mostlyai-engine 2.6.2__tar.gz → 2.7.1__tar.gz

{mostlyai_engine-2.6.2 → mostlyai_engine-2.7.1}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.5
 Name: mostlyai-engine
-Version: 2.6.2
+Version: 2.7.1
 Summary: Synthetic Data Engine
 Project-URL: homepage, https://github.com/mostly-ai/mostlyai-engine
 Project-URL: repository, https://github.com/mostly-ai/mostlyai-engine
@@ -27,7 +27,7 @@ Requires-Dist: accelerate>=1.5.0
 Requires-Dist: datasets>=3.0.0
 Requires-Dist: huggingface-hub[hf-xet]>=0.30.2
 Requires-Dist: joblib>=1.4.2
-Requires-Dist: json-repair>=0.47.0
+Requires-Dist: json-repair>=0.60.1
 Requires-Dist: numpy>=2.0.0
 Requires-Dist: opacus>=1.6.0
 Requires-Dist: pandas>=2.2.0
@@ -35,17 +35,16 @@ Requires-Dist: peft>=0.18.2
 Requires-Dist: psutil<6,>=5.9.5
 Requires-Dist: pyarrow>=16.0.0
 Requires-Dist: scikit-learn>=1.4.0
-Requires-Dist: setuptools<81.0.0,>=77.0.3
 Requires-Dist: tokenizers>=0.21.1
-Requires-Dist: torch<2.12.0,>=2.11.0
+Requires-Dist: torch<2.14.0,>=2.13.0
 Requires-Dist: torchaudio<2.12.0,>=2.11.0
-Requires-Dist: torchvision<0.27.0,>=0.26.0
-Requires-Dist: transformers>=5.5.1
-Requires-Dist: xgrammar<1.0.0,>=0.1.32
+Requires-Dist: torchvision<0.29.0,>=0.28.0
+Requires-Dist: transformers>=5.5.3
+Requires-Dist: xgrammar<1.0.0,>=0.2.1
 Provides-Extra: gpu
 Requires-Dist: bitsandbytes==0.42.0; (sys_platform == 'darwin') and extra == 'gpu'
 Requires-Dist: bitsandbytes>=0.45.5; (sys_platform == 'linux') and extra == 'gpu'
-Requires-Dist: vllm==0.20.0; (sys_platform == 'linux' or sys_platform == 'darwin') and extra == 'gpu'
+Requires-Dist: vllm==0.28.0; (sys_platform == 'linux') and extra == 'gpu'
 Description-Content-Type: text/markdown
 
 # Synthetic Data Engine 💎

{mostlyai_engine-2.6.2 → mostlyai_engine-2.7.1}/mostlyai/engine/__init__.py

@@ -34,7 +34,7 @@ __all__ = [
     "TabularARGN",
     "LanguageModel",
 ]
-__version__ = "2.6.2"
+__version__ = "2.7.1"
 
 # suppress specific warning related to os.fork() in multi-threaded processes
 warnings.filterwarnings("ignore", category=DeprecationWarning, message=".*multi-threaded.*fork.*")

{mostlyai_engine-2.6.2 → mostlyai_engine-2.7.1}/mostlyai/engine/_language/engine/vllm_engine.py

@@ -14,9 +14,52 @@
 
 from __future__ import annotations
 
+import inspect
+import re
+import textwrap
 import time
 from os import PathLike
 
+# Workaround for vLLM / Triton kernel warmup issue (https://github.com/vllm-project/vllm/issues/49920)
+# In Triton 3.7.x, `re.search(r"^def\s+\w+\s*\(", src, re.MULTILINE)` in `triton/runtime/jit.py:JITCallable.__init__`
+# assumes `def` is at the start of a line. When vLLM >= 0.28.0 warms up multi-decorated kernels
+# (e.g. `minimax_m3/common/ops/index_topk.py` wrapped with `@triton.heuristics`), the regex fails,
+# raising `AttributeError: 'NoneType' object has no attribute 'start'`.
+# TODO: Remove once upstream vLLM/Triton safely handles multi-decorated kernel source inspection.
+try:
+    import triton.runtime.jit
+
+    _orig_jit_callable_init = triton.runtime.jit.JITCallable.__init__
+
+    def _patched_jit_callable_init(self, fn):
+        try:
+            _orig_jit_callable_init(self, fn)
+        except AttributeError as e:
+            if "'NoneType' object has no attribute 'start'" in str(e):
+                self.fn = fn
+                self.signature = inspect.signature(fn)
+                self.raw_src, self.starting_line_number = inspect.getsourcelines(fn)
+                self._fn_name = triton.runtime.jit.get_full_name(fn)
+                self._hash_lock = triton.runtime.jit.threading.RLock()
+                src = textwrap.dedent("".join(self.raw_src))
+                m = re.search(r"(?:^|\n)\s*def\s+[\w_]+\s*\(", src)
+                if m:
+                    def_pos = src.find("def", m.start())
+                    src = src[def_pos:]
+                else:
+                    lines = src.splitlines()
+                    def_lines = [i for i, line in enumerate(lines) if "def " in line]
+                    if def_lines:
+                        src = "\n".join(lines[def_lines[0] :])
+                self._src = src
+                self.hash = None
+            else:
+                raise
+
+    triton.runtime.jit.JITCallable.__init__ = _patched_jit_callable_init
+except Exception:  # noqa: BLE001, S110
+    pass
+
 import torch
 from peft import PeftConfig
 from pydantic import BaseModel
@@ -88,7 +131,6 @@ class VLLMEngine(LanguageEngine):
             enable_lora=True,
             dtype=torch.bfloat16 if is_bf16_supported(device) else torch.float16,
             # enforce_eager=True,  # results in big slowdown, but is needed when running pytest locally
-            swap_space=0,
             disable_log_stats=True,
             tensor_parallel_size=torch.cuda.device_count(),
             gpu_memory_utilization=get_dynamic_gpu_memory_utilization(),

