diff --git a/.github/workflows/quality.yml b/.github/workflows/quality.yml
index c1f5f0c1..75777260 100644
--- a/.github/workflows/quality.yml
+++ b/.github/workflows/quality.yml
@@ -49,7 +49,7 @@ jobs:
         run: ollama pull llama3.2:1b
 
       - name: Run Tests
-        run: uv run -m pytest -v test -n auto --dist loadscope
+        run: uv run -m pytest -v test
       - name: Send failure message tests
         if: failure()  # This step will only run if a previous step failed
         run: echo "Tests failed. Please verify that tests are working locally."
diff --git a/docs/examples/context/contexts_with_sampling.py b/docs/examples/context/contexts_with_sampling.py
new file mode 100644
index 00000000..d760ca2a
--- /dev/null
+++ b/docs/examples/context/contexts_with_sampling.py
@@ -0,0 +1,51 @@
+from mellea.backends.types import ModelOption
+from mellea.stdlib.sampling.base import RejectionSamplingStrategy
+from mellea.stdlib.session import start_session
+
+# You can retrieve context information when using SamplingStrategies
+# and validation.
+
+m = start_session()
+
+# We want the full SamplingResult.
+res = m.instruct(
+    "Write a sentence.",
+    requirements=["be funny", "be formal", "start the sentence with the letter w"],
+    strategy=RejectionSamplingStrategy(loop_budget=3),
+    return_sampling_results=True,
+)
+
+print()
+print("Printing result of `Writing a sentence`.")
+print(f"Result: {res.success}")
+print(f"Result Output: {res.result}")
+print()
+
+# We can also look at the context for the chosen result and
+# any other results that weren't chosen.
+# (This prompt tends to take 2 attempts. If it only takes one, try re-running it.)
+print(f"Total Generation Attempts: {len(res.sample_generations)}")
+print()
+
+print(f"Getting index of another result.")
+index = 0  # Just choose the first one.
+
+print(
+    "If the below is the same output, try re-running this program to get multiple attempts."
+)
+print(f"Different attempted output: {res.sample_generations[index]}")
+print()
+
+# We can see the context that created this output.
+gen_ctx = res.sample_contexts[index]
+print(f"Previous step in generating this result was: {gen_ctx.previous_node.node_data}")
+print()
+
+# We can also see what the validation context looked like.
+req, val_result = res.sample_validations[index][0]
+print(
+    f"Getting context when evaluating the above output against Req({req.description})."
+)
+val_ctx = val_result.context
+
+print(f"Output of the validation for this requirement: {val_ctx.node_data}")
diff --git a/docs/examples/safety.py/guardian.py b/docs/examples/safety/guardian.py
similarity index 100%
rename from docs/examples/safety.py/guardian.py
rename to docs/examples/safety/guardian.py
diff --git a/docs/examples/safety.py/guardian_huggingface.py b/docs/examples/safety/guardian_huggingface.py
similarity index 100%
rename from docs/examples/safety.py/guardian_huggingface.py
rename to docs/examples/safety/guardian_huggingface.py
diff --git a/docs/examples/safety.py/repair_with_guardian.py b/docs/examples/safety/repair_with_guardian.py
similarity index 100%
rename from docs/examples/safety.py/repair_with_guardian.py
rename to docs/examples/safety/repair_with_guardian.py
diff --git a/mellea/backends/huggingface.py b/mellea/backends/huggingface.py
index c5d9b0db..99058e7d 100644
--- a/mellea/backends/huggingface.py
+++ b/mellea/backends/huggingface.py
@@ -206,7 +206,7 @@ def generate_from_context(
             mot = self._generate_from_context_alora(
                 action, ctx, _format=format, model_options=model_opts
             )
-            return mot, ctx.add(mot)
+            return mot, ctx.add(action).add(mot)
         else:
             mot = self._generate_from_context_standard(
                 action,
@@ -512,6 +512,14 @@ def generate_from_raw(
                 "The raw endpoint does not support tool calling at the moment."
             )
 
