[None][feat] Cosmos3 reasoner only support #15117
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,82 @@ | ||
| # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. | ||
| # SPDX-License-Identifier: Apache-2.0 | ||
| # | ||
| # Licensed under the Apache License, Version 2.0 (the "License"); | ||
| # you may not use this file except in compliance with the License. | ||
| # You may obtain a copy of the License at | ||
| # | ||
| # http://www.apache.org/licenses/LICENSE-2.0 | ||
| # | ||
| # Unless required by applicable law or agreed to in writing, software | ||
| # distributed under the License is distributed on an "AS IS" BASIS, | ||
| # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. | ||
|
|
||
| # Adapted from upstream transformers Cosmos3OmniConfig: | ||
| # https://github.com/huggingface/transformers/blob/main/src/transformers/models/cosmos3_omni/configuration_cosmos3_omni.py | ||
| # | ||
| # Workaround until TRT-LLM upgrades to a transformers release that registers cosmos3_omni natively. | ||
|
|
||
| import os | ||
|
|
||
| from huggingface_hub.dataclasses import strict | ||
| from transformers.configuration_utils import PreTrainedConfig | ||
| from transformers.models.auto.configuration_auto import CONFIG_MAPPING, AutoConfig | ||
|
|
||
|
|
||
| @strict | ||
| class Cosmos3Config(PreTrainedConfig): | ||
| model_type = "cosmos3" | ||
| sub_configs = {"vision_config": AutoConfig, "text_config": AutoConfig} | ||
| keys_to_ignore_at_inference = ["past_key_values"] | ||
|
|
||
| text_config: dict | PreTrainedConfig | None = None | ||
| vision_config: dict | PreTrainedConfig | None = None | ||
| image_token_id: int = 151655 | ||
| video_token_id: int = 151656 | ||
| vision_start_token_id: int = 151652 | ||
| vision_end_token_id: int = 151653 | ||
| tie_word_embeddings: bool = False | ||
|
|
||
| def __post_init__(self, **kwargs): | ||
| if isinstance(self.vision_config, dict): | ||
| model_type = self.vision_config.pop("model_type", "qwen3_vl_vision") | ||
| if model_type == "qwen3_vl": | ||
| model_type = "qwen3_vl_vision" | ||
| self.vision_config = CONFIG_MAPPING[model_type](**self.vision_config) | ||
| elif self.vision_config is None: | ||
| self.vision_config = CONFIG_MAPPING["qwen3_vl_vision"]() | ||
|
|
||
| if isinstance(self.text_config, dict): | ||
| model_type = self.text_config.get("model_type", "qwen3_vl_text") | ||
| self.text_config = CONFIG_MAPPING[model_type](**self.text_config) | ||
| elif self.text_config is None: | ||
| self.text_config = CONFIG_MAPPING["qwen3_vl_text"]() | ||
|
|
||
| super().__post_init__(**kwargs) | ||
|
|
||
| @classmethod | ||
| def from_dict(cls, config_dict, **kwargs): | ||
| config = super().from_dict(config_dict, **kwargs) | ||
| # PreTrainedConfig.from_pretrained / from_dict hydrate declared fields from | ||
| # config.json only. ``_name_or_path`` is runtime metadata (not stored in the | ||
| # JSON) and is not set to the checkpoint directory automatically. Cosmos3 | ||
| # needs that path to locate ``transformer/`` and ``vision_encoder/``; TRT-LLM | ||
| # does not pass it separately on ModelConfig, so set it when callers provide | ||
| # one via kwargs. | ||
| name_or_path = kwargs.get("_name_or_path") or kwargs.get("name_or_path") | ||
| if name_or_path and ( | ||
| not getattr(config, "_name_or_path", None) or len(str(config._name_or_path)) < 2 | ||
| ): | ||
| config._name_or_path = os.fspath(name_or_path) | ||
| return config | ||
|
|
||
| @classmethod | ||
| def from_pretrained(cls, pretrained_model_name_or_path, **kwargs): | ||
| config = super().from_pretrained(pretrained_model_name_or_path, **kwargs) | ||
| # Same as from_dict above: copy the resolved checkpoint identifier onto the | ||
| # config so Cosmos3Model can find the unified checkpoint root. | ||
