KenLM 5cc905bc2d214efa7de2db56a9a672b749a95591

37 files changed, 800 insertions, 487 deletions

diff --git a/klm/lm/bhiksha.cc b/klm/lm/bhiksha.cc
index 088ea98d..c8a18dfd 100644
--- a/klm/lm/bhiksha.cc
+++ b/klm/lm/bhiksha.cc
@@ -1,4 +1,6 @@
 #include "lm/bhiksha.hh"
+
+#include "lm/binary_format.hh"
 #include "lm/config.hh"
 #include "util/file.hh"
 #include "util/exception.hh"
@@ -15,11 +17,11 @@ DontBhiksha::DontBhiksha(const void * /*base*/, uint64_t /*max_offset*/, uint64_
 const uint8_t kArrayBhikshaVersion = 0;
 
 // TODO: put this in binary file header instead when I change the binary file format again.  
-void ArrayBhiksha::UpdateConfigFromBinary(int fd, Config &config) {
-  uint8_t version;
-  uint8_t configured_bits;
-  util::ReadOrThrow(fd, &version, 1);
-  util::ReadOrThrow(fd, &configured_bits, 1);
+void ArrayBhiksha::UpdateConfigFromBinary(const BinaryFormat &file, uint64_t offset, Config &config) {
+  uint8_t buffer[2];
+  file.ReadForConfig(buffer, 2, offset);
+  uint8_t version = buffer[0];
+  uint8_t configured_bits = buffer[1];
   if (version != kArrayBhikshaVersion) UTIL_THROW(FormatLoadException, "This file has sorted array compression version " << (unsigned) version << " but the code expects version " << (unsigned)kArrayBhikshaVersion);
   config.pointer_bhiksha_bits = configured_bits;
 }
@@ -87,9 +89,6 @@ void ArrayBhiksha::FinishedLoading(const Config &config) {
   *(head_write++) = config.pointer_bhiksha_bits;
 }
 
-void ArrayBhiksha::LoadedBinary() {
-}
-
 } // namespace trie
 } // namespace ngram
 } // namespace lm
diff --git a/klm/lm/bhiksha.hh b/klm/lm/bhiksha.hh
index 8ff88654..350571a6 100644
--- a/klm/lm/bhiksha.hh
+++ b/klm/lm/bhiksha.hh
@@ -24,6 +24,7 @@
 namespace lm {
 namespace ngram {
 struct Config;
+class BinaryFormat;
 
 namespace trie {
 
@@ -31,7 +32,7 @@ class DontBhiksha {
   public:
     static const ModelType kModelTypeAdd = static_cast<ModelType>(0);
 
-    static void UpdateConfigFromBinary(int /*fd*/, Config &/*config*/) {}
+    static void UpdateConfigFromBinary(const BinaryFormat &, uint64_t, Config &/*config*/) {}
 
     static uint64_t Size(uint64_t /*max_offset*/, uint64_t /*max_next*/, const Config &/*config*/) { return 0; }
 
@@ -53,8 +54,6 @@ class DontBhiksha {
 
     void FinishedLoading(const Config &/*config*/) {}
 
-    void LoadedBinary() {}
-
     uint8_t InlineBits() const { return next_.bits; }
 
   private:
@@ -65,7 +64,7 @@ class ArrayBhiksha {
   public:
     static const ModelType kModelTypeAdd = kArrayAdd;
 
-    static void UpdateConfigFromBinary(int fd, Config &config);
+    static void UpdateConfigFromBinary(const BinaryFormat &file, uint64_t offset, Config &config);
 
     static uint64_t Size(uint64_t max_offset, uint64_t max_next, const Config &config);
 
@@ -93,8 +92,6 @@ class ArrayBhiksha {
 
     void FinishedLoading(const Config &config);
 
-    void LoadedBinary();
-
     uint8_t InlineBits() const { return next_inline_.bits; }
 
   private:
diff --git a/klm/lm/binary_format.cc b/klm/lm/binary_format.cc
index 39c4a9b6..9c744b13 100644
--- a/klm/lm/binary_format.cc
+++ b/klm/lm/binary_format.cc
@@ -8,11 +8,15 @@
 #include <cstring>
 #include <limits>
 #include <string>
+#include <cstdlib>
 
 #include <stdint.h>
 
 namespace lm {
 namespace ngram {
+
+const char *kModelNames[6] = {"probing hash tables", "probing hash tables with rest costs", "trie", "trie with quantization", "trie with array-compressed pointers", "trie with quantization and array-compressed pointers"};
+
 namespace {
 const char kMagicBeforeVersion[] = "mmap lm http://kheafield.com/code format version";
 const char kMagicBytes[] = "mmap lm http://kheafield.com/code format version 5\n\0";
@@ -57,8 +61,6 @@ struct Sanity {
   }
 };
 
-const char *kModelNames[6] = {"probing hash tables", "probing hash tables with rest costs", "trie", "trie with quantization", "trie with array-compressed pointers", "trie with quantization and array-compressed pointers"};
-
 std::size_t TotalHeaderSize(unsigned char order) {
   return ALIGN8(sizeof(Sanity) + sizeof(FixedWidthParameters) + sizeof(uint64_t) * order);
 }
@@ -80,83 +82,6 @@ void WriteHeader(void *to, const Parameters &params) {
 
 } // namespace
 
-uint8_t *SetupJustVocab(const Config &config, uint8_t order, std::size_t memory_size, Backing &backing) {
-  if (config.write_mmap) {
-    std::size_t total = TotalHeaderSize(order) + memory_size;
-    backing.file.reset(util::CreateOrThrow(config.write_mmap));
-    if (config.write_method == Config::WRITE_MMAP) {
-      backing.vocab.reset(util::MapZeroedWrite(backing.file.get(), total), total, util::scoped_memory::MMAP_ALLOCATED);
-    } else {
-      util::ResizeOrThrow(backing.file.get(), 0);
-      util::MapAnonymous(total, backing.vocab);
-    }
-    strncpy(reinterpret_cast<char*>(backing.vocab.get()), kMagicIncomplete, TotalHeaderSize(order));
-    return reinterpret_cast<uint8_t*>(backing.vocab.get()) + TotalHeaderSize(order);
-  } else {
-    util::MapAnonymous(memory_size, backing.vocab);
-    return reinterpret_cast<uint8_t*>(backing.vocab.get());
-  }
-}
-
-uint8_t *GrowForSearch(const Config &config, std::size_t vocab_pad, std::size_t memory_size, Backing &backing) {
-  std::size_t adjusted_vocab = backing.vocab.size() + vocab_pad;
-  if (config.write_mmap) {
-    // Grow the file to accomodate the search, using zeros.
-    try {
-      util::ResizeOrThrow(backing.file.get(), adjusted_vocab + memory_size);
-    } catch (util::ErrnoException &e) {
-      e << " for file " << config.write_mmap;
-      throw e;
-    }
-
-    if (config.write_method == Config::WRITE_AFTER) {
-      util::MapAnonymous(memory_size, backing.search);
-      return reinterpret_cast<uint8_t*>(backing.search.get());
-    }
-    // mmap it now.
-    // We're skipping over the header and vocab for the search space mmap.  mmap likes page aligned offsets, so some arithmetic to round the offset down.
-    std::size_t page_size = util::SizePage();
-    std::size_t alignment_cruft = adjusted_vocab % page_size;
-    backing.search.reset(util::MapOrThrow(alignment_cruft + memory_size, true, util::kFileFlags, false, backing.file.get(), adjusted_vocab - alignment_cruft), alignment_cruft + memory_size, util::scoped_memory::MMAP_ALLOCATED);
-    return reinterpret_cast<uint8_t*>(backing.search.get()) + alignment_cruft;
-  } else {
-    util::MapAnonymous(memory_size, backing.search);
-    return reinterpret_cast<uint8_t*>(backing.search.get());
-  }
-}
-
-void FinishFile(const Config &config, ModelType model_type, unsigned int search_version, const std::vector<uint64_t> &counts, std::size_t vocab_pad, Backing &backing) {
-  if (!config.write_mmap) return;
-  switch (config.write_method) {
-    case Config::WRITE_MMAP:
-      util::SyncOrThrow(backing.vocab.get(), backing.vocab.size());
-      util::SyncOrThrow(backing.search.get(), backing.search.size());
-      break;
-    case Config::WRITE_AFTER:
-      util::SeekOrThrow(backing.file.get(), 0);
-      util::WriteOrThrow(backing.file.get(), backing.vocab.get(), backing.vocab.size());
-      util::SeekOrThrow(backing.file.get(), backing.vocab.size() + vocab_pad);
-      util::WriteOrThrow(backing.file.get(), backing.search.get(), backing.search.size());
-      util::FSyncOrThrow(backing.file.get());
-      break;
-  }
-  // header and vocab share the same mmap.  The header is written here because we know the counts.
-  Parameters params = Parameters();
-  params.counts = counts;
-  params.fixed.order = counts.size();
-  params.fixed.probing_multiplier = config.probing_multiplier;
-  params.fixed.model_type = model_type;
-  params.fixed.has_vocabulary = config.include_vocab;
-  params.fixed.search_version = search_version;
-  WriteHeader(backing.vocab.get(), params);
-  if (config.write_method == Config::WRITE_AFTER) {
-    util::SeekOrThrow(backing.file.get(), 0);
-    util::WriteOrThrow(backing.file.get(), backing.vocab.get(), TotalHeaderSize(counts.size()));
-  }
-}
-
-namespace detail {
-
 bool IsBinaryFormat(int fd) {
   const uint64_t size = util::SizeFile(fd);
   if (size == util::kBadSize || (size <= static_cast<uint64_t>(sizeof(Sanity)))) return false;
@@ -169,21 +94,21 @@ bool IsBinaryFormat(int fd) {
   }
   Sanity reference_header = Sanity();
   reference_header.SetToReference();
-  if (!memcmp(memory.get(), &reference_header, sizeof(Sanity))) return true;
-  if (!memcmp(memory.get(), kMagicIncomplete, strlen(kMagicIncomplete))) {
+  if (!std::memcmp(memory.get(), &reference_header, sizeof(Sanity))) return true;
+  if (!std::memcmp(memory.get(), kMagicIncomplete, strlen(kMagicIncomplete))) {
     UTIL_THROW(FormatLoadException, "This binary file did not finish building");
   }
-  if (!memcmp(memory.get(), kMagicBeforeVersion, strlen(kMagicBeforeVersion))) {
+  if (!std::memcmp(memory.get(), kMagicBeforeVersion, strlen(kMagicBeforeVersion))) {
     char *end_ptr;
     const char *begin_version = static_cast<const char*>(memory.get()) + strlen(kMagicBeforeVersion);
-    long int version = strtol(begin_version, &end_ptr, 10);
+    long int version = std::strtol(begin_version, &end_ptr, 10);
     if ((end_ptr != begin_version) && version != kMagicVersion) {
       UTIL_THROW(FormatLoadException, "Binary file has version " << version << " but this implementation expects version " << kMagicVersion << " so you'll have to use the ARPA to rebuild your binary");
     }
 
     OldSanity old_sanity = OldSanity();
     old_sanity.SetToReference();
-    UTIL_THROW_IF(!memcmp(memory.get(), &old_sanity, sizeof(OldSanity)), FormatLoadException, "Looks like this is an old 32-bit format.  The old 32-bit format has been removed so that 64-bit and 32-bit files are exchangeable.");
+    UTIL_THROW_IF(!std::memcmp(memory.get(), &old_sanity, sizeof(OldSanity)), FormatLoadException, "Looks like this is an old 32-bit format.  The old 32-bit format has been removed so that 64-bit and 32-bit files are exchangeable.");
     UTIL_THROW(FormatLoadException, "File looks like it should be loaded with mmap, but the test values don't match.  Try rebuilding the binary format LM using the same code revision, compiler, and architecture");
   }
   return false;
@@ -208,44 +133,164 @@ void MatchCheck(ModelType model_type, unsigned int search_version, const Paramet
   UTIL_THROW_IF(search_version != params.fixed.search_version, FormatLoadException, "The binary file has " << kModelNames[params.fixed.model_type] << " version " << params.fixed.search_version << " but this code expects " << kModelNames[params.fixed.model_type] << " version " << search_version);
 }
 
-void SeekPastHeader(int fd, const Parameters &params) {
-  util::SeekOrThrow(fd, TotalHeaderSize(params.counts.size()));
+const std::size_t kInvalidSize = static_cast<std::size_t>(-1);
+
+BinaryFormat::BinaryFormat(const Config &config) 
+  : write_method_(config.write_method), write_mmap_(config.write_mmap), load_method_(config.load_method),
+    header_size_(kInvalidSize), vocab_size_(kInvalidSize), vocab_string_offset_(kInvalidOffset) {}
+
+void BinaryFormat::InitializeBinary(int fd, ModelType model_type, unsigned int search_version, Parameters &params) {
+  file_.reset(fd);
+  write_mmap_ = NULL; // Ignore write requests; this is already in binary format.
+  ReadHeader(fd, params);
+  MatchCheck(model_type, search_version, params);
+  header_size_ = TotalHeaderSize(params.counts.size());
+}
+
+void BinaryFormat::ReadForConfig(void *to, std::size_t amount, uint64_t offset_excluding_header) const {
+  assert(header_size_ != kInvalidSize);
+  util::PReadOrThrow(file_.get(), to, amount, offset_excluding_header + header_size_);
 }
 
-uint8_t *SetupBinary(const Config &config, const Parameters &params, uint64_t memory_size, Backing &backing) {
-  const uint64_t file_size = util::SizeFile(backing.file.get());
+void *BinaryFormat::LoadBinary(std::size_t size) {
+  assert(header_size_ != kInvalidSize);
+  const uint64_t file_size = util::SizeFile(file_.get());
   // The header is smaller than a page, so we have to map the whole header as well.
-  std::size_t total_map = util::CheckOverflow(TotalHeaderSize(params.counts.size()) + memory_size);
-  if (file_size != util::kBadSize && static_cast<uint64_t>(file_size) < total_map)
-    UTIL_THROW(FormatLoadException, "Binary file has size " << file_size << " but the headers say it should be at least " << total_map);
+  uint64_t total_map = static_cast<uint64_t>(header_size_) + static_cast<uint64_t>(size);
+  UTIL_THROW_IF(file_size != util::kBadSize && file_size < total_map, FormatLoadException, "Binary file has size " << file_size << " but the headers say it should be at least " << total_map);
 
