langfun 0.1.1.dev20240826__py3-none-any.whl → 0.1.1.dev202408282153__py3-none-any.whl

langfun/core/concurrent.py

@@ -13,17 +13,30 @@
 # limitations under the License.
 """Utility library for handling concurrency in langfun."""
 
+import abc
 import collections
 import concurrent.futures
 import dataclasses
+import io
 import random
+import sys
 import threading
 import time
 from typing import Any, Callable, Iterable, Iterator, Literal, Sequence, Tuple, Type, Union
 
 from langfun.core import component
+from langfun.core import text_formatting
 import pyglove as pg
-from tqdm import auto as tqdm
+
+
+progress_bar: Literal['tqdm', 'console', None] = None
+
+try:
+  from tqdm import auto as tqdm   # pylint: disable=g-import-not-at-top
+  progress_bar = 'tqdm'
+except ImportError as e:
+  progress_bar = 'console'
+  tqdm = None
 
 
 def with_context_access(func: Callable[..., Any]) -> Callable[..., Any]:
@@ -142,7 +155,6 @@ def with_retry(
         attempt = 1
       return base_interval() * (2 ** (attempt - 1))
 
-    wait_interval = None
     wait_intervals = []
     errors = []
     while True:
@@ -356,17 +368,17 @@ class ProgressBar:
     label: str | None
     total: int
     color: str | None = None
-    postfix: dict[str, str] | None = None
+    status: dict[str, Any] | None = None
 
   @dataclasses.dataclass
   class Update:
     """Progress bar update."""
     bar_id: int
     delta: int
-    postfix: Union[dict[str, str], str, None] = None
+    status: Union[dict[str, Any], str, None] = None
     color: str | None = None
 
-  _progress_bars: dict[int, tqdm.tqdm] = {}
+  _progress_bars: dict[int, '_ProgressControl'] = {}
   _install_requests: list[tuple[int, Settings]] = []
   _updates: collections.deque[Update] = collections.deque()
   _uninstall_requests: list[int] = []
@@ -378,11 +390,11 @@ class ProgressBar:
       label: str | None,
       total: int,
       color: str | None = None,
-      postfix: dict[str, str] | None = None,
+      status: dict[str, Any] | None = None,
       ) -> int:
     """Installs a progress bar and returns a reference id."""
     with cls._lock:
-      settings = ProgressBar.Settings(label, total, color, postfix)
+      settings = ProgressBar.Settings(label, total, color, status)
       bar_id = id(settings)
       cls._install_requests.append((bar_id, settings))
       return bar_id
@@ -392,15 +404,17 @@ class ProgressBar:
       cls,
       bar_id: int,
       delta: int = 0,
-      postfix: Union[dict[str, str], str, None] = None,
+      status: Union[dict[str, Any], str, None] = None,
       color: str | None = None,
       refresh: bool = True,
       ) -> None:
     """Report the progress for a label."""
+    if status is not None and not isinstance(status, (str, dict)):
+      raise ValueError(f'Unsupported status: {status}')
     with cls._lock:
       cls._updates.append(
           ProgressBar.Update(
-              bar_id=bar_id, delta=delta, postfix=postfix, color=color,
+              bar_id=bar_id, delta=delta, status=status, color=color,
           )
       )
     if refresh:
@@ -422,11 +436,11 @@ class ProgressBar:
       # Process install requests.
       if cls._install_requests:
         for bar_id, settings in cls._install_requests:
-          cls._progress_bars[bar_id] = tqdm.tqdm(
+          cls._progress_bars[bar_id] = _progress_control(
               total=settings.total,
-              desc=settings.label,
-              colour=settings.color,
-              postfix=settings.postfix)
+              label=settings.label,
+              color=settings.color,
+              status=settings.status)
         cls._install_requests.clear()
 
       # Process updates.
@@ -441,15 +455,11 @@ class ProgressBar:
         if update.delta > 0:
           bar.update(update.delta)
 
