODEL_TENSOR]:
+        result = self.get_type_and_name(key, try_suffixes = try_suffixes)
+        if result is None:
+            return None
+        return result[0]
+
+    def __getitem__(self, key: str) -> str:
+        try:
+            return self.mapping[key][1]
+        except KeyError:
+            raise KeyError(key)
+
+    def __contains__(self, key: str) -> bool:
+        return key in self.mapping
+
+    def __repr__(self) -> str:
+        return repr(self.mapping)
+
+def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> TensorNameMap:
+    return TensorNameMap(arch, n_blocks)
 
 class TokenType(IntEnum):
     NORMAL       = 1
@@ -388,15 +423,21 @@ class GGUFValueType(IntEnum):
 
 
 class GGUFWriter:
-    def __init__(self, path: str, arch: str, use_temp_file = True):
+    fout: BufferedWriter
+    arch: str
+    offset_tensor = 0
+    data_alignment = GGUF_DEFAULT_ALIGNMENT
+    kv_data = b""
+    kv_data_count = 0
+    ti_data = b""
+    ti_data_count = 0
+    use_temp_file: bool
+    temp_file: Optional[tempfile.SpooledTemporaryFile[bytes]] = None
+    tensors: List[Tuple[np.ndarray[Any, Any], int]]
+
+    def __init__(self, path: Union[os.PathLike[str], str], arch: str, use_temp_file = True):
         self.fout = open(path, "wb")
         self.arch = arch
-        self.offset_tensor = 0
-        self.data_alignment = GGUF_DEFAULT_ALIGNMENT
-        self.kv_data = b""
-        self.kv_data_count = 0
-        self.ti_data = b""
-        self.ti_data_count = 0
         self.add_architecture()
         self.use_temp_file = use_temp_file
         self.tensors = []
@@ -470,14 +511,27 @@ class GGUFWriter:
         self.add_key(key)
         self.add_val(val, GGUFValueType.STRING)
 
-    def add_array(self, key: str, val: list):
-        if not isinstance(val, list):
-            raise ValueError("Value must be a list for array type")
+    def add_array(self, key: str, val: Sequence[Any]):
+        if not isinstance(val, Sequence):
+            raise ValueError("Value must be a sequence for array type")
 
         self.add_key(key)
         self.add_val(val, GGUFValueType.ARRAY)
 
-    def add_val(self: str, val: Any, vtype: GGUFValueType = None, add_vtype: bool = True):
+    _simple_value_packing = {
+        GGUFValueType.UINT8:   "<B",
+        GGUFValueType.INT8:    "<b",
+        GGUFValueType.UINT16:  "<H",
+        GGUFValueType.INT16:   "<h",
+        GGUFValueType.UINT32:  "<I",
+        GGUFValueType.INT32:   "<i",
+        GGUFValueType.FLOAT32: "<f",
+        GGUFValueType.UINT64:  "<Q",
+        GGUFValueType.INT64:   "<q",
+        GGUFValueType.FLOAT64: "<d",
+        GGUFValueType.BOOL:    "?" ,
+    }
+    def add_val(self, val: Any, vtype: Optional[GGUFValueType] = None, add_vtype: bool = True):
         if vtype is None:
             vtype = GGUFValueType.get_type(val)
 
@@ -485,47 +539,29 @@ class GGUFWriter:
             self.kv_data += struct.pack("<I", vtype)
             self.kv_data_count += 1
 