-  util::MapRead(config.load_method, backing.file.get(), 0, total_map, backing.search);
+  util::MapRead(load_method_, file_.get(), 0, util::CheckOverflow(total_map), mapping_);
 
-  if (config.enumerate_vocab && !params.fixed.has_vocabulary)
-    UTIL_THROW(FormatLoadException, "The decoder requested all the vocabulary strings, but this binary file does not have them.  You may need to rebuild the binary file with an updated version of build_binary.");
+  vocab_string_offset_ = total_map;
+  return reinterpret_cast<uint8_t*>(mapping_.get()) + header_size_;
+}
+
+void *BinaryFormat::SetupJustVocab(std::size_t memory_size, uint8_t order) {
+  vocab_size_ = memory_size;
+  if (!write_mmap_) {
+    header_size_ = 0;
+    util::MapAnonymous(memory_size, memory_vocab_);
+    return reinterpret_cast<uint8_t*>(memory_vocab_.get());
+  }
+  header_size_ = TotalHeaderSize(order);
+  std::size_t total = util::CheckOverflow(static_cast<uint64_t>(header_size_) + static_cast<uint64_t>(memory_size));
+  file_.reset(util::CreateOrThrow(write_mmap_));
+  // some gccs complain about uninitialized variables even though all enum values are covered.
+  void *vocab_base = NULL;
+  switch (write_method_) {
+    case Config::WRITE_MMAP:
+      mapping_.reset(util::MapZeroedWrite(file_.get(), total), total, util::scoped_memory::MMAP_ALLOCATED);
+      vocab_base = mapping_.get();
+      break;
+    case Config::WRITE_AFTER:
+      util::ResizeOrThrow(file_.get(), 0);
+      util::MapAnonymous(total, memory_vocab_);
+      vocab_base = memory_vocab_.get();
+      break;
+  }
+  strncpy(reinterpret_cast<char*>(vocab_base), kMagicIncomplete, header_size_);
+  return reinterpret_cast<uint8_t*>(vocab_base) + header_size_;
+}
 
-  // Seek to vocabulary words
-  util::SeekOrThrow(backing.file.get(), total_map);
-  return reinterpret_cast<uint8_t*>(backing.search.get()) + TotalHeaderSize(params.counts.size());
+void *BinaryFormat::GrowForSearch(std::size_t memory_size, std::size_t vocab_pad, void *&vocab_base) {
+  assert(vocab_size_ != kInvalidSize);
+  vocab_pad_ = vocab_pad;
+  std::size_t new_size = header_size_ + vocab_size_ + vocab_pad_ + memory_size;
+  vocab_string_offset_ = new_size;
+  if (!write_mmap_ || write_method_ == Config::WRITE_AFTER) {
+    util::MapAnonymous(memory_size, memory_search_);
+    assert(header_size_ == 0 || write_mmap_);
+    vocab_base = reinterpret_cast<uint8_t*>(memory_vocab_.get()) + header_size_;
+    return reinterpret_cast<uint8_t*>(memory_search_.get());
+  }
+
+  assert(write_method_ == Config::WRITE_MMAP);
+  // Also known as total size without vocab words.
+  // Grow the file to accomodate the search, using zeros.
+  // According to man mmap, behavior is undefined when the file is resized
+  // underneath a mmap that is not a multiple of the page size.  So to be
+  // safe, we'll unmap it and map it again.
+  mapping_.reset();
+  util::ResizeOrThrow(file_.get(), new_size);
+  void *ret;
+  MapFile(vocab_base, ret);
+  return ret;
 }
 
-void ComplainAboutARPA(const Config &config, ModelType model_type) {
-  if (config.write_mmap || !config.messages) return;
-  if (config.arpa_complain == Config::ALL) {
-    *config.messages << "Loading the LM will be faster if you build a binary file." << std::endl;
-  } else if (config.arpa_complain == Config::EXPENSIVE &&
-             (model_type == TRIE || model_type == QUANT_TRIE || model_type == ARRAY_TRIE || model_type == QUANT_ARRAY_TRIE)) {
-    *config.messages << "Building " << kModelNames[model_type] << " from ARPA is expensive.  Save time by building a binary format." << std::endl;
+void BinaryFormat::WriteVocabWords(const std::string &buffer, void *&vocab_base, void *&search_base) {
+  // Checking Config's include_vocab is the responsibility of the caller.
+  assert(header_size_ != kInvalidSize && vocab_size_ != kInvalidSize);
+  if (!write_mmap_) {
+    // Unchanged base.
+    vocab_base = reinterpret_cast<uint8_t*>(memory_vocab_.get());
+    search_base = reinterpret_cast<uint8_t*>(memory_search_.get());
+    return;
+  }
+  if (write_method_ == Config::WRITE_MMAP) {
+    mapping_.reset();
+  }
+  util::SeekOrThrow(file_.get(), VocabStringReadingOffset());
+  util::WriteOrThrow(file_.get(), &buffer[0], buffer.size());
+  if (write_method_ == Config::WRITE_MMAP) {
+    MapFile(vocab_base, search_base);
+  } else {
+    vocab_base = reinterpret_cast<uint8_t*>(memory_vocab_.get()) + header_size_;
+    search_base = reinterpret_cast<uint8_t*>(memory_search_.get());
+  }
+}
+
+void BinaryFormat::FinishFile(const Config &config, ModelType model_type, unsigned int search_version, const std::vector<uint64_t> &counts) {
+  if (!write_mmap_) return;
+  switch (write_method_) {
+    case Config::WRITE_MMAP:
+      util::SyncOrThrow(mapping_.get(), mapping_.size());
+      break;
+    case Config::WRITE_AFTER:
+      util::SeekOrThrow(file_.get(), 0);
+      util::WriteOrThrow(file_.get(), memory_vocab_.get(), memory_vocab_.size());
+      util::SeekOrThrow(file_.get(), header_size_ + vocab_size_ + vocab_pad_);
+      util::WriteOrThrow(file_.get(), memory_search_.get(), memory_search_.size());
+      util::FSyncOrThrow(file_.get());
+      break;
+  }
+  // header and vocab share the same mmap.
+  Parameters params = Parameters();
+  memset(&params, 0, sizeof(Parameters));
+  params.counts = counts;
+  params.fixed.order = counts.size();
+  params.fixed.probing_multiplier = config.probing_multiplier;
+  params.fixed.model_type = model_type;
+  params.fixed.has_vocabulary = config.include_vocab;
+  params.fixed.search_version = search_version;
+  switch (write_method_) {
+    case Config::WRITE_MMAP:
+      WriteHeader(mapping_.get(), params);
+      util::SyncOrThrow(mapping_.get(), mapping_.size());
+      break;
+    case Config::WRITE_AFTER:
+      {
+        std::vector<uint8_t> buffer(TotalHeaderSize(counts.size()));
+        WriteHeader(&buffer[0], params);
+        util::SeekOrThrow(file_.get(), 0);
+        util::WriteOrThrow(file_.get(), &buffer[0], buffer.size());
+      }
+      break;
   }
 }
 
-} // namespace detail
+void BinaryFormat::MapFile(void *&vocab_base, void *&search_base) {
+  mapping_.reset(util::MapOrThrow(vocab_string_offset_, true, util::kFileFlags, false, file_.get()), vocab_string_offset_, util::scoped_memory::MMAP_ALLOCATED);
+  vocab_base = reinterpret_cast<uint8_t*>(mapping_.get()) + header_size_;
+  search_base = reinterpret_cast<uint8_t*>(mapping_.get()) + header_size_ + vocab_size_ + vocab_pad_;
+}
 
 bool RecognizeBinary(const char *file, ModelType &recognized) {
   util::scoped_fd fd(util::OpenReadOrThrow(file));
-  if (!detail::IsBinaryFormat(fd.get())) return false;
+  if (!IsBinaryFormat(fd.get())) {
+    return false;
+  }
   Parameters params;
-  detail::ReadHeader(fd.get(), params);
+  ReadHeader(fd.get(), params);
   recognized = params.fixed.model_type;
   return true;
 }
diff --git a/klm/lm/binary_format.hh b/klm/lm/binary_format.hh
index bf699d5f..f33f88d7 100644
--- a/klm/lm/binary_format.hh
+++ b/klm/lm/binary_format.hh
@@ -17,6 +17,8 @@
 namespace lm {
 namespace ngram {
 
+extern const char *kModelNames[6];
+
 /*Inspect a file to determine if it is a binary lm.  If not, return false.  
  * If so, return true and set recognized to the type.  This is the only API in
  * this header designed for use by decoder authors.  
@@ -42,67 +44,63 @@ struct Parameters {
   std::vector<uint64_t> counts;
 };
 
-struct Backing {
-  // File behind memory, if any.  
-  util::scoped_fd file;
-  // Vocabulary lookup table.  Not to be confused with the vocab words themselves.  
-  util::scoped_memory vocab;
-  // Raw block of memory backing the language model data structures
-  util::scoped_memory search;
-};
-
-// Create just enough of a binary file to write vocabulary to it.  
-uint8_t *SetupJustVocab(const Config &config, uint8_t order, std::size_t memory_size, Backing &backing);
-// Grow the binary file for the search data structure and set backing.search, returning the memory address where the search data structure should begin.  
-uint8_t *GrowForSearch(const Config &config, std::size_t vocab_pad, std::size_t memory_size, Backing &backing);
-
-// Write header to binary file.  This is done last to prevent incomplete files
-// from loading.   
-void FinishFile(const Config &config, ModelType model_type, unsigned int search_version, const std::vector<uint64_t> &counts,  std::size_t vocab_pad, Backing &backing);
+class BinaryFormat {
+  public:
+    explicit BinaryFormat(const Config &config);
+
+    // Reading a binary file:
+    // Takes ownership of fd
+    void InitializeBinary(int fd, ModelType model_type, unsigned int search_version, Parameters &params);
+    // Used to read parts of the file to update the config object before figuring out full size.
+    void ReadForConfig(void *to, std::size_t amount, uint64_t offset_excluding_header) const;
+    // Actually load the binary file and return a pointer to the beginning of the search area.
+    void *LoadBinary(std::size_t size);
+
+    uint64_t VocabStringReadingOffset() const {
+      assert(vocab_string_offset_ != kInvalidOffset);
+      return vocab_string_offset_;
+    }
 
-namespace detail {
+    // Writing a binary file or initializing in RAM from ARPA:
+    // Size for vocabulary.
+    void *SetupJustVocab(std::size_t memory_size, uint8_t order);
+    // Warning: can change the vocaulary base pointer.
+    void *GrowForSearch(std::size_t memory_size, std::size_t vocab_pad, void *&vocab_base);
+    // Warning: can change vocabulary and search base addresses.
+    void WriteVocabWords(const std::string &buffer, void *&vocab_base, void *&search_base);
+    // Write the header at the beginning of the file.
+    void FinishFile(const Config &config, ModelType model_type, unsigned int search_version, const std::vector<uint64_t> &counts);
+
+  private:
+    void MapFile(void *&vocab_base, void *&search_base);
+
+    // Copied from configuration.
+    const Config::WriteMethod write_method_;
+    const char *write_mmap_;
+    util::LoadMethod load_method_;
+
+    // File behind memory, if any.  
+    util::scoped_fd file_;
+
+    // If there is a file involved, a single mapping.
+    util::scoped_memory mapping_;
+
+    // If the data is only in memory, separately allocate each because the trie
+    // knows vocab's size before it knows search's size (because SRILM might
+    // have pruned).
+    util::scoped_memory memory_vocab_, memory_search_;
+
+    // Memory ranges.  Note that these may not be contiguous and may not all
+    // exist.
+    std::size_t header_size_, vocab_size_, vocab_pad_;
+    // aka end of search.
+    uint64_t vocab_string_offset_;
+
+    static const uint64_t kInvalidOffset = (uint64_t)-1;
+};
 
 bool IsBinaryFormat(int fd);
 
-void ReadHeader(int fd, Parameters &params);
-
-void MatchCheck(ModelType model_type, unsigned int search_version, const Parameters &params);
-
-void SeekPastHeader(int fd, const Parameters &params);
-
-uint8_t *SetupBinary(const Config &config, const Parameters &params, uint64_t memory_size, Backing &backing);
-
-void ComplainAboutARPA(const Config &config, ModelType model_type);
-
-} // namespace detail
-
-template <class To> void LoadLM(const char *file, const Config &config, To &to) {
-  Backing &backing = to.MutableBacking();
-  backing.file.reset(util::OpenReadOrThrow(file));
-
-  try {
-    if (detail::IsBinaryFormat(backing.file.get())) {
-      Parameters params;
-      detail::ReadHeader(backing.file.get(), params);
-      detail::MatchCheck(To::kModelType, To::kVersion, params);
-      // Replace the run-time configured probing_multiplier with the one in the file.  
-      Config new_config(config);
-      new_config.probing_multiplier = params.fixed.probing_multiplier;
-      detail::SeekPastHeader(backing.file.get(), params);
-      To::UpdateConfigFromBinary(backing.file.get(), params.counts, new_config);
-      uint64_t memory_size = To::Size(params.counts, new_config);
-      uint8_t *start = detail::SetupBinary(new_config, params, memory_size, backing);
-      to.InitializeFromBinary(start, params, new_config, backing.file.get());
-    } else {
-      detail::ComplainAboutARPA(config, To::kModelType);
-      to.InitializeFromARPA(file, config);
-    }
-  } catch (util::Exception &e) {
-    e << " File: " << file;
-    throw;
-  }
-}
-
 } // namespace ngram
 } // namespace lm
 #endif // LM_BINARY_FORMAT__
diff --git a/klm/lm/build_binary_main.cc b/klm/lm/build_binary_main.cc
index ab2c0c32..15b421e9 100644
--- a/klm/lm/build_binary_main.cc
+++ b/klm/lm/build_binary_main.cc
@@ -52,6 +52,7 @@ void Usage(const char *name, const char *default_mem) {
 "-a compresses pointers using an array of offsets.  The parameter is the\n"
 "   maximum number of bits encoded by the array.  Memory is minimized subject\n"
 "   to the maximum, so pick 255 to minimize memory.\n\n"
+"-h print this help message.\n\n"
 "Get a memory estimate by passing an ARPA file without an output file name.\n";
   exit(1);
 }
@@ -104,12 +105,15 @@ int main(int argc, char *argv[]) {
 
   const char *default_mem = util::GuessPhysicalMemory() ? "80%" : "1G";
 
