renderers 0.1.10.dev10__tar.gz → 0.1.10.dev13__tar.gz

{renderers-0.1.10.dev10 → renderers-0.1.10.dev13}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: renderers
-Version: 0.1.10.dev10
+Version: 0.1.10.dev13
 Summary: Chat template renderers — deterministic message-to-token conversion for LLM training
 License-Expression: Apache-2.0
 License-File: LICENSE
@@ -56,7 +56,7 @@ next_prompt_ids = r.bridge_to_next_turn(
 )
 ```
 
-Hand-coded renderers ship for `qwen3`, `qwen3-vl`, `qwen3.5`, `qwen3.6`, `glm-5`, `glm-5.1`, `glm-4.5`, `minimax-m2`, `deepseek-v3`, `deepseek-r1`, `kimi-k2`, `kimi-k2.5` / `kimi-k2.6`, `laguna-xs.2`, `laguna-xs-2.1`, `laguna-s-2.1`, `laguna-m.1`, `nemotron-3`, `nemotron-3-ultra`, `nemotron-3.5`, `llama-3`, `gpt-oss`, `hy3`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
+Hand-coded renderers ship for `qwen3`, `qwen3-vl`, `qwen3.5`, `qwen3.6`, `gemma4`, `glm-5`, `glm-5.1`, `glm-4.5`, `minimax-m2`, `deepseek-v3`, `deepseek-r1`, `kimi-k2`, `kimi-k2.5` / `kimi-k2.6`, `laguna-xs.2`, `laguna-xs-2.1`, `laguna-s-2.1`, `laguna-m.1`, `nemotron-3`, `nemotron-3-ultra`, `nemotron-3.5`, `llama-3`, `gpt-oss`, `hy3`, `inkling` / `inkling-small`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper. `qwen3-vl`, `qwen3.5`, `qwen3.6`, `gemma4`, `kimi-k2.5` / `kimi-k2.6`, and the Inkling checkpoints are multimodal (Inkling handles both image **and** audio).
 
 ## API
 
@@ -159,7 +159,7 @@ renderer = create_renderer(
 
 Discriminated union: every per-renderer config is a variant of `RendererConfig`, dispatched on the `name` field. Bogus combinations (e.g. `add_vision_id` under `name="qwen3"`) error at construction with a `pydantic.ValidationError`. Downstream pydantic configs (prime-rl orchestrator, verifiers `ClientConfig`) hold a single field typed as `RendererConfig` and inherit the same strict-per-variant validation.
 
-When `chat_template_kwargs` is passed with `config=None` / `AutoRendererConfig`, renderers first resolves the concrete renderer from the model name, then validates those kwargs against that renderer's config. `Auto + unknown model + chat_template_kwargs` fails loudly; use an explicit typed config or explicit `DefaultRendererConfig` for opaque fallback templates.
+When `chat_template_kwargs` is passed with `config=None` / `AutoRendererConfig`, renderers first resolves the concrete renderer from the model name, then validates each key against that renderer's explicit template-kwarg allowlist. Renderer-only fields such as `image_cache_max` must be passed through the typed config instead. `Auto + unknown model + chat_template_kwargs` fails loudly; use an explicit typed config or explicit `DefaultRendererConfig` for opaque fallback templates.
 
 One shared behaviour flag lives on typed renderer configs: `thinking_retention`, an optional bridge-policy override. Leave it unset to derive bridge behaviour from the chat template and its renderer-exposed kwargs.
 
@@ -175,7 +175,7 @@ Fallback for unsupported text-only models. Wraps `apply_chat_template` and accep
 
 ## Roadmap
 
-- **VLM expansion.** `ImagePart` support exists for Qwen3-VL and Qwen3.5-family multimodal templates. Remaining work: video support, broader VLM coverage, and more RL validation.
+- **VLM expansion.** `ImagePart` support exists for Qwen3-VL, Qwen3.5-family, Gemma 4, and Kimi K2.5 / K2.6 multimodal templates. Remaining work: audio/video support, broader VLM coverage, and more RL validation. Gemma 4 image preprocessing requires a Transformers release that provides `Gemma4Processor`.
 - **Patched chat templates.** Some shipped templates re-tokenize history or normalize JSON in ways that break token identity. Plan: a `use_patched` opt-in per renderer that renders the same surface form while avoiding known-bad patterns. (Auto-stripping thinking from past turns is *not* one of these — that's intended template behaviour the renderer reproduces; use `thinking_retention` to override it.)
 
 ## Testing

{renderers-0.1.10.dev10 → renderers-0.1.10.dev13}/README.md

@@ -40,7 +40,7 @@ next_prompt_ids = r.bridge_to_next_turn(
 )
 ```
 