-        if vtype == GGUFValueType.UINT8:
-            self.kv_data += struct.pack("<B", val)
-        elif vtype == GGUFValueType.INT8:
-            self.kv_data += struct.pack("<b", val)
-        elif vtype == GGUFValueType.UINT16:
-            self.kv_data += struct.pack("<H", val)
-        elif vtype == GGUFValueType.INT16:
-            self.kv_data += struct.pack("<h", val)
-        elif vtype == GGUFValueType.UINT32:
-            self.kv_data += struct.pack("<I", val)
-        elif vtype == GGUFValueType.INT32:
-            self.kv_data += struct.pack("<i", val)
-        elif vtype == GGUFValueType.FLOAT32:
-            self.kv_data += struct.pack("<f", val)
-        elif vtype == GGUFValueType.UINT64:
-            self.kv_data += struct.pack("<Q", val)
-        elif vtype == GGUFValueType.INT64:
-            self.kv_data += struct.pack("<q", val)
-        elif vtype == GGUFValueType.FLOAT64:
-            self.kv_data += struct.pack("<d", val)
-        elif vtype == GGUFValueType.BOOL:
-            self.kv_data += struct.pack("?", val)
+        pack_fmt = self._simple_value_packing.get(vtype)
+        if pack_fmt is not None:
+            self.kv_data += struct.pack(pack_fmt, val)
         elif vtype == GGUFValueType.STRING:
             encoded_val = val.encode("utf8") if isinstance(val, str) else val
             self.kv_data += struct.pack("<Q", len(encoded_val))
             self.kv_data += encoded_val
-        elif vtype == GGUFValueType.ARRAY:
-            ltype = set([GGUFValueType.get_type(item) for item in val])
-            assert len(ltype) == 1, "All items in a GGUF array should be of the same type"
-            self.kv_data += struct.pack("<I", list(ltype)[0])
+        elif vtype == GGUFValueType.ARRAY and isinstance(val, Sequence) and len(val) > 0:
+            ltype = GGUFValueType.get_type(val[0])
+            if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]):
+                raise ValueError("All items in a GGUF array should be of the same type")
+            self.kv_data += struct.pack("<I", ltype)
             self.kv_data += struct.pack("<Q", len(val))
             for item in val:
                 self.add_val(item, add_vtype=False)
         else:
-            raise ValueError("Invalid GGUF metadata value type")
+            raise ValueError("Invalid GGUF metadata value type or value")
 
     @staticmethod
     def ggml_pad(x: int, n: int) -> int:
         return ((x + n - 1) // n) * n
 
-    def add_tensor_info(self, name: str, tensor_shape: np.ndarray, tensor_dtype: np.dtype, tensor_nbytes: int, raw_dtype: Optional[GGMLQuantizationType] = None):
+    def add_tensor_info(self, name: str, tensor_shape: Sequence[int], tensor_dtype: Union[np.dtype[np.float16], np.dtype[np.float32]], tensor_nbytes: int, raw_dtype: Optional[GGMLQuantizationType] = None):
         assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now"
 
         encoded_name = name.encode("utf8")
@@ -544,16 +580,18 @@ class GGUFWriter:
         self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment)
         self.ti_data_count += 1
 
-    def add_tensor(self, name: str, tensor: np.ndarray, raw_shape: Optional[np.ndarray] = None, raw_dtype: Optional[GGMLQuantizationType] = None):
-        if self.use_temp_file and not hasattr(self, "temp_file"):
-            self.temp_file = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256*1024*1024)
-            self.temp_file.seek(0)
+    def add_tensor(self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Optional[Sequence[int]] = None, raw_dtype: Optional[GGMLQuantizationType] = None):
+        if self.use_temp_file and self.temp_file is None:
+            fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256*1024*1024)
+            fp.seek(0)
+            self.temp_file = fp
 
-        self.add_tensor_info(name, raw_shape if raw_shape is not None else tensor.shape, tensor.dtype, tensor.nbytes, raw_dtype = raw_dtype)
+        shape: Sequence[int] = raw_shape if raw_shape is not None else tensor.shape
+        self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype = raw_dtype)
 
         pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes
 
-        if not self.use_temp_file:
+        if  self.temp_file is None:
             self.tensors.append((tensor, pad))
             return
 
@@ -562,25 +600,22 @@ class GGUFWriter:
         if pad != 0:
             self.temp_file.write(bytes([0] * pad))
 
-    def write_tensor_data(self, tensor: np.ndarray):
-        pad = GGUFWriter.ggml_pad(self.fout.tell(), self.data_alignment) - self.fout.tell()
+    def write_padding(self, fp: BinaryIO, n: int, align: Optional[int] = None):
+        pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n
         if pad != 0:
-            self.fout.write(bytes([0] * pad))
+            fp.write(bytes([0] * pad))
 
+    def write_tensor_data(self, tensor: np.ndarray[Any, Any]):
+        self.write_padding(self.fout, self.fout.tell())
         tensor.tofile(self.fout)
-
-        pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes
-        if pad != 0:
-            self.fout.write(bytes([0] * pad))
+        self.write_padding(self.fout, tensor.nbytes)
 
     def write_tensors_to_file(self):
         self.write_ti_data_to_file()
 
