vllm-cpu 0.8.5.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/model_executor/models/nemotron_nas.py

@@ -0,0 +1,446 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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.
+"""Inference-only deci model compatible with HuggingFace weights."""
+from typing import Iterable, Optional, Set, Tuple, Type, Union
+
+import torch
+from torch import nn
+from transformers import LlamaConfig
+
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group
+from vllm.model_executor.layers.layernorm import RMSNorm
+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 (
+    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
+from vllm.model_executor.model_loader.weight_utils import (
+    default_weight_loader, maybe_remap_kv_scale_name)
+from vllm.model_executor.models.llama import LlamaAttention, LlamaMLP
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from .interfaces import HasNoOps, SupportsLoRA, SupportsPP
+from .utils import (AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
+
+
+def _ffn_mult_to_intermediate_size(ffn_mult: float, n_embd: int) -> int:
+    # DeciLM-specific code
+    intermediate_size = int(2 * ffn_mult * n_embd / 3)
+    return _find_multiple(intermediate_size, 256)
+
+
+def _find_multiple(n: int, k: int) -> int:
+    # DeciLM-specific code
+    if n % k == 0:
+        return n
+    return n + k - (n % k)
+
+
+class DeciLMDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: LlamaConfig,
+        layer_idx: int,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        block_config = config.block_configs[layer_idx]
+        self._is_no_op_attention = block_config.attention.no_op
+        self._is_no_op_ffn = block_config.ffn.no_op
+
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None and getattr(
+                config, "original_max_position_embeddings", None):
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        # Support abacusai/Smaug-72B-v0.1 with attention_bias
+        # Support internlm/internlm-7b with bias
+        attention_bias = getattr(config, "attention_bias", False) or getattr(
+            config, "bias", False)
+        bias_o_proj = attention_bias
+        # support internlm/internlm3-8b with qkv_bias
+        if hasattr(config, "qkv_bias"):
+            attention_bias = config.qkv_bias
+
+        if not self._is_no_op_attention:
+            num_kv_heads = (config.num_attention_heads //
+                            block_config.attention.n_heads_in_group)
+            self.self_attn = LlamaAttention(
+                config=config,
+                hidden_size=self.hidden_size,
+                num_heads=config.num_attention_heads,
+                num_kv_heads=num_kv_heads,
+                rope_theta=rope_theta,
+                rope_scaling=rope_scaling,
+                max_position_embeddings=max_position_embeddings,
+                quant_config=quant_config,
+                bias=attention_bias,
+                bias_o_proj=bias_o_proj,
+                cache_config=cache_config,
+                prefix=f"{prefix}.self_attn",
+            )
+            self.input_layernorm = RMSNorm(config.hidden_size,
+                                           eps=config.rms_norm_eps)
+
+        if not self._is_no_op_ffn:
+            ffn_mult = block_config.ffn.ffn_mult
+            intermediate_size = _ffn_mult_to_intermediate_size(
+                ffn_mult, config.hidden_size)
+
+            self.mlp = LlamaMLP(
+                hidden_size=self.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                bias=getattr(config, "mlp_bias", False),
+                prefix=f"{prefix}.mlp",
+            )
+            self.post_attention_layernorm = RMSNorm(config.hidden_size,
+                                                    eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+
+        if self._is_no_op_attention:
+            pass
+        else:
+            if (residual is None):
+                residual = hidden_states
+                hidden_states = self.input_layernorm(hidden_states)
+            else:
+                hidden_states, residual = self.input_layernorm(
+                    hidden_states, residual)
+            hidden_states = self.self_attn(
+                positions=positions,
+                hidden_states=hidden_states,
+            )
+
+        # Fully Connected
+        if not self._is_no_op_ffn:
+            hidden_states, residual = self.post_attention_layernorm(
+                hidden_states, residual)
+            hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+@support_torch_compile
+class DeciModel(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        vllm_config: VllmConfig,
+        prefix: str = "",
+        layer_type: Type[DeciLMDecoderLayer] = DeciLMDecoderLayer,
+    ):
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+
+        self.config = config
+        self.quant_config = quant_config
+        self.padding_idx = config.pad_token_id
+        lora_vocab = ((lora_config.lora_extra_vocab_size *
+                       (lora_config.max_loras or 1)) if lora_config else 0)
+        self.vocab_size = config.vocab_size + lora_vocab
+        self.org_vocab_size = config.vocab_size
+        if get_pp_group().is_first_rank or (config.tie_word_embeddings
+                                            and get_pp_group().is_last_rank):
+            self.embed_tokens = VocabParallelEmbedding(
+                self.vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                quant_config=quant_config,
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        def get_layer(prefix: str):
+            layer_idx = int(prefix.rsplit(".", 1)[1])
+            return layer_type(
+                config,