-        if isinstance(update.postfix, str):
-          bar.set_postfix_str(update.postfix, refresh=False)
-        elif isinstance(update.postfix, dict):
-          bar.set_postfix(update.postfix, refresh=False)
-        elif update.postfix is not None:
-          raise ValueError(f'Unsupported postfix: {update.postfix}')
+        if update.status is not None:
+          bar.set_status(update.status)
 
         if update.color is not None:
-          bar.colour = update.color
+          bar.set_color(update.color)
         updated_bars.add(bar)
 
       # Refresh each updated bar just once.
@@ -603,7 +613,7 @@ def concurrent_map(
         if len(error_text) >= 64:
           error_text = error_text[:64] + '...'
         status['LastError'] = error_text
-      ProgressBar.update(bar_id, delta=1, postfix=status)
+      ProgressBar.update(bar_id, delta=1, status=status)
 
   try:
     if ordered:
@@ -729,5 +739,122 @@ class ExecutorPool:
       raise ValueError(f'Unsupported value: {maybe_executor}.')
 
 
+class _ProgressControl(pg.Object):
+  """Abstract progress control."""
+  # Disable symbolic comparison so the hash is based on object address.
+  use_symbolic_comparison = False
+
+  total: int
+  label: str | None
+  color: str | None
+  status: str | dict[str, Any] | None
+
+  def set_color(self, color: str | None):
+    with pg.notify_on_change(False):
+      self.rebind(color=color)
+
+  def set_status(self, status: str | dict[str, Any] | None):
+    with pg.notify_on_change(False):
+      self.rebind(status=status)
+
+  @abc.abstractmethod
+  def update(self, delta):
+    """Update progress."""
+
+  @abc.abstractmethod
+  def refresh(self) -> None:
+    """Refresh progress bar."""
+
+
+class _TqdmProgressControl(_ProgressControl):
+  """Tqdm-based progress control."""
+
+  def _on_bound(self):
+    super()._on_bound()
+    assert tqdm is not None
+    self._tqdm = tqdm.tqdm(
+        total=self.total,
+        desc=self.label,
+        colour=self.color,
+        postfix=self.status,
+    )
+
+  def update(self, delta: int) -> None:
+    self._tqdm.update(delta)
+
+  def refresh(self):
+    self._tqdm.set_description(self.label, refresh=False)
+    if isinstance(self.status, str):
+      self._tqdm.set_postfix_str(self.status, refresh=False)
+    else:
+      self._tqdm.set_postfix(self.status, refresh=False)
+    self._tqdm.colour = self.color
+    self._tqdm.refresh()
+
+
+class _ConsoleProgressControl(_ProgressControl):
+  """Simple progress control by printing the status to the console."""
+
+  def _on_bound(self):
+    super()._on_bound()
+    self._progress = 0
+
+  def update(self, delta: int) -> None:
+    self._progress += delta
+
+  def refresh(self):
+    s = io.StringIO()
+    if self.label is not None:
+      s.write(text_formatting.colored(self.label, 'red', styles=['bold']))
+      s.write(': ')
+    s.write(
+        text_formatting.colored(
+            '%d%% (%d/%d)' %
+            (
+                self._progress * 100 // self.total,
+                self._progress,
+                self.total,
+            ),
+            color=self.color or 'green'
+        )
+    )
+    if self.status is not None:
+      status = repr(self.status) if isinstance(
+          self.status, dict) else self.status
+      s.write(f' : {status}')
+    sys.stderr.write(s.getvalue() + '\n')
+
+
+class _NoopProgressControl(_ProgressControl):
+  """No-op progress control."""
+
+  def update(self, delta: int) -> None:
+    pass
+
+  def refresh(self) -> None:
+    pass
+
+
+def _progress_control(
+    total: int,
+    label: str | None,
+    color: str | None,
+    status: str | dict[str, Any] | None,
+) -> _ProgressControl:
+  """Creates a process control."""
+  if progress_bar == 'tqdm':
+    if not tqdm:
+      raise RuntimeError(
+          'Please install package "tqdm" to use `tqdm` progress bar.'
+      )
+    return _TqdmProgressControl(total, label, color, status)
+  elif progress_bar == 'console':
+    return _ConsoleProgressControl(total, label, color, status)
+  elif progress_bar is None:
+    return _NoopProgressControl(total, label, color, status)
+  else:
+    raise ValueError(f'Unsupported progress bar type: {progress_bar}')
+
+
 # The global executor pool based on resource IDs.
 _executor_pool = ExecutorPool()

langfun/core/concurrent_test.py

@@ -233,6 +233,57 @@ class ProgressTest(unittest.TestCase):
     self.assertIs(p.last_error, job2.error)
 