-Hand-coded renderers ship for `qwen3`, `qwen3-vl`, `qwen3.5`, `qwen3.6`, `glm-5`, `glm-5.1`, `glm-4.5`, `minimax-m2`, `deepseek-v3`, `deepseek-r1`, `kimi-k2`, `kimi-k2.5` / `kimi-k2.6`, `laguna-xs.2`, `laguna-xs-2.1`, `laguna-s-2.1`, `laguna-m.1`, `nemotron-3`, `nemotron-3-ultra`, `nemotron-3.5`, `llama-3`, `gpt-oss`, `hy3`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
+Hand-coded renderers ship for `qwen3`, `qwen3-vl`, `qwen3.5`, `qwen3.6`, `gemma4`, `glm-5`, `glm-5.1`, `glm-4.5`, `minimax-m2`, `deepseek-v3`, `deepseek-r1`, `kimi-k2`, `kimi-k2.5` / `kimi-k2.6`, `laguna-xs.2`, `laguna-xs-2.1`, `laguna-s-2.1`, `laguna-m.1`, `nemotron-3`, `nemotron-3-ultra`, `nemotron-3.5`, `llama-3`, `gpt-oss`, `hy3`, `inkling` / `inkling-small`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper. `qwen3-vl`, `qwen3.5`, `qwen3.6`, `gemma4`, `kimi-k2.5` / `kimi-k2.6`, and the Inkling checkpoints are multimodal (Inkling handles both image **and** audio).
 
 ## API
 
@@ -143,7 +143,7 @@ renderer = create_renderer(
 
 Discriminated union: every per-renderer config is a variant of `RendererConfig`, dispatched on the `name` field. Bogus combinations (e.g. `add_vision_id` under `name="qwen3"`) error at construction with a `pydantic.ValidationError`. Downstream pydantic configs (prime-rl orchestrator, verifiers `ClientConfig`) hold a single field typed as `RendererConfig` and inherit the same strict-per-variant validation.
 
-When `chat_template_kwargs` is passed with `config=None` / `AutoRendererConfig`, renderers first resolves the concrete renderer from the model name, then validates those kwargs against that renderer's config. `Auto + unknown model + chat_template_kwargs` fails loudly; use an explicit typed config or explicit `DefaultRendererConfig` for opaque fallback templates.
+When `chat_template_kwargs` is passed with `config=None` / `AutoRendererConfig`, renderers first resolves the concrete renderer from the model name, then validates each key against that renderer's explicit template-kwarg allowlist. Renderer-only fields such as `image_cache_max` must be passed through the typed config instead. `Auto + unknown model + chat_template_kwargs` fails loudly; use an explicit typed config or explicit `DefaultRendererConfig` for opaque fallback templates.
 
 One shared behaviour flag lives on typed renderer configs: `thinking_retention`, an optional bridge-policy override. Leave it unset to derive bridge behaviour from the chat template and its renderer-exposed kwargs.
 
@@ -159,7 +159,7 @@ Fallback for unsupported text-only models. Wraps `apply_chat_template` and accep
 
 ## Roadmap
 
-- **VLM expansion.** `ImagePart` support exists for Qwen3-VL and Qwen3.5-family multimodal templates. Remaining work: video support, broader VLM coverage, and more RL validation.
+- **VLM expansion.** `ImagePart` support exists for Qwen3-VL, Qwen3.5-family, Gemma 4, and Kimi K2.5 / K2.6 multimodal templates. Remaining work: audio/video support, broader VLM coverage, and more RL validation. Gemma 4 image preprocessing requires a Transformers release that provides `Gemma4Processor`.
 - **Patched chat templates.** Some shipped templates re-tokenize history or normalize JSON in ways that break token identity. Plan: a `use_patched` opt-in per renderer that renders the same surface form while avoiding known-bad patterns. (Auto-stripping thinking from past turns is *not* one of these — that's intended template behaviour the renderer reproduces; use `thinking_retention` to override it.)
 
 ## Testing

{renderers-0.1.10.dev10 → renderers-0.1.10.dev13}/docs/renderer-config.md

@@ -19,8 +19,10 @@ construction.
 
