Skip to content

vllm.model_executor.models.nemotron_nas

Inference-only deci model compatible with HuggingFace weights.

DeciLMAttention

Bases: LlamaAttention

Source code in vllm/model_executor/models/nemotron_nas.py 
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
class DeciLMAttention(LlamaAttention):

    def __init__(
        self,
        config: LlamaConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
        bias_o_proj: bool = False,
        cache_config: Optional[CacheConfig] = None,
        prefix: str = "",
        attn_type: str = AttentionType.DECODER,
    ) -> None:
        super().__init__(config, hidden_size, num_heads, num_kv_heads,
                         rope_theta, rope_scaling, max_position_embeddings,
                         quant_config, bias, bias_o_proj, cache_config, prefix,
                         attn_type)

    def _init_rotary_emb(self, config, rope_scaling: Optional[dict[str, Any]],
                         quant_config: Optional[QuantizationConfig]) -> None:
        # Enables YARN for Mistral and LLaMA4 derivatives.
        is_neox_style = True
        if hasattr(config, "position_embedding_type"):
            is_neox_style = config.position_embedding_type not in [
                "mistral_yarn", "rope_llama4"
            ]

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=self.max_position_embeddings,
            base=self.rope_theta,
            rope_scaling=rope_scaling,
            is_neox_style=is_neox_style,
            partial_rotary_factor=self.partial_rotary_factor)

__init__ 

__init__(
    config: LlamaConfig,
    hidden_size: int,
    num_heads: int,
    num_kv_heads: int,
    rope_theta: float = 10000,
    rope_scaling: Optional[dict[str, Any]] = None,
    max_position_embeddings: int = 8192,
    quant_config: Optional[QuantizationConfig] = None,
    bias: bool = False,
    bias_o_proj: bool = False,
    cache_config: Optional[CacheConfig] = None,
    prefix: str = "",
    attn_type: str = DECODER,
) -> None
Source code in vllm/model_executor/models/nemotron_nas.py 
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
def __init__(
    self,
    config: LlamaConfig,
    hidden_size: int,
    num_heads: int,
    num_kv_heads: int,
    rope_theta: float = 10000,
    rope_scaling: Optional[dict[str, Any]] = None,
    max_position_embeddings: int = 8192,
    quant_config: Optional[QuantizationConfig] = None,
    bias: bool = False,
    bias_o_proj: bool = False,
    cache_config: Optional[CacheConfig] = None,
    prefix: str = "",
    attn_type: str = AttentionType.DECODER,
) -> None:
    super().__init__(config, hidden_size, num_heads, num_kv_heads,
                     rope_theta, rope_scaling, max_position_embeddings,
                     quant_config, bias, bias_o_proj, cache_config, prefix,
                     attn_type)

_init_rotary_emb 

_init_rotary_emb(
    config,
    rope_scaling: Optional[dict[str, Any]],
    quant_config: Optional[QuantizationConfig],
) -> None
Source code in vllm/model_executor/models/nemotron_nas.py 
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
def _init_rotary_emb(self, config, rope_scaling: Optional[dict[str, Any]],
                     quant_config: Optional[QuantizationConfig]) -> None:
    # Enables YARN for Mistral and LLaMA4 derivatives.
    is_neox_style = True
    if hasattr(config, "position_embedding_type"):
        is_neox_style = config.position_embedding_type not in [
            "mistral_yarn", "rope_llama4"
        ]

    self.rotary_emb = get_rope(
        self.head_dim,
        rotary_dim=self.head_dim,
        max_position=self.max_position_embeddings,
        base=self.rope_theta,
        rope_scaling=rope_scaling,
        is_neox_style=is_neox_style,
        partial_rotary_factor=self.partial_rotary_factor)

DeciLMDecoderLayer

Bases: Module

Source code in vllm/model_executor/models/nemotron_nas.py 
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
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 = DeciLMAttention(
                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

_is_no_op_attention instance-attribute 

_is_no_op_attention = no_op

_is_no_op_ffn instance-attribute 

_is_no_op_ffn = no_op

hidden_size instance-attribute 

hidden_size = hidden_size

input_layernorm instance-attribute 

input_layernorm = RMSNorm(hidden_size, eps=rms_norm_eps)

mlp instance-attribute 

mlp = LlamaMLP(
    hidden_size=hidden_size,
    intermediate_size=intermediate_size,
    hidden_act=hidden_act,
    quant_config=quant_config,
    bias=getattr(config, "mlp_bias", False),
    prefix=f"{prefix}.mlp",
)

post_attention_layernorm instance-attribute 

post_attention_layernorm = RMSNorm(
    hidden_size, eps=rms_norm_eps
)

self_attn instance-attribute 

self_attn = DeciLMAttention(
    config=config,
    hidden_size=hidden_size,
    num_heads=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",
)