+  if (argc == 2 && !strcmp(argv[1], "--help"))
+    Usage(argv[0], default_mem);
+
   try {
     bool quantize = false, set_backoff_bits = false, bhiksha = false, set_write_method = false, rest = false;
     lm::ngram::Config config;
     config.building_memory = util::ParseSize(default_mem);
     int opt;
-    while ((opt = getopt(argc, argv, "q:b:a:u:p:t:T:m:S:w:sir:")) != -1) {
+    while ((opt = getopt(argc, argv, "q:b:a:u:p:t:T:m:S:w:sir:h")) != -1) {
       switch(opt) {
         case 'q':
           config.prob_bits = ParseBitCount(optarg);
@@ -161,6 +165,7 @@ int main(int argc, char *argv[]) {
           ParseFileList(optarg, config.rest_lower_files);
           config.rest_function = Config::REST_LOWER;
           break;
+        case 'h': // help
         default:
           Usage(argv[0], default_mem);
       }
@@ -186,6 +191,7 @@ int main(int argc, char *argv[]) {
       config.write_mmap = argv[optind + 2];
     } else {
       Usage(argv[0], default_mem);
+      return 1;
     }
     if (!strcmp(model_type, "probing")) {
       if (!set_write_method) config.write_method = Config::WRITE_AFTER;
diff --git a/klm/lm/builder/corpus_count.cc b/klm/lm/builder/corpus_count.cc
index aea93ad1..ccc06efc 100644
--- a/klm/lm/builder/corpus_count.cc
+++ b/klm/lm/builder/corpus_count.cc
@@ -238,12 +238,17 @@ void CorpusCount::Run(const util::stream::ChainPosition &position) {
   const WordIndex end_sentence = vocab.Lookup("</s>");
   Writer writer(NGram::OrderFromSize(position.GetChain().EntrySize()), position, dedupe_mem_.get(), dedupe_mem_size_);
   uint64_t count = 0;
-  StringPiece delimiters("\0\t\r ", 4);
+  bool delimiters[256];
+  memset(delimiters, 0, sizeof(delimiters));
+  const char kDelimiterSet[] = "\0\t\n\r ";
+  for (const char *i = kDelimiterSet; i < kDelimiterSet + sizeof(kDelimiterSet); ++i) {
+    delimiters[static_cast<unsigned char>(*i)] = true;
+  }
   try {
     while(true) {
       StringPiece line(from_.ReadLine());
       writer.StartSentence();
-      for (util::TokenIter<util::AnyCharacter, true> w(line, delimiters); w; ++w) {
+      for (util::TokenIter<util::BoolCharacter, true> w(line, delimiters); w; ++w) {
         WordIndex word = vocab.Lookup(*w);
         UTIL_THROW_IF(word <= 2, FormatLoadException, "Special word " << *w << " is not allowed in the corpus.  I plan to support models containing <unk> in the future.");
         writer.Append(word);
diff --git a/klm/lm/builder/lmplz_main.cc b/klm/lm/builder/lmplz_main.cc
index c87abdb8..2563deed 100644
--- a/klm/lm/builder/lmplz_main.cc
+++ b/klm/lm/builder/lmplz_main.cc
@@ -33,7 +33,10 @@ int main(int argc, char *argv[]) {
     po::options_description options("Language model building options");
     lm::builder::PipelineConfig pipeline;
 
+    std::string text, arpa;
+
     options.add_options()
+      ("help", po::bool_switch(), "Show this help message")
       ("order,o", po::value<std::size_t>(&pipeline.order)
 #if BOOST_VERSION >= 104200
          ->required()
@@ -47,8 +50,13 @@ int main(int argc, char *argv[]) {
       ("vocab_estimate", po::value<lm::WordIndex>(&pipeline.vocab_estimate)->default_value(1000000), "Assume this vocabulary size for purposes of calculating memory in step 1 (corpus count) and pre-sizing the hash table")
       ("block_count", po::value<std::size_t>(&pipeline.block_count)->default_value(2), "Block count (per order)")
       ("vocab_file", po::value<std::string>(&pipeline.vocab_file)->default_value(""), "Location to write vocabulary file")
-      ("verbose_header", po::bool_switch(&pipeline.verbose_header), "Add a verbose header to the ARPA file that includes information such as token count, smoothing type, etc.");
-    if (argc == 1) {
+      ("verbose_header", po::bool_switch(&pipeline.verbose_header), "Add a verbose header to the ARPA file that includes information such as token count, smoothing type, etc.")
+      ("text", po::value<std::string>(&text), "Read text from a file instead of stdin")
+      ("arpa", po::value<std::string>(&arpa), "Write ARPA to a file instead of stdout");
+    po::variables_map vm;
+    po::store(po::parse_command_line(argc, argv, options), vm);
+
+    if (argc == 1 || vm["help"].as<bool>()) {
       std::cerr << 
         "Builds unpruned language models with modified Kneser-Ney smoothing.\n\n"
         "Please cite:\n"
@@ -66,12 +74,17 @@ int main(int argc, char *argv[]) {
         "setting the temporary file location (-T) and sorting memory (-S) is recommended.\n\n"
         "Memory sizes are specified like GNU sort: a number followed by a unit character.\n"
         "Valid units are \% for percentage of memory (supported platforms only) and (in\n"
-        "increasing powers of 1024): b, K, M, G, T, P, E, Z, Y.  Default is K (*1024).\n\n";
+        "increasing powers of 1024): b, K, M, G, T, P, E, Z, Y.  Default is K (*1024).\n";
+      uint64_t mem = util::GuessPhysicalMemory();
+      if (mem) {
+        std::cerr << "This machine has " << mem << " bytes of memory.\n\n";
+      } else {
+        std::cerr << "Unable to determine the amount of memory on this machine.\n\n";
+      } 
       std::cerr << options << std::endl;
       return 1;
     }
-    po::variables_map vm;
-    po::store(po::parse_command_line(argc, argv, options), vm);
+
     po::notify(vm);
 
     // required() appeared in Boost 1.42.0.
@@ -92,9 +105,17 @@ int main(int argc, char *argv[]) {
     initial.adder_out.block_count = 2;
     pipeline.read_backoffs = initial.adder_out;
 
+    util::scoped_fd in(0), out(1);
+    if (vm.count("text")) {
+      in.reset(util::OpenReadOrThrow(text.c_str()));
+    }
+    if (vm.count("arpa")) {
+      out.reset(util::CreateOrThrow(arpa.c_str()));
+    }
+
     // Read from stdin
     try {
-      lm::builder::Pipeline(pipeline, 0, 1);
+      lm::builder::Pipeline(pipeline, in.release(), out.release());
     } catch (const util::MallocException &e) {
       std::cerr << e.what() << std::endl;
       std::cerr << "Try rerunning with a more conservative -S setting than " << vm["memory"].as<std::string>() << std::endl;
diff --git a/klm/lm/builder/pipeline.cc b/klm/lm/builder/pipeline.cc
index b89ea6ba..44a2313c 100644
--- a/klm/lm/builder/pipeline.cc
+++ b/klm/lm/builder/pipeline.cc
@@ -226,6 +226,7 @@ void CountText(int text_file /* input */, int vocab_file /* output */, Master &m
 
   util::stream::Sort<SuffixOrder, AddCombiner> sorter(chain, config.sort, SuffixOrder(config.order), AddCombiner());
   chain.Wait(true);
+  std::cerr << "Unigram tokens " << token_count << " types " << type_count << std::endl;
   std::cerr << "=== 2/5 Calculating and sorting adjusted counts ===" << std::endl;
   master.InitForAdjust(sorter, type_count);
 }
diff --git a/klm/lm/facade.hh b/klm/lm/facade.hh
index 8b186017..de1551f1 100644
--- a/klm/lm/facade.hh
+++ b/klm/lm/facade.hh
@@ -16,19 +16,28 @@ template <class Child, class StateT, class VocabularyT> class ModelFacade : publ
     typedef StateT State;
     typedef VocabularyT Vocabulary;
 
-    // Default Score function calls FullScore.  Model can override this.  
-    float Score(const State &in_state, const WordIndex new_word, State &out_state) const {
-      return static_cast<const Child*>(this)->FullScore(in_state, new_word, out_state).prob;
-    }
-
     /* Translate from void* to State */
-    FullScoreReturn FullScore(const void *in_state, const WordIndex new_word, void *out_state) const {
+    FullScoreReturn BaseFullScore(const void *in_state, const WordIndex new_word, void *out_state) const {
       return static_cast<const Child*>(this)->FullScore(
           *reinterpret_cast<const State*>(in_state),
           new_word,
           *reinterpret_cast<State*>(out_state));
     }
-    float Score(const void *in_state, const WordIndex new_word, void *out_state) const {
+
+    FullScoreReturn BaseFullScoreForgotState(const WordIndex *context_rbegin, const WordIndex *context_rend, const WordIndex new_word, void *out_state) const {
+      return static_cast<const Child*>(this)->FullScoreForgotState(
+          context_rbegin,
+          context_rend,
+          new_word,
+          *reinterpret_cast<State*>(out_state));
+    }
+
+    // Default Score function calls FullScore.  Model can override this.  
+    float Score(const State &in_state, const WordIndex new_word, State &out_state) const {
+      return static_cast<const Child*>(this)->FullScore(in_state, new_word, out_state).prob;
+    }
+
+    float BaseScore(const void *in_state, const WordIndex new_word, void *out_state) const {
       return static_cast<const Child*>(this)->Score(
           *reinterpret_cast<const State*>(in_state),
           new_word,
diff --git a/klm/lm/filter/arpa_io.hh b/klm/lm/filter/arpa_io.hh
index 5b31620b..602b5b31 100644
--- a/klm/lm/filter/arpa_io.hh
+++ b/klm/lm/filter/arpa_io.hh
@@ -14,7 +14,6 @@
 #include <string>
 #include <vector>
 
-#include <err.h>
 #include <string.h>
 #include <stdint.h>
 
diff --git a/klm/lm/filter/count_io.hh b/klm/lm/filter/count_io.hh
index 97c0fa25..d992026f 100644
--- a/klm/lm/filter/count_io.hh
+++ b/klm/lm/filter/count_io.hh
@@ -5,20 +5,18 @@
 #include <iostream>
 #include <string>
 
-#include <err.h>
-
+#include "util/fake_ofstream.hh"
+#include "util/file.hh"
 #include "util/file_piece.hh"
 
 namespace lm {
 
 class CountOutput : boost::noncopyable {
   public:
-    explicit CountOutput(const char *name) : file_(name, std::ios::out) {}
+    explicit CountOutput(const char *name) : file_(util::CreateOrThrow(name)) {}
 
     void AddNGram(const StringPiece &line) {
-      if (!(file_ << line << '\n')) {
-        err(3, "Writing counts file failed");
-      }
+      file_ << line << '\n';
     }
 
     template <class Iterator> void AddNGram(const Iterator &begin, const Iterator &end, const StringPiece &line) {
@@ -30,7 +28,7 @@ class CountOutput : boost::noncopyable {
     }
 
   private:
-    std::fstream file_;
+    util::FakeOFStream file_;
 };
 
 class CountBatch {
diff --git a/klm/lm/filter/filter_main.cc b/klm/lm/filter/filter_main.cc
index 1736bc40..82fdc1ef 100644
--- a/klm/lm/filter/filter_main.cc
+++ b/klm/lm/filter/filter_main.cc
@@ -6,6 +6,7 @@
 #endif
 #include "lm/filter/vocab.hh"
 #include "lm/filter/wrapper.hh"
+#include "util/exception.hh"
 #include "util/file_piece.hh"
 
 #include <boost/ptr_container/ptr_vector.hpp>
@@ -157,92 +158,96 @@ template <class Format> void DispatchFilterModes(const Config &config, std::istr
 } // namespace lm
 
 int main(int argc, char *argv[]) {
-  if (argc < 4) {
-    lm::DisplayHelp(argv[0]);
-    return 1;
-  }
+  try {
+    if (argc < 4) {
+      lm::DisplayHelp(argv[0]);
+      return 1;
+    }
 
-  // I used to have boost::program_options, but some users didn't want to compile boost.
-  lm::Config config;
-  config.mode = lm::MODE_UNSET;
-  for (int i = 1; i < argc - 2; ++i) {
-    const char *str = argv[i];
-    if (!std::strcmp(str, "copy")) {
-      config.mode = lm::MODE_COPY;
-    } else if (!std::strcmp(str, "single")) {
-      config.mode = lm::MODE_SINGLE;
-    } else if (!std::strcmp(str, "multiple")) {
-      config.mode = lm::MODE_MULTIPLE;
-    } else if (!std::strcmp(str, "union")) {
-      config.mode = lm::MODE_UNION;
-    } else if (!std::strcmp(str, "phrase")) {
-      config.phrase = true;
-    } else if (!std::strcmp(str, "context")) {
-      config.context = true;
-    } else if (!std::strcmp(str, "arpa")) {
-      config.format = lm::FORMAT_ARPA;
-    } else if (!std::strcmp(str, "raw")) {
-      config.format = lm::FORMAT_COUNT;
+    // I used to have boost::program_options, but some users didn't want to compile boost.
+    lm::Config config;
+    config.mode = lm::MODE_UNSET;
+    for (int i = 1; i < argc - 2; ++i) {
+      const char *str = argv[i];
+      if (!std::strcmp(str, "copy")) {
+        config.mode = lm::MODE_COPY;
+      } else if (!std::strcmp(str, "single")) {
+        config.mode = lm::MODE_SINGLE;
+      } else if (!std::strcmp(str, "multiple")) {
+        config.mode = lm::MODE_MULTIPLE;
+      } else if (!std::strcmp(str, "union")) {
+        config.mode = lm::MODE_UNION;
+      } else if (!std::strcmp(str, "phrase")) {
+        config.phrase = true;
+      } else if (!std::strcmp(str, "context")) {
+        config.context = true;
+      } else if (!std::strcmp(str, "arpa")) {
+        config.format = lm::FORMAT_ARPA;
+      } else if (!std::strcmp(str, "raw")) {
+        config.format = lm::FORMAT_COUNT;
 #ifndef NTHREAD
-    } else if (!std::strncmp(str, "threads:", 8)) {
-      config.threads = boost::lexical_cast<size_t>(str + 8);
-      if (!config.threads) {
-        std::cerr << "Specify at least one thread." << std::endl;
+      } else if (!std::strncmp(str, "threads:", 8)) {
+        config.threads = boost::lexical_cast<size_t>(str + 8);
+        if (!config.threads) {
+          std::cerr << "Specify at least one thread." << std::endl;
+          return 1;
+        }
+      } else if (!std::strncmp(str, "batch_size:", 11)) {
+        config.batch_size = boost::lexical_cast<size_t>(str + 11);
+        if (config.batch_size < 5000) {
+          std::cerr << "Batch size must be at least one and should probably be >= 5000" << std::endl;
+          if (!config.batch_size) return 1;
+        }
+#endif
+      } else {
+        lm::DisplayHelp(argv[0]);
         return 1;
       }
-    } else if (!std::strncmp(str, "batch_size:", 11)) {
-      config.batch_size = boost::lexical_cast<size_t>(str + 11);
-      if (config.batch_size < 5000) {
-        std::cerr << "Batch size must be at least one and should probably be >= 5000" << std::endl;
-        if (!config.batch_size) return 1;
-      }
-#endif
-    } else {
+    }
+
+    if (config.mode == lm::MODE_UNSET) {
       lm::DisplayHelp(argv[0]);
       return 1;
     }
-  }
-  
-  if (config.mode == lm::MODE_UNSET) {
-    lm::DisplayHelp(argv[0]);
-    return 1;
-  }
 