{mostlyai_engine-2.6.2 → mostlyai_engine-2.7.1}/mostlyai/engine/_language/xgrammar_utils.py

@@ -32,14 +32,17 @@ JSON_NULL = "null"
 
 
 def prepend_grammar_root_with_space(grammar: str) -> str:
-    # XGrammar always starts with "{" when enforcing JSON Schema
-    # training is done on strings like ` {...} {...}`
-    # later, generation is done with prompt like ` {...}`
-    # so the first output token must be a space
-    start_of_grammar = 'root ::= "{"'
-    start_of_grammar_with_space = 'root ::= " {"'
-    assert start_of_grammar in grammar
-    return grammar.replace(start_of_grammar, start_of_grammar_with_space)
+    # JSON-schema grammars start at "{". Training strings look like ` {...} {...}`
+    # and generation prompts look like ` {...}`, so the first output token must be a space.
+    # xgrammar <=0.2.3 emits `root ::= "{"`. xgrammar 0.2.7 emits `root ::= (("{"`.
+    replacements = (
+        ('root ::= "{"', 'root ::= " {"'),
+        ('root ::= (("{"', 'root ::= ((" {"'),
+    )
+    for start_of_grammar, start_of_grammar_with_space in replacements:
+        if start_of_grammar in grammar:
+            return grammar.replace(start_of_grammar, start_of_grammar_with_space)
+    raise AssertionError("unrecognized xgrammar root rule")
 
 
 def ensure_seed_can_be_tokenized(seed_data: pd.DataFrame, tokenizer: PreTrainedTokenizerBase) -> pd.DataFrame:

{mostlyai_engine-2.6.2 → mostlyai_engine-2.7.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "mostlyai-engine"
-version = "2.6.2"
+version = "2.7.1"
 description = "Synthetic Data Engine"
 requires-python = ">=3.11,<3.14"
 readme = "README.md"
@@ -23,32 +23,31 @@ classifiers = [
 ]
 
 dependencies = [
-    "setuptools>=77.0.3,<81.0.0",  # vllm 0.20 caps setuptools for Python > 3.11
     "numpy>=2.0.0",
     "pandas>=2.2.0",
     "pyarrow>=16.0.0",
     "joblib>=1.4.2",
     "scikit-learn>=1.4.0",
     "psutil>=5.9.5,<6",  # upgrade when colab psutil is updated
-    "tokenizers>=0.21.1",  # vllm 0.20 lower bound
-    "transformers>=5.5.1",  # vllm 0.20 forbids 5.0–5.4 and 5.5.0; >=5.5.1 needed for newer HF model types (e.g. gemma4)
+    "tokenizers>=0.21.1",  # vllm lower bound
+    "transformers>=5.5.3",  # aligned with vllm 0.28
     "datasets>=3.0.0",
     "accelerate>=1.5.0",
     "peft>=0.18.2",  # transformers 5.7+ checks min PEFT in model.add_adapter (integrations/peft.py)
     "huggingface-hub[hf-xet]>=0.30.2",
     "opacus>=1.6.0",
-    "xgrammar>=0.1.32,<1.0.0",  # aligned with vllm 0.20
-    "json-repair>=0.47.0",
-    "torch>=2.11.0,<2.12.0",
+    "xgrammar>=0.2.1,<1.0.0",  # aligned with vllm 0.28
+    "json-repair>=0.60.1",
+    "torch>=2.13.0,<2.14.0",
     "torchaudio>=2.11.0,<2.12.0",
-    "torchvision>=0.26.0,<0.27.0"
+    "torchvision>=0.28.0,<0.29.0",
 ]
 
 [project.optional-dependencies]
 gpu = [
     "bitsandbytes==0.42.0; sys_platform == 'darwin'",
     "bitsandbytes>=0.45.5; sys_platform == 'linux'",
-    "vllm==0.20.0; sys_platform == 'linux' or sys_platform == 'darwin'",
+    "vllm==0.28.0; sys_platform == 'linux'",
 ]
 
 [dependency-groups]
@@ -66,7 +65,7 @@ docs = [
     "mkdocs-material>=9.0",
     "mkdocs-llmstxt>=0.2",
     "griffe>=1.0",
-    "pymdown-extensions>=10.0",
+    "pymdown-extensions>=11.0",
     "griffe-fieldz>=0.2",
     "black>=26.3.1",
 ]