chore: vLLM, kernels, and flashinfer upgrades

README.md

@@ -21,14 +21,14 @@ Read detailed usage below, or jump to the documentation with [Getting Started](h
 uv pip install "nemo-safe-synthesizer[cu129,engine]" \
   --index https://flashinfer.ai/whl/cu129 \
   --index https://download.pytorch.org/whl/cu129 \
-  --index https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129 \
+  --index https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129 \
   --index-strategy unsafe-best-match
 
 # With pip:
 pip install "nemo-safe-synthesizer[cu129,engine]" \
   --extra-index-url https://download.pytorch.org/whl/cu129 \
   --extra-index-url https://flashinfer.ai/whl/cu129 \
-  --extra-index-url https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129
+  --extra-index-url https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129
 ```
 
 Or install from source:

docs/dev-notes/posts/introducing-nemo-safe-synthesizer.md

@@ -87,7 +87,7 @@ Each score maps to concrete remediation guidance in the documentation. The [Prod
 pip install "nemo-safe-synthesizer[cu129,engine]" \
   --extra-index-url https://download.pytorch.org/whl/cu129 \
   --extra-index-url https://flashinfer.ai/whl/cu129 \
-  --extra-index-url https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129
+  --extra-index-url https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129
 ```
 
 The quickest way to run your first pipeline is the CLI:

docs/tutorials/differential-privacy.ipynb

@@ -39,10 +39,10 @@
     "# SPDX-License-Identifier: Apache-2.0\n",
     "\n",
     "if command -v uv > /dev/null 2>&1; then\n",
-    "    uv pip install \"nemo-safe-synthesizer[engine,cu129]\" --index https://flashinfer.ai/whl/cu129 --index https://download.pytorch.org/whl/cu129 --index https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129 --index-strategy unsafe-best-match\n",
+    "    uv pip install \"nemo-safe-synthesizer[engine,cu129]\" --index https://flashinfer.ai/whl/cu129 --index https://download.pytorch.org/whl/cu129 --index https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129 --index-strategy unsafe-best-match\n",
     "    uv pip install kagglehub\n",
     "else\n",
-    "    pip install \"nemo-safe-synthesizer[engine,cu129]\" --extra-index-url https://flashinfer.ai/whl/cu129 --extra-index-url https://download.pytorch.org/whl/cu129 --extra-index-url https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129\n",
+    "    pip install \"nemo-safe-synthesizer[engine,cu129]\" --extra-index-url https://flashinfer.ai/whl/cu129 --extra-index-url https://download.pytorch.org/whl/cu129 --extra-index-url https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129\n",
     "    pip install kagglehub\n",
     "fi\n"
    ]

docs/tutorials/safe-synthesizer-101.ipynb

@@ -45,10 +45,10 @@
         "# SPDX-License-Identifier: Apache-2.0\n",
         "\n",
         "if command -v uv > /dev/null 2>&1; then\n",
-        "    uv pip install \"nemo-safe-synthesizer[engine,cu129]\" --index https://flashinfer.ai/whl/cu129 --index https://download.pytorch.org/whl/cu129 --index https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129 --index-strategy unsafe-best-match\n",
+        "    uv pip install \"nemo-safe-synthesizer[engine,cu129]\" --index https://flashinfer.ai/whl/cu129 --index https://download.pytorch.org/whl/cu129 --index https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129 --index-strategy unsafe-best-match\n",
         "    uv pip install datasets\n",
         "else\n",
-        "    pip install \"nemo-safe-synthesizer[engine,cu129]\" --extra-index-url https://flashinfer.ai/whl/cu129 --extra-index-url https://download.pytorch.org/whl/cu129 --extra-index-url https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129\n",
+        "    pip install \"nemo-safe-synthesizer[engine,cu129]\" --extra-index-url https://flashinfer.ai/whl/cu129 --extra-index-url https://download.pytorch.org/whl/cu129 --extra-index-url https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129\n",
         "    pip install datasets\n",
         "fi\n"
       ]
@@ -217,7 +217,7 @@
   ],
   "metadata": {
     "kernelspec": {
-      "display_name": "Python 3",
+      "display_name": ".venv",
       "language": "python",
       "name": "python3"
     },
@@ -231,7 +231,7 @@
       "name": "python",
       "nbconvert_exporter": "python",
       "pygments_lexer": "ipython3",
-      "version": "3.13.9"
+      "version": "3.13.12"
     }
   },
   "nbformat": 4,

docs/tutorials/time-series-financial-transactions.ipynb

@@ -57,9 +57,9 @@
         "# SPDX-License-Identifier: Apache-2.0\n",
         "\n",
         "if command -v uv > /dev/null 2>&1; then\n",
-        "    uv pip install \"nemo-safe-synthesizer[engine,cu129]\" --index https://flashinfer.ai/whl/cu129 --index https://download.pytorch.org/whl/cu129 --index https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129 --index-strategy unsafe-best-match\n",
+        "    uv pip install \"nemo-safe-synthesizer[engine,cu129]\" --index https://flashinfer.ai/whl/cu129 --index https://download.pytorch.org/whl/cu129 --index https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129 --index-strategy unsafe-best-match\n",
         "else\n",
-        "    pip install \"nemo-safe-synthesizer[engine,cu129]\" --extra-index-url https://flashinfer.ai/whl/cu129 --extra-index-url https://download.pytorch.org/whl/cu129 --extra-index-url https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129\n",
+        "    pip install \"nemo-safe-synthesizer[engine,cu129]\" --extra-index-url https://flashinfer.ai/whl/cu129 --extra-index-url https://download.pytorch.org/whl/cu129 --extra-index-url https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129\n",
         "fi"
       ]
     },