-  if (config.phrase && config.mode != lm::MODE_UNION && config.mode != lm::MODE_MULTIPLE) {
-    std::cerr << "Phrase constraint currently only works in multiple or union mode.  If you really need it for single, put everything on one line and use union." << std::endl;
-    return 1;
-  }
+    if (config.phrase && config.mode != lm::MODE_UNION && config.mode != lm::MODE_MULTIPLE) {
+      std::cerr << "Phrase constraint currently only works in multiple or union mode.  If you really need it for single, put everything on one line and use union." << std::endl;
+      return 1;
+    }
 
-  bool cmd_is_model = true;
-  const char *cmd_input = argv[argc - 2];
-  if (!strncmp(cmd_input, "vocab:", 6)) {
-    cmd_is_model = false;
-    cmd_input += 6;
-  } else if (!strncmp(cmd_input, "model:", 6)) {
-    cmd_input += 6;
-  } else if (strchr(cmd_input, ':')) {
-    errx(1, "Specify vocab: or model: before the input file name, not \"%s\"", cmd_input);
-  } else {
-    std::cerr << "Assuming that " << cmd_input << " is a model file" << std::endl;
-  }
-  std::ifstream cmd_file;
-  std::istream *vocab;
-  if (cmd_is_model) {
-    vocab = &std::cin;
-  } else {
-    cmd_file.open(cmd_input, std::ios::in);
-    if (!cmd_file) {
-      err(2, "Could not open input file %s", cmd_input);
+    bool cmd_is_model = true;
+    const char *cmd_input = argv[argc - 2];
+    if (!strncmp(cmd_input, "vocab:", 6)) {
+      cmd_is_model = false;
+      cmd_input += 6;
+    } else if (!strncmp(cmd_input, "model:", 6)) {
+      cmd_input += 6;
+    } else if (strchr(cmd_input, ':')) {
+      std::cerr << "Specify vocab: or model: before the input file name, not " << cmd_input << std::endl;
+      return 1;
+    } else {
+      std::cerr << "Assuming that " << cmd_input << " is a model file" << std::endl;
+    }
+    std::ifstream cmd_file;
+    std::istream *vocab;
+    if (cmd_is_model) {
+      vocab = &std::cin;
+    } else {
+      cmd_file.open(cmd_input, std::ios::in);
+      UTIL_THROW_IF(!cmd_file, util::ErrnoException, "Failed to open " << cmd_input);
+      vocab = &cmd_file;
     }
-    vocab = &cmd_file;
-  }
 
-  util::FilePiece model(cmd_is_model ? util::OpenReadOrThrow(cmd_input) : 0, cmd_is_model ? cmd_input : NULL, &std::cerr);
+    util::FilePiece model(cmd_is_model ? util::OpenReadOrThrow(cmd_input) : 0, cmd_is_model ? cmd_input : NULL, &std::cerr);
 
-  if (config.format == lm::FORMAT_ARPA) {
-    lm::DispatchFilterModes<lm::ARPAFormat>(config, *vocab, model, argv[argc - 1]);
-  } else if (config.format == lm::FORMAT_COUNT) {
-    lm::DispatchFilterModes<lm::CountFormat>(config, *vocab, model, argv[argc - 1]);
+    if (config.format == lm::FORMAT_ARPA) {
+      lm::DispatchFilterModes<lm::ARPAFormat>(config, *vocab, model, argv[argc - 1]);
+    } else if (config.format == lm::FORMAT_COUNT) {
+      lm::DispatchFilterModes<lm::CountFormat>(config, *vocab, model, argv[argc - 1]);
+    }
+    return 0;
+  } catch (const std::exception &e) {
+    std::cerr << e.what() << std::endl;
+    return 1;
   }
-  return 0;
 }
diff --git a/klm/lm/filter/format.hh b/klm/lm/filter/format.hh
index 7f945b0d..7d8c28db 100644
--- a/klm/lm/filter/format.hh
+++ b/klm/lm/filter/format.hh
@@ -1,5 +1,5 @@
 #ifndef LM_FILTER_FORMAT_H__
-#define LM_FITLER_FORMAT_H__
+#define LM_FILTER_FORMAT_H__
 
 #include "lm/filter/arpa_io.hh"
 #include "lm/filter/count_io.hh"
diff --git a/klm/lm/filter/phrase.cc b/klm/lm/filter/phrase.cc
index 1bef2a3f..e2946b14 100644
--- a/klm/lm/filter/phrase.cc
+++ b/klm/lm/filter/phrase.cc
@@ -48,21 +48,21 @@ unsigned int ReadMultiple(std::istream &in, Substrings &out) {
   return sentence_id + sentence_content;
 }
 
-namespace detail { const StringPiece kEndSentence("</s>"); }
-
 namespace {
-
 typedef unsigned int Sentence;
 typedef std::vector<Sentence> Sentences;
+} // namespace
 
-class Vertex;
+namespace detail { 
+
+const StringPiece kEndSentence("</s>");
 
 class Arc {
   public:
     Arc() {}
 
     // For arcs from one vertex to another.  
-    void SetPhrase(Vertex &from, Vertex &to, const Sentences &intersect) {
+    void SetPhrase(detail::Vertex &from, detail::Vertex &to, const Sentences &intersect) {
       Set(to, intersect);
       from_ = &from;
     }
@@ -71,7 +71,7 @@ class Arc {
      * aligned).  These have no from_ vertex; it implictly matches every
      * sentence.  This also handles when the n-gram is a substring of a phrase. 
      */
-    void SetRight(Vertex &to, const Sentences &complete) {
+    void SetRight(detail::Vertex &to, const Sentences &complete) {
       Set(to, complete);
       from_ = NULL;
     }
@@ -97,11 +97,11 @@ class Arc {
     void LowerBound(const Sentence to);
 
   private:
-    void Set(Vertex &to, const Sentences &sentences);
+    void Set(detail::Vertex &to, const Sentences &sentences);
 
     const Sentence *current_;
     const Sentence *last_;
-    Vertex *from_;
+    detail::Vertex *from_;
 };
 
 struct ArcGreater : public std::binary_function<const Arc *, const Arc *, bool> {
@@ -183,7 +183,13 @@ void Vertex::LowerBound(const Sentence to) {
   }
 }
 
-void BuildGraph(const Substrings &phrase, const std::vector<Hash> &hashes, Vertex *const vertices, Arc *free_arc) {
+} // namespace detail
+
+namespace {
+
+void BuildGraph(const Substrings &phrase, const std::vector<Hash> &hashes, detail::Vertex *const vertices, detail::Arc *free_arc) {
+  using detail::Vertex;
+  using detail::Arc;
   assert(!hashes.empty());
 
   const Hash *const first_word = &*hashes.begin();
@@ -231,17 +237,29 @@ void BuildGraph(const Substrings &phrase, const std::vector<Hash> &hashes, Verte
 
 namespace detail {
 
-} // namespace detail
+// Here instead of header due to forward declaration.
+ConditionCommon::ConditionCommon(const Substrings &substrings) : substrings_(substrings) {}
 
-bool Union::Evaluate() {
+// Rest of the variables are temporaries anyway
+ConditionCommon::ConditionCommon(const ConditionCommon &from) : substrings_(from.substrings_) {}
+
+ConditionCommon::~ConditionCommon() {}
+
+detail::Vertex &ConditionCommon::MakeGraph() {
   assert(!hashes_.empty());
-  // Usually there are at most 6 words in an n-gram, so stack allocation is reasonable.  
-  Vertex vertices[hashes_.size()];
+  vertices_.clear();
+  vertices_.resize(hashes_.size());
+  arcs_.clear();
   // One for every substring.  
-  Arc arcs[((hashes_.size() + 1) * hashes_.size()) / 2];
-  BuildGraph(substrings_, hashes_, vertices, arcs);
-  Vertex &last_vertex = vertices[hashes_.size() - 1];
+  arcs_.resize(((hashes_.size() + 1) * hashes_.size()) / 2);
+  BuildGraph(substrings_, hashes_, &*vertices_.begin(), &*arcs_.begin());
+  return vertices_[hashes_.size() - 1];
+}
+
+} // namespace detail
 
+bool Union::Evaluate() {
+  detail::Vertex &last_vertex = MakeGraph();
   unsigned int lower = 0;
   while (true) {
     last_vertex.LowerBound(lower);
@@ -252,14 +270,7 @@ bool Union::Evaluate() {
 }
 
 template <class Output> void Multiple::Evaluate(const StringPiece &line, Output &output) {
-  assert(!hashes_.empty());
-  // Usually there are at most 6 words in an n-gram, so stack allocation is reasonable.  
-  Vertex vertices[hashes_.size()];
-  // One for every substring.  
-  Arc arcs[((hashes_.size() + 1) * hashes_.size()) / 2];
-  BuildGraph(substrings_, hashes_, vertices, arcs);
-  Vertex &last_vertex = vertices[hashes_.size() - 1];
-
+  detail::Vertex &last_vertex = MakeGraph();
   unsigned int lower = 0;
   while (true) {
     last_vertex.LowerBound(lower);
diff --git a/klm/lm/filter/phrase.hh b/klm/lm/filter/phrase.hh
index b4edff41..e8e85835 100644
--- a/klm/lm/filter/phrase.hh
+++ b/klm/lm/filter/phrase.hh
@@ -103,11 +103,33 @@ template <class Iterator> void MakeHashes(Iterator i, const Iterator &end, std::
   }
 }
 
+class Vertex;
+class Arc;
+
+class ConditionCommon {
+  protected:
+    ConditionCommon(const Substrings &substrings);
+    ConditionCommon(const ConditionCommon &from);
+
+    ~ConditionCommon();
+
+    detail::Vertex &MakeGraph();
+
+    // Temporaries in PassNGram and Evaluate to avoid reallocation.
+    std::vector<Hash> hashes_;
+
+  private:
+    std::vector<detail::Vertex> vertices_;
+    std::vector<detail::Arc> arcs_;
+
+    const Substrings &substrings_;
+};
+
 } // namespace detail
 
-class Union {
+class Union : public detail::ConditionCommon {
   public:
-    explicit Union(const Substrings &substrings) : substrings_(substrings) {}
+    explicit Union(const Substrings &substrings) : detail::ConditionCommon(substrings) {}
 
     template <class Iterator> bool PassNGram(const Iterator &begin, const Iterator &end) {
       detail::MakeHashes(begin, end, hashes_);
@@ -116,23 +138,19 @@ class Union {
 
   private:
     bool Evaluate();
-
-    std::vector<Hash> hashes_;
-
-    const Substrings &substrings_;
 };
 
-class Multiple {
+class Multiple : public detail::ConditionCommon {
   public:
-    explicit Multiple(const Substrings &substrings) : substrings_(substrings) {}
+    explicit Multiple(const Substrings &substrings) : detail::ConditionCommon(substrings) {}
 
     template <class Iterator, class Output> void AddNGram(const Iterator &begin, const Iterator &end, const StringPiece &line, Output &output) {
       detail::MakeHashes(begin, end, hashes_);
       if (hashes_.empty()) {
         output.AddNGram(line);
-        return;
+      } else {
+        Evaluate(line, output);
       }
-      Evaluate(line, output);
     }
 
     template <class Output> void AddNGram(const StringPiece &ngram, const StringPiece &line, Output &output) {
@@ -143,10 +161,6 @@ class Multiple {
 
   private:
     template <class Output> void Evaluate(const StringPiece &line, Output &output);
-
-    std::vector<Hash> hashes_;
-
-    const Substrings &substrings_;
 };
 