+        if self._model.device.type == "mps":
+            # TODO: Remove this when we are able to update the torch package.
+            #       Test this by ensuring all outputs from this call are populated when running on mps.
+            #       https://github.com/pytorch/pytorch/pull/157727
+            FancyLogger.get_logger().warning(
+                "utilizing device mps with a `generate_from_raw` request; you may see issues when submitting batches of prompts to a huggingface backend; ensure all ModelOutputThunks have non-empty values."
+            )
+
         model_opts = self._simplify_and_merge(model_options)
         seed = model_opts.get(ModelOption.SEED, None)
         if seed is not None:
diff --git a/mellea/backends/litellm.py b/mellea/backends/litellm.py
index 9af431b0..54b8f3c2 100644
--- a/mellea/backends/litellm.py
+++ b/mellea/backends/litellm.py
@@ -270,8 +270,9 @@ def _generate_from_chat_context_standard(
             [OpenAIBackend.message_to_openai_message(m) for m in messages]
         )
 
+        extra_params: dict[str, Any] = {}
         if _format is not None:
-            response_format = {
+            extra_params["response_format"] = {
                 "type": "json_schema",
                 "json_schema": {
                     "name": _format.__name__,
@@ -279,8 +280,6 @@ def _generate_from_chat_context_standard(
                     "strict": True,
                 },
             }
-        else:
-            response_format = {"type": "text"}
 
         thinking = model_opts.get(ModelOption.THINKING, None)
         if type(thinking) is bool and thinking:
@@ -304,9 +303,9 @@ def _generate_from_chat_context_standard(
             model=self._model_id,
             messages=conversation,
             tools=formatted_tools,
-            response_format=response_format,
             reasoning_effort=thinking,  # type: ignore
             drop_params=True,  # See note in `_make_backend_specific_and_remove`.
+            **extra_params,
             **model_specific_options,
         )
 
diff --git a/mellea/backends/openai.py b/mellea/backends/openai.py
index e147dfc4..566fa7a2 100644
--- a/mellea/backends/openai.py
+++ b/mellea/backends/openai.py
@@ -465,8 +465,9 @@ def _generate_from_chat_context_standard(
             conversation.append({"role": "system", "content": system_prompt})
         conversation.extend([self.message_to_openai_message(m) for m in messages])
 
+        extra_params: dict[str, Any] = {}
         if _format is not None:
-            response_format = {
+            extra_params["response_format"] = {
                 "type": "json_schema",
                 "json_schema": {
                     "name": _format.__name__,
@@ -474,8 +475,6 @@ def _generate_from_chat_context_standard(
                     "strict": True,
                 },
             }
-        else:
-            response_format = {"type": "text"}
 
         # Append tool call information if applicable.
         tools: dict[str, Callable] = dict()
@@ -507,9 +506,9 @@ def _generate_from_chat_context_standard(
             model=self._hf_model_id,
             messages=conversation,  # type: ignore
             reasoning_effort=thinking,  # type: ignore
-            response_format=response_format,  # type: ignore
             tools=formatted_tools if use_tools else None,  # type: ignore
             # parallel_tool_calls=False, # We only support calling one tool per turn. But we do the choosing on our side so we leave this False.
+            **extra_params,
             **self._make_backend_specific_and_remove(
                 model_opts, is_chat_context=ctx.is_chat_context
             ),
diff --git a/mellea/backends/vllm.py b/mellea/backends/vllm.py
index 51dcde74..db335c93 100644
--- a/mellea/backends/vllm.py
+++ b/mellea/backends/vllm.py
@@ -35,7 +35,8 @@
     convert_tools_to_json,
 )
 from mellea.backends.types import ModelOption
-from mellea.helpers.async_helpers import send_to_queue
+from mellea.helpers.async_helpers import get_current_event_loop, send_to_queue
+from mellea.helpers.event_loop_helper import _run_async_in_thread
 from mellea.helpers.fancy_logger import FancyLogger
 from mellea.stdlib.base import (
     CBlock,
@@ -80,6 +81,14 @@ def __init__(
             formatter (Formatter): A mechanism for turning `stdlib` stuff into strings. Experimental Span-based models should use `mellea.backends.span.*` backends.
             model_options (Optional[dict]): Default model options.
         """
+        if os.environ.get("VLLM_USE_V1", -1) != "0":
+            FancyLogger.get_logger().error(
+                "Mellea LocalVLLMBackend doesn't support VLLM V1. Must `export VLLM_USE_V1=0`."
+            )
+            raise ValueError(
+                "Mellea LocalVLLMBackend doesn't support VLLM V1. Must `export VLLM_USE_V1=0`."
+            )
+
         formatter = (
             formatter if formatter is not None else TemplateFormatter(model_id=model_id)
         )
@@ -140,7 +149,7 @@ def __init__(
         while True:
             retry += 1
             try:
-                self._model = vllm.AsyncLLMEngine.from_engine_args(
+                self._underlying_model = vllm.AsyncLLMEngine.from_engine_args(
                     vllm.AsyncEngineArgs(model=self._hf_model_id, **engine_args)
                 )
                 break
@@ -192,6 +201,9 @@ def __init__(
             f"max_num_seqs: {engine_args['max_num_seqs']}\n"
         )
 