docs/user-guide/getting-started.md

@@ -37,7 +37,7 @@ indexes outside PyPI. You must pass the extra index URLs shown below.
         pip install "nemo-safe-synthesizer[cu129,engine]" \
           --extra-index-url https://download.pytorch.org/whl/cu129 \
           --extra-index-url https://flashinfer.ai/whl/cu129 \
-          --extra-index-url https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129
+          --extra-index-url https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129
         ```
 
     === "uv"
@@ -46,7 +46,7 @@ indexes outside PyPI. You must pass the extra index URLs shown below.
         uv pip install "nemo-safe-synthesizer[cu129,engine]" \
           --index https://flashinfer.ai/whl/cu129 \
           --index https://download.pytorch.org/whl/cu129 \
-          --index https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129 \
+          --index https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129 \
           --index-strategy unsafe-best-match
         ```
 

docs/user-guide/troubleshooting.md

@@ -71,22 +71,21 @@ configuration reference.
 `uv sync` fails with an error mentioning incompatible `transformers` and
 `vllm` requirements.
 
-Safe Synthesizer requires `transformers>=5.6,<6`. vLLM 0.20.0 accepts
-transformers v5, but excludes several early 5.x releases that are not
-compatible with its runtime. Keep vLLM's exclusions intact and resolve to a
-newer transformers v5 release.
+Safe Synthesizer requires `transformers>=5.12,<5.12.1` with vLLM 0.24.0.
+Keep vLLM's constraints intact so the resolver selects the tested
+Transformers/vLLM pairing.
 
 ```toml
 [project]
 dependencies = [
-  "transformers>=5.6,<6",
-  "vllm==0.20.0",
+  "transformers>=5.12,<5.12.1",
+  "vllm==0.24.0",
 ]
 ```
 
 If you've vendored or copied parts of `pyproject.toml` into another project,
 avoid adding a broad `transformers>=5.0,<6` override for vLLM. That can erase
-vLLM's explicit exclusions and allow incompatible early v5 releases.
+vLLM's explicit constraints and allow incompatible Transformers releases.
 
 ### Slow Tokenizer Warning
 

pyproject.toml

@@ -113,8 +113,8 @@ engine = [
 cpu = [
   "accelerate>=1.1.0",
   "bitsandbytes>=0.46.1",
-  "flashinfer-python==0.6.8.post1; sys_platform=='linux'",
-  "flashinfer-cubin==0.6.8.post1; sys_platform=='linux'",
+  "flashinfer-python==0.6.13; sys_platform=='linux'",
+  "flashinfer-cubin==0.6.13; sys_platform=='linux'",
   "gliner",
   "kernels>=0.12.1,<0.15",
   "peft>=0.18.0",
@@ -127,19 +127,19 @@ cpu = [
   "torchvision==0.26.0; sys_platform == 'darwin'",
   "torchvision==0.26.0+cpu; sys_platform == 'linux'",
   "torchao==0.17.0",
-  "transformers>=5.6,<6",
+  "transformers==5.12.0",
   "triton>=2.0.0; sys_platform=='linux'",
   "trl>=0.23.0",
-  "vllm==0.20.0; sys_platform=='linux'",
+  "vllm==0.24.0; sys_platform=='linux'",
   "xgrammar>=0.2.0; sys_platform=='linux'",
 ]
 
 cu129 = [
   "accelerate>=1.1.0",
   "bitsandbytes>=0.46.1",
-  "flashinfer-python==0.6.8.post1; sys_platform == 'linux'",
-  "flashinfer-cubin==0.6.8.post1; sys_platform == 'linux'",
-  "flashinfer-jit-cache==0.6.8.post1+cu129; sys_platform == 'linux'",
+  "flashinfer-python==0.6.13; sys_platform == 'linux'",
+  "flashinfer-cubin==0.6.13; sys_platform == 'linux'",
+  "flashinfer-jit-cache==0.6.13+cu129; sys_platform == 'linux'",
   "gliner",
   "kernels>=0.12.1,<0.15",
   "nvidia-cublas-cu12; sys_platform == 'linux'",
@@ -152,10 +152,10 @@ cu129 = [
   "torchaudio==2.11.0+cu129; sys_platform == 'linux'",
   "torchvision==0.26.0+cu129; sys_platform == 'linux'",
   "torchao==0.17.0+cu129; sys_platform == 'linux' and platform_machine == 'x86_64'",
-  "transformers>=5.6,<6",
+  "transformers>=5.12,<5.12.1",
   "triton>=2.0.0; sys_platform == 'linux'",
   "trl>=0.23.0",
-  "vllm==0.20.0+cu129; sys_platform == 'linux'",
+  "vllm==0.24.0+cu129; sys_platform == 'linux'",
   "xgrammar>=0.2.0; sys_platform == 'linux'",
 ]
 