 } // namespace phrase
diff --git a/klm/lm/filter/phrase_table_vocab_main.cc b/klm/lm/filter/phrase_table_vocab_main.cc
new file mode 100644
index 00000000..e0f47d89
--- /dev/null
+++ b/klm/lm/filter/phrase_table_vocab_main.cc
@@ -0,0 +1,165 @@
+#include "util/fake_ofstream.hh"
+#include "util/file_piece.hh"
+#include "util/murmur_hash.hh"
+#include "util/pool.hh"
+#include "util/string_piece.hh"
+#include "util/string_piece_hash.hh"
+#include "util/tokenize_piece.hh"
+
+#include <boost/unordered_map.hpp>
+#include <boost/unordered_set.hpp>
+
+#include <cstddef>
+#include <vector>
+
+namespace {
+
+struct MutablePiece {
+  mutable StringPiece behind;
+  bool operator==(const MutablePiece &other) const {
+    return behind == other.behind;
+  }
+};
+
+std::size_t hash_value(const MutablePiece &m) {
+  return hash_value(m.behind);
+}
+
+class InternString {
+  public:
+    const char *Add(StringPiece str) {
+      MutablePiece mut;
+      mut.behind = str;
+      std::pair<boost::unordered_set<MutablePiece>::iterator, bool> res(strs_.insert(mut));
+      if (res.second) {
+        void *mem = backing_.Allocate(str.size() + 1);
+        memcpy(mem, str.data(), str.size());
+        static_cast<char*>(mem)[str.size()] = 0;
+        res.first->behind = StringPiece(static_cast<char*>(mem), str.size());
+      }
+      return res.first->behind.data();
+    }
+
+  private:
+    util::Pool backing_;
+    boost::unordered_set<MutablePiece> strs_;
+};
+
+class TargetWords {
+  public:
+    void Introduce(StringPiece source) {
+      vocab_.resize(vocab_.size() + 1);
+      std::vector<unsigned int> temp(1, vocab_.size() - 1);
+      Add(temp, source);
+    }
+
+    void Add(const std::vector<unsigned int> &sentences, StringPiece target) {
+      if (sentences.empty()) return;
+      interns_.clear();
+      for (util::TokenIter<util::SingleCharacter, true> i(target, ' '); i; ++i) {
+        interns_.push_back(intern_.Add(*i));
+      }
+      for (std::vector<unsigned int>::const_iterator i(sentences.begin()); i != sentences.end(); ++i) {
+        boost::unordered_set<const char *> &vocab = vocab_[*i];
+        for (std::vector<const char *>::const_iterator j = interns_.begin(); j != interns_.end(); ++j) {
+          vocab.insert(*j);
+        }
+      }
+    }
+
+    void Print() const {
+      util::FakeOFStream out(1);
+      for (std::vector<boost::unordered_set<const char *> >::const_iterator i = vocab_.begin(); i != vocab_.end(); ++i) {
+        for (boost::unordered_set<const char *>::const_iterator j = i->begin(); j != i->end(); ++j) {
+          out << *j << ' ';
+        }
+        out << '\n';
+      }
+    }
+
+  private:
+    InternString intern_;
+
+    std::vector<boost::unordered_set<const char *> > vocab_;
+
+    // Temporary in Add.
+    std::vector<const char *> interns_;
+};
+
+class Input {
+  public:
+    explicit Input(std::size_t max_length) 
+      : max_length_(max_length), sentence_id_(0), empty_() {}
+
+    void AddSentence(StringPiece sentence, TargetWords &targets) {
+      canonical_.clear();
+      starts_.clear();
+      starts_.push_back(0);
+      for (util::TokenIter<util::AnyCharacter, true> i(sentence, StringPiece("\0 \t", 3)); i; ++i) {
+        canonical_.append(i->data(), i->size());
+        canonical_ += ' ';
+        starts_.push_back(canonical_.size());
+      }
+      targets.Introduce(canonical_);
+      for (std::size_t i = 0; i < starts_.size() - 1; ++i) {
+        std::size_t subtract = starts_[i];
+        const char *start = &canonical_[subtract];
+        for (std::size_t j = i + 1; j < std::min(starts_.size(), i + max_length_ + 1); ++j) {
+          map_[util::MurmurHash64A(start, &canonical_[starts_[j]] - start - 1)].push_back(sentence_id_);
+        }
+      }
+      ++sentence_id_;
+    }
+
+    // Assumes single space-delimited phrase with no space at the beginning or end.
+    const std::vector<unsigned int> &Matches(StringPiece phrase) const {
+      Map::const_iterator i = map_.find(util::MurmurHash64A(phrase.data(), phrase.size()));
+      return i == map_.end() ? empty_ : i->second;
+    }
+
+  private:
+    const std::size_t max_length_;
+
+    // hash of phrase is the key, array of sentences is the value.
+    typedef boost::unordered_map<uint64_t, std::vector<unsigned int> > Map;
+    Map map_;
+
+    std::size_t sentence_id_;
+    
+    // Temporaries in AddSentence.
+    std::string canonical_;
+    std::vector<std::size_t> starts_;
+
+    const std::vector<unsigned int> empty_;
+};
+
+} // namespace
+
+int main(int argc, char *argv[]) {
+  if (argc != 2) {
+    std::cerr << "Expected source text on the command line" << std::endl;
+    return 1;
+  }
+  Input input(7);
+  TargetWords targets;
+  try {
+    util::FilePiece inputs(argv[1], &std::cerr);
+    while (true)
+      input.AddSentence(inputs.ReadLine(), targets);
+  } catch (const util::EndOfFileException &e) {}
+
+  util::FilePiece table(0, NULL, &std::cerr);
+  StringPiece line;
+  const StringPiece pipes("|||");
+  while (true) {
+    try {
+      line = table.ReadLine();
+    } catch (const util::EndOfFileException &e) { break; }
+    util::TokenIter<util::MultiCharacter> it(line, pipes);
+    StringPiece source(*it);
+    if (!source.empty() && source[source.size() - 1] == ' ')
+      source.remove_suffix(1);
+    targets.Add(input.Matches(source), *++it);
+  }
+  targets.Print();
+}
diff --git a/klm/lm/filter/vocab.cc b/klm/lm/filter/vocab.cc
index 7ee4e84b..011ab599 100644
--- a/klm/lm/filter/vocab.cc
+++ b/klm/lm/filter/vocab.cc
@@ -4,7 +4,6 @@
 #include <iostream>
 
 #include <ctype.h>
-#include <err.h>
 
 namespace lm {
 namespace vocab {
diff --git a/klm/lm/filter/wrapper.hh b/klm/lm/filter/wrapper.hh
index 90b07a08..eb657501 100644
--- a/klm/lm/filter/wrapper.hh
+++ b/klm/lm/filter/wrapper.hh
@@ -39,17 +39,15 @@ template <class FilterT> class ContextFilter {
     explicit ContextFilter(Filter &backend) : backend_(backend) {}
 
     template <class Output> void AddNGram(const StringPiece &ngram, const StringPiece &line, Output &output) {
-      pieces_.clear();
-      // TODO: this copy could be avoided by a lookahead iterator.
-      std::copy(util::TokenIter<util::SingleCharacter, true>(ngram, ' '), util::TokenIter<util::SingleCharacter, true>::end(), std::back_insert_iterator<std::vector<StringPiece> >(pieces_));
-      backend_.AddNGram(pieces_.begin(), pieces_.end() - !pieces_.empty(), line, output);
+      // Find beginning of string or last space.
+      const char *last_space;
+      for (last_space = ngram.data() + ngram.size() - 1; last_space > ngram.data() && *last_space != ' '; --last_space) {}
+      backend_.AddNGram(StringPiece(ngram.data(), last_space - ngram.data()), line, output);
     }
 
     void Flush() const {}
 
   private:
-    std::vector<StringPiece> pieces_;
-
     Filter backend_;
 };
 
diff --git a/klm/lm/model.cc b/klm/lm/model.cc
index a26654a6..a5a16bf8 100644
--- a/klm/lm/model.cc
+++ b/klm/lm/model.cc
@@ -34,23 +34,17 @@ template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT
   if (static_cast<std::size_t>(start - static_cast<uint8_t*>(base)) != goal_size) UTIL_THROW(FormatLoadException, "The data structures took " << (start - static_cast<uint8_t*>(base)) << " but Size says they should take " << goal_size);
 }
 
-template <class Search, class VocabularyT> GenericModel<Search, VocabularyT>::GenericModel(const char *file, const Config &config) {
-  LoadLM(file, config, *this);
-
-  // g++ prints warnings unless these are fully initialized.
-  State begin_sentence = State();
-  begin_sentence.length = 1;
-  begin_sentence.words[0] = vocab_.BeginSentence();
-  typename Search::Node ignored_node;
-  bool ignored_independent_left;
-  uint64_t ignored_extend_left;
-  begin_sentence.backoff[0] = search_.LookupUnigram(begin_sentence.words[0], ignored_node, ignored_independent_left, ignored_extend_left).Backoff();
-  State null_context = State();
-  null_context.length = 0;
-  P::Init(begin_sentence, null_context, vocab_, search_.Order());
+namespace {
+void ComplainAboutARPA(const Config &config, ModelType model_type) {
+  if (config.write_mmap || !config.messages) return;
+  if (config.arpa_complain == Config::ALL) {
+    *config.messages << "Loading the LM will be faster if you build a binary file." << std::endl;
+  } else if (config.arpa_complain == Config::EXPENSIVE &&
+             (model_type == TRIE || model_type == QUANT_TRIE || model_type == ARRAY_TRIE || model_type == QUANT_ARRAY_TRIE)) {
+    *config.messages << "Building " << kModelNames[model_type] << " from ARPA is expensive.  Save time by building a binary format." << std::endl;
+  }
 }
 
-namespace {
 void CheckCounts(const std::vector<uint64_t> &counts) {
   UTIL_THROW_IF(counts.size() > KENLM_MAX_ORDER, FormatLoadException, "This model has order " << counts.size() << " but KenLM was compiled to support up to " << KENLM_MAX_ORDER << ".  " << KENLM_ORDER_MESSAGE);
   if (sizeof(uint64_t) > sizeof(std::size_t)) {
@@ -59,18 +53,45 @@ void CheckCounts(const std::vector<uint64_t> &counts) {
     }
   }
 }
+
 } // namespace
 
-template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT>::InitializeFromBinary(void *start, const Parameters &params, const Config &config, int fd) {
-  CheckCounts(params.counts);
-  SetupMemory(start, params.counts, config);
-  vocab_.LoadedBinary(params.fixed.has_vocabulary, fd, config.enumerate_vocab);
-  search_.LoadedBinary();
+template <class Search, class VocabularyT> GenericModel<Search, VocabularyT>::GenericModel(const char *file, const Config &init_config) : backing_(init_config) {
+  util::scoped_fd fd(util::OpenReadOrThrow(file));
+  if (IsBinaryFormat(fd.get())) {
+    Parameters parameters;
+    int fd_shallow = fd.release();
+    backing_.InitializeBinary(fd_shallow, kModelType, kVersion, parameters);
+    CheckCounts(parameters.counts);
+
+    Config new_config(init_config);
+    new_config.probing_multiplier = parameters.fixed.probing_multiplier;
+    Search::UpdateConfigFromBinary(backing_, parameters.counts, VocabularyT::Size(parameters.counts[0], new_config), new_config);
+    UTIL_THROW_IF(new_config.enumerate_vocab && !parameters.fixed.has_vocabulary, FormatLoadException, "The decoder requested all the vocabulary strings, but this binary file does not have them.  You may need to rebuild the binary file with an updated version of build_binary.");
+
+    SetupMemory(backing_.LoadBinary(Size(parameters.counts, new_config)), parameters.counts, new_config);
+    vocab_.LoadedBinary(parameters.fixed.has_vocabulary, fd_shallow, new_config.enumerate_vocab, backing_.VocabStringReadingOffset());
+  } else {
+    ComplainAboutARPA(init_config, kModelType);
+    InitializeFromARPA(fd.release(), file, init_config);
+  }
+
+  // g++ prints warnings unless these are fully initialized.
+  State begin_sentence = State();
+  begin_sentence.length = 1;
+  begin_sentence.words[0] = vocab_.BeginSentence();
+  typename Search::Node ignored_node;
+  bool ignored_independent_left;
+  uint64_t ignored_extend_left;
+  begin_sentence.backoff[0] = search_.LookupUnigram(begin_sentence.words[0], ignored_node, ignored_independent_left, ignored_extend_left).Backoff();
+  State null_context = State();
+  null_context.length = 0;
+  P::Init(begin_sentence, null_context, vocab_, search_.Order());
 }
 
-template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT>::InitializeFromARPA(const char *file, const Config &config) {
-  // Backing file is the ARPA.  Steal it so we can make the backing file the mmap output if any.
-  util::FilePiece f(backing_.file.release(), file, config.ProgressMessages());
+template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT>::InitializeFromARPA(int fd, const char *file, const Config &config) {
+  // Backing file is the ARPA.
+  util::FilePiece f(fd, file, config.ProgressMessages());
   try {
     std::vector<uint64_t> counts;
     // File counts do not include pruned trigrams that extend to quadgrams etc.   These will be fixed by search_.
@@ -81,13 +102,17 @@ template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT
 
     std::size_t vocab_size = util::CheckOverflow(VocabularyT::Size(counts[0], config));
     // Setup the binary file for writing the vocab lookup table.  The search_ is responsible for growing the binary file to its needs.
-    vocab_.SetupMemory(SetupJustVocab(config, counts.size(), vocab_size, backing_), vocab_size, counts[0], config);
+    vocab_.SetupMemory(backing_.SetupJustVocab(vocab_size, counts.size()), vocab_size, counts[0], config);
 
-    if (config.write_mmap) {
+    if (config.write_mmap && config.include_vocab) {
       WriteWordsWrapper wrap(config.enumerate_vocab);
       vocab_.ConfigureEnumerate(&wrap, counts[0]);
       search_.InitializeFromARPA(file, f, counts, config, vocab_, backing_);
-      wrap.Write(backing_.file.get(), backing_.vocab.size() + vocab_.UnkCountChangePadding() + Search::Size(counts, config));
+      void *vocab_rebase, *search_rebase;
+      backing_.WriteVocabWords(wrap.Buffer(), vocab_rebase, search_rebase);
+      // Due to writing at the end of file, mmap may have relocated data.  So remap.
+      vocab_.Relocate(vocab_rebase);
+      search_.SetupMemory(reinterpret_cast<uint8_t*>(search_rebase), counts, config);
     } else {
       vocab_.ConfigureEnumerate(config.enumerate_vocab, counts[0]);
       search_.InitializeFromARPA(file, f, counts, config, vocab_, backing_);
@@ -99,18 +124,13 @@ template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT
       search_.UnknownUnigram().backoff = 0.0;
       search_.UnknownUnigram().prob = config.unknown_missing_logprob;
     }
-    FinishFile(config, kModelType, kVersion, counts, vocab_.UnkCountChangePadding(), backing_);
+    backing_.FinishFile(config, kModelType, kVersion, counts);
   } catch (util::Exception &e) {
     e << " Byte: " << f.Offset();
     throw;
   }
 }
 
-template <class Search, class VocabularyT> void GenericModel<Search, VocabularyT>::UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &counts, Config &config) {
-  util::AdvanceOrThrow(fd, VocabularyT::Size(counts[0], config));
-  Search::UpdateConfigFromBinary(fd, counts, config);
-}
-
 template <class Search, class VocabularyT> FullScoreReturn GenericModel<Search, VocabularyT>::FullScore(const State &in_state, const WordIndex new_word, State &out_state) const {
   FullScoreReturn ret = ScoreExceptBackoff(in_state.words, in_state.words + in_state.length, new_word, out_state);
   for (const float *i = in_state.backoff + ret.ngram_length - 1; i < in_state.backoff + in_state.length; ++i) {
diff --git a/klm/lm/model.hh b/klm/lm/model.hh
index 60f55110..e75da93b 100644
--- a/klm/lm/model.hh
+++ b/klm/lm/model.hh
@@ -67,7 +67,7 @@ template <class Search, class VocabularyT> class GenericModel : public base::Mod
     FullScoreReturn FullScoreForgotState(const WordIndex *context_rbegin, const WordIndex *context_rend, const WordIndex new_word, State &out_state) const;
 