+                layer_idx,
+                cache_config,
+                quant_config=quant_config,
+                prefix=prefix,
+            )
+
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            get_layer,
+            prefix=f"{prefix}.layers",
+        )
+        if get_pp_group().is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer()
+
+        self.make_empty_intermediate_tensors = (
+            make_empty_intermediate_tensors_factory(
+                ["hidden_states", "residual"], config.hidden_size))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor],
+        positions: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+
+        kv_cache_index = 0
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            if not layer._is_no_op_attention:
+                hidden_states, residual = layer(positions, hidden_states,
+                                                residual)
+                kv_cache_index += 1
+            else:
+                hidden_states, residual = layer(positions, hidden_states,
+                                                residual)
+
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            (".qkv_proj", ".q_proj", "q"),
+            (".qkv_proj", ".k_proj", "k"),
+            (".qkv_proj", ".v_proj", "v"),
+            (".gate_up_proj", ".gate_proj", 0),
+            (".gate_up_proj", ".up_proj", 1),
+        ]
+        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
+            if ("rotary_emb.cos_cached" in name
+                    or "rotary_emb.sin_cached" in name):
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                continue
+            if self.quant_config is not None and (
+                    scale_name := self.quant_config.get_cache_scale(name)):
+                # Loading kv cache quantization scales
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
+                                 loaded_weight[0])
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_name)
+                continue
+            if "scale" in name:
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+
+                if is_pp_missing_parameter(name, self):
+                    continue
+
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+
+                if is_pp_missing_parameter(name, self):
+                    continue
+
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
+
+class DeciLMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, HasNoOps):
+    packed_modules_mapping = {
+        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+        "gate_up_proj": ["gate_proj", "up_proj"],
+    }
+
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "qkv_proj",
+        "o_proj",
+        "gate_up_proj",
+        "down_proj",
+        "embed_tokens",
+        "lm_head",
+    ]
+    embedding_modules = {
+        "embed_tokens": "input_embeddings",
+        "lm_head": "output_embeddings",
+    }
+    embedding_padding_modules = ["lm_head"]
+
+    # Mistral/Llama models can also be loaded with --load-format mistral
+    # from consolidated.safetensors checkpoints
+    mistral_mapping = {
+        "layers": "model.layers",
+        "attention": "self_attn",
+        "wq": "q_proj",
+        "wk": "k_proj",
+        "wv": "v_proj",
+        "wo": "o_proj",
+        "attention_norm": "input_layernorm",
+        "feed_forward": "mlp",
+        "w1": "gate_proj",
+        "w2": "down_proj",
+        "w3": "up_proj",
+        "ffn_norm": "post_attention_layernorm",
+        "tok_embeddings": "model.embed_tokens",
+        "output": "lm_head",
+        "norm": "model.norm",
+    }
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+        self.config = config
+        self.lora_config = lora_config
+
+        self.model = self._init_model(vllm_config=vllm_config,
+                                      prefix=maybe_prefix(prefix, "model"))
+
+        if get_pp_group().is_last_rank:
+            self.unpadded_vocab_size = config.vocab_size
+            if lora_config:
+                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
+            self.lm_head = ParallelLMHead(
+                self.unpadded_vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                padding_size=(
+                    DEFAULT_VOCAB_PADDING_SIZE
+                    # We need bigger padding if using lora for kernel
+                    # compatibility
+                    if not lora_config else
+                    lora_config.lora_vocab_padding_size),
+                quant_config=quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
+            )
+            if config.tie_word_embeddings:
+                self.lm_head = self.lm_head.tie_weights(
+                    self.model.embed_tokens)
+
+            logit_scale = getattr(config, "logit_scale", 1.0)
+            self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                    config.vocab_size,
+                                                    logit_scale)
+        else:
+            self.lm_head = PPMissingLayer()
+
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors)
+
+    def _init_model(self, vllm_config: VllmConfig, prefix: str = ""):
+        return DeciModel(vllm_config=vllm_config, prefix=prefix)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        model_output = self.model(input_ids, positions, intermediate_tensors,
+                                  inputs_embeds)
+        return model_output
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        logits = self.logits_processor(self.lm_head, hidden_states,
+                                       sampling_metadata)
+        return logits
+
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=(["lm_head."]