| if not getattr(config, "_name_or_path", None) or len(str(config._name_or_path)) < 2: | ||
| config._name_or_path = os.fspath(pretrained_model_name_or_path) | ||
| return config |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,67 @@ | ||
| # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. | ||
| # SPDX-License-Identifier: Apache-2.0 | ||
|
|
||
| import re | ||
|
|
||
| from tensorrt_llm._torch.models.checkpoints.hf.qwen3vl_weight_mapper import Qwen3VLHfWeightMapper | ||
| from tensorrt_llm._torch.models.modeling_utils import register_mapper | ||
|
|
||
|
|
||
| @register_mapper("HF", "Cosmos3ForConditionalGeneration") | ||
| class Cosmos3HfWeightMapper(Qwen3VLHfWeightMapper): | ||
| """ | ||
| Cosmos3 unified checkpoints store the Reasoner LLM with the old `model.` prefix | ||
| stripped off and the ViT under flat `blocks.*` / `merger.*` / `patch_embed.*` / | ||
| `pos_embed.*` / `deepstack_merger_list.*`. Re-target both to the nested Qwen3-VL | ||
| layout (`model.language_model.*` and `model.visual.*`). Newer checkpoints also | ||
| use Diffusers-style attention names; map them back to Qwen3-VL module names. | ||
| """ | ||
|
|
||
| KEYS_TO_DROP = ( | ||
| # Generator (image / video diffusion) MoT expert + cross-modal projections # codespell:ignore | ||
| r"\.add_q_proj\.", | ||
| r"\.add_k_proj\.", | ||
| r"\.add_v_proj\.", | ||
| r"\.to_add_out\.", | ||
| r"\.norm_added_q\.", | ||
| r"\.norm_added_k\.", | ||
| r"moe_gen", | ||
| r"^proj_out\.", | ||
| r"^proj_in\.", | ||
| r"^time_embedder\.", | ||
| # Sound tower | ||
| r"^audio_proj_out\.", | ||
| r"^audio_proj_in\.", | ||
| r"^audio_modality_embed$", | ||
| # Action tower | ||
| r"^action_proj_out\.", | ||
| r"^action_proj_in\.", | ||
| r"^action_modality_embed$", | ||
| ) | ||
|
|
||
| def __init__(self): | ||
| super().__init__() | ||
|
|
||
| self.prefix_params_map = { | ||
| r"^(layers\.|embed_tokens\.|norm\.)": r"model.language_model.\1", | ||
| r"^(blocks\.|merger\.|patch_embed\.|pos_embed\.|deepstack_merger_list\.)": r"model.visual.\1", | ||
| } | ||
| self.attn_params_map = { | ||
| r"(.*)\.self_attn\.to_q\.(.*)": r"\1.self_attn.q_proj.\2", | ||
| r"(.*)\.self_attn\.to_k\.(.*)": r"\1.self_attn.k_proj.\2", | ||
| r"(.*)\.self_attn\.to_v\.(.*)": r"\1.self_attn.v_proj.\2", | ||
| r"(.*)\.self_attn\.to_out\.(.*)": r"\1.self_attn.o_proj.\2", | ||
| r"(.*)\.self_attn\.norm_q\.(.*)": r"\1.self_attn.q_norm.\2", | ||
| r"(.*)\.self_attn\.norm_k\.(.*)": r"\1.self_attn.k_norm.\2", | ||
| } | ||
|
|
||
| @classmethod | ||
| def should_drop_checkpoint_key(cls, key: str) -> bool: | ||
| return any(re.search(pattern, key) for pattern in cls.KEYS_TO_DROP) | ||
|
|
||
| def preprocess_weights(self, weights: dict) -> dict: | ||
| weights = { | ||
| key: value for key, value in weights.items() if not self.should_drop_checkpoint_key(key) | ||
| } | ||
| weights = self.rename_by_params_map(self.prefix_params_map, weights) | ||
| return self.rename_by_params_map(self.attn_params_map, weights) |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,105 @@ | ||
| # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. | ||
| # SPDX-License-Identifier: Apache-2.0 | ||
|
|
||
| import os | ||
| from typing import Dict, Tuple | ||
|
|
||
| import safetensors.torch | ||
| import torch | ||
| from transformers import PretrainedConfig | ||
|
|
||
| from ...inputs import ( | ||
| ContentFormat, | ||
| MultimodalPlaceholderMetadata, | ||
| MultimodalPlaceholderPlacement, | ||
| register_input_processor, | ||
| support_multimodal_disaggregated, | ||
| ) | ||
| from ..model_config import ModelConfig | ||
| from .checkpoints.base_weight_mapper import BaseWeightMapper | ||
| from .checkpoints.hf.cosmos3_weight_mapper import Cosmos3HfWeightMapper | ||
| from .modeling_qwen3vl import ( | ||
| Qwen3VisionModel, | ||