__init__ 

__init__(
    config: LlamaConfig,
    layer_idx: int,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/nemotron_nas.py 
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
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 = DeciLMAttention(
            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)

forward 

forward(
    positions: Tensor,
    hidden_states: Tensor,
    residual: Optional[Tensor],
) -> tuple[Tensor, Tensor]
Source code in vllm/model_executor/models/nemotron_nas.py 
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
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

DeciLMForCausalLM

Bases: Module, SupportsLoRA, SupportsPP, HasNoOps

Source code in vllm/model_executor/models/nemotron_nas.py 
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
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
    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)

config instance-attribute 

config = config

embedding_modules class-attribute instance-attribute 

embedding_modules = {
    "embed_tokens": "input_embeddings",
    "lm_head": "output_embeddings",
}

embedding_padding_modules class-attribute instance-attribute 

embedding_padding_modules = ['lm_head']

lm_head instance-attribute 

lm_head = ParallelLMHead(
    unpadded_vocab_size,
    hidden_size,
    org_num_embeddings=vocab_size,
    padding_size=DEFAULT_VOCAB_PADDING_SIZE
    if not lora_config
    else lora_vocab_padding_size,
    quant_config=quant_config,
    prefix=maybe_prefix(prefix, "lm_head"),
)

logits_processor instance-attribute 

logits_processor = LogitsProcessor(
    unpadded_vocab_size, vocab_size, logit_scale
)

lora_config instance-attribute 

lora_config = lora_config

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

mistral_mapping class-attribute instance-attribute 

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",
}

model instance-attribute 

model = _init_model(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "model"),
)

packed_modules_mapping class-attribute instance-attribute 

packed_modules_mapping = {
    "qkv_proj": ["q_proj", "k_proj", "v_proj"],
    "gate_up_proj": ["gate_proj", "up_proj"],
}

unpadded_vocab_size instance-attribute 

unpadded_vocab_size = vocab_size

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/nemotron_nas.py 
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
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)

_init_model 

_init_model(vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/nemotron_nas.py 
451
452
def _init_model(self, vllm_config: VllmConfig, prefix: str = ""):
    return DeciModel(vllm_config=vllm_config, prefix=prefix)

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]
Source code in vllm/model_executor/models/nemotron_nas.py 
468
469
470
471
472
473
474
475
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

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/nemotron_nas.py 
457
458
459
460
461
462
463
464
465
466
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

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/nemotron_nas.py 
454
455
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.model.get_input_embeddings(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/nemotron_nas.py 
477
478
479
480
481
482
483
484
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)

DeciModel

Bases: Module

Source code in vllm/model_executor/models/nemotron_nas.py 
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
@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

config instance-attribute 

config = config

embed_tokens instance-attribute 

embed_tokens = VocabParallelEmbedding(
    vocab_size,
    hidden_size,
    org_num_embeddings=vocab_size,
    quant_config=quant_config,
)

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors_factory(
        ["hidden_states", "residual"], hidden_size
    )
)

norm instance-attribute 

norm = RMSNorm(hidden_size, eps=rms_norm_eps)

org_vocab_size instance-attribute 

org_vocab_size = vocab_size

padding_idx instance-attribute 

padding_idx = pad_token_id

quant_config instance-attribute 

quant_config = quant_config

vocab_size instance-attribute 

vocab_size = vocab_size + lora_vocab

__init__ 

__init__(
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    layer_type: type[
        DeciLMDecoderLayer
    ] = DeciLMDecoderLayer,
)
Source code in vllm/model_executor/models/nemotron_nas.py 
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
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))

forward 

forward(
    input_ids: Optional[Tensor],
    positions: Tensor,
    intermediate_tensors: Optional[IntermediateTensors],
    inputs_embeds: Optional[Tensor] = None,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/nemotron_nas.py 
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
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

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/nemotron_nas.py 
268
269
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.embed_tokens(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/nemotron_nas.py 
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
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

_ffn_mult_to_intermediate_size 

_ffn_mult_to_intermediate_size(
    ffn_mult: float, n_embd: int
) -> int
Source code in vllm/model_executor/models/nemotron_nas.py 
55
56
57
58
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)

_find_multiple 

_find_multiple(n: int, k: int) -> int
Source code in vllm/model_executor/models/nemotron_nas.py 
61
62
63
64
65
def _find_multiple(n: int, k: int) -> int:
    # DeciLM-specific code
    if n % k == 0:
        return n
    return n + k - (n % k)