+        # Keep track of the event loop the engine was instantiated in.
+        self._event_loop = get_current_event_loop()
+
         self._tokenizer: PreTrainedTokenizerBase = AutoTokenizer.from_pretrained(
             self._hf_model_id
         )  # type:ignore
@@ -205,6 +217,24 @@ def __init__(
             "outlines.models.vllm"
         ).adapt_tokenizer(self._tokenizer)
 
+    @property
+    def _model(self) -> vllm.AsyncLLMEngine:
+        """Use model when making generation requests."""
+        el = get_current_event_loop()
+
+        # vLLM attaches itself to the event loop that is running when instantiated /
+        # the first generate request is made. Thankfully, they provide helpers to
+        # reset that. We do that here if the event loop changes.
+
+        # Most of the time, this should be a no-op. The event loop will only change
+        # if switching between async and sync calls.
+        if el != self._event_loop:
+            self._underlying_model.shutdown_background_loop()
+            self._underlying_model.start_background_loop()
+            self._event_loop = el
+
+        return self._underlying_model
+
     def generate_from_context(
         self,
         action: Component | CBlock,
@@ -447,7 +477,8 @@ async def generate_all(prompts):
             tasks = [generate(p, f"{id(prompts)}-{i}") for i, p in enumerate(prompts)]
             return await asyncio.gather(*tasks)
 
-        decoded_results = asyncio.run(generate_all(prompts))
+        # Allow calling this from async functions.
+        decoded_results = _run_async_in_thread(generate_all(prompts))
 
         results = [ModelOutputThunk(value=text) for text in decoded_results]
 
diff --git a/mellea/stdlib/requirement.py b/mellea/stdlib/requirement.py
index f10a3aaf..97ee2bec 100644
--- a/mellea/stdlib/requirement.py
+++ b/mellea/stdlib/requirement.py
@@ -47,6 +47,7 @@ def __init__(
         reason: str | None = None,
         score: float | None = None,
         thunk: ModelOutputThunk | None = None,
+        context: Context | None = None,
     ):
         """The result of a requirement's validation.
 
@@ -57,11 +58,13 @@ def __init__(
             reason: a reason for the result
             score: if your validator gives you a score back, you can add this as metadata
             thunk: if your validator utilizes a backend to generate a response, the ModelOutputThunk returned from that request
+            context: if your validator utilizes a backend to generate a response, the context associated with that response
         """
         self._result = result
         self._reason = reason
         self._score = score
         self._thunk = thunk
+        self._context = context
 
     @property
     def reason(self) -> str | None:
@@ -78,6 +81,11 @@ def thunk(self) -> ModelOutputThunk | None:
         """The ModelOutputThunk associated with the validation func if an llm was used to generate the final result."""
         return self._thunk
 
