deprecate timeout

Signed-off-by: NickLucche <nlucches@redhat.com>

docs/models/supported_models.md

@@ -439,6 +439,7 @@ th {
 | `Mamba2ForCausalLM` | Mamba2 | `mistralai/Mamba-Codestral-7B-v0.1`, etc. | | ✅︎ |
 | `MiMoForCausalLM` | MiMo | `XiaomiMiMo/MiMo-7B-RL`, etc. | ✅︎ | ✅︎ |
 | `MiMoV2FlashForCausalLM` | MiMoV2Flash | `XiaomiMiMo/MiMo-V2-Flash`, etc. | | ✅︎ |
+| `MiMoV2ProForCausalLM` | MiMoV2Pro | `XiaomiMiMo/MiMo-V2.5-Pro`, etc. | | ✅︎ |
 | `MiniCPMForCausalLM` | MiniCPM | `openbmb/MiniCPM-2B-sft-bf16`, `openbmb/MiniCPM-2B-dpo-bf16`, `openbmb/MiniCPM-S-1B-sft`, etc. | ✅︎ | ✅︎ |
 | `MiniCPM3ForCausalLM` | MiniCPM3 | `openbmb/MiniCPM3-4B`, etc. | ✅︎ | ✅︎ |
 | `MiniMaxForCausalLM` | MiniMax-Text | `MiniMaxAI/MiniMax-Text-01-hf`, etc. | | |
@@ -590,6 +591,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `LlavaNextVideoForConditionalGeneration` | LLaVA-NeXT-Video | T + V | `llava-hf/LLaVA-NeXT-Video-7B-hf`, etc. | | ✅︎ |
 | `LlavaOnevisionForConditionalGeneration` | LLaVA-Onevision | T + I<sup>+</sup> + V<sup>+</sup> | `llava-hf/llava-onevision-qwen2-7b-ov-hf`, `llava-hf/llava-onevision-qwen2-0.5b-ov-hf`, etc. | | ✅︎ |
 | `MiDashengLMModel` | MiDashengLM | T + A<sup>+</sup> | `mispeech/midashenglm-7b` | | ✅︎ |
+| `MiMoV2OmniForCausalLM` | MiMo-V2.5-Omni | T + I<sup>E+</sup> + V<sup>E+</sup> + A<sup>+</sup> | `XiaomiMiMo/MiMo-V2.5-Omni` | | ✅︎ |
 | `MiniCPMO` | MiniCPM-O | T + I<sup>E+</sup> + V<sup>E+</sup> + A<sup>E+</sup> | `openbmb/MiniCPM-o-2_6`, etc. | ✅︎ | ✅︎ |
 | `MiniCPMV` | MiniCPM-V | T + I<sup>E+</sup> + V<sup>E+</sup> | `openbmb/MiniCPM-V-2` (see note), `openbmb/MiniCPM-Llama3-V-2_5`, `openbmb/MiniCPM-V-2_6`, `openbmb/MiniCPM-V-4`, `openbmb/MiniCPM-V-4_5`, etc. | ✅︎ | |
 | `MiniMaxVL01ForConditionalGeneration` | MiniMax-VL | T + I<sup>E+</sup> | `MiniMaxAI/MiniMax-VL-01`, etc. | | ✅︎ |

tests/models/registry.py

@@ -594,6 +594,9 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "MiMoV2FlashForCausalLM": _HfExamplesInfo(
         "XiaomiMiMo/MiMo-V2-Flash", trust_remote_code=True
     ),
+    "MiMoV2ProForCausalLM": _HfExamplesInfo(
+        "XiaomiMiMo/MiMo-V2.5-Pro", trust_remote_code=True, is_available_online=False
+    ),
     "Dots1ForCausalLM": _HfExamplesInfo("rednote-hilab/dots.llm1.inst"),
 }
 