 
+class ProgressControlTest(unittest.TestCase):
+
+  def test_noop(self):
+    concurrent.progress_bar = None
+    ctrl = concurrent._progress_control(100, 'noop', 'blue', None)
+    self.assertIsInstance(ctrl, concurrent._NoopProgressControl)
+    string_io = io.StringIO()
+    with contextlib.redirect_stderr(string_io):
+      ctrl.update(1)
+      ctrl.refresh()
+    self.assertEqual(string_io.getvalue(), '')
+    concurrent.progress_bar = 'tqdm'
+
+  def test_console(self):
+    concurrent.progress_bar = 'console'
+    ctrl = concurrent._progress_control(100, 'foo', 'blue', None)
+    self.assertIsInstance(ctrl, concurrent._ConsoleProgressControl)
+    string_io = io.StringIO()
+    with contextlib.redirect_stderr(string_io):
+      ctrl.set_status('bar')
+      ctrl.update(10)
+      ctrl.refresh()
+    self.assertEqual(
+        string_io.getvalue(),
+        '\x1b[1m\x1b[31mfoo\x1b[0m: \x1b[34m10% (10/100)\x1b[0m : bar\n'
+    )
+    concurrent.progress_bar = 'tqdm'
+
+  def test_tqdm(self):
+    concurrent.progress_bar = 'tqdm'
+    string_io = io.StringIO()
+    with contextlib.redirect_stderr(string_io):
+      ctrl = concurrent._progress_control(100, 'foo', 'blue', None)
+      self.assertIsInstance(ctrl, concurrent._TqdmProgressControl)
+      ctrl.update(10)
+      ctrl.refresh()
+    self.assertIn('10/100', string_io.getvalue())
+
+    tqdm = concurrent.tqdm
+    concurrent.tqdm = None
+    with self.assertRaisesRegex(RuntimeError, 'install package "tqdm"'):
+      _ = concurrent._progress_control(100, 'foo', 'blue', None)
+    concurrent.tqdm = tqdm
+
+  def test_unsupported(self):
+    concurrent.progress_bar = 'unknown'
+    with self.assertRaisesRegex(ValueError, 'Unsupported progress bar type'):
+      _ = concurrent._progress_control(100, 'foo', 'blue', None)
+    concurrent.progress_bar = 'tqdm'
+
+
 class ProgressBarTest(unittest.TestCase):
 
   def test_multithread_support(self):
@@ -241,13 +292,12 @@ class ProgressBarTest(unittest.TestCase):
       bar_id = concurrent.ProgressBar.install(None, 5)
       def fun(x):
         del x
-        concurrent.ProgressBar.update(bar_id, 1, postfix=None)
+        concurrent.ProgressBar.update(bar_id, 1, status=None)
 
       for _ in concurrent.concurrent_execute(fun, range(5)):
         concurrent.ProgressBar.refresh()
       concurrent.ProgressBar.uninstall(bar_id)
     output_str = string_io.getvalue()
-    print(output_str)
     self.assertIn('100%', output_str)
     self.assertIn('5/5', output_str)
 
