Add MiniMax H3 integration

LICENSE

@@ -19,6 +19,11 @@ FlashDreams is open-source software. Its licensing posture is:
         is licensed under the Zlib license. Full text:
         LICENSES/Zlib.txt.
 
+  * MiniMax H3 codec and conditioning ports under
+    integrations_v2/minimax_h3/impl/ retain their upstream Apache-2.0
+    attribution. Full text: LICENSES/Apache-2.0.txt; source details
+    are recorded in THIRD-PARTY-NOTICES.
+
 Each first-party source file carries an inline SPDX license
 identifier; the REUSE 3.3 manifest at REUSE.toml fills the gaps for
 files (configuration, assets, generated outputs) that cannot

NOTICE

@@ -15,6 +15,10 @@ LICENSES/:
     cudaraster/framework/3rdparty/lodepng/{lodepng.h,lodepng.cpp}
         Zlib          (see LICENSES/Zlib.txt)
 
+The MiniMax/HuggingFace H3 codec and conditioning ports under
+integrations_v2/minimax_h3/impl/ retain their upstream Apache-2.0
+attribution (see LICENSES/Apache-2.0.txt).
+
 Third-party software attributions, source-level redistribution
 disclosures, and the full per-dependency license inventory are
 documented in THIRD-PARTY-NOTICES at the repository root.

REUSE.toml

@@ -66,6 +66,21 @@ SPDX-FileCopyrightText = [
 ]
 SPDX-License-Identifier = "Apache-2.0"
 
+# --- MiniMax/HuggingFace inference mathematics adapted to native APIs. ---
+[[annotations]]
+path = [
+    "integrations_v2/minimax_h3/impl/video_vae.py",
+    "integrations_v2/minimax_h3/impl/audio_encoder.py",
+    "integrations_v2/minimax_h3/impl/conditioning.py",
+    "integrations_v2/minimax_h3/impl/layout.py",
+]
+precedence = "aggregate"
+SPDX-FileCopyrightText = [
+    "Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.",
+    "Copyright 2026 The MiniMax and HuggingFace Teams. All rights reserved.",
+]
+SPDX-License-Identifier = "Apache-2.0"
+
 # --- Generated protobuf stubs. The .proto sources are authoritative; the
 #     *_pb2.py / *_pb2.pyi / *_pb2_grpc.py outputs inherit the same license
 #     because the regeneration script (compile_protos.sh) is project-owned. ---

THIRD-PARTY-NOTICES

@@ -147,6 +147,30 @@ The following third-party source is physically present in this repository
 (not just consumed at runtime). Each file retains its original copyright
 notice inline; license texts are reproduced under LICENSES/.
 
+--------------------------------------------------------------------------------
+MiniMax H3 native codecs and conditioning
+--------------------------------------------------------------------------------
+
+Path:     integrations_v2/minimax_h3/impl/{video_vae,audio_encoder,conditioning,layout}.py
+License:  Apache-2.0 (see LICENSES/Apache-2.0.txt)
+          Copyright 2026 The MiniMax and HuggingFace Teams. All rights reserved.
+Upstream: https://github.com/huggingface/diffusers/tree/175fe6b2419a01db9c2ceabd01ec37d2c0305fc2/src/diffusers
+
+The inference mathematics are adapted from the MiniMax H3 implementation.
+FlashDreams replaces pipeline blocks, model mixins, attention processors,
+checkpoint loading, and runtime orchestration with its native APIs. The
+audio port contains only reference encoding, not generated-audio decoding.
+Diffusers is not a runtime dependency of this integration. Model weights
+are not redistributed. Native checkpoint assets are pinned to
+MiniMaxAI/MiniMax-H3 revision 42ed227ee7df40d41602854ae760620d6eb651fe.
+
+The integration dynamically uses Transformers and Accelerate (Apache-2.0;
+https://github.com/huggingface/transformers and
+https://github.com/huggingface/accelerate), PyAV (BSD-3-Clause;
+https://github.com/PyAV-Org/PyAV), and Pillow (MIT-CMU;
+https://github.com/python-pillow/Pillow), alongside the shared runtime
+dependencies listed above. These dependencies are not vendored.
+
 --------------------------------------------------------------------------------
 HPG-2011 NVIDIA CUDA Rasterizer port
 --------------------------------------------------------------------------------
@@ -176,8 +200,7 @@ License:  Zlib (see LICENSES/Zlib.txt)
 Upstream: https://lodev.org/lodepng/
 
 LodePNG is a PNG codec implementation embedded by the upstream
-cudaraster framework. It is the only non-NVIDIA source physically
-redistributed in this repository. The inline notice in lodepng.h
+cudaraster framework. The inline notice in lodepng.h
 preserves the upstream copyright notice as required by the Zlib
 license; LICENSES/Zlib.txt reproduces the license text in full.
 

apps/t2v/t2v/application.py

@@ -180,6 +180,15 @@ def create_session(self, session_desc: SessionDesc) -> ISession:
         if prompt is not None and (not isinstance(prompt, str) or not prompt.strip()):
             raise ValueError("A session prompt must be non-empty text.")
         self._validate_frame_size(session_desc, pipeline)
+        cache_kwargs = self._cache_initialization_kwargs(session_desc)
+        if cache_kwargs:
+            return self.session_type(
+                pipeline,
+                prompt,
+                session_desc,
+                config.total_blocks,
+                cache_init_kwargs=cache_kwargs,
+            )
         return self.session_type(pipeline, prompt, session_desc, config.total_blocks)
 
     def close(self) -> None:
@@ -193,6 +202,13 @@ def close(self) -> None:
 
     ## Integration hooks
 
+    def _cache_initialization_kwargs(self, session_desc: SessionDesc) -> dict[str, Any]:
+        """Return model-specific request inputs retained across session resets.
+
+        Standard text, image, height, and width inputs remain framework-owned.
+        """
+        return {}
+
     def _configure_argument_parser(self, parser: argparse.ArgumentParser) -> None:
         """Add arguments this integration takes beyond the shared ones."""
 

apps/t2v/t2v/session.py

@@ -3,7 +3,7 @@
 
 """One text-to-video rollout: a prompt in, a chunk of frames per step out."""
 