@@ -1069,6 +1072,9 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     "MiDashengLMModel": _HfExamplesInfo(
         "mispeech/midashenglm-7b", trust_remote_code=True
     ),
+    "MiMoV2OmniForCausalLM": _HfExamplesInfo(
+        "XiaomiMiMo/MiMo-V2.5-Omni", trust_remote_code=True, is_available_online=False
+    ),
     "MiniCPMO": _HfExamplesInfo(
         "openbmb/MiniCPM-o-2_6",
         trust_remote_code=True,
@@ -1552,6 +1558,18 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
         trust_remote_code=True,
         speculative_model="XiaomiMiMo/MiMo-7B-RL",
     ),
+    "MiMoV2MTPModel": _HfExamplesInfo(
+        "XiaomiMiMo/MiMo-V2.5-Pro",
+        trust_remote_code=True,
+        speculative_model="XiaomiMiMo/MiMo-V2.5-Pro",
+        is_available_online=False,
+    ),
+    "MiMoV2OmniMTPModel": _HfExamplesInfo(
+        "XiaomiMiMo/MiMo-V2.5-Omni",
+        trust_remote_code=True,
+        speculative_model="XiaomiMiMo/MiMo-V2.5-Omni",
+        is_available_online=False,
+    ),
     "NemotronHMTPModel": _HfExamplesInfo(
         "nvidia/Nemotron-Super-Placeholder",
         speculative_model="nvidia/Nemotron-Super-Placeholder",

tests/multimodal/test_registry.py

@@ -5,6 +5,8 @@ Unit tests for MultiModalRegistry.supports_multimodal_inputs and
 Qwen2.5-VL visual component loading behavior.
 """
 
+from types import SimpleNamespace
+
 import pytest
 
 from vllm.multimodal import MULTIMODAL_REGISTRY
@@ -32,3 +34,17 @@ def test_supports_multimodal_inputs(model_id, limit_mm_per_prompt, expected):
         limit_mm_per_prompt=limit_mm_per_prompt,
     )
     assert MULTIMODAL_REGISTRY.supports_multimodal_inputs(ctx.model_config) is expected
+
+
+def test_create_processor_error_uses_served_model_name():
+    model_config = SimpleNamespace(
+        is_multimodal_model=False,
+        model="/path/to/model/weights",
+        served_model_name="friendly-model-name",
+    )
+
+    with pytest.raises(
+        ValueError,
+        match="friendly-model-name is not a multimodal model",
+    ):
+        MULTIMODAL_REGISTRY.create_processor(model_config)

vllm/config/model.py

@@ -521,6 +521,7 @@ class ModelConfig:
         if dict_overrides:
             self._apply_dict_overrides(hf_config, dict_overrides)
         self.hf_text_config = get_hf_text_config(self.hf_config)
+        self.model_arch_config = self.get_model_arch_config()
         self.attention_chunk_size = getattr(
             self.hf_text_config, "attention_chunk_size", None
         )
@@ -528,7 +529,6 @@ class ModelConfig:
         self.hf_image_processor_config = get_hf_image_processor_config(
             self.model, hf_token=self.hf_token, revision=self.revision
         )
-        self.model_arch_config = self.get_model_arch_config()
 
         architectures = self.architectures
         registry = self.registry

vllm/config/speculative.py

@@ -34,6 +34,7 @@ logger = init_logger(__name__)
 MTPModelTypes = Literal[
     "deepseek_mtp",
     "mimo_mtp",
+    "mimo_v2_mtp",
     "glm4_moe_mtp",
     "glm4_moe_lite_mtp",
     "glm_ocr_mtp",
@@ -332,6 +333,48 @@ class SpeculativeConfig:
                 }
             )
 
