freesolo-flash-dev 1.4.19__tar.gz → 1.4.20__tar.gz

{freesolo_flash_dev-1.4.19 → freesolo_flash_dev-1.4.20}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: freesolo-flash-dev
-Version: 1.4.19
+Version: 1.4.20
 Summary: Flash — managed LoRA post-training (SFT/GRPO) for Freesolo environments, driven by the `flash` CLI
 Project-URL: Homepage, https://github.com/freesolo-co/flash
 Project-URL: Repository, https://github.com/freesolo-co/flash

freesolo_flash_dev-1.4.20/flash/content/processor.py

@@ -0,0 +1,37 @@
+"""Normalize chat-template metadata from a checkpoint's own processor and tokenizer."""
+
+from typing import Any
+
+
+def inherit_tokenizer_chat_template(processor: Any) -> None:
+    """Support checkpoints that publish their template only in tokenizer_config.json.
+
+    Transformers loads that template onto the tokenizer but not AutoProcessor.
+    Preserve explicit processor templates and never substitute another model's tokenizer.
+    Missing templates remain missing so the normal request validation can reject them.
+    """
+    if getattr(processor, "chat_template", None) is None:
+        tokenizer = getattr(processor, "tokenizer", None)
+        template = getattr(tokenizer, "chat_template", None)
+        if template is not None:
+            processor.chat_template = template
+
+
+def processor_template_kwargs(processor: Any, kwargs: dict[str, Any]) -> dict[str, Any]:
+    """Keep rendering options out of Transformers' legacy processor-kwargs path.
+
+    Serving renders text only: vLLM processes the decoded images separately. Options
+    unused by the selected Jinja template (often preserve_thinking) must not be
+    interpreted as image/tokenization options. Keep tools, an explicit HF argument.
+    Use HF's cached introspection so the classification matches its warning check.
+    """
+    from transformers.utils.chat_template_utils import _get_template_variables
+
+    template = getattr(processor, "chat_template", None)
+    if isinstance(template, dict):
+        template = template.get("default")
+    if not isinstance(template, str):
+        # Let the processor report missing/ambiguous template metadata as usual.
+        return kwargs
+    variables = _get_template_variables(template)
+    return {key: value for key, value in kwargs.items() if key == "tools" or key in variables}

{freesolo_flash_dev-1.4.19 → freesolo_flash_dev-1.4.20}/flash/serve/runtime/prompt.py

@@ -9,6 +9,7 @@ from collections import OrderedDict
 from dataclasses import dataclass
 from typing import Any
 
+from flash.content.processor import processor_template_kwargs
 from flash.content.reasoning_normalization import messages_for_chat_template
 from flash.serve.request.tool_calls import detached_template_messages, tools_active, tools_wire
 
@@ -222,7 +223,7 @@ class PromptPreparer:
                     template_messages,
                     tokenize=False,
                     add_generation_prompt=True,
-                    **kwargs,
+                    **processor_template_kwargs(self._processor, kwargs),
                 )
             except ServingRuntimeError:
                 raise

{freesolo_flash_dev-1.4.19 → freesolo_flash_dev-1.4.20}/flash/serving/src/engine/lora_engine.py

@@ -17,6 +17,7 @@ from functools import cached_property
 from pathlib import Path
 from typing import Any
 
+from flash.content.processor import processor_template_kwargs
 from flash.content.structured_options import vllm_structured_kwargs
 
 # Adapter-cache paths and token accounting live in engine_support.py, alongside
@@ -519,7 +520,7 @@ class _LoraEngineImpl(TextPromptTokenizer):
                     messages_for_chat_template(detached_template_messages(template_messages)),
                     tokenize=False,
                     add_generation_prompt=True,
-                    **ctk,
+                    **processor_template_kwargs(processor, ctk),
                 )
             )
         except BaseException:

{freesolo_flash_dev-1.4.19 → freesolo_flash_dev-1.4.20}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "freesolo-flash-dev"
-version = "1.4.19"
+version = "1.4.20"
 description ="Flash — managed LoRA post-training (SFT/GRPO) for Freesolo environments, driven by the `flash` CLI"
 # The training and serving stack is validated on Python 3.11 and 3.12.
 requires-python = ">=3.11,<3.13"
@@ -210,7 +210,7 @@ artifacts = ["flash/py.typed"]
 # auto-published by .github/workflows/publish-dev.yml: a push to `dev` whose version isn't yet on
 # PyPI publishes it; pushes that don't change it find it already published and no-op. So cutting a
 # dev-channel release = bump this (and [project].version) and merge to `dev`.
-version = "1.4.19"
+version = "1.4.20"
 
 [tool.pytest.ini_options]
 pythonpath = ["."]