+                           if self.config.tie_word_embeddings else None),
+        )
+        return loader.load_weights(weights)

vllm/model_executor/models/nvlm_d.py

@@ -0,0 +1,212 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# adapted from https://huggingface.co/nvidia/NVLM-D-72B/blob/main/modeling_nvlm_d.py
+# --------------------------------------------------------
+# NVLM-D
+# Copyright (c) 2024 NVIDIA
+# Licensed under Apache 2.0 License [see LICENSE for details]
+# --------------------------------------------------------
+from collections.abc import Mapping, Sequence
+from typing import Optional
+
+import torch
+import torch.nn as nn
+from transformers import PretrainedConfig
+
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.inputs import MultiModalDataDict, MultiModalKwargs
+from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
+                                   MultiModalDataItems)
+from vllm.multimodal.processing import (PromptReplacement, PromptUpdate,
+                                        PromptUpdateDetails)
+
+from .intern_vit import InternVisionModel
+from .internvl import (BaseInternVLProcessingInfo, BaseInternVLProcessor,
+                       InternVLChatModel, InternVLDummyInputsBuilder,
+                       InternVLMultiModalProcessor)
+
+IMG_PAD = "<|vision_pad|>"
+
+
+class NVLMProcessor(BaseInternVLProcessor):
+
+    @property
+    def image_token_id(self) -> int:
+        return self.tokenizer.get_vocab()[IMG_PAD]
+
+    def get_image_repl(
+        self,
+        feature_size: int,
+        num_patches: Optional[int],
+    ) -> PromptUpdateDetails[str]:
+        if num_patches is None:
+            raise NotImplementedError("Embedding inputs are not supported")
+
+        tile_pos_identifiers = [f"<tile_{i}>" for i in range(1, num_patches)]
+        if self.use_thumbnail:
+            tile_pos_identifiers += ["<tile_global_thumbnail>"]
+
+        context_size = feature_size // num_patches
+        features = "".join(identifier + IMG_PAD * context_size
+                           for identifier in tile_pos_identifiers)
+
+        # We include the start and end as well because "<Image><tile" is
+        # tokenized as ["<Image", "><", "tile"], resulting in assertion error
+        # when trying to find "<tile" as a subsequence of "<Image><tile"
+        repl = "<Image>" + features + "</Image>"
+
+        return PromptUpdateDetails.select_text(repl, IMG_PAD)
+
+
+class NVLMProcessingInfo(BaseInternVLProcessingInfo):
+
+    def get_hf_processor(
+        self,
+        *,
+        min_dynamic_patch: Optional[int] = None,
+        max_dynamic_patch: Optional[int] = None,
+        dynamic_image_size: Optional[bool] = None,
+        **kwargs: object,
+    ) -> NVLMProcessor:
+        if min_dynamic_patch is not None:
+            kwargs["min_dynamic_patch"] = min_dynamic_patch
+        if max_dynamic_patch is not None:
+            kwargs["max_dynamic_patch"] = max_dynamic_patch
+        if dynamic_image_size is not None:
+            kwargs["dynamic_image_size"] = dynamic_image_size
+
+        return self.ctx.init_processor(
+            NVLMProcessor,
+            config=self.get_hf_config(),
+            tokenizer=self.get_tokenizer(),
+            **kwargs,
+        )
+
+
+class NVLMDummyInputsBuilder(InternVLDummyInputsBuilder[NVLMProcessingInfo]):
+
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        # The newline is necessary to separate ">" of the current item
+        # and "<" of the next item
+        return "<image>\n" * num_images
+
+    def get_dummy_mm_data(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> MultiModalDataDict:
+        target_width, target_height = \
+            self.info.get_image_size_with_most_features()
+        num_images = mm_counts.get("image", 0)
+
+        return {
+            "image":
+            self._get_dummy_images(width=target_width,
+                                   height=target_height,
+                                   num_images=num_images)
+        }
+
+
+class NVLMMultiModalProcessor(InternVLMultiModalProcessor[NVLMProcessingInfo]):
+
+    def _get_prompt_updates(
+        self,
+        mm_items: MultiModalDataItems,
+        hf_processor_mm_kwargs: Mapping[str, object],
+        out_mm_kwargs: MultiModalKwargs,
+    ) -> Sequence[PromptUpdate]:
+        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
+
+        if "image_num_patches" in out_mm_kwargs:
+            image_num_patches = out_mm_kwargs["image_num_patches"]
+            assert isinstance(image_num_patches, torch.Tensor)
+            image_num_patches = image_num_patches.tolist()
+        elif "image_embeds" in out_mm_kwargs:
+            # TODO: Use image size information in dictionary embedding inputs
+            # to compute num_patches (similar to Qwen2-VL)
+            image_num_patches = [None] * len(out_mm_kwargs["image_embeds"])
+        else:
+            image_num_patches = []
+
+        def get_replacement_nvlm(item_idx: int):
+            images = mm_items.get_items(
+                "image", (ImageEmbeddingItems, ImageProcessorItems))
+
+            if isinstance(images, ImageEmbeddingItems):
+                feature_size = images.get_feature_size(item_idx)
+            else:
+                image_size = images.get_image_size(item_idx)
+                feature_size = self.info.get_num_image_tokens(
+                    image_width=image_size.width,
+                    image_height=image_size.height,
+                    processor=hf_processor,
+                )
+
+            num_patches = image_num_patches[item_idx]
+            if num_patches is not None:
+                assert isinstance(num_patches, int)
+
+            repl = hf_processor.get_image_repl(feature_size, num_patches)
+
+            return PromptUpdateDetails.select_text(repl.full + "\n", IMG_PAD)
+
+        # See note in dummy data regarding why we have the extra newline
+        return [
+            PromptReplacement(
+                modality="image",
+                target="<image>\n",
+                replacement=get_replacement_nvlm,
+            )
+        ]
+
+
+@MULTIMODAL_REGISTRY.register_processor(NVLMMultiModalProcessor,
+                                        info=NVLMProcessingInfo,
+                                        dummy_inputs=NVLMDummyInputsBuilder)
+class NVLM_D_Model(InternVLChatModel):
+
+    def _init_mlp1(self, config: PretrainedConfig) -> nn.Sequential:
+        vit_hidden_size = config.vision_config.hidden_size
+        llm_intermediate_size = config.text_config.intermediate_size
+        llm_hidden_size = config.text_config.hidden_size
+
+        return nn.Sequential(
+            nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio)**2),
+            nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio)**2,
+                      llm_intermediate_size,
+                      bias=False),
+            nn.GELU(),
+            nn.Linear(llm_intermediate_size, llm_hidden_size, bias=False),
+        )
+
+    def _init_vision_model(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig],
+        *,
+        is_mono: bool,
+        prefix: str,
+    ):
+        if not is_mono:
+            vision_feature_layer = config.select_layer
+            if vision_feature_layer < 0:
+                num_hidden_layers = config.vision_config.num_hidden_layers \
+                    + vision_feature_layer + 1
+            else:
+                num_hidden_layers = vision_feature_layer + 1
+
+            # We added additional dummy heads to the original num of heads to
+            # make the number of heads divisible by 8.
+            return InternVisionModel(
+                config.vision_config,
+                quant_config=quant_config,
+                num_hidden_layers_override=num_hidden_layers,
+                num_dummy_heads=7,
+                prefix=prefix,
+            )
+        else:
+            msg = "Monolith mode is not applicable to NVLM_D"
+            raise NotImplementedError(msg)