@@ -180,8 +180,8 @@ conflicts = [
 
 
 override-dependencies = [
-    "flashinfer-python==0.6.8.post1; sys_platform != 'darwin'",
-    "flashinfer-cubin==0.6.8.post1; sys_platform != 'darwin'",
+    "flashinfer-python==0.6.13; sys_platform != 'darwin'",
+    "flashinfer-cubin==0.6.13; sys_platform != 'darwin'",
 ]
 
 environments = [
@@ -225,7 +225,7 @@ constraint-dependencies = [
   # here we specifiy the dep's source from which UV should attempt to resolve
   triton = { index = "pytorch-cu129", extra = "cu129" }
   vllm = [
-    { index = "vllm-v0-20-0-cu129", marker = "sys_platform == 'linux'", extra = "cu129" },
+    { index = "vllm-v0-24-0-cu129", marker = "sys_platform == 'linux'", extra = "cu129" },
   ]
 # for torch - we have a "cpu" dep set and "cu129" dep set, specified in the optional dependencies above
   torch = [
@@ -319,8 +319,8 @@ url = "https://flashinfer.ai/whl/cu129"
 explicit = true
 
 [[tool.uv.index]]
-name = "vllm-v0-20-0-cu129"
-url = "https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129"
+name = "vllm-v0-24-0-cu129"
+url = "https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129"
 explicit = true
 
 [[tool.uv.index]]

script/slurm/slurm_nss_matrix.sh

@@ -119,7 +119,7 @@ if [[ -n "${NSS_VERSION:-}" ]]; then
     uv pip install "nemo-safe-synthesizer[cu129,engine]==${NSS_VERSION}" \
         --index https://flashinfer.ai/whl/cu129 \
         --index https://download.pytorch.org/whl/cu129 \
-        --index https://wheels.vllm.ai/88d34c6409e9fb3c7b8ca0c04756f061d2099eb1/cu129 \
+        --index https://wheels.vllm.ai/ee0da84ab9e04ac7610e28580af62c365e898389/cu129 \
         --index-strategy unsafe-best-match
     NSS_RUN_CMD="${PYPI_VENV}/bin/safe-synthesizer"
     echo "[NSS SLURM] Using PyPI install: nemo-safe-synthesizer==${NSS_VERSION} on Python ${NSS_PYTHON_VERSION}"

src/nemo_safe_synthesizer/generation/processors.py

@@ -6,7 +6,7 @@
 from __future__ import annotations
 
 from abc import ABC, abstractmethod
-from typing import TYPE_CHECKING, Any
+from typing import Any, Protocol
 
 from ..config import SafeSynthesizerParameters
 from ..config.generate import ValidationParameters
@@ -21,9 +21,6 @@
 from ..llm.metadata import ModelMetadata
 from ..observability import get_logger
 
-if TYPE_CHECKING:
-    from transformers import PreTrainedTokenizerBase
-
 logger = get_logger(__name__)
 
 # Re-export the parsed-record types so existing imports of
@@ -39,6 +36,13 @@
 ]
 
 
+class EncodeOnlyTokenizer(Protocol):
+    """Tokenizer interface needed for per-record token counting."""
+
+    def encode(self, text: str, *args: Any, **kwargs: Any) -> list[int]:
+        """Return token IDs for ``text``."""
+
+
 class Processor(ABC):
     """Abstract class for processing text generation results from the LLM.
 
@@ -55,7 +59,7 @@ def __init__(
         self,
         schema: dict[str, Any],
         config: ValidationParameters,
-        tokenizer: PreTrainedTokenizerBase | None = None,
+        tokenizer: EncodeOnlyTokenizer | None = None,
     ):
         self.schema = schema
         self.config = config
@@ -166,7 +170,7 @@ def __init__(
         time_column: str | None,
         interval_seconds: int | None,
         time_format: str | None,
-        tokenizer: PreTrainedTokenizerBase | None = None,
+        tokenizer: EncodeOnlyTokenizer | None = None,
     ):
         super().__init__(schema=schema, config=config, tokenizer=tokenizer)
         if time_column is None:
@@ -232,7 +236,7 @@ def __init__(
         eos_token: str,
         group_by: str,
         order_by: str | None = None,
-        tokenizer: PreTrainedTokenizerBase | None = None,
+        tokenizer: EncodeOnlyTokenizer | None = None,
     ):
         super().__init__(schema=schema, config=config, tokenizer=tokenizer)
         self.group_by: list[str] = [group_by]
@@ -376,7 +380,7 @@ def create_processor(
     schema: dict[str, Any],
     metadata: ModelMetadata,
     config: SafeSynthesizerParameters,
-    tokenizer: PreTrainedTokenizerBase | None = None,
+    tokenizer: EncodeOnlyTokenizer | None = None,
 ) -> Processor:
     """Create the appropriate record processor for the current pipeline mode.