Add Cosmos3 Reasoner model (#43356)

Signed-off-by: Maciej Bala <mbala@nvidia.com>
Signed-off-by: MaciejBalaNV <mbala@nvidia.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Isotr0py <2037008807@qq.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
MaciejBalaNV
2026-05-28 09:43:55 -07:00
committed by GitHub
co-authored by Isotr0py Isotr0py mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Roger Wang
parent 1b5437cec8
commit 9aa131f944
10 changed files with 170 additions and 1 deletions
+1
View File
@@ -551,6 +551,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
| `ChameleonForConditionalGeneration` | Chameleon | T + I | `facebook/chameleon-7b`, etc. | | ✅︎ |
| `CheersForConditionalGeneration` | Cheers | T + I | `ai9stars/Cheers` | | ✅︎ |
| `Cohere2VisionForConditionalGeneration` | Command A Vision | T + I<sup>+</sup> | `CohereLabs/command-a-vision-07-2025`, etc. | | ✅︎ |
| `Cosmos3ForConditionalGeneration` | Cosmos3 (understanding tower) | T + I<sup>E+</sup> + V<sup>E+</sup> | `nvidia/Cosmos3-Nano` | | ✅︎ |
| `DeepseekVLV2ForCausalLM` | DeepSeek-VL2 | T + I<sup>+</sup> | `deepseek-ai/deepseek-vl2-tiny`, `deepseek-ai/deepseek-vl2-small`, `deepseek-ai/deepseek-vl2`, etc. | | ✅︎ |
| `DeepseekOCRForCausalLM` | DeepSeek-OCR | T + I<sup>+</sup> | `deepseek-ai/DeepSeek-OCR`, etc. | ✅︎ | ✅︎ |
| `DeepseekOCR2ForCausalLM` | DeepSeek-OCR-2 | T + I<sup>+</sup> | `deepseek-ai/DeepSeek-OCR-2`, etc. | ✅︎ | ✅︎ |
@@ -383,6 +383,26 @@ VLM_TEST_SETTINGS = {
max_tokens=8,
dtype="bfloat16",
),
"cosmos3": VLMTestInfo(
models=["nvidia/Cosmos3-Nano"],
test_type=(
VLMTestType.IMAGE,
VLMTestType.MULTI_IMAGE,
VLMTestType.VIDEO,
),
enforce_eager=False,
needs_video_metadata=True,
prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501
video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
num_logprobs=20,
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
patch_hf_runner=model_utils.qwen3_vl_patch_hf_runner,
image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
),
"deepseek_vl_v2": VLMTestInfo(
models=["Isotr0py/deepseek-vl2-tiny"], # model repo using dynamic module
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
+58
View File
@@ -16,6 +16,64 @@ from vllm.transformers_utils.config import try_get_safetensors_metadata
from ..registry import _MULTIMODAL_EXAMPLE_MODELS, HF_EXAMPLE_MODELS
def test_cosmos3_new_checkpoint_weights_mapper():
from vllm.model_executor.models.cosmos3 import Cosmos3ForConditionalGeneration
mapper = Cosmos3ForConditionalGeneration.hf_to_vllm_mapper
assert mapper.apply_list(
[
"layers.0.self_attn.to_q.weight",
"layers.0.self_attn.to_k.weight",
"layers.0.self_attn.to_v.weight",
"layers.0.self_attn.to_out.weight",
"layers.0.self_attn.norm_q.weight",
"layers.0.self_attn.norm_k.weight",
"embed_tokens.weight",
"norm.weight",
"lm_head.weight",
"blocks.0.attn.qkv.weight",
]
) == [
"language_model.model.layers.0.self_attn.q_proj.weight",
"language_model.model.layers.0.self_attn.k_proj.weight",
"language_model.model.layers.0.self_attn.v_proj.weight",
"language_model.model.layers.0.self_attn.o_proj.weight",
"language_model.model.layers.0.self_attn.q_norm.weight",
"language_model.model.layers.0.self_attn.k_norm.weight",
"language_model.model.embed_tokens.weight",
"language_model.model.norm.weight",
"language_model.lm_head.weight",
"visual.blocks.0.attn.qkv.weight",
]
assert (
mapper.apply_list(
[
"layers.0.self_attn.add_q_proj.weight",
"layers.0.self_attn.add_k_proj.weight",
"layers.0.self_attn.add_v_proj.weight",
"layers.0.self_attn.to_add_out.weight",
"layers.0.self_attn.norm_added_q.weight",
"layers.0.self_attn.norm_added_k.weight",
"layers.0.self_attn.q_proj_moe_gen.weight",
"layers.0.mlp_moe_gen.gate_up_proj.weight",
"norm_moe_gen.weight",
"proj_in.weight",
"proj_out.weight",
"time_embedder.linear_1.weight",
"audio_proj_in.weight",
"audio_proj_out.weight",
"action_proj_in.weight",
"action_proj_out.weight",
"audio_modality_embed",
"action_modality_embed",
]
)
== []
)
def create_repo_dummy_weights(repo: str) -> Iterable[tuple[str, torch.Tensor]]:
"""Create weights from safetensors checkpoint metadata"""
metadata = try_get_safetensors_metadata(repo)
+6
View File
@@ -846,6 +846,12 @@ _MULTIMODAL_EXAMPLE_MODELS = {
"Cohere2VisionForConditionalGeneration": _HfExamplesInfo(
"CohereLabs/command-a-vision-07-2025"
),
"Cosmos3ForConditionalGeneration": _HfExamplesInfo(
"nvidia/Cosmos3-Nano",
max_model_len=4096,
min_transformers_version="4.57",
is_available_online=False,
),
"DeepseekVLV2ForCausalLM": _HfExamplesInfo(
"deepseek-ai/deepseek-vl2-tiny",
extras={"fork": "Isotr0py/deepseek-vl2-tiny"},
@@ -177,7 +177,7 @@ class DefaultModelLoader(BaseModelLoader):
for pattern in allow_patterns:
hf_weights_files += glob.glob(os.path.join(hf_folder, pattern))
if len(hf_weights_files) > 0:
if pattern == "*.safetensors":
if pattern.endswith(".safetensors"):
use_safetensors = True
break
+73
View File
@@ -0,0 +1,73 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import regex
from vllm.config import VllmConfig
from vllm.model_executor.model_loader.default_loader import DefaultModelLoader
from vllm.model_executor.models.qwen3_vl import Qwen3VLForConditionalGeneration
from vllm.model_executor.models.utils import WeightsMapper
class Cosmos3ForConditionalGeneration(Qwen3VLForConditionalGeneration):