 ## Per-renderer configs
 
-Use `type(config).template_field_names()` to inspect the fields that mirror
-chat-template kwargs. Those fields are covered by parity tests against
+Use `type(config).template_field_names()` to inspect the explicit allowlist of
+fields accepted through `chat_template_kwargs`. Every renderer-specific field
+is classified as either a template field or a renderer-only field at class
+definition time. Template fields are covered by parity tests against
 `apply_chat_template` in `tests/test_renderer_config_parity.py`.
 
 | Renderer | Config class | Template fields | Renderer-only fields |
@@ -30,12 +32,14 @@ chat-template kwargs. Those fields are covered by parity tests against
 | Qwen3.5 | `Qwen35RendererConfig` | `enable_thinking`, `add_vision_id` | `image_cache_max` |
 | Qwen3.6 | `Qwen36RendererConfig` | `enable_thinking`, `add_vision_id`, `preserve_thinking` | `image_cache_max` |
 | Qwen3-VL | `Qwen3VLRendererConfig` | `add_vision_id` | `image_cache_max` |
+| Gemma 4 | `Gemma4RendererConfig` | `enable_thinking`, `preserve_thinking` | `image_cache_max` |
 | GLM-5 / 5.1 | `GLM5RendererConfig` / `GLM51RendererConfig` | `enable_thinking`, `clear_thinking` | - |
 | GLM-4.5 | `GLM45RendererConfig` | `enable_thinking` | - |
 | gpt-oss | `GptOssRendererConfig` | `reasoning_effort`, `conversation_start_date` | `use_system_prompt`, `knowledge_cutoff`, `model_identity`, `auto_drop_analysis` |
 | Hy3 | `Hy3RendererConfig` | `reasoning_effort`, `preserved_thinking`, `is_training`, `raw_last_assistant`, `fallback_strategy` | - |
 | Kimi K2 | `KimiK2RendererConfig` | - | `enable_thinking` |
 | Kimi K2.5 / 2.6 | `KimiK25RendererConfig` | `thinking` | `image_cache_max` |
+| Inkling / Inkling-Small | `InklingRendererConfig` | `reasoning_effort` | `image_cache_max`, `audio_cache_max` |
 | Laguna XS.2 | `LagunaXS2RendererConfig` | `enable_thinking`, `render_assistant_messages_raw` | - |
 | Laguna M.1 | `LagunaM1RendererConfig` | `enable_thinking`, `render_assistant_messages_raw` | - |
 | Laguna XS-2.1 | `LagunaXS21RendererConfig` | `enable_thinking` | - |
@@ -79,8 +83,10 @@ pool = create_renderer_pool(
 ```
 
 Renderers resolves auto configs before applying `chat_template_kwargs`, so the
-kwargs validate against the concrete renderer config. Unknown kwargs, or kwargs
-that conflict with an explicit `thinking_retention`, fail at construction.
+kwargs validate against the concrete renderer's template-field allowlist.
+Unknown kwargs, renderer-only fields, or kwargs that conflict with an explicit
+`thinking_retention` fail at construction. Renderer-only options remain valid
+when supplied through the typed config itself.
 
 Auto-resolution fails loudly for VLMs without an exact registered renderer.
 Text-only unknown models fall back to `DefaultRenderer`, unless
@@ -89,7 +95,10 @@ cannot implement selective bridge retention, so that combination raises.
 `AutoRendererConfig` with `chat_template_kwargs` also raises for unknown models,
 because renderers cannot validate those kwargs without a concrete renderer.
 Use an explicit model-specific config, or `DefaultRendererConfig(...)` when you
-intentionally want opaque `apply_chat_template` kwargs.
+intentionally want opaque `apply_chat_template` kwargs. Even for the default
+renderer, typed fields such as `tool_parser`, `reasoning_parser`, and
+`thinking_retention` must be passed through the config rather than through the
+opaque kwargs mapping.
 
 ## `thinking_retention`
 
@@ -131,6 +140,7 @@ the knobs its template actually exposes:
 | Qwen3 | `enable_thinking=False -> all`, else `tool_cycle` |
 | Qwen3.5 | `enable_thinking=False -> all`, else `tool_cycle` |
 | Qwen3.6 | `preserve_thinking=True -> all`; else `enable_thinking=False -> all`; else `tool_cycle` |
+| Gemma 4 | `preserve_thinking=True -> all`; else `enable_thinking=False -> all`; else `tool_cycle` |