@@ -255,12 +305,12 @@ class ProgressBarTest(unittest.TestCase):
     string_io = io.StringIO()
     with contextlib.redirect_stderr(string_io):
       bar_id = concurrent.ProgressBar.install(None, 4)
-      concurrent.ProgressBar.update(bar_id, 1, postfix=None)
-      concurrent.ProgressBar.update(bar_id, 1, postfix='hello')
-      concurrent.ProgressBar.update(bar_id, color='lightgreen')
-      concurrent.ProgressBar.update(bar_id, 2, postfix=dict(x=1))
-      with self.assertRaisesRegex(ValueError, 'Unsupported postfix'):
-        concurrent.ProgressBar.update(bar_id, 0, postfix=1)
+      concurrent.ProgressBar.update(bar_id, 1, status=None)
+      concurrent.ProgressBar.update(bar_id, 1, status='hello')
+      concurrent.ProgressBar.update(bar_id, color='green')
+      concurrent.ProgressBar.update(bar_id, 2, status=dict(x=1))
+      with self.assertRaisesRegex(ValueError, 'Unsupported status'):
+        concurrent.ProgressBar.update(bar_id, 0, status=1)
       concurrent.ProgressBar.uninstall(bar_id)
     self.assertIn('1/4', string_io.getvalue())
     self.assertIn('2/4', string_io.getvalue())

langfun/core/eval/base.py

@@ -242,7 +242,7 @@ class Evaluable(lf.Component):
       ):
         if show_progress:
           lf.concurrent.ProgressBar.update(
-              progress_bar, postfix='LOADING SAVED RESULTS...', color='yellow'
+              progress_bar, status='LOADING SAVED RESULTS...', color='yellow'
           )
         if self.try_load_result():
           run_status = 'CACHED'
@@ -271,7 +271,7 @@ class Evaluable(lf.Component):
         if should_save:
           if show_progress:
             lf.concurrent.ProgressBar.update(
-                progress_bar, postfix='SAVING RESULTS...', color='yellow'
+                progress_bar, status='SAVING RESULTS...', color='yellow'
             )
 
           # Save evaluation results.
@@ -284,7 +284,7 @@ class Evaluable(lf.Component):
       if show_progress:
         lf.concurrent.ProgressBar.update(
             progress_bar,
-            postfix=self._completion_status(run_status),
+            status=self._completion_status(run_status),
             color='green',
         )
     else:
@@ -340,7 +340,7 @@ class Evaluable(lf.Component):
               f'[#{leaf.index} - {leaf.node.id}]',
               total=leaf.node.num_examples if leaf.enabled else 0,
               color='cyan' if leaf.enabled else 'yellow',
-              postfix=None if leaf.enabled else 'SKIPPED.')
+              status=None if leaf.enabled else 'SKIPPED.')
 
         # Run leaf groups in parallel.
         try:
@@ -354,7 +354,7 @@ class Evaluable(lf.Component):
           # Save results for non-leaf nodes.
           lf.concurrent.ProgressBar.update(
               overview_bar,
-              postfix='SAVING RESULTS...',
+              status='SAVING RESULTS...',
               color='yellow')
 
           for node in self.nonleaf_nodes:
@@ -364,7 +364,7 @@ class Evaluable(lf.Component):
 
           if should_save and summary:
             lf.concurrent.ProgressBar.update(
-                overview_bar, postfix='FINALIZING SUMMARY...'
+                overview_bar, status='FINALIZING SUMMARY...'
             )
 
             summary.save(os.path.join(self.root_dir, Evaluable.SUMMARY_HTML))
@@ -378,7 +378,7 @@ class Evaluable(lf.Component):
           # Signal all task completed by making the bar green.
           lf.concurrent.ProgressBar.update(
               overview_bar,
-              postfix='COMPLETED',
+              status='COMPLETED',
               color='green')
 
         finally:
@@ -1261,7 +1261,7 @@ class Evaluation(Evaluable):
 
   def finalize(self) -> pg.Dict:
     """Finalizes the evaluation result."""
-    if self.cache:
+    if self.cache is not None:
       cache_stats = dict(
           use_cache=True,
           num_queries=self.cache.stats.num_queries,

langfun/core/eval/base_test.py

@@ -564,7 +564,7 @@ class SuiteTest(unittest.TestCase):
                 schema_fn='answer_schema()',
             ),
             cache_stats=dict(
-                use_cache=True, num_queries=4, num_hits=1, num_updates=3
+                use_cache=True, num_queries=4, num_hits=0, num_updates=4
             ),
             metrics=dict(
                 total=2,

langfun/core/langfunc.py

@@ -269,6 +269,9 @@ class LangFunc(
         # Send rendered text to LM.
         lm_output = self.lm(lm_input, cache_seed=cache_seed)
 
+        # Attach cache seed.
+        lm_input.metadata.cache_seed = cache_seed
+
         # Transform the output message.
         lm_output = self.transform_output(lm_output)
         lm_output.tag(message_lib.Message.TAG_LM_OUTPUT)

langfun/core/langfunc_test.py

@@ -94,11 +94,13 @@ class LangFuncCallTest(unittest.TestCase):
         )
     )
     self.assertEqual(r.tags, ['lm-response', 'lm-output'])
-    self.assertEqual(r.source, message.UserMessage('Hello'))
+    self.assertEqual(
+        r.source,
+        message.UserMessage('Hello', metadata=dict(cache_seed=0))
+    )
     self.assertEqual(r.source.tags, ['rendered', 'lm-input'])
 
     self.assertEqual(str(l), 'Hello')
-    print(repr(l))
     self.assertEqual(
         repr(l),
         "LangFunc(template_str='Hello', clean=True,"
@@ -114,7 +116,7 @@ class LangFuncCallTest(unittest.TestCase):
       self.assertEqual(l, 'Hello')
       self.assertEqual(l.natural_language_format(), 'Hello')
       self.assertEqual(l.render(), 'Hello')
-      r = l()
+      r = l(cache_seed=1)
       self.assertEqual(
           r,
           message.AIMessage(
@@ -123,6 +125,7 @@ class LangFuncCallTest(unittest.TestCase):
           )
       )
       self.assertEqual(r.tags, ['lm-response', 'lm-output'])
+      self.assertEqual(r.source.metadata.cache_seed, 1)
 
     self.assertEqual(str(l), 'Hello')
 

langfun/core/language_model.py

@@ -234,6 +234,7 @@ class LMCache(pg.Object):
     num_hit_expires: int = 0
     num_misses: int = 0
     num_updates: int = 0
+    num_deletes: int = 0
 
   @abc.abstractmethod
   def get(
@@ -251,6 +252,15 @@ class LMCache(pg.Object):
   ) -> None:
     """Puts the result of a prompt generated by a language model in cache."""
 
+  @abc.abstractmethod
+  def delete(
+      self,
+      lm: 'LanguageModel',
+      prompt: message_lib.Message,
+      seed: int,
+  ) -> bool:
+    """Deletes the result of a prompt generated by a language model in cache."""
+
   @property
   @abc.abstractmethod
   def stats(self) -> Stats:

langfun/core/llms/cache/base.py

@@ -60,13 +60,16 @@ class LMCacheBase(lf.LMCache):
       self, lm: lf.LanguageModel, prompt: lf.Message, seed: int
   ) -> lf.LMSamplingResult | None:
     """Gets the cached result of a prompt generated by a language model."""
-    entry = self._get(lm.model_id, self._key(lm, prompt, seed))
+    key = self._key(lm, prompt, seed)
+    entry = self._get(lm.model_id, key)
     self._stats.num_queries += 1
     if entry is None:
       self._stats.num_misses += 1
       return None
     if entry.expire is not None and entry.expire < datetime.datetime.now():
       self._stats.num_hit_expires += 1
+      self._stats.num_deletes += 1
+      assert self._delete(lm.model_id, key)
       return None
     self._stats.num_hits += 1
     return entry.result
@@ -86,6 +89,18 @@ class LMCacheBase(lf.LMCache):
     self._put(lm.model_id, self._key(lm, prompt, seed), entry)
     self._stats.num_updates += 1
 
+  def delete(
+      self,
+      lm: lf.LanguageModel,
+      prompt: lf.Message,
+      seed: int,
+  ) -> bool:
+    """Deletes the result of a prompt generated by a language model in cache."""
+    deleted = self._delete(lm.model_id, self._key(lm, prompt, seed))
+    if deleted:
+      self._stats.num_deletes += 1
+    return deleted
+
   @abc.abstractmethod
   def _get(self, model_id: str, key: str) -> LMCacheEntry | None:
     """Returns a LM cache entry associated with the key."""
@@ -94,6 +109,10 @@ class LMCacheBase(lf.LMCache):
   def _put(self, model_id: str, key: str, entry: LMCacheEntry) -> None:
     """Puts a LM cache entry associated with the key."""
 