{freesolo_flash_dev-1.4.19 → freesolo_flash_dev-1.4.20}/tests/serving/test_multimodal.py

@@ -1151,3 +1151,47 @@ def test_a_usable_tool_call_history_still_passes() -> None:
         {"role": "tool", "tool_call_id": "call_1", "content": "ok"},
     ]
     assert normalize_chat_messages(messages, supports_images=False, image_limit=4) == messages
+
+
+@pytest.mark.parametrize("runtime", ["hosted", "packaged"])
+@pytest.mark.parametrize("thinking", [False, True])
+@pytest.mark.parametrize("named_template", [False, True])
+def test_processor_render_does_not_route_unused_template_flags_to_processing(
+    runtime, thinking, named_template, caplog
+):
+    from transformers.processing_utils import ProcessorMixin
+
+    from flash.serve.runtime import EngineConfig, GenerationRequest
+    from flash.serve.runtime.prompt import PromptPreparer
+
+    # Use the real HF method: a stub accepting **kwargs cannot catch this warning.
+    # preserve_thinking is deliberately absent, as it is in some served templates.
+    template = (
+        "{{ enable_thinking }}|{{ custom }}|"
+        "{% for m in messages %}{% for c in m.content %}{{ c.type }};{% endfor %}{% endfor %}"
+    )
+    processor = types.SimpleNamespace(
+        chat_template={"default": template} if named_template else template,
+        tokenizer=types.SimpleNamespace(special_tokens_map={}),
+    )
+    processor.apply_chat_template = types.MethodType(ProcessorMixin.apply_chat_template, processor)
+    kwargs = {"custom": "kept", "preserve_thinking": False, "unused": "ignored"}
+    if runtime == "hosted":
+        engine = _engine_impl(processor=processor)
+        payload = types.SimpleNamespace(messages=_messages(), chat_template_kwargs=kwargs)
+        value = asyncio.run(engine._prepare_prompt_input(payload, thinking))
+
+        def close():
+            engine._close_prompt_images(value)
+    else:
+        preparer = PromptPreparer(EngineConfig(model=QWEN, image_limit=4), None, processor)
+        request = GenerationRequest(messages=_messages(), chat_template_kwargs=kwargs)
+        prepared = asyncio.run(preparer.prepare(request, thinking))
+        value, close = prepared.value, prepared.close
+    try:
+        assert value["prompt"].startswith(f"{thinking}|kept|")
+        assert "image;" in value["prompt"]
+        assert isinstance(value["multi_modal_data"]["image"], Image.Image)
+        assert "Kwargs passed to" not in caplog.text
+    finally:
+        close()

{freesolo_flash_dev-1.4.19 → freesolo_flash_dev-1.4.20}/tests/test_processor_chat_template.py

@@ -67,3 +67,32 @@ def test_serving_and_training_loaders_accept_inline_tokenizer_template(monkeypat
     assert loaded_tokenizer.eos_token == "<|endoftext|>"
     assert calls[0][0] == model
     assert calls[0][1]["revision"] == revision
+
+
+def test_referenced_preserve_thinking_remains_a_template_option(caplog):
+    from types import MethodType
+
+    from transformers.processing_utils import ProcessorMixin
+
+    from flash.content.processor import processor_template_kwargs
+
+    processor = SimpleNamespace(
+        chat_template="{{ enable_thinking }}|{{ preserve_thinking }}|{{ tools[0].function.name }}",
+        tokenizer=SimpleNamespace(special_tokens_map={}),
+    )
+    processor.apply_chat_template = MethodType(ProcessorMixin.apply_chat_template, processor)
+    kwargs = {
+        "enable_thinking": True,
+        "preserve_thinking": False,
+        "unused": 1,
+        "tools": [{"type": "function", "function": {"name": "lookup"}}],
+    }
+    messages = [{"role": "user", "content": [{"type": "text", "text": "hello"}]}]
+    original = processor.apply_chat_template(messages, tokenize=False, **kwargs)
+    assert "Kwargs passed to" in caplog.text
+    caplog.clear()
+    rendered = processor.apply_chat_template(
+        messages, tokenize=False, **processor_template_kwargs(processor, kwargs)
+    )
+    assert rendered == original == "True|False|lookup"
+    assert "Kwargs passed to" not in caplog.text

{freesolo_flash_dev-1.4.19 → freesolo_flash_dev-1.4.20}/uv.lock

@@ -1297,7 +1297,7 @@ wheels = [
 
 [[package]]
 name = "freesolo-flash"
-version = "1.4.19"
+version = "1.4.20"
 source = { editable = "." }
 
 [package.optional-dependencies]

freesolo_flash_dev-1.4.19/flash/content/processor.py

@@ -1,17 +0,0 @@
-"""Normalize chat-template metadata from a checkpoint's own processor and tokenizer."""
-
-from typing import Any
-
-
-def inherit_tokenizer_chat_template(processor: Any) -> None:
-    """Support checkpoints that publish their template only in tokenizer_config.json.
-
-    Transformers loads that template onto the tokenizer but not AutoProcessor.
-    Preserve explicit processor templates and never substitute another model's tokenizer.
-    Missing templates remain missing so the normal request validation can reject them.
-    """
-    if getattr(processor, "chat_template", None) is None:
-        tokenizer = getattr(processor, "tokenizer", None)
-        template = getattr(tokenizer, "chat_template", None)
-        if template is not None:
-            processor.chat_template = template