| Qwen3VisionModelBase, | ||
| Qwen3VLInputProcessorBase, | ||
| Qwen3VLModel, | ||
| ) | ||
| from .modeling_utils import register_auto_model, register_vision_encoder | ||
|
|
||
|
|
||
| def _get_cosmos3_model_paths(config: PretrainedConfig) -> Tuple[str, str, str]: | ||
| """Resolve unified Cosmos3 checkpoint paths from the omni config. | ||
|
|
||
| Unified Cosmos3 checkpoints use `transformer/` for LLM weights and | ||
| `vision_encoder/` for the vision tower. | ||
| """ | ||
| root_path = config._name_or_path | ||
|
|
||
| root_path = os.fspath(root_path) | ||
| llm_path = os.path.join(root_path, "transformer") | ||
| vision_path = os.path.join(root_path, "vision_encoder") | ||
|
|
||
| if not os.path.isdir(llm_path): | ||
| raise FileNotFoundError(f"Cosmos3 transformer weights not found under {llm_path}.") | ||
|
|
||
| return root_path, llm_path, vision_path | ||
|
|
||
|
|
||
| PLACEHOLDER_METADATA = MultimodalPlaceholderMetadata( | ||
| placeholder_map={ | ||
| "image": "<|vision_start|><|image_pad|><|vision_end|>", | ||
| "video": "<|vision_start|><|video_pad|><|vision_end|>", | ||
| }, | ||
| placeholder_placement=MultimodalPlaceholderPlacement.BEFORE_TEXT, | ||
| placeholders_separator="", | ||
| content_format=ContentFormat.STRING, | ||
| ) | ||
|
|
||
|
|
||
| @support_multimodal_disaggregated | ||
| @register_vision_encoder(Qwen3VisionModelBase, vlm_base_model=Qwen3VisionModel) | ||
| @register_auto_model("Cosmos3ForConditionalGeneration") | ||
| @register_input_processor( | ||
| Qwen3VLInputProcessorBase, model_type="cosmos3", placeholder_metadata=PLACEHOLDER_METADATA | ||
| ) | ||
| # cosmos3_omni is the backward-compat alias for cosmos3, remove it when checkpoints migrate to cosmos3 | ||
| @register_input_processor( | ||
| Qwen3VLInputProcessorBase, model_type="cosmos3_omni", placeholder_metadata=PLACEHOLDER_METADATA | ||
| ) | ||
| class Cosmos3Model(Qwen3VLModel): | ||
| def __init__(self, model_config: ModelConfig[PretrainedConfig], *args, **kwargs): | ||
| omni_config = model_config.pretrained_config | ||
| if omni_config is None: | ||
| raise ValueError( | ||
| "Cosmos3Model requires model_config.pretrained_config to resolve " | ||
| "the LLM and vision encoder checkpoint paths, but it was None." | ||
| ) | ||
| if not getattr(omni_config, "_name_or_path", None): | ||
| raise ValueError( | ||
| "Cosmos3Model requires model_config.pretrained_config._name_or_path to resolve " | ||
| "the LLM and vision encoder checkpoint paths, but it was None or empty." | ||
| ) | ||
|
|
||
| (self._checkpoint_root, self.llm_path, self._vision_encoder_path) = ( | ||
| _get_cosmos3_model_paths(omni_config) | ||
| ) | ||
|
|
||
| super().__init__(model_config, *args, **kwargs) | ||
|
|
||
| @property | ||
| def llm_checkpoint_dir(self) -> str: | ||
| """Return the directory of the LLM checkpoint (``transformer/`` subdir).""" | ||
| return self.llm_path | ||
|
|
||
| def load_weights(self, weights: Dict[str, torch.Tensor], weight_mapper: BaseWeightMapper): | ||
| vision_weights_file = os.path.join(self._vision_encoder_path, "model.safetensors") | ||
| if not os.path.isfile(vision_weights_file): | ||
| raise FileNotFoundError( | ||
| f"Cosmos3 vision encoder weights not found at {vision_weights_file}." | ||
| ) | ||
| weights.update(safetensors.torch.load_file(vision_weights_file)) | ||
|
|
||
| if not isinstance(weight_mapper, Cosmos3HfWeightMapper): | ||
| weight_mapper = Cosmos3HfWeightMapper() | ||
| weights = weight_mapper.preprocess_weights(weights) | ||
| super().load_weights(weights, weight_mapper) |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Non-blocking:
This PR seems to enable
trtllm-serve. What about our offline inference paths, such astrtllm-benchor the multimodal quickstart examples? are they supported as well?