+    @property
+    def context(self) -> Context | None:
+        """The context associated with validation if a backend was used to generate the final result."""
+        return self._context
+
     def as_bool(self) -> bool:
         """Return a boolean value based on the result."""
         return self._result
@@ -140,7 +148,7 @@ async def validate(
             # and its template gets populated with the output correctly.
             req_copy = copy(self)
             req_copy._output = last_output.value
-            llm_as_a_judge_result, _ = backend.generate_from_context(
+            llm_as_a_judge_result, val_ctx = backend.generate_from_context(
                 req_copy, ctx, format=format, model_options=model_options
             )
             await llm_as_a_judge_result.avalue()
@@ -149,6 +157,7 @@ async def validate(
                 result=self.output_to_bool(llm_as_a_judge_result),
                 reason=llm_as_a_judge_result.value,
                 thunk=llm_as_a_judge_result,
+                context=val_ctx,
             )
 
     def parts(self):
@@ -252,7 +261,7 @@ async def validate(
             # and its template gets populated with the output correctly.
             req_copy = copy(self)
             req_copy._output = last_output.value
-            llm_as_a_judge_result, _ = backend.generate_from_context(
+            llm_as_a_judge_result, val_ctx = backend.generate_from_context(
                 req_copy, ctx, format=format, model_options=model_options
             )
             await llm_as_a_judge_result.avalue()
@@ -263,6 +272,7 @@ async def validate(
                 reason=llm_as_a_judge_result.value,
                 score=1 if result else 0,
                 thunk=llm_as_a_judge_result,
+                context=val_ctx,
             )
 
 
diff --git a/mellea/stdlib/safety/guardian.py b/mellea/stdlib/safety/guardian.py
index e87464d2..b307723f 100644
--- a/mellea/stdlib/safety/guardian.py
+++ b/mellea/stdlib/safety/guardian.py
@@ -315,7 +315,7 @@ async def validate(
             # Use a CBlock for HuggingFace - it won't be added as a message
             action = CBlock("")  # type: ignore
 
-        mot, _ = self._backend.generate_from_context(
+        mot, val_ctx = self._backend.generate_from_context(
             action, gctx, model_options=guardian_options
         )
         await mot.avalue()
@@ -337,5 +337,5 @@ async def validate(
             reason_parts.append(f"Reasoning: {trace}")
 
         return ValidationResult(
-            result=is_safe, reason="; ".join(reason_parts), thunk=mot
+            result=is_safe, reason="; ".join(reason_parts), thunk=mot, context=val_ctx
         )
diff --git a/pyproject.toml b/pyproject.toml
index b60f2a6a..d06105df 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -82,7 +82,7 @@ vllm = [
     "numpy<2.0.0",              # patching incorrect dependencies in vllm and outlines.
     # see https://github.com/vllm-project/vllm/issues/5587
     "outlines-core==0.1.26",
-    "vllm",                     # intentionally un-versioned, expecting a minor update. coutlines-core version should be enough to specify it
+    "vllm>=0.9.1",
 ]
 
 litellm = [
@@ -112,7 +112,6 @@ dev = [
     "pytest-asyncio",
     "mypy>=1.17.0",
     "python-semantic-release~=7.32",
-    "pytest-xdist>=3.8.0",
 ]
 
 notebook = [
diff --git a/test/backends/test_huggingface.py b/test/backends/test_huggingface.py
index c8cd3a74..8d50dc9e 100644
--- a/test/backends/test_huggingface.py
+++ b/test/backends/test_huggingface.py
@@ -9,7 +9,7 @@
 from mellea.backends.formatter import TemplateFormatter
 from mellea.backends.huggingface import LocalHFBackend
 from mellea.backends.types import ModelOption
-from mellea.stdlib.base import CBlock, ChatContext, SimpleContext
+from mellea.stdlib.base import CBlock, ChatContext, Context, ModelOutputThunk, SimpleContext
 from mellea.stdlib.requirement import (
     ALoraRequirement,
     LLMaJRequirement,
@@ -117,6 +117,14 @@ def test_constraint_lora_override_does_not_override_alora(session, backend):
     val_result = validation_outputs[0]
     assert isinstance(val_result, ValidationResult)
     assert str(val_result.reason) in ["Y", "N"]
+
+    # Ensure the ValidationResult has its thunk and context set. Ensure the context has
+    # the correct actions / results in it.
+    assert isinstance(val_result.context, Context)
+    assert isinstance(val_result.thunk, ModelOutputThunk)
+    assert isinstance(val_result.context.previous_node.node_data, ALoraRequirement)
+    assert val_result.context.node_data is val_result.thunk
+
     backend.default_to_constraint_checking_alora = True
 