+        if (arch := hf_config.architectures[0]) in (
+            "MiMoV2ProForCausalLM",
+            "MiMoV2OmniForCausalLM",
+        ):
+            from vllm.model_executor.models.mimo_v2_mtp import (
+                _MIMO_V2_PRO_NUM_MTP_LAYERS,
+            )
+
+            mtp_arch_maps = {
+                "MiMoV2ProForCausalLM": "MiMoV2MTPModel",
+                "MiMoV2OmniForCausalLM": "MiMoV2OmniMTPModel",
+            }
+
+            hf_config.model_type = "mimo_v2_mtp"
+            # vLLM currently supports only the first MiMo-V2 MTP layer.
+            n_predict = _MIMO_V2_PRO_NUM_MTP_LAYERS
+            hf_config.update(
+                {
+                    "num_hidden_layers": 0,
+                    "n_predict": n_predict,
+                    "num_nextn_predict_layers": n_predict,
+                    "architectures": [mtp_arch_maps[arch]],
+                }
+            )
+
+        if hf_config.architectures[0] == "MiMoV2FlashForCausalLM":
+            from vllm.model_executor.models.mimo_v2_mtp import (
+                _MIMO_V2_FLASH_NUM_MTP_LAYERS,
+            )
+
+            hf_config.model_type = "mimo_v2_mtp"
+            # vLLM currently supports only the first MiMo-V2 MTP layer.
+            n_predict = _MIMO_V2_FLASH_NUM_MTP_LAYERS
+            hf_config.update(
+                {
+                    "num_hidden_layers": 0,
+                    "n_predict": n_predict,
+                    "num_nextn_predict_layers": n_predict,
+                    "architectures": ["MiMoV2MTPModel"],
+                }
+            )
+
         if hf_config.architectures[0] == "Glm4MoeForCausalLM":
             hf_config.model_type = "glm4_moe_mtp"
             n_predict = getattr(hf_config, "num_nextn_predict_layers", None)

vllm/distributed/kv_transfer/kv_connector/v1/nixl/scheduler.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Scheduler-side logic for the NIXL connector."""
 
+import os
 import threading
 import time
 from typing import TYPE_CHECKING, Any
@@ -88,6 +89,12 @@ class NixlConnectorScheduler:
         if vllm_config.scheduler_config.disable_hybrid_kv_cache_manager:
             logger.info("Hybrid Memory Allocator is enabled with NIXL")
 
+        if os.environ.get("VLLM_NIXL_ABORT_REQUEST_TIMEOUT") is not None:
+            logger.warning(
+                "VLLM_NIXL_ABORT_REQUEST_TIMEOUT is deprecated and will be "
+                "removed in release 0.22.0."
+            )
+
         # Background thread for handling new handshake requests.
         self._nixl_handshake_listener_t: threading.Thread | None = None
         self._stop_event = threading.Event()

vllm/model_executor/layers/linear.py

@@ -1120,7 +1120,12 @@ class QKVParallelLinear(ColumnParallelLinear):
             # Special case for Quantization.
             # If quantized, we need to adjust the offset and size to account
             # for the packing.
-            if (
+            if isinstance(param, BlockQuantScaleParameter):
+                weight_block_size = getattr(self, "weight_block_size", None)
+                shard_size, shard_offset = adjust_block_scale_shard(
+                    weight_block_size, shard_size, shard_offset
+                )
+            elif (
                 isinstance(param, (PackedColumnParameter, PackedvLLMParameter))
                 and param.packed_dim == param.output_dim
             ):

vllm/model_executor/layers/mamba/linear_attn.py

@@ -3,6 +3,7 @@
 
 import math
 from collections.abc import Callable
+from functools import partial
 
 import torch
 import torch.nn.functional as F
@@ -33,28 +34,46 @@ from vllm.v1.attention.backend import AttentionMetadata
 from vllm.v1.attention.backends.linear_attn import LinearAttentionMetadata
 
 
+@CustomOp.register("minimax_text01_rmsnorm_tp")
 class MiniMaxText01RMSNormTP(CustomOp):
-    name = "MiniMaxText01RMSNormTP"
-
-    def __init__(self, hidden_size: int, eps: float = 1e-6) -> None:
+    def __init__(
+        self,
+        hidden_size: int,
+        eps: float = 1e-6,
+        *,
+        weight_shard_world_size: int | None = None,
+        weight_shard_rank: int | None = None,
+    ) -> None:
         super().__init__()
         self.tp_world = get_tensor_model_parallel_world_size()
         self.tp_rank = get_tensor_model_parallel_rank()
-        self.weight = nn.Parameter(torch.ones(int(hidden_size / self.tp_world)))
+        self.weight_shard_world = weight_shard_world_size or self.tp_world
+        self.weight_shard_rank = (
+            self.tp_rank if weight_shard_rank is None else weight_shard_rank
+        )
 