-        pad = GGUFWriter.ggml_pad(self.fout.tell(), self.data_alignment) - self.fout.tell()
-        if pad != 0:
-            self.fout.write(bytes([0] * pad))
+        self.write_padding(self.fout, self.fout.tell())
 
-        if not self.use_temp_file:
+        if self.temp_file is None:
             for (currtensor, currpad) in self.tensors:
                 currtensor.tofile(self.fout)
                 if currpad != 0:
@@ -654,10 +689,6 @@ class GGUFWriter:
         self.add_bool(
             KEY_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
 
-    def add_tensor_data_layout(self, layout: str):
-        self.add_string(
-            KEY_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
-
     def add_head_count(self, count: int):
         self.add_uint32(
             KEY_ATTENTION_HEAD_COUNT.format(arch=self.arch), count)
@@ -695,16 +726,16 @@ class GGUFWriter:
     def add_tokenizer_model(self, model: str):
         self.add_string(KEY_TOKENIZER_MODEL, model)
 
-    def add_token_list(self, tokens: List):
+    def add_token_list(self, tokens: Union[Sequence[str], Sequence[bytes], Sequence[bytearray]]):
         self.add_array(KEY_TOKENIZER_LIST, tokens)
 
-    def add_token_merges(self, merges: List):
+    def add_token_merges(self, merges: Union[Sequence[str], Sequence[bytes], Sequence[bytearray]]):
         self.add_array(KEY_TOKENIZER_MERGES, merges)
 
-    def add_token_types(self, types: List[int]):
+    def add_token_types(self, types: Union[Sequence[TokenType], Sequence[int]]):
         self.add_array(KEY_TOKENIZER_TOKEN_TYPE, types)
 
-    def add_token_scores(self, scores: List[float]):
+    def add_token_scores(self, scores: Sequence[float]):
         self.add_array(KEY_TOKENIZER_SCORES, scores)
 
     def add_bos_token_id(self, id: int):
@@ -723,6 +754,84 @@ class GGUFWriter:
         self.add_uint32(KEY_TOKENIZER_PAD_ID, id)
 
 
+class SpecialVocab:
+    load_merges: bool = False
+    merges: List[str] = []
+    special_token_types: Tuple[str, ...] = tuple(('bos', 'eos', 'unk', 'sep', 'pad'))
+    special_token_ids: Dict[str, int] = {}
+
+    def __init__(self, path: Path, load_merges: bool = False, special_token_types: Optional[Tuple[str, ...]] = None):
+        self.special_token_ids = {}
+        self.load_merges = load_merges
+        if special_token_types is not None:
+            self.special_token_types = special_token_types
+        self.load(path)
+
+    def load(self, path: Path):
+        if not self.try_load_from_tokenizer_json(path):
+            self.try_load_from_config_json(path)
+
+    def try_load_from_tokenizer_json(self, path: Path) -> bool:
+        tokenizer_file = path / 'tokenizer.json'
+        if not tokenizer_file.is_file():
+            return False
+        with open(tokenizer_file, 'r', encoding = 'utf-8') as f:
+            tokenizer = json.load(f)
+        if self.load_merges:
+            merges = tokenizer.get('model', {}).get('merges')
+            if isinstance(merges, list) and len(merges) > 0 and isinstance(merges[0], str):
+                self.merges = merges
+        tokenizer_config_file = path / 'tokenizer_config.json'
+        added_tokens = tokenizer.get('added_tokens')
+        if added_tokens is None or not tokenizer_config_file.is_file():
+            return True
+        with open(tokenizer_config_file, 'r', encoding = 'utf-8') as f:
+            tokenizer_config = json.load(f)
+        for typ in self.special_token_types:
+            entry = tokenizer_config.get(f'{typ}_token')
+            if isinstance(entry, str):
+                tc_content = entry
+            elif isinstance(entry, dict):
+                entry_content = entry.get('content')
+                if not isinstance(entry_content, str):
+                    continue
+                tc_content = entry_content
+            else:
+                continue
+            for maybe_token_id in (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content):
+                if isinstance(maybe_token_id, int):
+                    self.special_token_ids[typ] = maybe_token_id
+                break
+        return True
+
+    def try_load_from_config_json(self, path: Path) -> bool:
+        config_file = path / 'config.json'
+        if not config_file.is_file():
+            return False
+        with open(config_file, 'r', encoding = 'utf-8') as f:
+            config = json.load(f)
+        for typ in self.special_token_types:
+            maybe_token_id = config.get(f'{typ}_token_id')
+            if isinstance(maybe_token_id, int):
+                self.special_token_ids[typ] = maybe_token_id
+        return True
+
+    def add_to_gguf(self, gw: GGUFWriter):
+        if len(self.merges) > 0:
+            print(f'gguf: Adding {len(self.merges)} merge(s).')
+            gw.add_token_merges(self.merges)
+        for typ, tokid in self.special_token_ids.items():
+            handler: Optional[Callable[[int], None]] = getattr(gw, f'add_{typ}_token_id', None)
+            if handler is None:
+                print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping')
+                continue
+            print(f'gguf: Setting special token type {typ} to {tokid}')
+            handler(tokid)
+
+    def __repr__(self):
+        return f'<SpecialVocab with {len(self.merges)} merges and special tokens {self.special_token_ids if self.special_token_ids else "unset"}>'
+
+
 # Example usage:
 if __name__ == "__main__":
     # Example usage with a file
diff --git a/gguf-py/gguf/py.typed b/gguf-py/gguf/py.typed
new file mode 100644
index 000000000..e69de29bb
diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml
index cc70e28b7..c66b069f9 100644
--- a/gguf-py/pyproject.toml
+++ b/gguf-py/pyproject.toml
@@ -5,6 +5,7 @@ description = "Write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [
     {include = "gguf"},
+    {include = "gguf/py.typed"},
 ]
 readme = "README.md"
 homepage = "https://ggml.ai"

