Fix: do not close file on mmap (#1017)

Calling `mmap.mmap` on Windows apparently resets the file offset of the
raw file object (and makes the BufferedReader return a *negative* file
offset).  For safetensors, avoid using the file offset after calling
mmap.  For GGML format, explicitly save and restore the offset.

Fixes #966.

convert.py

@@ -735,7 +735,7 @@ def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
     header: Dict[str, Dict[str, Any]] = json.loads(fp.read(header_size))
     # Use mmap for the actual data to avoid race conditions with the file offset.
     mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
-    byte_buf = mapped[fp.tell():]
+    byte_buf = mapped[8 + header_size:]
 
     def convert(info: Dict[str, Any]) -> LazyTensor:
         data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
@@ -761,7 +761,7 @@ def must_read(fp: IO[bytes], length: int) -> bytes:
     return ret
 
 
-def lazy_load_ggml_file(fp: IO[bytes], path: Path) -> ModelPlus:
+def lazy_load_ggml_file(fp: io.BufferedReader, path: Path) -> ModelPlus:
     magic = must_read(fp, 4)[::-1]
     if magic in (b'ggmf', b'ggjt'):
         version, = struct.unpack("i", must_read(fp, 4))
@@ -795,7 +795,9 @@ def lazy_load_ggml_file(fp: IO[bytes], path: Path) -> ModelPlus:
 
     model: LazyModel = {}
     # Use mmap for the actual data to avoid race conditions with the file offset.
+    off = fp.raw.tell()
     mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
+    fp.raw.seek(off) # needed on Windows
 
     def read_tensor() -> None:  # this is a function so that variables captured in `load` don't change
         shape_len, name_len, ftype = struct.unpack("iii", must_read(fp, 12))
@@ -949,8 +951,9 @@ class OutputFile:
 
         ndarrays = bounded_parallel_map(do_item, model.items(), concurrency=8)
         for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
-            size = ' x '.join(map(str, lazy_tensor.shape))
-            print(f"[{i+1}/{len(model)}] Writing tensor {name}, size {size}...")
+            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
+            padi = len(str(len(model)))
+            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type}")
             of.write_tensor_header(name, lazy_tensor.shape, lazy_tensor.data_type)
             ndarray.tofile(of.fout)
         of.fout.close()

examples/embedding/embedding.cpp

@@ -1,6 +1,8 @@
 #include "common.h"
 #include "llama.h"
 
+#include <ctime>
+
 int main(int argc, char ** argv) {
     gpt_params params;
     params.model = "models/llama-7B/ggml-model.bin";

examples/main/main.cpp

@@ -11,6 +11,7 @@
 #include <cmath>
 #include <cstdio>
 #include <cstring>
+#include <ctime>
 #include <fstream>
 #include <iostream>
 #include <string>

examples/perplexity/perplexity.cpp

@@ -2,6 +2,7 @@
 #include "llama.h"
 
 #include <cmath>
+#include <ctime>
 
 std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());

llama.cpp

@@ -9,6 +9,7 @@
 #include "ggml.h"
 
 #include <array>
+#include <ctime>
 #include <cinttypes>
 #include <fstream>
 #include <random>
@@ -261,12 +262,12 @@ static size_t checked_div(size_t a, size_t b) {
 }
 
 static std::string llama_format_tensor_shape(const std::vector<uint32_t> & ne) {
-    std::string ret = "[" + std::to_string(ne.at(0));
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5u", ne.at(0));
     for (size_t i = 1; i < ne.size(); i++) {
-        ret += " x " + std::to_string(ne.at(i));
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), " x %5u", ne.at(i));
     }
-    ret += "]";
-    return ret;
+    return buf;
 }
 
 static size_t llama_calc_tensor_size(const std::vector<uint32_t> & ne, enum ggml_type type) {
@@ -941,8 +942,8 @@ static void llama_model_load_internal(
         ml->ggml_ctx = ctx;
 
         model.tok_embeddings = ml->get_tensor("tok_embeddings.weight", {n_embd, n_vocab});
-        model.norm   = ml->get_tensor("norm.weight", {n_embd});
-        model.output = ml->get_tensor("output.weight", {n_embd, n_vocab});
+        model.norm           = ml->get_tensor("norm.weight",           {n_embd});
+        model.output         = ml->get_tensor("output.weight",         {n_embd, n_vocab});
 
         model.layers.resize(n_layer);
         for (uint32_t i = 0; i < n_layer; ++i) {
@@ -1569,7 +1570,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         tensor.data = read_data.addr;
         model_loader->load_data_for(tensor);
 
-        printf("[%zu/%zu] %36s - %s, type = %6s, ",
+        printf("[%4zu/%4zu] %36s - %16s, type = %6s, ",
                ++idx, model_loader->tensors_map.tensors.size(),
                tensor.name.c_str(), llama_format_tensor_shape(tensor.ne).c_str(),
                ggml_type_name(tensor.type));

llama_util.h

@@ -43,8 +43,12 @@
     } while (0)
 
 #ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
 __attribute__((format(printf, 1, 2)))
 #endif
+#endif
 static std::string format(const char * fmt, ...) {
     va_list ap, ap2;
     va_start(ap, fmt);
@@ -57,7 +61,7 @@ static std::string format(const char * fmt, ...) {
     va_end(ap2);
     va_end(ap);
     return std::string(buf.data(), size);
-};
+}
 
 struct llama_file {
     // use FILE * so we don't have to re-open the file to mmap
@@ -172,7 +176,6 @@ struct llama_mmap {
         flags |= MAP_POPULATE;
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
-        close(fd);
         if (addr == MAP_FAILED) {
             throw format("mmap failed: %s", strerror(errno));
         }