     /* Get the state for a context.  Don't use this if you can avoid it.  Use
-     * BeginSentenceState or EmptyContextState and extend from those.  If
+     * BeginSentenceState or NullContextState and extend from those.  If
      * you're only going to use this state to call FullScore once, use
      * FullScoreForgotState. 
      * To use this function, make an array of WordIndex containing the context
@@ -104,10 +104,6 @@ template <class Search, class VocabularyT> class GenericModel : public base::Mod
     }
 
   private:
-    friend void lm::ngram::LoadLM<>(const char *file, const Config &config, GenericModel<Search, VocabularyT> &to);
-
-    static void UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &counts, Config &config);
-
     FullScoreReturn ScoreExceptBackoff(const WordIndex *const context_rbegin, const WordIndex *const context_rend, const WordIndex new_word, State &out_state) const;
 
     // Score bigrams and above.  Do not include backoff.   
@@ -116,15 +112,11 @@ template <class Search, class VocabularyT> class GenericModel : public base::Mod
     // Appears after Size in the cc file.
     void SetupMemory(void *start, const std::vector<uint64_t> &counts, const Config &config);
 
-    void InitializeFromBinary(void *start, const Parameters &params, const Config &config, int fd);
-
-    void InitializeFromARPA(const char *file, const Config &config);
+    void InitializeFromARPA(int fd, const char *file, const Config &config);
 
     float InternalUnRest(const uint64_t *pointers_begin, const uint64_t *pointers_end, unsigned char first_length) const;
 
-    Backing &MutableBacking() { return backing_; }
-
-    Backing backing_;
+    BinaryFormat backing_;
     
     VocabularyT vocab_;
 
diff --git a/klm/lm/model_test.cc b/klm/lm/model_test.cc
index eb159094..7005b05e 100644
--- a/klm/lm/model_test.cc
+++ b/klm/lm/model_test.cc
@@ -360,10 +360,11 @@ BOOST_AUTO_TEST_CASE(quant_bhiksha_trie) {
   LoadingTest<QuantArrayTrieModel>();
 }
 
-template <class ModelT> void BinaryTest() {
+template <class ModelT> void BinaryTest(Config::WriteMethod write_method) {
   Config config;
   config.write_mmap = "test.binary";
   config.messages = NULL;
+  config.write_method = write_method;
   ExpectEnumerateVocab enumerate;
   config.enumerate_vocab = &enumerate;
 
@@ -406,6 +407,11 @@ template <class ModelT> void BinaryTest() {
   unlink("test_nounk.binary");
 }
 
+template <class ModelT> void BinaryTest() {
+  BinaryTest<ModelT>(Config::WRITE_MMAP);
+  BinaryTest<ModelT>(Config::WRITE_AFTER);
+}
+
 BOOST_AUTO_TEST_CASE(write_and_read_probing) {
   BinaryTest<ProbingModel>();
 }
diff --git a/klm/lm/ngram_query.hh b/klm/lm/ngram_query.hh
index dfcda170..ec2590f4 100644
--- a/klm/lm/ngram_query.hh
+++ b/klm/lm/ngram_query.hh
@@ -11,21 +11,25 @@
 #include <istream>
 #include <string>
 
+#include <math.h>
+
 namespace lm {
 namespace ngram {
 
 template <class Model> void Query(const Model &model, bool sentence_context, std::istream &in_stream, std::ostream &out_stream) {
-  std::cerr << "Loading statistics:\n";
-  util::PrintUsage(std::cerr);
   typename Model::State state, out;
   lm::FullScoreReturn ret;
   std::string word;
 
+  double corpus_total = 0.0;
+  uint64_t corpus_oov = 0;
+  uint64_t corpus_tokens = 0;
+
   while (in_stream) {
     state = sentence_context ? model.BeginSentenceState() : model.NullContextState();
     float total = 0.0;
     bool got = false;
-    unsigned int oov = 0;
+    uint64_t oov = 0;
     while (in_stream >> word) {
       got = true;
       lm::WordIndex vocab = model.GetVocabulary().Index(word);
@@ -33,6 +37,7 @@ template <class Model> void Query(const Model &model, bool sentence_context, std
       ret = model.FullScore(state, vocab, out);
       total += ret.prob;
       out_stream << word << '=' << vocab << ' ' << static_cast<unsigned int>(ret.ngram_length)  << ' ' << ret.prob << '\t';
+      ++corpus_tokens;
       state = out;
       char c;
       while (true) {
@@ -50,12 +55,14 @@ template <class Model> void Query(const Model &model, bool sentence_context, std
     if (sentence_context) {
       ret = model.FullScore(state, model.GetVocabulary().EndSentence(), out);
       total += ret.prob;
+      ++corpus_tokens;
       out_stream << "</s>=" << model.GetVocabulary().EndSentence() << ' ' << static_cast<unsigned int>(ret.ngram_length)  << ' ' << ret.prob << '\t';
     }
     out_stream << "Total: " << total << " OOV: " << oov << '\n';
+    corpus_total += total;
+    corpus_oov += oov;
   }
-  std::cerr << "After queries:\n";
-  util::PrintUsage(std::cerr);
+  out_stream << "Perplexity " << pow(10.0, -(corpus_total / static_cast<double>(corpus_tokens))) << std::endl;
 }
 
 template <class M> void Query(const char *file, bool sentence_context, std::istream &in_stream, std::ostream &out_stream) {
diff --git a/klm/lm/quantize.cc b/klm/lm/quantize.cc
index b58c3f3f..273ea398 100644
--- a/klm/lm/quantize.cc
+++ b/klm/lm/quantize.cc
@@ -38,13 +38,13 @@ const char kSeparatelyQuantizeVersion = 2;
 
 } // namespace
 
-void SeparatelyQuantize::UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &/*counts*/, Config &config) {
-  char version;
-  util::ReadOrThrow(fd, &version, 1);
-  util::ReadOrThrow(fd, &config.prob_bits, 1);
-  util::ReadOrThrow(fd, &config.backoff_bits, 1);
+void SeparatelyQuantize::UpdateConfigFromBinary(const BinaryFormat &file, uint64_t offset, Config &config) {
+  unsigned char buffer[3];
+  file.ReadForConfig(buffer, 3, offset);
+  char version = buffer[0];
+  config.prob_bits = buffer[1];
+  config.backoff_bits = buffer[2];
   if (version != kSeparatelyQuantizeVersion) UTIL_THROW(FormatLoadException, "This file has quantization version " << (unsigned)version << " but the code expects version " << (unsigned)kSeparatelyQuantizeVersion);
-  util::AdvanceOrThrow(fd, -3);
 }
 
 void SeparatelyQuantize::SetupMemory(void *base, unsigned char order, const Config &config) {
diff --git a/klm/lm/quantize.hh b/klm/lm/quantize.hh
index 8ce2378a..9d3a2f43 100644
--- a/klm/lm/quantize.hh
+++ b/klm/lm/quantize.hh
@@ -18,12 +18,13 @@ namespace lm {
 namespace ngram {
 
 struct Config;
+class BinaryFormat;
 
 /* Store values directly and don't quantize. */
 class DontQuantize {
   public:
     static const ModelType kModelTypeAdd = static_cast<ModelType>(0);
-    static void UpdateConfigFromBinary(int, const std::vector<uint64_t> &, Config &) {}
+    static void UpdateConfigFromBinary(const BinaryFormat &, uint64_t, Config &) {}
     static uint64_t Size(uint8_t /*order*/, const Config &/*config*/) { return 0; }
     static uint8_t MiddleBits(const Config &/*config*/) { return 63; }
     static uint8_t LongestBits(const Config &/*config*/) { return 31; }
@@ -136,7 +137,7 @@ class SeparatelyQuantize {
   public:
     static const ModelType kModelTypeAdd = kQuantAdd;
 
-    static void UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &counts, Config &config);
+    static void UpdateConfigFromBinary(const BinaryFormat &file, uint64_t offset, Config &config);
 
     static uint64_t Size(uint8_t order, const Config &config) {
       uint64_t longest_table = (static_cast<uint64_t>(1) << static_cast<uint64_t>(config.prob_bits)) * sizeof(float);
diff --git a/klm/lm/query_main.cc b/klm/lm/query_main.cc
index 27d3a1a5..bd4fde62 100644
--- a/klm/lm/query_main.cc
+++ b/klm/lm/query_main.cc
@@ -1,42 +1,65 @@
 #include "lm/ngram_query.hh"
 
+#ifdef WITH_NPLM
+#include "lm/wrappers/nplm.hh"
+#endif
+
+#include <stdlib.h>
+
+void Usage(const char *name) {
+  std::cerr << "KenLM was compiled with maximum order " << KENLM_MAX_ORDER << "." << std::endl;
+  std::cerr << "Usage: " << name << " [-n] lm_file" << std::endl;
+  std::cerr << "Input is wrapped in <s> and </s> unless -n is passed." << std::endl;
+  exit(1);
+}
+
 int main(int argc, char *argv[]) {
-  if (!(argc == 2 || (argc == 3 && !strcmp(argv[2], "null")))) {
-    std::cerr << "KenLM was compiled with maximum order " << KENLM_MAX_ORDER << "." << std::endl;
-    std::cerr << "Usage: " << argv[0] << " lm_file [null]" << std::endl;
-    std::cerr << "Input is wrapped in <s> and </s> unless null is passed." << std::endl;
-    return 1;
+  bool sentence_context = true;
+  const char *file = NULL;
+  for (char **arg = argv + 1; arg != argv + argc; ++arg) {
+    if (!strcmp(*arg, "-n")) {
+      sentence_context = false;
+    } else if (!strcmp(*arg, "-h") || !strcmp(*arg, "--help") || file) {
+      Usage(argv[0]);
+    } else {
+      file = *arg;
+    }
   }
+  if (!file) Usage(argv[0]);
   try {
-    bool sentence_context = (argc == 2);
     using namespace lm::ngram;
     ModelType model_type;
-    if (RecognizeBinary(argv[1], model_type)) {
+    if (RecognizeBinary(file, model_type)) {
       switch(model_type) {
         case PROBING:
-          Query<lm::ngram::ProbingModel>(argv[1], sentence_context, std::cin, std::cout);
+          Query<lm::ngram::ProbingModel>(file, sentence_context, std::cin, std::cout);
           break;
         case REST_PROBING:
-          Query<lm::ngram::RestProbingModel>(argv[1], sentence_context, std::cin, std::cout);
+          Query<lm::ngram::RestProbingModel>(file, sentence_context, std::cin, std::cout);
           break;
         case TRIE:
-          Query<TrieModel>(argv[1], sentence_context, std::cin, std::cout);
+          Query<TrieModel>(file, sentence_context, std::cin, std::cout);
           break;
         case QUANT_TRIE:
-          Query<QuantTrieModel>(argv[1], sentence_context, std::cin, std::cout);
+          Query<QuantTrieModel>(file, sentence_context, std::cin, std::cout);
           break;
         case ARRAY_TRIE:
-          Query<ArrayTrieModel>(argv[1], sentence_context, std::cin, std::cout);
+          Query<ArrayTrieModel>(file, sentence_context, std::cin, std::cout);
           break;
         case QUANT_ARRAY_TRIE:
-          Query<QuantArrayTrieModel>(argv[1], sentence_context, std::cin, std::cout);
+          Query<QuantArrayTrieModel>(file, sentence_context, std::cin, std::cout);
           break;
         default:
           std::cerr << "Unrecognized kenlm model type " << model_type << std::endl;
           abort();
       }
+#ifdef WITH_NPLM
+    } else if (lm::np::Model::Recognize(file)) {
+      lm::np::Model model(file);
+      Query(model, sentence_context, std::cin, std::cout);
+#endif
     } else {
-      Query<ProbingModel>(argv[1], sentence_context, std::cin, std::cout);
+      Query<ProbingModel>(file, sentence_context, std::cin, std::cout);
     }
     std::cerr << "Total time including destruction:\n";
     util::PrintUsage(std::cerr);
diff --git a/klm/lm/read_arpa.cc b/klm/lm/read_arpa.cc
index 9ea08798..fb8bbfa2 100644
--- a/klm/lm/read_arpa.cc
+++ b/klm/lm/read_arpa.cc
@@ -19,7 +19,7 @@
 
 namespace lm {
 
-// 1 for '\t', '\n', and ' '.  This is stricter than isspace.  
+// 1 for '\t', '\n', and ' '.  This is stricter than isspace.
 const bool kARPASpaces[256] = {0,0,0,0,0,0,0,0,0,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 
 namespace {
@@ -50,7 +50,7 @@ void ReadARPACounts(util::FilePiece &in, std::vector<uint64_t> &number) {
   // In general, ARPA files can have arbitrary text before "\data\"
   // But in KenLM, we require such lines to start with "#", so that
   // we can do stricter error checking
-  while (IsEntirelyWhiteSpace(line) || line.starts_with("#")) {
+  while (IsEntirelyWhiteSpace(line) || starts_with(line, "#")) {
     line = in.ReadLine();
   }
 