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from typing import Any
 
 from flashdreams.api_v2.loop import IModelLoop
@@ -25,6 +25,9 @@ class T2VModelState:
     image: Any = None
     """Optional first-frame tensor for image-conditioned generation."""
 
+    cache_init_kwargs: dict[str, Any] = field(default_factory=dict)
+    """Integration-specific request inputs reused when resetting the cache."""
+
     blocks_generated: int = 0
     cache: Any = None
 
@@ -91,6 +94,7 @@ def __init__(
         total_blocks: int,
         *,
         image: Any = None,
+        cache_init_kwargs: dict[str, Any] | None = None,
     ) -> None:
         """
         Args:
@@ -101,12 +105,20 @@ def __init__(
                 against what the model can produce.
             total_blocks: Blocks this rollout generates before it is finished.
             image: Optional first-frame tensor retained across session resets.
+            cache_init_kwargs: Additional request inputs, excluding standard
+                ``text``, ``image``, ``height``, and ``width`` arguments.
         """
         self._pipeline = pipeline
         self._prompt = prompt
         self._session_desc = session_desc
         self._total_blocks = total_blocks
         self._image = image
+        self._cache_init_kwargs = dict(cache_init_kwargs or {})
+        reserved = {"text", "image", "height", "width"} & self._cache_init_kwargs.keys()
+        if reserved:
+            raise ValueError(
+                f"Cache inputs cannot override framework arguments: {sorted(reserved)}"
+            )
 
     def init(self) -> None:
         """Encode the prompt and prepare the rollout's cache.
@@ -119,6 +131,7 @@ def init(self) -> None:
             session_desc=self._session_desc,
             total_blocks=self._total_blocks,
             image=self._image,
+            cache_init_kwargs=dict(self._cache_init_kwargs),
         )
         if state.prompt is not None:
             state.cache = _new_cache(state)
@@ -139,15 +152,22 @@ def _new_cache(state: T2VModelState) -> Any:
     """Encode the prompt into a cache for one rollout."""
     if state.prompt is None:
         raise RuntimeError("Cannot initialize a text-to-video cache without a prompt.")
+    reserved = {"text", "image", "height", "width"} & state.cache_init_kwargs.keys()
+    if reserved:
+        raise ValueError(
+            f"Cache inputs cannot override framework arguments: {sorted(reserved)}"
+        )
     if state.image is not None:
         return state.pipeline.initialize_cache(
             text=[state.prompt],
             image=state.image,
+            **state.cache_init_kwargs,
         )
     ratio = state.pipeline.decoder.spatial_compression_ratio
     return state.pipeline.initialize_cache(
         text=[state.prompt],
         image=None,
         height=state.session_desc.video_height // ratio,
         width=state.session_desc.video_width // ratio,
+        **state.cache_init_kwargs,
     )

apps/t2v/tests/test_session.py

@@ -22,6 +22,33 @@
 
 pytestmark = pytest.mark.ci_cpu
 
+
+def test_request_kwargs_survive_reset_and_cannot_override_standard_fields():
+    from t2v.session import T2VSession
+
+    pipeline = _stand_in()
+    session = T2VSession(
+        pipeline,
+        _PROMPT,
+        _session_desc(),
+        1,
+        cache_init_kwargs={"reference_paths": ("first", "last")},
+    )
+    session.init()
+    session.model_loop.reset()
+    assert len(pipeline.caches) == 2
+    assert pipeline.caches[0]["reference_paths"] == ("first", "last")
+    assert pipeline.caches[1]["reference_paths"] == ("first", "last")
+    with pytest.raises(ValueError, match="framework arguments"):
+        T2VSession(
+            pipeline,
+            _PROMPT,
+            _session_desc(),
+            1,
+            cache_init_kwargs={"text": ["override"]},
+        )
+
+
 _WIDTH = 128
 """Frame width the stand-in generates. Not square, so a transposed frame cannot
 pass unnoticed, and a whole number of latents across."""

flashdreams/flashdreams/accelerated/multi_head_attention/__init__.py

@@ -271,7 +271,7 @@ def compute_kv(
     def forward(
         self,
         x: Tensor,
-        kv_cache: KVCacheT,
+        kv_cache: KVCacheT | None = None,
         rope_freqs: Tensor | None = None,
     ) -> Tensor:
         """Apply the configured attention type to ``x`` and ``kv_cache``.
@@ -281,7 +281,8 @@ def forward(
 
         Args:
             x: Query tokens, shape ``[..., L, query_dim]``.
-            kv_cache: Streaming cache for self-attention or precomputed static
+            kv_cache: ``None`` for cacheless bidirectional self-attention,
+                streaming cache for self-attention or precomputed static
                 cache for cross-attention. A streaming cache must already be in
                 its current-chunk update phase.
             rope_freqs: Optional positional data. Before-cache RoPE expects the