-        self.weight.weight_loader = self.weight_loader
+        self.weight = nn.Parameter(torch.ones(hidden_size // self.weight_shard_world))
+        self.weight.weight_loader = partial(
+            self.weight_loader,
+            shard_world_size=self.weight_shard_world,
+            shard_rank=self.weight_shard_rank,
+        )
         self.variance_epsilon = eps
 
     @staticmethod
     def weight_loader(
         param: nn.Parameter,
         loaded_weight: torch.Tensor,
+        shard_world_size: int | None = None,
+        shard_rank: int | None = None,
     ) -> None:
-        tp_world = get_tensor_model_parallel_world_size()
-        tp_rank = get_tensor_model_parallel_rank()
+        if shard_world_size is None:
+            shard_world_size = get_tensor_model_parallel_world_size()
+        if shard_rank is None:
+            shard_rank = get_tensor_model_parallel_rank()
 
-        shard_size = loaded_weight.shape[0] // tp_world
-        shard = slice(tp_rank * shard_size, (tp_rank + 1) * shard_size)
+        shard_size = loaded_weight.shape[0] // shard_world_size
+        shard = slice(shard_rank * shard_size, (shard_rank + 1) * shard_size)
         param.data.copy_(loaded_weight[shard])
 
     def _forward(

vllm/model_executor/models/mimo_v2.py

@@ -6,6 +6,7 @@ from itertools import islice
 import torch
 from torch import nn
 
+from vllm.compilation.decorators import support_torch_compile
 from vllm.config import (
     CacheConfig,
     VllmConfig,
@@ -268,7 +269,7 @@ class MiMoV2Attention(nn.Module):
             self.total_num_heads * self.v_head_dim,
             hidden_size,
             bias=False,
-            quant_config=quant_config,
+            quant_config=quant_config if "mtp.layers" not in prefix else None,
             reduce_results=True,
             prefix=f"{prefix}.o_proj",
         )
@@ -440,6 +441,7 @@ class MiMoV2FlashDecoderLayer(nn.Module):
         return self.config.hybrid_layer_pattern[self.layer_id] == 1
 
 
+@support_torch_compile
 class MiMoV2Model(nn.Module):
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
@@ -603,7 +605,13 @@ class MiMoV2Model(nn.Module):
 
             if expert_matched:
                 continue
-
+            # Support fused qkv_proj checkpoint (Pro format)
+            if "qkv_proj" in name:
+                if name in params_dict:
+                    param = params_dict[name]
+                    loaded_weight = loaded_weight.chunk(tp_size, dim=0)[tp_rank]
+                    default_weight_loader(param, loaded_weight)
+                continue
             stacked_matched = False
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
@@ -662,6 +670,11 @@ class MiMoV2Model(nn.Module):
 
 
 class MiMoV2FlashForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
+    packed_modules_mapping = {
+        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+        "gate_up_proj": ["gate_proj", "up_proj"],
+    }
+
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         config = vllm_config.model_config.hf_config
@@ -718,3 +731,10 @@ class MiMoV2FlashForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
         return loader.load_weights(weights)
+
+
+class MiMoV2ProForCausalLM(MiMoV2FlashForCausalLM):
+    packed_modules_mapping = {
+        "qkv_proj": ["qkv_proj"],
+        "gate_up_proj": ["gate_proj", "up_proj"],
+    }