# Cosmos3 unified checkpoints store a Qwen3-VL understanding tower
# alongside a generation tower in a flat key layout. This mapper drops
# the generation tower weights and rewrites the understanding tower keys
# into the nested form expected by Qwen3VLForConditionalGeneration.
hf_to_vllm_mapper = WeightsMapper(
orig_to_new_regex={
regex.compile(
r"^(layers\.|embed_tokens\.|norm\.)(.+)$"
): r"language_model.model.\1\2",
regex.compile(
r"^(blocks\.|merger\.|patch_embed\.|pos_embed\.|deepstack_merger_list\.)"
): r"visual.\1",
regex.compile(r"^audio_modality_embed(?:\..*)?$"): None,
regex.compile(r"^action_modality_embed(?:\..*)?$"): None,
},
orig_to_new_substr={
"_moe_gen": None,
".add_q_proj.": None,
".add_k_proj.": None,
".add_v_proj.": None,
".to_add_out.": None,
".norm_added_q.": None,
".norm_added_k.": None,
".to_q.": ".q_proj.",
".to_k.": ".k_proj.",
".to_v.": ".v_proj.",
".to_out.": ".o_proj.",
".norm_q.": ".q_norm.",
".norm_k.": ".k_norm.",
},
orig_to_new_prefix={
"proj_in.": None,
"proj_out.": None,
"time_embedder.": None,
"audio_proj_in.": None,
"audio_proj_out.": None,
"action_proj_in.": None,
"action_proj_out.": None,
"lm_head.": "language_model.lm_head.",
},
)
allow_patterns_overrides = ["transformer/*.safetensors"]
"""
Cosmos3 checkpoint separates transformer weights and vision_encoder weights
into separate directories, as it's in diffusers checkpoint format.
Using secondary_weights here to load all necessary weights for
the Reasoner-only part.
"""
def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
super().__init__(vllm_config=vllm_config, prefix=prefix)
self.secondary_weights = [
DefaultModelLoader.Source(
model_or_path=vllm_config.model_config.model,
revision=vllm_config.model_config.revision,
prefix="",
allow_patterns_overrides=["vision_encoder/*.safetensors"],
),
]
+1
View File
@@ -369,6 +369,7 @@ _MULTIMODAL_MODELS = {
"cohere2_vision",
"Cohere2VisionForConditionalGeneration",
),
"Cosmos3ForConditionalGeneration": ("cosmos3", "Cosmos3ForConditionalGeneration"),
"DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
"DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
"DeepseekOCR2ForCausalLM": ("deepseek_ocr2", "DeepseekOCR2ForCausalLM"),
+1
View File
@@ -94,6 +94,7 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
colqwen3="ColQwen3Config",
ops_colqwen3="OpsColQwen3Config",
qwen3_vl_nemotron_embed="Qwen3VLNemotronEmbedConfig",
cosmos3_omni="Cosmos3Config",
deepseek_vl_v2="DeepseekVLV2Config",
deepseek_v32="DeepseekV3Config",
deepseek_v4="DeepseekV4Config",
@@ -25,6 +25,7 @@ _CLASS_TO_MODULE: dict[str, str] = {
"ColQwen3Config": "vllm.transformers_utils.configs.colqwen3",
"OpsColQwen3Config": "vllm.transformers_utils.configs.colqwen3",
"Qwen3VLNemotronEmbedConfig": "vllm.transformers_utils.configs.colqwen3",
"Cosmos3Config": "vllm.transformers_utils.configs.cosmos3",
"DeepseekVLV2Config": "vllm.transformers_utils.configs.deepseek_vl2",
"DeepseekV4Config": "vllm.transformers_utils.configs.deepseek_v4",
"DotsOCRConfig": "vllm.transformers_utils.configs.dotsocr",
@@ -95,6 +96,7 @@ __all__ = [
"ColQwen3Config",
"OpsColQwen3Config",
"Qwen3VLNemotronEmbedConfig",
"Cosmos3Config",
"DeepseekVLV2Config",
"DeepseekV3Config",
"DeepseekV4Config",
@@ -0,0 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers.models.qwen3_vl.configuration_qwen3_vl import Qwen3VLConfig
class Cosmos3Config(Qwen3VLConfig):
model_type = "cosmos3_omni"