@@ -58,7 +58,7 @@ void ReadARPACounts(util::FilePiece &in, std::vector<uint64_t> &number) {
     if ((line.size() >= 2) && (line.data()[0] == 0x1f) && (static_cast<unsigned char>(line.data()[1]) == 0x8b)) {
       UTIL_THROW(FormatLoadException, "Looks like a gzip file.  If this is an ARPA file, pipe " << in.FileName() << " through zcat.  If this already in binary format, you need to decompress it because mmap doesn't work on top of gzip.");
     }
-    if (static_cast<size_t>(line.size()) >= strlen(kBinaryMagic) && StringPiece(line.data(), strlen(kBinaryMagic)) == kBinaryMagic) 
+    if (static_cast<size_t>(line.size()) >= strlen(kBinaryMagic) && StringPiece(line.data(), strlen(kBinaryMagic)) == kBinaryMagic)
       UTIL_THROW(FormatLoadException, "This looks like a binary file but got sent to the ARPA parser.  Did you compress the binary file or pass a binary file where only ARPA files are accepted?");
     UTIL_THROW_IF(line.size() >= 4 && StringPiece(line.data(), 4) == "blmt", FormatLoadException, "This looks like an IRSTLM binary file.  Did you forget to pass --text yes to compile-lm?");
     UTIL_THROW_IF(line == "iARPA", FormatLoadException, "This looks like an IRSTLM iARPA file.  You need an ARPA file.  Run\n  compile-lm --text yes " << in.FileName() << " " << in.FileName() << ".arpa\nfirst.");
@@ -66,7 +66,7 @@ void ReadARPACounts(util::FilePiece &in, std::vector<uint64_t> &number) {
   }
   while (!IsEntirelyWhiteSpace(line = in.ReadLine())) {
     if (line.size() < 6 || strncmp(line.data(), "ngram ", 6)) UTIL_THROW(FormatLoadException, "count line \"" << line << "\"doesn't begin with \"ngram \"");
-    // So strtol doesn't go off the end of line.  
+    // So strtol doesn't go off the end of line.
     std::string remaining(line.data() + 6, line.size() - 6);
     char *end_ptr;
     unsigned int length = std::strtol(remaining.c_str(), &end_ptr, 10);
@@ -102,8 +102,8 @@ void ReadBackoff(util::FilePiece &in, Prob &/*weights*/) {
 }
 
 void ReadBackoff(util::FilePiece &in, float &backoff) {
-  // Always make zero negative.  
-  // Negative zero means that no (n+1)-gram has this n-gram as context.  
+  // Always make zero negative.
+  // Negative zero means that no (n+1)-gram has this n-gram as context.
   // Therefore the hypothesis state can be shorter.  Of course, many n-grams
   // are context for (n+1)-grams.  An algorithm in the data structure will go
   // back and set the backoff to positive zero in these cases.
@@ -150,7 +150,7 @@ void PositiveProbWarn::Warn(float prob) {
     case THROW_UP:
       UTIL_THROW(FormatLoadException, "Positive log probability " << prob << " in the model.  This is a bug in IRSTLM; you can set config.positive_log_probability = SILENT or pass -i to build_binary to substitute 0.0 for the log probability.  Error");
     case COMPLAIN:
-      std::cerr << "There's a positive log probability " << prob << " in the APRA file, probably because of a bug in IRSTLM.  This and subsequent entires will be mapepd to 0 log probability." << std::endl;
+      std::cerr << "There's a positive log probability " << prob << " in the APRA file, probably because of a bug in IRSTLM.  This and subsequent entires will be mapped to 0 log probability." << std::endl;
       action_ = SILENT;
       break;
     case SILENT:
diff --git a/klm/lm/search_hashed.cc b/klm/lm/search_hashed.cc
index 62275d27..354a56b4 100644
--- a/klm/lm/search_hashed.cc
+++ b/klm/lm/search_hashed.cc
@@ -204,9 +204,10 @@ template <class Build, class Activate, class Store> void ReadNGrams(
 namespace detail {
 
 template <class Value> uint8_t *HashedSearch<Value>::SetupMemory(uint8_t *start, const std::vector<uint64_t> &counts, const Config &config) {
-  std::size_t allocated = Unigram::Size(counts[0]);
-  unigram_ = Unigram(start, counts[0], allocated);
-  start += allocated;
+  unigram_ = Unigram(start, counts[0]);
+  start += Unigram::Size(counts[0]);
+  std::size_t allocated;
+  middle_.clear();
   for (unsigned int n = 2; n < counts.size(); ++n) {
     allocated = Middle::Size(counts[n - 1], config.probing_multiplier);
     middle_.push_back(Middle(start, allocated));
@@ -218,9 +219,21 @@ template <class Value> uint8_t *HashedSearch<Value>::SetupMemory(uint8_t *start,
   return start;
 }
 
-template <class Value> void HashedSearch<Value>::InitializeFromARPA(const char * /*file*/, util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, ProbingVocabulary &vocab, Backing &backing) {
-  // TODO: fix sorted.
-  SetupMemory(GrowForSearch(config, vocab.UnkCountChangePadding(), Size(counts, config), backing), counts, config);
+/*template <class Value> void HashedSearch<Value>::Relocate(uint8_t *start, const std::vector<uint64_t> &counts, const Config &config) {
+  unigram_ = Unigram(start, counts[0]);
+  start += Unigram::Size(counts[0]);
+  for (unsigned int n = 2; n < counts.size(); ++n) {
+    middle[n-2].Relocate(start);
+    start += Middle::Size(counts[n - 1], config.probing_multiplier)
+  }
+  longest_.Relocate(start);
+}*/
+
+template <class Value> void HashedSearch<Value>::InitializeFromARPA(const char * /*file*/, util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, ProbingVocabulary &vocab, BinaryFormat &backing) {
+  void *vocab_rebase;
+  void *search_base = backing.GrowForSearch(Size(counts, config), vocab.UnkCountChangePadding(), vocab_rebase);
+  vocab.Relocate(vocab_rebase);
+  SetupMemory(reinterpret_cast<uint8_t*>(search_base), counts, config);
 
   PositiveProbWarn warn(config.positive_log_probability);
   Read1Grams(f, counts[0], vocab, unigram_.Raw(), warn);
@@ -277,14 +290,6 @@ template <class Value> template <class Build> void HashedSearch<Value>::ApplyBui
   ReadEnd(f);
 }
 
-template <class Value> void HashedSearch<Value>::LoadedBinary() {
-  unigram_.LoadedBinary();
-  for (typename std::vector<Middle>::iterator i = middle_.begin(); i != middle_.end(); ++i) {
-    i->LoadedBinary();
-  }
-  longest_.LoadedBinary();
-}
-
 template class HashedSearch<BackoffValue>;
 template class HashedSearch<RestValue>;
 
diff --git a/klm/lm/search_hashed.hh b/klm/lm/search_hashed.hh
index 9d067bc2..8193262b 100644
--- a/klm/lm/search_hashed.hh
+++ b/klm/lm/search_hashed.hh
@@ -18,7 +18,7 @@ namespace util { class FilePiece; }
 
 namespace lm {
 namespace ngram {
-struct Backing;
+class BinaryFormat;
 class ProbingVocabulary;
 namespace detail {
 
@@ -72,7 +72,7 @@ template <class Value> class HashedSearch {
     static const unsigned int kVersion = 0;
 
     // TODO: move probing_multiplier here with next binary file format update.
-    static void UpdateConfigFromBinary(int, const std::vector<uint64_t> &, Config &) {}
+    static void UpdateConfigFromBinary(const BinaryFormat &, const std::vector<uint64_t> &, uint64_t, Config &) {}
 
     static uint64_t Size(const std::vector<uint64_t> &counts, const Config &config) {
       uint64_t ret = Unigram::Size(counts[0]);
@@ -84,9 +84,7 @@ template <class Value> class HashedSearch {
 
     uint8_t *SetupMemory(uint8_t *start, const std::vector<uint64_t> &counts, const Config &config);
 
-    void InitializeFromARPA(const char *file, util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, ProbingVocabulary &vocab, Backing &backing);
-
-    void LoadedBinary();
+    void InitializeFromARPA(const char *file, util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, ProbingVocabulary &vocab, BinaryFormat &backing);
 
     unsigned char Order() const {
       return middle_.size() + 2;
@@ -148,7 +146,7 @@ template <class Value> class HashedSearch {
       public:
         Unigram() {}
 
-        Unigram(void *start, uint64_t count, std::size_t /*allocated*/) :
+        Unigram(void *start, uint64_t count) :
           unigram_(static_cast<typename Value::Weights*>(start))
 #ifdef DEBUG
          ,  count_(count)
@@ -168,8 +166,6 @@ template <class Value> class HashedSearch {
 
         typename Value::Weights &Unknown() { return unigram_[0]; }
 
-        void LoadedBinary() {}
-
         // For building.
         typename Value::Weights *Raw() { return unigram_; }
 
diff --git a/klm/lm/search_trie.cc b/klm/lm/search_trie.cc
index 1b0d9b26..4a88194e 100644
--- a/klm/lm/search_trie.cc
+++ b/klm/lm/search_trie.cc
@@ -253,11 +253,6 @@ class FindBlanks {
       ++counts_.back();
     }
 
-    // Unigrams wrote one past.
-    void Cleanup() {
-      --counts_[0];
-    }
-
     const std::vector<uint64_t> &Counts() const {
       return counts_;
     }
@@ -310,8 +305,6 @@ template <class Quant, class Bhiksha> class WriteEntries {
       typename Quant::LongestPointer(quant_, longest_.Insert(words[order_ - 1])).Write(reinterpret_cast<const Prob*>(words + order_)->prob);
     }
 
-    void Cleanup() {}
-
   private:
     RecordReader *contexts_;
     const Quant &quant_;
@@ -385,14 +378,14 @@ template <class Doing> void RecursiveInsert(const unsigned char total_order, con
   util::ErsatzProgress progress(unigram_count + 1, progress_out, message);
   WordIndex unigram = 0;
   std::priority_queue<Gram> grams;
-  grams.push(Gram(&unigram, 1));
+  if (unigram_count) grams.push(Gram(&unigram, 1));
   for (unsigned char i = 2; i <= total_order; ++i) {
     if (input[i-2]) grams.push(Gram(reinterpret_cast<const WordIndex*>(input[i-2].Data()), i));
   }
 
   BlankManager<Doing> blank(total_order, doing);
 
-  while (true) {
+  while (!grams.empty()) {
     Gram top = grams.top();
     grams.pop();
     unsigned char order = top.end - top.begin;
@@ -400,8 +393,7 @@ template <class Doing> void RecursiveInsert(const unsigned char total_order, con
       blank.Visit(&unigram, 1, doing.UnigramProb(unigram));
       doing.Unigram(unigram);
       progress.Set(unigram);
-      if (++unigram == unigram_count + 1) break;
-      grams.push(top);
+      if (++unigram < unigram_count) grams.push(top);
     } else {
       if (order == total_order) {
         blank.Visit(top.begin, order, reinterpret_cast<const Prob*>(top.end)->prob);
@@ -414,8 +406,6 @@ template <class Doing> void RecursiveInsert(const unsigned char total_order, con
       if (++reader) grams.push(top);
     }
   }
-  assert(grams.empty());
-  doing.Cleanup();
 }
 
 void SanityCheckCounts(const std::vector<uint64_t> &initial, const std::vector<uint64_t> &fixed) {
@@ -469,7 +459,7 @@ void PopulateUnigramWeights(FILE *file, WordIndex unigram_count, RecordReader &c
 
 } // namespace
 
-template <class Quant, class Bhiksha> void BuildTrie(SortedFiles &files, std::vector<uint64_t> &counts, const Config &config, TrieSearch<Quant, Bhiksha> &out, Quant &quant, const SortedVocabulary &vocab, Backing &backing) {
+template <class Quant, class Bhiksha> void BuildTrie(SortedFiles &files, std::vector<uint64_t> &counts, const Config &config, TrieSearch<Quant, Bhiksha> &out, Quant &quant, SortedVocabulary &vocab, BinaryFormat &backing) {
   RecordReader inputs[KENLM_MAX_ORDER - 1];
   RecordReader contexts[KENLM_MAX_ORDER - 1];
 
@@ -498,7 +488,10 @@ template <class Quant, class Bhiksha> void BuildTrie(SortedFiles &files, std::ve
 
   sri.ObtainBackoffs(counts.size(), unigram_file.get(), inputs);
 
-  out.SetupMemory(GrowForSearch(config, vocab.UnkCountChangePadding(), TrieSearch<Quant, Bhiksha>::Size(fixed_counts, config), backing), fixed_counts, config);
+  void *vocab_relocate;
+  void *search_base = backing.GrowForSearch(TrieSearch<Quant, Bhiksha>::Size(fixed_counts, config), vocab.UnkCountChangePadding(), vocab_relocate);
+  vocab.Relocate(vocab_relocate);
+  out.SetupMemory(reinterpret_cast<uint8_t*>(search_base), fixed_counts, config);
 
   for (unsigned char i = 2; i <= counts.size(); ++i) {
     inputs[i-2].Rewind();
@@ -524,6 +517,8 @@ template <class Quant, class Bhiksha> void BuildTrie(SortedFiles &files, std::ve
   {
     WriteEntries<Quant, Bhiksha> writer(contexts, quant, unigrams, out.middle_begin_, out.longest_, counts.size(), sri);
     RecursiveInsert(counts.size(), counts[0], inputs, config.ProgressMessages(), "Writing trie", writer);
+    // Write the last unigram entry, which is the end pointer for the bigrams.  
+    writer.Unigram(counts[0]);
   }
 
   // Do not disable this error message or else too little state will be returned.  Both WriteEntries::Middle and returning state based on found n-grams will need to be fixed to handle this situation.
@@ -579,15 +574,7 @@ template <class Quant, class Bhiksha> uint8_t *TrieSearch<Quant, Bhiksha>::Setup
   return start + Longest::Size(Quant::LongestBits(config), counts.back(), counts[0]);
 }
 
-template <class Quant, class Bhiksha> void TrieSearch<Quant, Bhiksha>::LoadedBinary() {
-  unigram_.LoadedBinary();
-  for (Middle *i = middle_begin_; i != middle_end_; ++i) {
-    i->LoadedBinary();
-  }
-  longest_.LoadedBinary();
-}
-
-template <class Quant, class Bhiksha> void TrieSearch<Quant, Bhiksha>::InitializeFromARPA(const char *file, util::FilePiece &f, std::vector<uint64_t> &counts, const Config &config, SortedVocabulary &vocab, Backing &backing) {
+template <class Quant, class Bhiksha> void TrieSearch<Quant, Bhiksha>::InitializeFromARPA(const char *file, util::FilePiece &f, std::vector<uint64_t> &counts, const Config &config, SortedVocabulary &vocab, BinaryFormat &backing) {
   std::string temporary_prefix;
   if (config.temporary_directory_prefix) {
     temporary_prefix = config.temporary_directory_prefix;
diff --git a/klm/lm/search_trie.hh b/klm/lm/search_trie.hh
index 60be416b..299262a5 100644
--- a/klm/lm/search_trie.hh
+++ b/klm/lm/search_trie.hh
@@ -11,18 +11,19 @@
 #include "util/file_piece.hh"
 
 #include <vector>
+#include <cstdlib>
 
 #include <assert.h>
 
 namespace lm {
 namespace ngram {
-struct Backing;
+class BinaryFormat;
 class SortedVocabulary;
 namespace trie {
 
 template <class Quant, class Bhiksha> class TrieSearch;
 class SortedFiles;
-template <class Quant, class Bhiksha> void BuildTrie(SortedFiles &files, std::vector<uint64_t> &counts, const Config &config, TrieSearch<Quant, Bhiksha> &out, Quant &quant, const SortedVocabulary &vocab, Backing &backing);
+template <class Quant, class Bhiksha> void BuildTrie(SortedFiles &files, std::vector<uint64_t> &counts, const Config &config, TrieSearch<Quant, Bhiksha> &out, Quant &quant, SortedVocabulary &vocab, BinaryFormat &backing);
 
 template <class Quant, class Bhiksha> class TrieSearch {
   public:
@@ -38,11 +39,11 @@ template <class Quant, class Bhiksha> class TrieSearch {
 
     static const unsigned int kVersion = 1;
 