vllm/model_executor/models/mimo_v2_mtp.py

@@ -0,0 +1,373 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""Inference-only MiMo-V2 MTP (Multi-Token Prediction) draft model.
+
+Supports both MiMo-V2-Pro and MiMo-V2-Flash checkpoints.
+
+Checkpoint weight layout (model.mtp.layers.{idx}.*):
+  enorm            - RMSNorm for token embeddings
+  hnorm            - RMSNorm for previous hidden states
+  eh_proj          - ReplicatedLinear(hidden*2 -> hidden)
+  input_layernorm  - pre-attention RMSNorm
+  self_attn.*      - attention weights; format differs by variant:
+                       Pro:   fused qkv_proj  [Q;K;V] concatenated
+                       Flash: separate q_proj, k_proj, v_proj
+  pre_mlp_layernorm - post-attention / pre-MLP RMSNorm
+  mlp.*            - dense MLP (gate_proj / up_proj / down_proj)
+  final_layernorm  - norm applied before logit computation
+"""
+
+from collections.abc import Iterable
+
+import torch
+import torch.nn as nn
+from transformers import PretrainedConfig
+
+from vllm.config import VllmConfig
+from vllm.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+)
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import ReplicatedLinear
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.sequence import IntermediateTensors
+
+from .interfaces import (
+    MultiModalEmbeddings,
+    SupportsMultiModal,
+    _require_is_multimodal,
+)
+from .mimo_v2 import MiMoV2Attention, MiMoV2MLP
+from .utils import _merge_multimodal_embeddings, maybe_prefix
+
+# MiMo-V2 checkpoints contain multiple MTP layers, but vLLM currently supports
+# only the first layer and only one speculative token.
+_MIMO_V2_PRO_NUM_MTP_LAYERS = 1
+_MIMO_V2_FLASH_NUM_MTP_LAYERS = 1
+
+
+class MiMoV2MTPLayer(nn.Module):
+    """Single MTP predictor layer for MiMo-V2 (Pro and Flash).
+
+    Mirrors the single-layer MiMo-V2 nextn reference implementation.
+    """
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        prefix: str,
+        quant_config: QuantizationConfig | None = None,
+    ) -> None:
+        super().__init__()
+
+        # Predictor head components
+        self.enorm = RMSNorm(config.hidden_size, eps=config.layernorm_epsilon)
+        self.hnorm = RMSNorm(config.hidden_size, eps=config.layernorm_epsilon)
+        self.eh_proj = ReplicatedLinear(
+            config.hidden_size * 2, config.hidden_size, bias=False
+        )
+
+        # MTP uses the SWA attention configuration
+        # implementation.
+        swa_rope_theta = getattr(
+            config,
+            "swa_rope_theta",
+            getattr(config, "rope_theta", 1000000),
+        )
+        sliding_window_size = getattr(config, "sliding_window_size", -1)
+
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.layernorm_epsilon)
+        self.self_attn = MiMoV2Attention(
+            hidden_size=config.hidden_size,
+            num_heads=config.swa_num_attention_heads,
+            num_kv_heads=config.swa_num_key_value_heads,
+            head_dim=config.swa_head_dim,
+            v_head_dim=getattr(config, "swa_v_head_dim", None),
+            v_scale=getattr(config, "attention_value_scale", None),
+            sliding_window_size=sliding_window_size,
+            attention_bias=config.attention_bias,
+            add_swa_attention_sink_bias=getattr(
+                config, "add_swa_attention_sink_bias", False
+            ),
+            layer_id=0,
+            rope_theta=swa_rope_theta,
+            max_position_embeddings=getattr(config, "max_position_embeddings", 32768),
+            quant_config=quant_config,
+            partial_rotary_factor=getattr(config, "partial_rotary_factor", 1.0),
+            prefix=f"{prefix}.self_attn",
+        )
+        self.pre_mlp_layernorm = RMSNorm(
+            config.hidden_size, eps=config.layernorm_epsilon
+        )
+        self.mlp = MiMoV2MLP(
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            prefix=f"{prefix}.mlp",
+        )
+        self.final_layernorm = RMSNorm(config.hidden_size, eps=config.layernorm_epsilon)
+
+    def forward(
+        self,
+        inputs_embeds: torch.Tensor,
+        positions: torch.Tensor,
+        previous_hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        # Combine token embedding and previous hidden state
+        h, _ = self.eh_proj(
+            torch.cat(
+                [self.enorm(inputs_embeds), self.hnorm(previous_hidden_states)], dim=-1
+            )
+        )
+
+        # Transformer block with fused residual norms
+        residual = h
+        h = self.input_layernorm(h)
+        h = self.self_attn(positions=positions, hidden_states=h)
+        h, residual = self.pre_mlp_layernorm(h, residual)
+        h = self.mlp(h)
+        h = h + residual
+
+        return self.final_layernorm(h)
+
+
+class _MiMoV2MTPLayers(nn.Module):
+    """Thin wrapper so parameter paths match checkpoint: model.mtp.layers.*"""
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        num_mtp_layers: int,
+        quant_config: QuantizationConfig | None,
+        prefix: str,
+    ) -> None:
+        super().__init__()
+        self.layers = nn.ModuleDict(
+            {
+                str(i): MiMoV2MTPLayer(
+                    config=config,
+                    prefix=f"{prefix}.{i}",
+                    quant_config=quant_config,
+                )
+                for i in range(num_mtp_layers)
+            }
+        )
+
+
+class MiMoV2MultiTokenPredictor(nn.Module):
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        spec_cfg = vllm_config.speculative_config
+        assert spec_cfg is not None
+        if spec_cfg.num_speculative_tokens != 1:
+            raise ValueError(
+                "MiMo-V2 MTP in vLLM only supports num_speculative_tokens=1."
+            )
+        num_mtp_layers = 1
+
+        self.num_mtp_layers = num_mtp_layers
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+
+        self.mtp = _MiMoV2MTPLayers(
+            config=config,
+            num_mtp_layers=num_mtp_layers,
+            quant_config=vllm_config.quant_config,
+            prefix=maybe_prefix(prefix, "mtp.layers"),