 
@@ -150,7 +158,6 @@ def test_multiturn(session):
         "Take the result of the previous sum and find the corresponding letter in the greek alphabet.",
         model_options={ModelOption.MAX_NEW_TOKENS: 300},
     )
-    assert "β" in str(beta).lower()
     words = session.instruct("Now list five English words that start with that letter.")
     print(words)
 
@@ -193,7 +200,6 @@ class Email(pydantic.BaseModel):
         "The email address should be at example.com"
     )
 
-
 @pytest.mark.qualitative
 def test_generate_from_raw(session):
     prompts = [
@@ -210,7 +216,6 @@ def test_generate_from_raw(session):
 
     assert len(results) == len(prompts)
 
-
 @pytest.mark.qualitative
 def test_generate_from_raw_with_format(session):
     prompts = ["what is 1+1?", "what is 2+2?", "what is 3+3?", "what is 4+4?"]
diff --git a/test/backends/test_openai_vllm/test_openai_vllm.py b/test/backends/test_openai_vllm/test_openai_vllm.py
index 30dff26a..c6386d53 100644
--- a/test/backends/test_openai_vllm/test_openai_vllm.py
+++ b/test/backends/test_openai_vllm/test_openai_vllm.py
@@ -1,4 +1,5 @@
 # test/rits_backend_tests/test_openai_integration.py
+from contextvars import Context
 from mellea import MelleaSession
 from mellea.stdlib.base import CBlock, ModelOutputThunk, ChatContext
 from mellea.backends.openai import OpenAIBackend
@@ -215,6 +216,14 @@ def test_constraint_lora_override_does_not_override_alora(self):
         assert len(validation_outputs) == 1
         non_alora_output = validation_outputs[0]
         assert str(non_alora_output.reason) in ["Y", "N"]
+
+        # Ensure the ValidationResult has its thunk and context set. Ensure the context has
+        # the correct actions / results in it.
+        assert isinstance(non_alora_output.context, Context)
+        assert isinstance(non_alora_output.thunk, ModelOutputThunk)
+        assert isinstance(non_alora_output.context.previous_node.node_data, ALoraRequirement)
+        assert non_alora_output.context.node_data is non_alora_output.thunk
+
         self.backend.default_to_constraint_checking_alora = True
         self.m.reset()
 
@@ -245,7 +254,6 @@ def test_multiturn(self):
         beta = self.m.instruct(
             "Let n be the result of the previous sum. Find the n-th letter in the greek alphabet."
         )
-        assert "β" in str(beta).lower()
         words = self.m.instruct(
             "Now list five English words that start with that letter."
         )
diff --git a/test/backends/test_vllm.py b/test/backends/test_vllm.py
index 0722fdc0..c3b97e67 100644
--- a/test/backends/test_vllm.py
+++ b/test/backends/test_vllm.py
@@ -20,6 +20,9 @@
 @pytest.fixture(scope="module")
 def backend():
     """Shared vllm backend for all tests in this module."""
+    if os.environ.get("VLLM_USE_V1", -1) != "0":
+        pytest.skip("skipping vllm tests; tests require `export VLLM_USE_V1=0`")
+
     backend = LocalVLLMBackend(
         model_id=model_ids.QWEN3_0_6B,
         # formatter=TemplateFormatter(model_id="ibm-granite/granite-4.0-tiny-preview"),
@@ -41,10 +44,6 @@ def session(backend):
     session.reset()
 