 | GLM-5 / 5.1 | `clear_thinking=False -> all`; else `enable_thinking=False -> all`; else `tool_cycle` |
 | GLM-4.5 | `enable_thinking=False -> all`, else `tool_cycle` |
 | gpt-oss | `auto_drop_analysis=False -> all`, else `tool_cycle` |
@@ -139,7 +149,7 @@ the knobs its template actually exposes:
 | Nemotron-3 / 3.5 | `truncate_history_thinking=False -> all`; else `enable_thinking=False -> all`; else `tool_cycle` |
 | DeepSeek R1 | `template` |
 | MiniMax M2 | `tool_cycle` |
-| DeepSeek V3, Qwen3-VL, Kimi K2, Laguna XS.2 / M.1 / XS-2.1 / S-2.1, Llama 3 | `all` |
+| DeepSeek V3, Qwen3-VL, Kimi K2, Laguna XS.2 / M.1 / XS-2.1 / S-2.1, Llama 3, Inkling | `all` |
 | PrimeIntellect Qwen3 | `all` |
 
 Config construction raises when an explicit template knob directly contradicts

{renderers-0.1.10.dev10 → renderers-0.1.10.dev13}/pyproject.toml

@@ -19,6 +19,10 @@ dependencies = [
     "openai>=1.108.1",
     "tiktoken",
     "jinja2",
+    # Keep this floor compatible with prime-rl's transformers pin. Inkling's
+    # tokenizer and text-only renderer work on older releases; image/audio
+    # processing fails lazily with an upgrade message when InklingProcessor is
+    # unavailable (native support starts in transformers 5.14).
     "transformers>=4.50.0",
     # Used by GptOssRenderer to render and parse harmony tokens. Vendoring
     # OpenAI's reference implementation keeps us byte-identical with vLLM

{renderers-0.1.10.dev10 → renderers-0.1.10.dev13}/renderers/__init__.py

@@ -50,8 +50,10 @@ from renderers.configs import (
     GLM45RendererConfig,
     GLM51RendererConfig,
     GLM5RendererConfig,
+    Gemma4RendererConfig,
     GptOssRendererConfig,
     Hy3RendererConfig,
+    InklingRendererConfig,
     KimiK25RendererConfig,
     KimiK2RendererConfig,
     LagunaM1RendererConfig,
@@ -90,7 +92,9 @@ _LAZY_RENDERERS: dict[str, str] = {
     "GLM51Renderer": "renderers.glm5",
     "GLM5Renderer": "renderers.glm5",
     "GptOssRenderer": "renderers.gpt_oss",
+    "Gemma4Renderer": "renderers.gemma4",
     "Hy3Renderer": "renderers.hy3",
+    "InklingRenderer": "renderers.inkling",
     "KimiK25Renderer": "renderers.kimi_k25",
     "KimiK2Renderer": "renderers.kimi_k2",
     "LagunaM1Renderer": "renderers.laguna_xs2",
@@ -142,11 +146,15 @@ __all__ = [
     "GLM51RendererConfig",
     "GLM5Renderer",
     "GLM5RendererConfig",
+    "Gemma4Renderer",
+    "Gemma4RendererConfig",
     "GptOssRenderer",
     "GptOssRendererConfig",
     "Hy3Renderer",
     "Hy3RendererConfig",
     "ImagePart",
+    "InklingRenderer",
+    "InklingRendererConfig",
     "KimiK25Renderer",
     "KimiK25RendererConfig",
     "KimiK2Renderer",

{renderers-0.1.10.dev10 → renderers-0.1.10.dev13}/renderers/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.1.10.dev10'
-__version_tuple__ = version_tuple = (0, 1, 10, 'dev10')
+__version__ = version = '0.1.10.dev13'
+__version_tuple__ = version_tuple = (0, 1, 10, 'dev13')
 
 __commit_id__ = commit_id = None

{renderers-0.1.10.dev10 → renderers-0.1.10.dev13}/renderers/base.py

@@ -1035,6 +1035,14 @@ MODEL_RENDERER_MAP: dict[str, str] = {
     "Qwen/Qwen3-VL-4B-Instruct": "qwen3-vl",
     "Qwen/Qwen3-VL-8B-Instruct": "qwen3-vl",
     "Qwen/Qwen3-VL-30B-A3B-Instruct": "qwen3-vl",
+    # Gemma 4 instruction checkpoints share Google's canonical turn/tool
+    # grammar and dynamic Gemma4Processor image expansion. E2B/E4B omit the
+    # disabled-thinking empty-channel prefill used by the 26B/31B revision;
+    # Gemma4Renderer detects that small template variant per tokenizer.
+    "google/gemma-4-E2B-it": "gemma4",
+    "google/gemma-4-E4B-it": "gemma4",
+    "google/gemma-4-26B-A4B-it": "gemma4",
+    "google/gemma-4-31B-it": "gemma4",
     # GLM-5 family (GLM-4.7 reuses the GLM-5 template).