+        )
+
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        previous_hidden_states: torch.Tensor,
+        inputs_embeds: torch.Tensor | None = None,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor:
+        assert spec_step_idx == 0, "MiMo-V2 MTP only supports one speculative token."
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_input_ids(input_ids)
+        return self.mtp.layers[str(spec_step_idx)](
+            inputs_embeds, positions, previous_hidden_states
+        )
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        lm_head: ParallelLMHead,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor:
+        assert spec_step_idx == 0, "MiMo-V2 MTP only supports one speculative token."
+        return self.logits_processor(lm_head, hidden_states)
+
+
+class MiMoV2MTP(nn.Module):
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
+        super().__init__()
+        self.config = vllm_config.model_config.hf_config
+        self.model = MiMoV2MultiTokenPredictor(
+            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
+        )
+        self.lm_head = ParallelLMHead(
+            self.config.vocab_size,
+            self.config.hidden_size,
+            prefix=maybe_prefix(prefix, "lm_head"),
+        )
+
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.embed_input_ids(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor | None,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        intermediate_tensors: IntermediateTensors | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor:
+        assert spec_step_idx == 0, "MiMo-V2 MTP only supports one speculative token."
+        return self.model(
+            input_ids, positions, hidden_states, inputs_embeds, spec_step_idx
+        )
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor | None:
+        assert spec_step_idx == 0, "MiMo-V2 MTP only supports one speculative token."
+        return self.model.compute_logits(hidden_states, self.lm_head, spec_step_idx)
+
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+        tp_rank = get_tensor_model_parallel_rank()
+        tp_size = get_tensor_model_parallel_world_size()
+
+        stacked_params_mapping = [
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+            # Flash format: separate projections → fused qkv_proj
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+        ]
+
+        params_dict = dict(self.named_parameters())
+        loaded_params: set[str] = set()
+
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name:
+                continue
+
+            # Only load MTP-related weights, shared embeddings, and lm_head
+            if (
+                "model.mtp" not in name
+                and "model.embed_tokens" not in name
+                and not name.startswith("lm_head")
+            ):
+                continue
+
+            # Support fused qkv_proj checkpoint (Pro format).
+            # The checkpoint is stored pre-sharded for TP=8 as
+            # [Q_rank0, K_rank0, V_rank0, Q_rank1, ...], so splitting along
+            # dim 0 with chunk(tp_size) gives each rank its Q+K+V slice for
+            # both the FP8 weight and the block weight_scale_inv. This matches
+            # how the main model loads the same layout.
+            if "qkv_proj" in name:
+                if name in params_dict:
+                    param = params_dict[name]
+                    loaded_weight = loaded_weight.chunk(tp_size, dim=0)[tp_rank]
+                    default_weight_loader(param, loaded_weight)
+                    loaded_params.add(name)
+                continue
+
+            # gate_proj/up_proj → gate_up_proj stacking (both formats);
+            # Flash: q_proj/k_proj/v_proj → qkv_proj merging.
+            stacked_matched = False
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name_rewritten = name.replace(weight_name, param_name)
+                if (
+                    name_rewritten.endswith(".bias")
+                    and name_rewritten not in params_dict
+                ):
+                    continue
+                if name_rewritten not in params_dict:
+                    continue
+                param = params_dict[name_rewritten]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight, shard_id)
+                loaded_params.add(name_rewritten)
+                stacked_matched = True
+                break
+
+            if stacked_matched:
+                continue
+
+            if name.endswith(".bias") and name not in params_dict:
+                continue
+            if name not in params_dict:
+                continue
+
+            param = params_dict[name]
+            # attention_sink_bias is head-parallel; slice by tp
+            if "attention_sink_bias" in name:
+                total_heads = loaded_weight.shape[0]
+                heads_per_rank = total_heads // tp_size
+                loaded_weight = loaded_weight.narrow(
+                    0, tp_rank * heads_per_rank, heads_per_rank
+                )
+
+            weight_loader = getattr(param, "weight_loader", default_weight_loader)
+            weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+
+        return loaded_params
+
+
+class MiMoV2OmniMTP(MiMoV2MTP, SupportsMultiModal):
+    def embed_input_ids(
+        self,
+        input_ids: torch.Tensor,
+        multimodal_embeddings: MultiModalEmbeddings | None = None,
+        *,
+        is_multimodal: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        inputs_embeds = self._embed_text_input_ids(
+            input_ids,
+            self.model.embed_input_ids,
+            is_multimodal=is_multimodal,
+        )
+
+        if multimodal_embeddings is None or len(multimodal_embeddings) == 0:
+            return inputs_embeds
+
+        is_multimodal = _require_is_multimodal(is_multimodal)
+
+        inputs_embeds = _merge_multimodal_embeddings(
+            inputs_embeds=inputs_embeds,
+            multimodal_embeddings=multimodal_embeddings,
+            is_multimodal=is_multimodal,
+        )
+
+        return inputs_embeds