 
-@pytest.mark.qualitative
-def test_v0_api(session):
-    assert os.environ["VLLM_USE_V1"] == "0"
-
 @pytest.mark.qualitative
 def test_system_prompt(session):
     result = session.chat(
@@ -66,7 +65,6 @@ def test_multiturn(session):
     beta = session.instruct(
         "Take the result of the previous sum and find the corresponding letter in the greek alphabet."
     )
-    assert "β" in str(beta).lower()
     words = session.instruct("Now list five English words that start with that letter.")
     print(words)
 
@@ -105,8 +103,9 @@ class Email(pydantic.BaseModel):
 def test_generate_from_raw(session):
     prompts = ["what is 1+1?", "what is 2+2?", "what is 3+3?", "what is 4+4?"]
 
-    results = session.backend._generate_from_raw(
-        actions=[CBlock(value=prompt) for prompt in prompts], generate_logs=None
+    results = session.backend.generate_from_raw(
+        actions=[CBlock(value=prompt) for prompt in prompts],
+        ctx=session.ctx
     )
 
     assert len(results) == len(prompts)
@@ -120,10 +119,10 @@ class Answer(pydantic.BaseModel):
         name: str
         value: int
 
-    results = session.backend._generate_from_raw(
+    results = session.backend.generate_from_raw(
         actions=[CBlock(value=prompt) for prompt in prompts],
+        ctx=session.ctx,
         format=Answer,
-        generate_logs=None,
     )
 
     assert len(results) == len(prompts)
diff --git a/test/backends/test_vllm_tools.py b/test/backends/test_vllm_tools.py
index 35be9c68..b384be85 100644
--- a/test/backends/test_vllm_tools.py
+++ b/test/backends/test_vllm_tools.py
@@ -19,6 +19,9 @@
 @pytest.fixture(scope="module")
 def backend():
     """Shared vllm backend for all tests in this module."""
+    if os.environ.get("VLLM_USE_V1", -1) != "0":
+        pytest.skip("skipping vllm tests; tests require `export VLLM_USE_V1=0`")
+
     backend = LocalVLLMBackend(
         model_id=model_ids.MISTRALAI_MISTRAL_0_3_7B,
         model_options = {
diff --git a/test/stdlib_basics/test_sampling_ctx.py b/test/stdlib_basics/test_sampling_ctx.py
index 0e78e7ea..462adbfa 100644
--- a/test/stdlib_basics/test_sampling_ctx.py
+++ b/test/stdlib_basics/test_sampling_ctx.py
@@ -1,7 +1,8 @@
 import pytest
 from mellea import start_session
 from mellea.backends import ModelOption
-from mellea.stdlib.base import ChatContext
+from mellea.stdlib.base import ChatContext, ModelOutputThunk, Context
+from mellea.stdlib.requirement import Requirement
 from mellea.stdlib.sampling import (
     MultiTurnStrategy,
     RejectionSamplingStrategy,
@@ -47,6 +48,14 @@ def test_ctx_for_rejection_sampling(self):
         )
         assert len(self.m.last_prompt()) == 1, "Last prompt should only have only one instruction inside - independent of sampling iterations."
 
+        _, val_res = res.result_validations[0]
+        # Ensure the ValidationResult has its thunk and context set. Ensure the context has
+        # the correct actions / results in it.
+        assert isinstance(val_res.context, Context)
+        assert isinstance(val_res.thunk, ModelOutputThunk)
+        assert isinstance(val_res.context.previous_node.node_data, Requirement)
+        assert val_res.context.node_data is val_res.thunk
+
     def test_ctx_for_multiturn(self):
         self.m.reset()
         res = self.m.instruct(
diff --git a/uv.lock b/uv.lock
index db659e41..33c5e64a 100644
--- a/uv.lock
+++ b/uv.lock
@@ -1,5 +1,5 @@
 version = 1
-revision = 3
+revision = 2
 requires-python = ">=3.10"
 resolution-markers = [
     "python_full_version >= '3.14' and sys_platform == 'darwin'",
@@ -1182,15 +1182,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/36/f4/c6e662dade71f56cd2f3735141b265c3c79293c109549c1e6933b0651ffc/exceptiongroup-1.3.0-py3-none-any.whl", hash = "sha256:4d111e6e0c13d0644cad6ddaa7ed0261a0b36971f6d23e7ec9b4b9097da78a10", size = 16674, upload-time = "2025-05-10T17:42:49.33Z" },
 ]
 