+  @abc.abstractmethod
+  def _delete(self, model_id: str, key: str) -> bool:
+    """Deletes a LM cache entry associated with the key."""
+
   def _sym_clone(self, deep: bool, memo: Any = None) -> 'LMCacheBase':
     v = super()._sym_clone(deep, memo)
     v._stats = self._stats  # pylint: disable=protected-access
@@ -102,4 +121,4 @@ class LMCacheBase(lf.LMCache):
 
 def default_key(lm: lf.LanguageModel, prompt: lf.Message, seed: int) -> Any:
   """Default key for LM cache."""
-  return (prompt.text, lm.sampling_options.cache_key(), seed)
+  return (prompt.text_with_modality_hash, lm.sampling_options.cache_key(), seed)

langfun/core/llms/cache/in_memory.py

@@ -99,6 +99,13 @@ class InMemory(base.LMCacheBase):
     """Puts a LM cache entry associated with the key."""
     self._cache[model_id][key] = entry
 
+  def _delete(self, model_id: str, key: str) -> bool:
+    """Deletes a LM cache entry associated with the key."""
+    model_cache = self._cache.get(model_id, None)
+    if model_cache is None:
+      return False
+    return model_cache.pop(key, None) is not None
+
   def reset(self, model_id: str | None = None) -> None:
     """Resets the cache."""
     if model_id is not None:

langfun/core/llms/cache/in_memory_test.py

@@ -148,6 +148,50 @@ class InMemoryLMCacheTest(unittest.TestCase):
     self.assertIs(copy.deepcopy(cache)._cache, cache._cache)
     self.assertIs(copy.deepcopy(cache)._stats, cache._stats)
 
+    self.assertFalse(
+        cache.delete(fake.StaticResponse('hi'), lf.UserMessage('c'), seed=0)
+    )
+    self.assertFalse(cache.delete(lm, lf.UserMessage('c'), seed=1))
+    self.assertFalse(cache.delete(lm, lf.UserMessage('d'), seed=0))
+    self.assertTrue(cache.delete(lm, lf.UserMessage('c'), seed=0))
+    self.assertEqual(
+        list(cache.keys('StaticSequence')),
+        [
+            ('a', (None, None, 1, 40, None, None), 0),
+            ('a', (None, None, 1, 40, None, None), 1),
+            ('b', (None, None, 1, 40, None, None), 0),
+        ],
+    )
+    self.assertEqual(cache.stats.num_deletes, 1)
+
+  def test_cache_with_modalities(self):
+
+    class CustomModality(lf.Modality):
+      content: str
+
+      def to_bytes(self):
+        return self.content.encode()
+
+    cache = in_memory.InMemory()
+    lm = fake.StaticSequence(['1', '2', '3', '4', '5', '6'], cache=cache)
+    lm(lf.UserMessage('hi <<[[image]]>>', image=CustomModality('foo')))
+    lm(lf.UserMessage('hi <<[[image]]>>', image=CustomModality('bar')))
+    self.assertEqual(
+        list(cache.keys()),
+        [
+            (
+                'hi <<[[image]]>><image>acbd18db</image>',
+                (None, None, 1, 40, None, None),
+                0,
+            ),
+            (
+                'hi <<[[image]]>><image>37b51d19</image>',
+                (None, None, 1, 40, None, None),
+                0,
+            ),
+        ],
+    )
+
   def test_ttl(self):
     cache = in_memory.InMemory(ttl=1)
     lm = fake.StaticSequence(['1', '2', '3'], cache=cache)
@@ -160,6 +204,7 @@ class InMemoryLMCacheTest(unittest.TestCase):
     self.assertEqual(cache.stats.num_hits, 1)
     self.assertEqual(cache.stats.num_hit_expires, 1)
     self.assertEqual(cache.stats.num_misses, 1)
+    self.assertEqual(cache.stats.num_deletes, 1)
 
   def test_different_sampling_options(self):
     cache = in_memory.InMemory()