vllm/model_executor/models/minimax_m2.py

@@ -35,6 +35,7 @@ from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.distributed import (
     get_pp_group,
+    get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
 )
 from vllm.model_executor.layers.attention import Attention
@@ -220,9 +221,21 @@ class MiniMaxM2Attention(nn.Module):
         self.q_norm = MiniMaxText01RMSNormTP(
             self.head_dim * self.total_num_heads, eps=rms_norm_eps
         )
-        self.k_norm = MiniMaxText01RMSNormTP(
-            self.head_dim * self.total_num_kv_heads, eps=rms_norm_eps
-        )
+        if self.total_num_kv_heads >= tp_size:
+            self.k_norm = MiniMaxText01RMSNormTP(
+                self.head_dim * self.total_num_kv_heads, eps=rms_norm_eps
+            )
+        else:
+            # KV heads are replicated across TP ranks; shard k_norm weight by
+            # total_num_kv_heads rather than tp_size to avoid incorrect sharding.
+            num_kv_head_replicas = tp_size // self.total_num_kv_heads
+            self.k_norm = MiniMaxText01RMSNormTP(
+                self.head_dim * self.total_num_kv_heads,
+                eps=rms_norm_eps,
+                weight_shard_world_size=self.total_num_kv_heads,
+                weight_shard_rank=get_tensor_model_parallel_rank()
+                // num_kv_head_replicas,
+            )
 