From 35092fb54712d032860f3976a6fc1ae1f84a4a28 Mon Sep 17 00:00:00 2001
From: Gilad S <giladgd@users.noreply.github.com>
Date: Wed, 30 Aug 2023 11:40:12 +0300
Subject: [PATCH 16/21] docs : add `node-llama-cpp` to `README.md` (#2885)

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index a880fd29f..d727b0554 100644
--- a/README.md
+++ b/README.md
@@ -107,7 +107,7 @@ as the main playground for developing new features for the [ggml](https://github
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
-- Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
+- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp), [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)

From 950929442070874d45561d2a4b68b010457767de Mon Sep 17 00:00:00 2001
From: alonfaraj <alonfaraj@gmail.com>
Date: Wed, 30 Aug 2023 12:42:51 +0300
Subject: [PATCH 17/21] make : add test and update CI (#2897)

* build ci: run make test

* makefile:
- add all
- add test

* enable tests/test-tokenizer-0-llama

* fix path to model

* remove gcc-8 from macos build test

* Update Makefile

* Update Makefile
---
 .github/workflows/build.yml | 12 ++++++++++++
 Makefile                    | 17 +++++++++++++++++
 2 files changed, 29 insertions(+)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 10320ad1f..20fd8c2b5 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -41,6 +41,12 @@ jobs:
         run: |
           CC=gcc-8 make
 
+      - name: Test
+        id: make_test
+        run: |
+          CC=gcc-8 make tests
+          make test
+
   ubuntu-latest-cmake:
     runs-on: ubuntu-latest
 
@@ -157,6 +163,12 @@ jobs:
         run: |
           make
 
+      - name: Test
+        id: make_test
+        run: |
+          make tests
+          make test
+
   macOS-latest-cmake:
     runs-on: macos-latest
 
diff --git a/Makefile b/Makefile
index c8b8a92d7..bd2d92869 100644
--- a/Makefile
+++ b/Makefile
@@ -6,6 +6,23 @@ TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-dou
 
 default: $(BUILD_TARGETS)
 
+test:
+	@echo "Running tests..."
+	@for test_target in $(TEST_TARGETS); do \
+		if [ "$$test_target" = "tests/test-tokenizer-0-llama" ]; then \
+			./$$test_target $(CURDIR)/models/ggml-vocab-llama.gguf; \
+		elif [ "$$test_target" = "tests/test-tokenizer-0-falcon" ]; then \
+			continue; \
+		elif [ "$$test_target" = "tests/test-tokenizer-1" ]; then \
+			continue; \
+		else \
+			./$$test_target; \
+		fi; \
+	done
+	@echo "All tests have been run."
+
+all: $(BUILD_TARGETS) $(TEST_TARGETS)
+
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif

From 0d1c706181cd31e7f368dd14eeb16c1a2569e4df Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?M=2E=20Yusuf=20Sar=C4=B1g=C3=B6z?= <yusufsarigoz@gmail.com>
Date: Wed, 30 Aug 2023 12:47:40 +0300
Subject: [PATCH 18/21] gguf : add workflow for Pypi publishing (#2896)

* gguf : add workflow for Pypi publishing

* gguf : add workflow for Pypi publishing

* fix trailing whitespace
---
 .github/workflows/gguf-publish.yml | 43 ++++++++++++++++++++++++++++++
 gguf-py/README.md                  | 23 +++++++++++++---
 2 files changed, 63 insertions(+), 3 deletions(-)
 create mode 100644 .github/workflows/gguf-publish.yml

diff --git a/.github/workflows/gguf-publish.yml b/.github/workflows/gguf-publish.yml
new file mode 100644
index 000000000..a6289e335
--- /dev/null
+++ b/.github/workflows/gguf-publish.yml
@@ -0,0 +1,43 @@
+# This workflow will upload a Python Package using Twine when a GGUF release is created
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+# See `gguf-py/README.md` for how to make a release.
+
+# This workflow uses actions that are not certified by GitHub.
+# They are provided by a third-party and are governed by
+# separate terms of service, privacy policy, and support
+# documentation.
+
+name: Upload Python Package
+
+on:
+  workflow_dispatch:
+  push:
+    # Pattern matched against refs/tags
+    tags:
+      - 'gguf-v*'           # Push events to every version tag
+
+
+jobs:
+  deploy:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python
+      uses: actions/setup-python@v2
+      with:
+        python-version: '3.9.x'
+    - name: Install dependencies
+      run: |
+        cd gguf-py
+        python -m pip install poetry
+        poetry install
+
+    - name: Build package
+      run: poetry build
+    - name: Publish package
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        password: ${{ secrets.PYPI_API_TOKEN }}
diff --git a/gguf-py/README.md b/gguf-py/README.md
index 03ad306ec..ffe25c495 100644
--- a/gguf-py/README.md
+++ b/gguf-py/README.md
@@ -27,8 +27,25 @@ In this case, upgrade Pip to the latest:
 pip install --upgrade pip
 ```
 
-## Publishing
-To publish the package, you need to have `twine` and `build` installed:
+## Automatic publishing with CI
+
+There's a GitHub workflow to make a release automatically upon creation of tags in a specified format.
+
+1. Bump the version in `pyproject.toml`.
+2. Create a tag named `gguf-vx.x.x` where `x.x.x` is the semantic version number.
+
+```sh
+git tag -a gguf-v1.0.0 -m "Version 1.0 release"
+```
+
+3. Push the tags.
+
+```sh
+git push origin --tags
+```
+
+## Manual publishing
+If you want to publish the package manually for any reason, you need to have `twine` and `build` installed:
 
 ```sh
 pip install build twine
@@ -36,7 +53,7 @@ pip install build twine
 
 Then, folow these steps to release a new version:
 