     "zai-org/GLM-5": "glm-5",
     "zai-org/GLM-5-FP8": "glm-5",
@@ -1085,6 +1093,10 @@ MODEL_RENDERER_MAP: dict[str, str] = {
     # GPT-OSS.
     "openai/gpt-oss-20b": "gpt-oss",
     "openai/gpt-oss-120b": "gpt-oss",
+    # Thinking Machines Inkling checkpoints share byte-identical tokenizer,
+    # chat-template, and processor assets (vision + audio; transformers >= 5.14).
+    "thinkingmachines/Inkling": "inkling",
+    "thinkingmachines/Inkling-Small": "inkling",
     # Tencent Hunyuan Hy3 (295B-A21B MoE). The FP8 checkpoint shares the same
     # tokenizer and chat template. Hy3-preview is deliberately unmapped: it
     # ships an older, incompatible template (un-suffixed special tokens,
@@ -1107,6 +1119,10 @@ MULTIMODAL_MODELS: dict[str, set[str]] = {
     "Qwen/Qwen3-VL-4B-Instruct": {"image"},
     "Qwen/Qwen3-VL-8B-Instruct": {"image"},
     "Qwen/Qwen3-VL-30B-A3B-Instruct": {"image"},
+    "google/gemma-4-E2B-it": {"image"},
+    "google/gemma-4-E4B-it": {"image"},
+    "google/gemma-4-26B-A4B-it": {"image"},
+    "google/gemma-4-31B-it": {"image"},
     # Qwen3.5 is itself a VLM family (HF tag ``image-text-to-text``,
     # processor class ``Qwen3VLProcessor``) — same vision tokens and
     # image-processor as Qwen3-VL, with a different tool-call format.
@@ -1129,6 +1145,8 @@ MULTIMODAL_MODELS: dict[str, set[str]] = {
     # ``grid_thws``.
     "moonshotai/Kimi-K2.5": {"image"},
     "moonshotai/Kimi-K2.6": {"image"},
+    "thinkingmachines/Inkling": {"image", "audio"},
+    "thinkingmachines/Inkling-Small": {"image", "audio"},
 }
 
 
@@ -1326,7 +1344,9 @@ def _populate_registry():
     from renderers.glm5 import GLM5Renderer, GLM51Renderer
     from renderers.glm45 import GLM45Renderer
     from renderers.gpt_oss import GptOssRenderer
+    from renderers.gemma4 import Gemma4Renderer
     from renderers.hy3 import Hy3Renderer
+    from renderers.inkling import InklingRenderer
     from renderers.kimi_k2 import KimiK2Renderer
     from renderers.kimi_k25 import KimiK25Renderer
     from renderers.laguna_s21 import LagunaS21Renderer
@@ -1354,6 +1374,7 @@ def _populate_registry():
             "qwen3": Qwen3Renderer,
             "prime-qwen3": PrimeQwen3Renderer,
             "qwen3-vl": Qwen3VLRenderer,
+            "gemma4": Gemma4Renderer,
             "qwen3.5": Qwen35Renderer,
             "qwen3.6": Qwen36Renderer,
             "glm-5": GLM5Renderer,
@@ -1363,6 +1384,7 @@ def _populate_registry():
             "deepseek-v3": DeepSeekV3Renderer,
             "deepseek-r1": DeepSeekR1Renderer,
             "hy3": Hy3Renderer,
+            "inkling": InklingRenderer,
             "kimi-k2": KimiK2Renderer,
             "kimi-k2.5": KimiK25Renderer,
             "laguna-xs.2": LagunaXS2Renderer,
@@ -1469,12 +1491,31 @@ def _merge_chat_template_kwargs(
         return config
     if not isinstance(chat_template_kwargs, Mapping):
         raise TypeError("chat_template_kwargs must be a mapping.")
+    kwargs = dict(chat_template_kwargs)
+    config_cls = type(config)
+    allowed = config_cls.template_field_names()
+    if config_cls._allow_opaque_template_kwargs:
+        reserved = frozenset(config_cls.model_fields) - allowed - {"name"}
+        unsupported = frozenset(kwargs) & reserved
+    else:
+        unsupported = frozenset(kwargs) - allowed
+    if unsupported:
+        allowed_text = (
+            "opaque Jinja kwargs"
+            if config_cls._allow_opaque_template_kwargs
+            else ", ".join(sorted(allowed)) or "(none)"
+        )
+        raise ValueError(
+            f"Unsupported chat_template_kwargs for {config.name!r}: "
+            f"{sorted(unsupported)}. Allowed: {allowed_text}. Pass "
+            "renderer-internal options through the typed config instead."
+        )
     data: dict[str, Any] = {"name": config.name}