     def forward(
         self,

vllm/model_executor/models/registry.py

@@ -171,7 +171,8 @@ _TEXT_GENERATION_MODELS = {
     "MptForCausalLM": ("mpt", "MPTForCausalLM"),
     "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
     "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
-    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
+    "MiMoV2FlashForCausalLM": ("mimo_v2", "MiMoV2FlashForCausalLM"),
+    "MiMoV2ProForCausalLM": ("mimo_v2", "MiMoV2ProForCausalLM"),
     "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
     "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
     "NemotronHPuzzleForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
@@ -468,6 +469,7 @@ _MULTIMODAL_MODELS = {
     ),
     "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),
     "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
+    "MiMoV2OmniForCausalLM": ("mimo_v2_omni", "MiMoV2OmniForCausalLM"),
     "MiniMaxVL01ForConditionalGeneration": (
         "minimax_vl_01",
         "MiniMaxVL01ForConditionalGeneration",
@@ -570,6 +572,8 @@ _MULTIMODAL_MODELS = {
 _SPECULATIVE_DECODING_MODELS = {
     "ExtractHiddenStatesModel": ("extract_hidden_states", "ExtractHiddenStatesModel"),
     "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
+    "MiMoV2MTPModel": ("mimo_v2_mtp", "MiMoV2MTP"),
+    "MiMoV2OmniMTPModel": ("mimo_v2_mtp", "MiMoV2OmniMTP"),
     "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
     "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
     "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),

vllm/multimodal/registry.py

@@ -207,7 +207,8 @@ class MultiModalRegistry:
         Create a multi-modal processor for a specific model and tokenizer.
         """
         if not model_config.is_multimodal_model:
-            raise ValueError(f"{model_config.model} is not a multimodal model")
+            model_name = model_config.served_model_name or model_config.model
+            raise ValueError(f"{model_name} is not a multimodal model")
 
         model_cls = self._get_model_cls(model_config)
         factories = model_cls._processor_factory

vllm/transformers_utils/config.py

@@ -66,6 +66,13 @@ MISTRAL_CONFIG_NAME = "params.json"
 
 logger = init_logger(__name__)
 
+if Version(version("transformers")) < Version("5.0.0"):
+    logger.warning(
+        "Support for Transformers v4 is deprecated. The Transformers v4 codepath will "
+        "become unmaintained in vLLM v0.22.0 and will be removed in vLLM v0.24.0. "
+        "Please upgrade to Transformers v5: pip install --upgrade transformers"
+    )
+
 
 class LazyConfigDict(dict):
     def __getitem__(self, key):

vllm/transformers_utils/configs/mimo_v2_omni.py

@@ -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

vllm/transformers_utils/model_arch_config_convertor.py

@@ -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,

vllm/transformers_utils/processors/__init__.py

@@ -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",

vllm/v1/core/kv_cache_utils.py

@@ -1381,7 +1381,9 @@ def unify_hybrid_kv_cache_specs(kv_cache_spec: dict[str, KVCacheSpec]):
                     block_size=spec.block_size,
                     num_kv_heads=spec.num_kv_heads,
                     head_size=spec.head_size,
+                    head_size_v=spec.head_size_v,
                     dtype=spec.dtype,
+                    kv_quant_mode=spec.kv_quant_mode,
                     sliding_window=spec.sliding_window,
                     page_size_padded=spec.page_size_padded,
                 )

vllm/v1/spec_decode/llm_base_proposer.py

@@ -1344,6 +1344,7 @@ class SpecDecodeBaseProposer:
                 "Exaone4_5_ForConditionalGeneration",
                 "GlmOcrForConditionalGeneration",
                 "HunYuanVLForConditionalGeneration",
+                "MiMoV2OmniForCausalLM",
                 "Qwen2_5_VLForConditionalGeneration",
                 "Qwen3_5ForConditionalGeneration",
                 "Qwen3_5MoeForConditionalGeneration",