-1. Update the version in `pyproject.toml`.
+1. Bump the version in `pyproject.toml`.
 2. Build the package:
 
 ```sh

From c90d135eb433cf0d40fb95e46a48d1391d2352b5 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Wed, 30 Aug 2023 12:52:46 +0300
Subject: [PATCH 19/21] examples : fix underscore in beam-search + .gitignore
 (close #2900)

---
 .gitignore                                                    | 3 +++
 Makefile                                                      | 4 ++--
 examples/CMakeLists.txt                                       | 2 +-
 examples/{beam_search => beam-search}/CMakeLists.txt          | 4 ++--
 .../beam_search.cpp => beam-search/beam-search.cpp}           | 0
 5 files changed, 8 insertions(+), 5 deletions(-)
 rename examples/{beam_search => beam-search}/CMakeLists.txt (78%)
 rename examples/{beam_search/beam_search.cpp => beam-search/beam-search.cpp} (100%)

diff --git a/.gitignore b/.gitignore
index 54ea2b522..8b5f45a2d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -42,6 +42,9 @@ models-mnt
 /gguf-llama-simple
 /libllama.so
 /llama-bench
+/baby-llama
+/beam-search
+/save-load-state
 build-info.h
 arm_neon.h
 compile_commands.json
diff --git a/Makefile b/Makefile
index bd2d92869..b750540fe 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,5 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam_search tests/test-c.o
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam-search tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1
@@ -446,7 +446,7 @@ llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o co
 baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-beam_search: examples/beam_search/beam_search.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index 94b785224..6e65eb087 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -25,7 +25,7 @@ else()
     add_subdirectory(simple)
     add_subdirectory(embd-input)
     add_subdirectory(llama-bench)
-    add_subdirectory(beam_search)
+    add_subdirectory(beam-search)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()
diff --git a/examples/beam_search/CMakeLists.txt b/examples/beam-search/CMakeLists.txt
similarity index 78%
rename from examples/beam_search/CMakeLists.txt
rename to examples/beam-search/CMakeLists.txt
index b29e01092..e44a74975 100644
--- a/examples/beam_search/CMakeLists.txt
+++ b/examples/beam-search/CMakeLists.txt
@@ -1,5 +1,5 @@
-set(TARGET beam_search)
-add_executable(${TARGET} beam_search.cpp)
+set(TARGET beam-search)
+add_executable(${TARGET} beam-search.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/beam_search/beam_search.cpp b/examples/beam-search/beam-search.cpp
similarity index 100%
rename from examples/beam_search/beam_search.cpp
rename to examples/beam-search/beam-search.cpp

From b532a69b2fd08067f34f32f37a2fd9b37678a34a Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Wed, 30 Aug 2023 13:29:40 +0300
Subject: [PATCH 20/21] convert.py : use dir name to name the llama

---
 convert.py | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/convert.py b/convert.py
index 448b6f0f3..a7f4c2d75 100755
--- a/convert.py
+++ b/convert.py
@@ -811,10 +811,12 @@ class OutputFile:
 
     def add_meta_arch(self, params: Params) -> None:
         name = "LLaMA"
+
+        # TODO: better logic to determine model name
         if (params.n_ctx == 4096):
             name = "LLaMA v2"
-            if params.path_model:
-                name = str(params.path_model.parent).split('/')[-1]
+        elif params.path_model:
+            name = str(params.path_model.parent).split('/')[-1]
 
         self.gguf.add_name                (name)
         self.gguf.add_context_length      (params.n_ctx)
@@ -839,8 +841,7 @@ class OutputFile:
         tokens = []
         scores = []
         toktypes = []
-        # NOTE: `all_tokens` returns the the base vocabulary and added tokens
-        # TODO: add special tokens?
+        # NOTE: `all_tokens` returns the base vocabulary and added tokens
         for text, score, toktype in vocab.all_tokens():
             tokens.append(text)
             scores.append(score)

From 71d6975559acfd6c8407a4ef8275a9979c737765 Mon Sep 17 00:00:00 2001
From: Henri Vasserman <henv@hot.ee>
Date: Wed, 30 Aug 2023 19:14:53 +0300
Subject: [PATCH 21/21] [Docker] fix tools.sh argument passing. (#2884)

* [Docker] fix tools.sh argument passing.

This should allow passing multiple arguments to containers with
the full image that are using the tools.sh frontend.

Fix from https://github.com/ggerganov/llama.cpp/issues/2535#issuecomment-1697091734
---
 .devops/tools.sh | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)

diff --git a/.devops/tools.sh b/.devops/tools.sh
index 2787c21fe..9d999315f 100755
--- a/.devops/tools.sh
+++ b/.devops/tools.sh
@@ -7,15 +7,12 @@ arg1="$1"
 # Shift the arguments to remove the first one
 shift
 
-# Join the remaining arguments into a single string
-arg2="$@"
-
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
-    python3 ./convert.py "$arg2"
+    python3 ./convert.py "$@"
 elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
-    ./quantize "$arg2"
+    ./quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
-    ./main "$arg2"
+    ./main "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -27,7 +24,7 @@ elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
         fi
     done
 elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
-    ./server "$arg2"
+    ./server "$@"
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "