Merge branch 'dottxt-ai:main' into multimodal-input-handling-via-chat-interface

Author: laitifranz

.github/workflows/tests.yml

@@ -53,7 +53,7 @@ jobs:
       run: |
         rm -f .coverage*
         uv run coverage erase
-        uv run python -m coverage run --branch --parallel-mode -m pytest -x -m 'not api_call'
+        uv run python -m coverage run --branch --source=outlines --parallel-mode -m pytest -x -m 'not api_call'
     - name: Upload coverage data
       uses: actions/upload-artifact@v4
       with:
@@ -86,37 +86,18 @@ jobs:
         with:
           name: coverage-data
 
-      - name: Determine base branch for comparison
-        id: base-branch
-        run: |
-          if [[ "${{ github.event_name }}" == "pull_request" ]]; then
-            # For PRs, use the remote tracking branch
-            COMPARE_BRANCH="origin/${{ github.base_ref }}"
-            echo "COMPARE_BRANCH=$COMPARE_BRANCH" >> $GITHUB_ENV
-          else
-            # For push events, compare against the parent commit
-            COMPARE_BRANCH="${{ github.event.before }}"
-            echo "COMPARE_BRANCH=$COMPARE_BRANCH" >> $GITHUB_ENV
-          fi
-          echo "Using $COMPARE_BRANCH for coverage comparison"
-
-      - name: Fetch base branch for coverage diff
-        run: |
-          git fetch --no-tags --prune origin ${COMPARE_BRANCH#origin/}
-
       - name: Combine coverage & fail if it's <100%.
         run: |
           python -m coverage combine
           python -m coverage html --skip-covered --skip-empty
           python -m coverage xml
-          diff-cover coverage.xml --compare-branch=$COMPARE_BRANCH --markdown-report=coverage.md --fail-under=100 || (cat coverage.md >> $GITHUB_STEP_SUMMARY && exit 1)
+          python -m coverage report --fail-under=100 || (python -m coverage report && exit 1)
 
       - name: Upload HTML report if check failed.
         uses: actions/upload-artifact@v4
         with:
           name: html-report
           path: htmlcov
-          # TODO FIXME: This is only using the last run
           overwrite: true
         if: ${{ failure() }}
 

outlines/backends/llguidance.py

@@ -37,9 +37,6 @@ def __init__(
             The name of the tensor library used by the model
 
         """
-        if tensor_library_name not in SUPPORTED_TENSOR_LIBRARIES:
-            raise TypeError(f"Unsupported tensor library: {tensor_library_name}")
-
         self.is_first_token = True
         self.grammar = grammar
         self.llg_tokenizer = llg_tokenizer

outlines/backends/outlines_core.py

@@ -70,15 +70,15 @@ def _setup(self, batch_size: int, vocab_size: int) -> None:
             self.allocate_token_bitmask = allocate_token_bitmask
             self.bias_logits = self._bias_logits_numpy
 
-        elif self.tensor_library_name == "mlx":
+        elif self.tensor_library_name == "mlx": # pragma: no cover
             from outlines_core.kernels.mlx import (
                 allocate_token_bitmask
             )
 
             self.allocate_token_bitmask = allocate_token_bitmask
             self.bias_logits = self._bias_logits_mlx
 
-        else:
+        else: # pragma: no cover
             raise ValueError(
                 f"Unsupported tensor library: {self.tensor_library_name}"
             )
@@ -179,7 +179,13 @@ def process_logits(
         else:
             for i in range(batch_size):
                 last_token_id = self.tensor_adapter.to_scalar(input_ids[i][-1]) # type: ignore
-                if not self._guides[i].is_finished():
+                # This circumvents issue #227 in outlines_core
+                # Ideally, we would be able to advance all the times as the final
+                # state would accept the eos token leading to itself
+                if (
+                    not self._guides[i].is_finished()
+                    or self._guides[i].accepts_tokens([last_token_id])
+                ):
                     self._guides[i].advance(
                         token_id=last_token_id,
                         return_tokens=False
@@ -211,13 +217,13 @@ def __init__(self, model: SteerableModel):
             eos_token_id = tokenizer.eos_token_id
             eos_token = tokenizer.eos_token
             token_to_str = tokenizer.convert_token_to_string
-        elif isinstance(model, MLXLM):
+        elif isinstance(model, MLXLM): # pragma: no cover
             tokenizer = model.mlx_tokenizer # type: ignore
             vocabulary = tokenizer.get_vocab()
             eos_token_id = tokenizer.eos_token_id
             eos_token = tokenizer.eos_token
             token_to_str = lambda token: tokenizer.convert_tokens_to_string([token]) # type: ignore
-        else:
+        else: # pragma: no cover
             raise ValueError(f"Unsupported model type: {type(model)}")
 