     for field_name in config.__pydantic_fields_set__:
         data[field_name] = getattr(config, field_name)
     data.update(getattr(config, "model_extra", None) or {})
-    data.update(dict(chat_template_kwargs))
-    return type(config).model_validate(data)
+    data.update(kwargs)
+    return config_cls.model_validate(data)
 
 
 def _resolve_renderer_config(

{renderers-0.1.10.dev10 → renderers-0.1.10.dev13}/renderers/client.py

@@ -428,6 +428,7 @@ def _build_mm_features(
     (``MultiModalData``) is already framework-agnostic and does not need
     to change. Don't pre-build the abstraction with one engine in tree.
     """
+    from renderers.gemma4 import Gemma4Renderer
     from renderers.qwen3_vl import Qwen3VLRenderer
     from renderers.qwen35 import Qwen35Renderer
 
@@ -443,6 +444,8 @@ def _build_mm_features(
     # the family default and matches every Qwen-VL processor in tree.
     if issubclass(renderer_cls, (Qwen3VLRenderer, Qwen35Renderer)):
         return _build_qwen_vl_features(mm_data, spatial_merge_size=2)
+    if issubclass(renderer_cls, Gemma4Renderer):
+        return _build_gemma4_features(mm_data)
 
     raise NotImplementedError(
         f"Multimodal serialization not implemented for {renderer_cls.__name__}. "
@@ -450,6 +453,69 @@ def _build_mm_features(
     )
 
 
+def _build_gemma4_features(mm_data: MultiModalData) -> dict[str, Any]:
+    """vLLM features payload for Gemma 4 image inputs.
+
+    Hugging Face names the position field ``image_position_ids`` while
+    vLLM's Gemma 4 processor schema calls it ``pixel_position_ids``. Keep
+    renderer output faithful to the HF processor and translate at this
+    engine-specific boundary.
+    """
+    try:
+        import torch
+        from transformers.feature_extraction_utils import BatchFeature
+        from vllm.entrypoints.scale_out.token_in_token_out.mm_serde import (
+            encode_mm_kwargs_item,
+        )
+        from vllm.multimodal.inputs import (
+            MultiModalFieldConfig,
+            MultiModalKwargsItems,
+        )
+    except ImportError as exc:
+        raise RuntimeError(
+            "Gemma 4 multimodal generate via /inference/v1/generate requires "
+            "a vLLM release with Gemma 4 support and `torch`."
+        ) from exc
+
+    out: dict[str, Any] = {
+        "mm_hashes": {},
+        "mm_placeholders": {},
+        "kwargs_data": {},
+    }
+    image_items = mm_data.mm_items.get("image") or []
+    if image_items:
+        pixel_values = torch.cat(
+            [torch.as_tensor(item["pixel_values"]) for item in image_items], dim=0
+        )
+        pixel_position_ids = torch.cat(
+            [torch.as_tensor(item["image_position_ids"]) for item in image_items],
+            dim=0,
+        )
+        hf_inputs = BatchFeature(
+            data={
+                "pixel_values": pixel_values,
+                "pixel_position_ids": pixel_position_ids,
+            }
+        )
+        field_config = {
+            "pixel_values": MultiModalFieldConfig.batched("image"),
+            "pixel_position_ids": MultiModalFieldConfig.batched("image"),
+        }
+        kwargs_items = MultiModalKwargsItems.from_hf_inputs(hf_inputs, field_config)
+        out["kwargs_data"]["image"] = [
+            encode_mm_kwargs_item(item) for item in kwargs_items["image"]
+        ]
+        out["mm_hashes"]["image"] = list(mm_data.mm_hashes.get("image") or [])
+        out["mm_placeholders"]["image"] = [
+            {"offset": placeholder.offset, "length": placeholder.length}
+            for placeholder in mm_data.mm_placeholders.get("image") or []
+        ]
+
+    if not any(out["kwargs_data"].values()):
+        out["kwargs_data"] = None
+    return out
+
+
 def _build_qwen_vl_features(
     mm_data: MultiModalData, *, spatial_merge_size: int
 ) -> dict[str, Any]: