[Model] Add MiMo-V2.5 support (#40967)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: zjy0516 <zhujiangyun@inferact.ai>
Co-authored-by: yasong <yasong.wang@inferact.ai>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Copilot <copilot@github.com>
This commit is contained in:
Isotr0py
2026-04-27 13:26:51 +00:00
committed by GitHub
co-authored by Jee Jee Li zjy0516 zjy0516 yasong Claude Opus 4.6 Copilot
parent f8ac0c7cf0
commit c245d35ff4
16 changed files with 4737 additions and 5 deletions
@@ -0,0 +1,65 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers import PretrainedConfig
class Mimo_VLVisionConfig(PretrainedConfig):
model_type = "mimovl"
base_config_key = "vision_config"
def __init__(
self,
depth=28,
hidden_size=1280,
hidden_act="silu",
intermediate_size=4608,
num_heads=32,
in_channels=3,
patch_size=16,
spatial_merge_size=2,
temporal_patch_size=2,
tokens_per_second=2,
window_size=128,
out_hidden_size=2048,
fullatt_block_indexes=None,
initializer_range=0.02,
kv_channels=64, # HACK
qk_channels=64,
num_query_groups=4,
num_key_value_heads=8,
vit_window_attn_types=None,
visual_token_window_size=64,
**kwargs,
):
super().__init__(**kwargs)
self.depth = depth
self.hidden_size = hidden_size
self.hidden_act = hidden_act
self.intermediate_size = intermediate_size
self.num_heads = num_heads
# Support GQA: if num_key_value_heads is not provided,
# default to num_heads (MHA)
if num_key_value_heads is None:
num_key_value_heads = num_heads
self.num_key_value_heads = num_key_value_heads
self.in_channels = in_channels
self.patch_size = patch_size
self.spatial_merge_size = spatial_merge_size
self.temporal_patch_size = temporal_patch_size
self.tokens_per_second = tokens_per_second
self.window_size = window_size
self.fullatt_block_indexes = (
fullatt_block_indexes
if fullatt_block_indexes is not None
else [7, 15, 23, 31]
)
self.out_hidden_size = out_hidden_size
self.initializer_range = initializer_range
self.kv_channels = kv_channels
self.qk_channels = qk_channels
self.num_query_groups = num_query_groups
self.vit_window_attn_types = vit_window_attn_types or [-1] * depth
self.visual_token_window_size = visual_token_window_size
@@ -445,6 +445,37 @@ class MimoMTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
return getattr(self.hf_text_config, "num_nextn_predict_layers", 0)
def _strip_mimo_v2_attention_chunk_size(
hf_config: PretrainedConfig, hf_text_config: PretrainedConfig
) -> None:
# MiMo-V2-Flash's config.json sets `attention_chunk_size=128` but the
# architecture does not actually use chunked local attention. Leaving it
# set makes vLLM disable the hybrid KV cache manager
for cfg in (hf_text_config, hf_config):
if cfg is not None and hasattr(cfg, "attention_chunk_size"):
delattr(cfg, "attention_chunk_size")
class MimoV2ModelArchConfigConvertor(ModelArchConfigConvertorBase):
def __init__(self, hf_config: PretrainedConfig, hf_text_config: PretrainedConfig):
super().__init__(hf_config, hf_text_config)
_strip_mimo_v2_attention_chunk_size(hf_config, hf_text_config)
class MimoV2MTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
def __init__(self, hf_config: PretrainedConfig, hf_text_config: PretrainedConfig):
super().__init__(hf_config, hf_text_config)
_strip_mimo_v2_attention_chunk_size(hf_config, hf_text_config)
def get_num_hidden_layers(self) -> int:
n = getattr(self.hf_text_config, "num_nextn_predict_layers", None)
if n is not None:
return n
# Fall back to n_predict set by hf_config_override
n = getattr(self.hf_text_config, "n_predict", None)
return n if n is not None else 0
class GLM4MoeMTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
def get_num_hidden_layers(self) -> int:
return getattr(self.hf_text_config, "num_nextn_predict_layers", 0)
@@ -511,6 +542,10 @@ MODEL_ARCH_CONFIG_CONVERTORS = {
"qwen3_next_mtp": Qwen3NextMTPModelArchConfigConvertor,
"qwen3_5_mtp": Qwen3_5MTPModelArchConfigConvertor,
"mimo_mtp": MimoMTPModelArchConfigConvertor,
"mimo_v2_pro": MimoV2ModelArchConfigConvertor,
"mimo_v2_flash": MimoV2ModelArchConfigConvertor,
"mimo_v2_mtp": MimoV2MTPModelArchConfigConvertor,
"mimo_v2_omni_mtp": MimoV2MTPModelArchConfigConvertor,
"glm4_moe_mtp": GLM4MoeMTPModelArchConfigConvertor,
"glm_ocr_mtp": GLM4MoeMTPModelArchConfigConvertor,
"ernie_mtp": ErnieMTPModelArchConfigConvertor,
@@ -27,6 +27,7 @@ __all__ = [
"IsaacProcessor",
"KimiAudioProcessor",
"KimiK25Processor",
"MiMoOmniProcessor",
"MistralCommonPixtralProcessor",
"MistralCommonVoxtralProcessor",
"NanoNemotronVLProcessor",
@@ -57,6 +58,7 @@ _CLASS_TO_MODULE: dict[str, str] = {
"IsaacProcessor": "vllm.transformers_utils.processors.isaac",
"KimiAudioProcessor": "vllm.transformers_utils.processors.kimi_audio",
"KimiK25Processor": "vllm.transformers_utils.processors.kimi_k25",
"MiMoOmniProcessor": "vllm.transformers_utils.processors.mimo_v2_omni",
"MistralCommonPixtralProcessor": "vllm.transformers_utils.processors.pixtral",
"MistralCommonVoxtralProcessor": "vllm.transformers_utils.processors.voxtral",
"NanoNemotronVLProcessor": "vllm.transformers_utils.processors.nano_nemotron_vl",
File diff suppressed because it is too large Load Diff