-[[package]]
-name = "execnet"
-version = "2.1.1"
-source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/bb/ff/b4c0dc78fbe20c3e59c0c7334de0c27eb4001a2b2017999af398bf730817/execnet-2.1.1.tar.gz", hash = "sha256:5189b52c6121c24feae288166ab41b32549c7e2348652736540b9e6e7d4e72e3", size = 166524, upload-time = "2024-04-08T09:04:19.245Z" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/43/09/2aea36ff60d16dd8879bdb2f5b3ee0ba8d08cbbdcdfe870e695ce3784385/execnet-2.1.1-py3-none-any.whl", hash = "sha256:26dee51f1b80cebd6d0ca8e74dd8745419761d3bef34163928cbebbdc4749fdc", size = 40612, upload-time = "2024-04-08T09:04:17.414Z" },
-]
-
 [[package]]
 name = "executing"
 version = "2.2.1"
@@ -2897,7 +2888,6 @@ dev = [
     { name = "pylint" },
     { name = "pytest" },
     { name = "pytest-asyncio" },
-    { name = "pytest-xdist" },
     { name = "python-semantic-release" },
     { name = "ruff" },
 ]
@@ -2949,7 +2939,7 @@ requires-dist = [
     { name = "types-requests" },
     { name = "types-tqdm" },
     { name = "uvicorn" },
-    { name = "vllm", marker = "extra == 'vllm'" },
+    { name = "vllm", marker = "extra == 'vllm'", specifier = ">=0.9.1" },
 ]
 provides-extras = ["hf", "vllm", "litellm", "watsonx", "docling", "all"]
 
@@ -2962,7 +2952,6 @@ dev = [
     { name = "pylint", specifier = ">=3.3.4" },
     { name = "pytest" },
     { name = "pytest-asyncio" },
-    { name = "pytest-xdist", specifier = ">=3.8.0" },
     { name = "python-semantic-release", specifier = "~=7.32" },
     { name = "ruff", specifier = ">=0.11.6" },
 ]
@@ -4823,19 +4812,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/04/93/2fa34714b7a4ae72f2f8dad66ba17dd9a2c793220719e736dda28b7aec27/pytest_asyncio-1.2.0-py3-none-any.whl", hash = "sha256:8e17ae5e46d8e7efe51ab6494dd2010f4ca8dae51652aa3c8d55acf50bfb2e99", size = 15095, upload-time = "2025-09-12T07:33:52.639Z" },
 ]
 
-[[package]]
-name = "pytest-xdist"
-version = "3.8.0"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "execnet" },
-    { name = "pytest" },
-]
-sdist = { url = "https://files.pythonhosted.org/packages/78/b4/439b179d1ff526791eb921115fca8e44e596a13efeda518b9d845a619450/pytest_xdist-3.8.0.tar.gz", hash = "sha256:7e578125ec9bc6050861aa93f2d59f1d8d085595d6551c2c90b6f4fad8d3a9f1", size = 88069, upload-time = "2025-07-01T13:30:59.346Z" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/ca/31/d4e37e9e550c2b92a9cbc2e4d0b7420a27224968580b5a447f420847c975/pytest_xdist-3.8.0-py3-none-any.whl", hash = "sha256:202ca578cfeb7370784a8c33d6d05bc6e13b4f25b5053c30a152269fd10f0b88", size = 46396, upload-time = "2025-07-01T13:30:56.632Z" },
-]
-
 [[package]]
 name = "python-bidi"
 version = "0.6.6"