-    static void UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &counts, Config &config) {
-      Quant::UpdateConfigFromBinary(fd, counts, config);
-      util::AdvanceOrThrow(fd, Quant::Size(counts.size(), config) + Unigram::Size(counts[0]));
+    static void UpdateConfigFromBinary(const BinaryFormat &file, const std::vector<uint64_t> &counts, uint64_t offset, Config &config) {
+      Quant::UpdateConfigFromBinary(file, offset, config);
       // Currently the unigram pointers are not compresssed, so there will only be a header for order > 2.
-      if (counts.size() > 2) Bhiksha::UpdateConfigFromBinary(fd, config);
+      if (counts.size() > 2)
+        Bhiksha::UpdateConfigFromBinary(file, offset + Quant::Size(counts.size(), config) + Unigram::Size(counts[0]), config);
     }
 
     static uint64_t Size(const std::vector<uint64_t> &counts, const Config &config) {
@@ -59,9 +60,7 @@ template <class Quant, class Bhiksha> class TrieSearch {
 
     uint8_t *SetupMemory(uint8_t *start, const std::vector<uint64_t> &counts, const Config &config);
 
-    void LoadedBinary();
-
-    void InitializeFromARPA(const char *file, util::FilePiece &f, std::vector<uint64_t> &counts, const Config &config, SortedVocabulary &vocab, Backing &backing);
+    void InitializeFromARPA(const char *file, util::FilePiece &f, std::vector<uint64_t> &counts, const Config &config, SortedVocabulary &vocab, BinaryFormat &backing);
 
     unsigned char Order() const {
       return middle_end_ - middle_begin_ + 2;
@@ -102,14 +101,14 @@ template <class Quant, class Bhiksha> class TrieSearch {
     }
 
   private:
-    friend void BuildTrie<Quant, Bhiksha>(SortedFiles &files, std::vector<uint64_t> &counts, const Config &config, TrieSearch<Quant, Bhiksha> &out, Quant &quant, const SortedVocabulary &vocab, Backing &backing);
+    friend void BuildTrie<Quant, Bhiksha>(SortedFiles &files, std::vector<uint64_t> &counts, const Config &config, TrieSearch<Quant, Bhiksha> &out, Quant &quant, SortedVocabulary &vocab, BinaryFormat &backing);
 
-    // Middles are managed manually so we can delay construction and they don't have to be copyable.  
+    // Middles are managed manually so we can delay construction and they don't have to be copyable.
     void FreeMiddles() {
       for (const Middle *i = middle_begin_; i != middle_end_; ++i) {
         i->~Middle();
       }
-      free(middle_begin_);
+      std::free(middle_begin_);
     }
 
     typedef trie::BitPackedMiddle<Bhiksha> Middle;
diff --git a/klm/lm/state.hh b/klm/lm/state.hh
index a6b9accb..543df37c 100644
--- a/klm/lm/state.hh
+++ b/klm/lm/state.hh
@@ -102,7 +102,7 @@ struct ChartState {
   }
 
   bool operator<(const ChartState &other) const {
-    return Compare(other) == -1;
+    return Compare(other) < 0;
   }
 
   void ZeroRemaining() {
diff --git a/klm/lm/trie.hh b/klm/lm/trie.hh
index 9ea3c546..d858ab5e 100644
--- a/klm/lm/trie.hh
+++ b/klm/lm/trie.hh
@@ -62,8 +62,6 @@ class Unigram {
       return unigram_;
     }
     
-    void LoadedBinary() {}
-
     UnigramPointer Find(WordIndex word, NodeRange &next) const {
       UnigramValue *val = unigram_ + word;
       next.begin = val->next;
@@ -108,8 +106,6 @@ template <class Bhiksha> class BitPackedMiddle : public BitPacked {
 
     void FinishedLoading(uint64_t next_end, const Config &config);
 
-    void LoadedBinary() { bhiksha_.LoadedBinary(); }
-
     util::BitAddress Find(WordIndex word, NodeRange &range, uint64_t &pointer) const;
 
     util::BitAddress ReadEntry(uint64_t pointer, NodeRange &range) {
@@ -138,14 +134,9 @@ class BitPackedLongest : public BitPacked {
       BaseInit(base, max_vocab, quant_bits);
     }
 
-    void LoadedBinary() {}
-
     util::BitAddress Insert(WordIndex word);
 
     util::BitAddress Find(WordIndex word, const NodeRange &node) const;
-
-  private:
-    uint8_t quant_bits_;
 };
 
 } // namespace trie
diff --git a/klm/lm/trie_sort.cc b/klm/lm/trie_sort.cc
index dc542bb3..126d43ab 100644
--- a/klm/lm/trie_sort.cc
+++ b/klm/lm/trie_sort.cc
@@ -50,6 +50,10 @@ class PartialViewProxy {
     const void *Data() const { return inner_.Data(); }
     void *Data() { return inner_.Data(); }
 
+    friend void swap(PartialViewProxy first, PartialViewProxy second) {
+      std::swap_ranges(reinterpret_cast<char*>(first.Data()), reinterpret_cast<char*>(first.Data()) + first.attention_size_, reinterpret_cast<char*>(second.Data()));
+    }
+
   private:
     friend class util::ProxyIterator<PartialViewProxy>;
 
diff --git a/klm/lm/value_build.cc b/klm/lm/value_build.cc
index 6124f8da..3ec3dce2 100644
--- a/klm/lm/value_build.cc
+++ b/klm/lm/value_build.cc
@@ -9,6 +9,7 @@ namespace ngram {
 template <class Model> LowerRestBuild<Model>::LowerRestBuild(const Config &config, unsigned int order, const typename Model::Vocabulary &vocab) {
   UTIL_THROW_IF(config.rest_lower_files.size() != order - 1, ConfigException, "This model has order " << order << " so there should be " << (order - 1) << " lower-order models for rest cost purposes.");
   Config for_lower = config;
+  for_lower.write_mmap = NULL;
   for_lower.rest_lower_files.clear();
 
   // Unigram models aren't supported, so this is a custom loader.  
diff --git a/klm/lm/virtual_interface.hh b/klm/lm/virtual_interface.hh
index 17f064b2..7a3e2379 100644
--- a/klm/lm/virtual_interface.hh
+++ b/klm/lm/virtual_interface.hh
@@ -125,10 +125,13 @@ class Model {
     void NullContextWrite(void *to) const { memcpy(to, null_context_memory_, StateSize()); }
 
     // Requires in_state != out_state
-    virtual float Score(const void *in_state, const WordIndex new_word, void *out_state) const = 0;
+    virtual float BaseScore(const void *in_state, const WordIndex new_word, void *out_state) const = 0;
 
     // Requires in_state != out_state
-    virtual FullScoreReturn FullScore(const void *in_state, const WordIndex new_word, void *out_state) const = 0;
+    virtual FullScoreReturn BaseFullScore(const void *in_state, const WordIndex new_word, void *out_state) const = 0;
+
+    // Prefer to use FullScore.  The context words should be provided in reverse order.
+    virtual FullScoreReturn BaseFullScoreForgotState(const WordIndex *context_rbegin, const WordIndex *context_rend, const WordIndex new_word, void *out_state) const = 0;
 
     unsigned char Order() const { return order_; }
 
diff --git a/klm/lm/vocab.cc b/klm/lm/vocab.cc
index fd7f96dc..7f0878f4 100644
--- a/klm/lm/vocab.cc
+++ b/klm/lm/vocab.cc
@@ -32,7 +32,8 @@ const uint64_t kUnknownHash = detail::HashForVocab("<unk>", 5);
 // Sadly some LMs have <UNK>.  
 const uint64_t kUnknownCapHash = detail::HashForVocab("<UNK>", 5);
 
-void ReadWords(int fd, EnumerateVocab *enumerate, WordIndex expected_count) {
+void ReadWords(int fd, EnumerateVocab *enumerate, WordIndex expected_count, uint64_t offset) {
+  util::SeekOrThrow(fd, offset);
   // Check that we're at the right place by reading <unk> which is always first.
   char check_unk[6];
   util::ReadOrThrow(fd, check_unk, 6);
@@ -80,11 +81,6 @@ void WriteWordsWrapper::Add(WordIndex index, const StringPiece &str) {
   buffer_.push_back(0);
 }
 
-void WriteWordsWrapper::Write(int fd, uint64_t start) {
-  util::SeekOrThrow(fd, start);
-  util::WriteOrThrow(fd, buffer_.data(), buffer_.size());
-}
-
 SortedVocabulary::SortedVocabulary() : begin_(NULL), end_(NULL), enumerate_(NULL) {}
 
 uint64_t SortedVocabulary::Size(uint64_t entries, const Config &/*config*/) {
@@ -100,6 +96,12 @@ void SortedVocabulary::SetupMemory(void *start, std::size_t allocated, std::size
   saw_unk_ = false;
 }
 
+void SortedVocabulary::Relocate(void *new_start) {
+  std::size_t delta = end_ - begin_;
+  begin_ = reinterpret_cast<uint64_t*>(new_start) + 1;
+  end_ = begin_ + delta;
+}
+
 void SortedVocabulary::ConfigureEnumerate(EnumerateVocab *to, std::size_t max_entries) {
   enumerate_ = to;
   if (enumerate_) {
@@ -147,11 +149,11 @@ void SortedVocabulary::FinishedLoading(ProbBackoff *reorder_vocab) {
   bound_ = end_ - begin_ + 1;
 }
 
-void SortedVocabulary::LoadedBinary(bool have_words, int fd, EnumerateVocab *to) {
+void SortedVocabulary::LoadedBinary(bool have_words, int fd, EnumerateVocab *to, uint64_t offset) {
   end_ = begin_ + *(reinterpret_cast<const uint64_t*>(begin_) - 1);
   SetSpecial(Index("<s>"), Index("</s>"), 0);
   bound_ = end_ - begin_ + 1;
-  if (have_words) ReadWords(fd, to, bound_);
+  if (have_words) ReadWords(fd, to, bound_, offset);
 }
 
 namespace {
@@ -179,6 +181,11 @@ void ProbingVocabulary::SetupMemory(void *start, std::size_t allocated, std::siz
   saw_unk_ = false;
 }
 
+void ProbingVocabulary::Relocate(void *new_start) {
+  header_ = static_cast<detail::ProbingVocabularyHeader*>(new_start);
+  lookup_.Relocate(static_cast<uint8_t*>(new_start) + ALIGN8(sizeof(detail::ProbingVocabularyHeader)));
+}
+
 void ProbingVocabulary::ConfigureEnumerate(EnumerateVocab *to, std::size_t /*max_entries*/) {
   enumerate_ = to;
   if (enumerate_) {
@@ -206,12 +213,11 @@ void ProbingVocabulary::InternalFinishedLoading() {
   SetSpecial(Index("<s>"), Index("</s>"), 0);
 }
 
-void ProbingVocabulary::LoadedBinary(bool have_words, int fd, EnumerateVocab *to) {
+void ProbingVocabulary::LoadedBinary(bool have_words, int fd, EnumerateVocab *to, uint64_t offset) {
   UTIL_THROW_IF(header_->version != kProbingVocabularyVersion, FormatLoadException, "The binary file has probing version " << header_->version << " but the code expects version " << kProbingVocabularyVersion << ".  Please rerun build_binary using the same version of the code.");
-  lookup_.LoadedBinary();
   bound_ = header_->bound;
   SetSpecial(Index("<s>"), Index("</s>"), 0);
-  if (have_words) ReadWords(fd, to, bound_);
+  if (have_words) ReadWords(fd, to, bound_, offset);
 }
 
 void MissingUnknown(const Config &config) throw(SpecialWordMissingException) {
diff --git a/klm/lm/vocab.hh b/klm/lm/vocab.hh
index 226ae438..074b74d8 100644
--- a/klm/lm/vocab.hh
+++ b/klm/lm/vocab.hh
@@ -36,7 +36,7 @@ class WriteWordsWrapper : public EnumerateVocab {
     
     void Add(WordIndex index, const StringPiece &str);
 
-    void Write(int fd, uint64_t start);
+    const std::string &Buffer() const { return buffer_; }
 
   private:
     EnumerateVocab *inner_;
@@ -71,6 +71,8 @@ class SortedVocabulary : public base::Vocabulary {
     // Everything else is for populating.  I'm too lazy to hide and friend these, but you'll only get a const reference anyway.
     void SetupMemory(void *start, std::size_t allocated, std::size_t entries, const Config &config);
 
+    void Relocate(void *new_start);
+
     void ConfigureEnumerate(EnumerateVocab *to, std::size_t max_entries);
 
     WordIndex Insert(const StringPiece &str);
@@ -83,15 +85,13 @@ class SortedVocabulary : public base::Vocabulary {
 
     bool SawUnk() const { return saw_unk_; }
 
-    void LoadedBinary(bool have_words, int fd, EnumerateVocab *to);
+    void LoadedBinary(bool have_words, int fd, EnumerateVocab *to, uint64_t offset);
 
   private:
     uint64_t *begin_, *end_;
 
     WordIndex bound_;
 
-    WordIndex highest_value_;
-
     bool saw_unk_;
 
     EnumerateVocab *enumerate_;
@@ -140,6 +140,8 @@ class ProbingVocabulary : public base::Vocabulary {
     // Everything else is for populating.  I'm too lazy to hide and friend these, but you'll only get a const reference anyway.
     void SetupMemory(void *start, std::size_t allocated, std::size_t entries, const Config &config);
 
+    void Relocate(void *new_start);
+
     void ConfigureEnumerate(EnumerateVocab *to, std::size_t max_entries);
 
     WordIndex Insert(const StringPiece &str);
@@ -152,7 +154,7 @@ class ProbingVocabulary : public base::Vocabulary {
 
     bool SawUnk() const { return saw_unk_; }
 
-    void LoadedBinary(bool have_words, int fd, EnumerateVocab *to);
+    void LoadedBinary(bool have_words, int fd, EnumerateVocab *to, uint64_t offset);
 
   private:
     void InternalFinishedLoading();