         self.eos_token_id = eos_token_id

outlines/backends/xgrammar.py

@@ -39,7 +39,7 @@ def _setup(self, batch_size: int, vocab_size: int) -> None:
         """Setup the logits processor for a new generation."""
         if self.tensor_library_name == "torch":
             self._bias_logits = self._bias_logits_torch
-        elif self.tensor_library_name == "mlx":
+        elif self.tensor_library_name == "mlx": # pragma: no cover
             self._bias_logits = self._bias_logits_mlx
         else: # pragma: no cover
             raise ValueError(
@@ -101,7 +101,7 @@ def process_logits(
             self.is_first_token = False
         else:
             for i in range(batch_size):
-                if not self._matchers[i].is_terminated():
+                if not self._matchers[i].is_terminated(): # pragma: no cover
                     last_token_id = self.tensor_adapter.to_scalar(
                         input_ids[i][-1] # type: ignore
                     )
@@ -125,7 +125,7 @@ def __init__(self, model: SteerableModel):
 
         if isinstance(model, Transformers):
             tokenizer = model.hf_tokenizer
-        elif isinstance(model, MLXLM):
+        elif isinstance(model, MLXLM): # pragma: no cover
             tokenizer = model.mlx_tokenizer._tokenizer
         else: # pragma: no cover
             raise ValueError(

outlines/caching.py

@@ -61,7 +61,7 @@ def get_cache():
         cache_dir = outlines_cache_dir
     elif xdg_cache_home:  # pragma: no cover
         cache_dir = os.path.join(xdg_cache_home, ".cache", "outlines")
-    elif home_dir != "/":
+    elif home_dir != "/": # pragma: no cover
         cache_dir = os.path.join(home_dir, ".cache", "outlines")
     else:  # pragma: no cover
         # home_dir may be / inside a docker container without existing user

outlines/models/transformers.py

@@ -244,11 +244,29 @@ def __init__(
             and isinstance(model, FlaxPreTrainedModel)
         ):
             self.tensor_library_name = "jax"
+            warnings.warn("""
+                Support for `jax` has been deprecated and will be removed in
+                version 1.4.0 of Outlines. Please use `torch` instead.
+                Transformers models using `jax` do not support structured
+                generation.
+                """,
+                DeprecationWarning,
+                stacklevel=2,
+            )
         elif (
             TFPreTrainedModel is not None
             and isinstance(model, TFPreTrainedModel)
         ):
             self.tensor_library_name = "tensorflow"
+            warnings.warn("""
+                Support for `tensorflow` has been deprecated and will be removed in
+                version 1.4.0 of Outlines. Please use `torch` instead.
+                Transformers models using `tensorflow` do not support structured
+                generation.
+                """,
+                DeprecationWarning,
+                stacklevel=2,
+            )
         else:
             self.tensor_library_name = "torch"
 

outlines/processors/base_logits_processor.py

@@ -28,8 +28,8 @@ def __init__(self, tensor_library_name: str):
         ----------
         tensor_library_name
             The name of the library to use to manipulate tensors. Possible
-            values are "jax", "mlx", "numpy", "tensorflow" and "torch". You
-            must choose the library that your model is using.
+            values are "mlx", "numpy" and "torch". You must choose the library
+            that your model is using.
         """
         # Temporary fix as torch raises a warning that can cause can an error
         # with python 3.12.
@@ -52,7 +52,7 @@ def reset(self):
         needs to be reset for a new generation.
 
         """
-        pass
+        pass # pragma: no cover
 
     @abstractmethod
     def process_logits(

outlines/processors/tensor_adapters/__init__.py

@@ -2,25 +2,19 @@
 
 from typing import Union
 
-from .jax import JAXTensorAdapter
 from .mlx import MLXTensorAdapter
 from .numpy import NumpyTensorAdapter
-from .tensorflow import TensorFlowTensorAdapter
 from .torch import TorchTensorAdapter
 
 
 tensor_adapters = {
-    "jax": JAXTensorAdapter,
     "mlx": MLXTensorAdapter,
     "numpy": NumpyTensorAdapter,
-    "tensorflow": TensorFlowTensorAdapter,
     "torch": TorchTensorAdapter,
 }
 
 TensorAdapterImplementation = Union[
-    JAXTensorAdapter,
     MLXTensorAdapter,
     NumpyTensorAdapter,
-    TensorFlowTensorAdapter,
     TorchTensorAdapter,
 ]