9efdddca28 · tduncan/Nvidia-Optimization-Stack-VLLM

vllm/model_executor/models/arcee.py

+6 -2

@@ -45,6 +45,7 @@ from .utils import (

45 is_pp_missing_parameter,

46 make_empty_intermediate_tensors_factory,

47 make_layers,

48 + maybe_prefix,

49 )

50

51

@@ -367,7 +368,10 @@ class ArceeForCausalLM(

368 self.config = config

369

370 # Initialize the inner Transformer model (ArceeModel)

370 - self.model = ArceeModel(vllm_config=vllm_config, prefix=f"{prefix}.model")

371 + self.model = ArceeModel(

372 + vllm_config=vllm_config,

373 + prefix=maybe_prefix(prefix, "model"),

374 + )

375 # On the last pipeline stage, set up the LM head and logits processor

376 if get_pp_group().is_last_rank:

377 # Determine vocabulary size (including any LoRA extra tokens

@@ -378,7 +382,7 @@ class ArceeForCausalLM(

382 config.hidden_size,

383 quant_config=vllm_config.quant_config,

384 bias=getattr(config, "lm_head_bias", False),

381 - prefix=f"{prefix}.lm_head",

385 + prefix=maybe_prefix(prefix, "lm_head"),

386 )

387 if config.tie_word_embeddings:

388 # Tie output weights with input embedding matrix

vllm/model_executor/models/aria.py

+1 -1

@@ -540,7 +540,7 @@ class AriaForConditionalGeneration(nn.Module, SupportsMultiModal):

540 self.vision_tower = AriaVisionTransformer(

541 config.vision_config,

542 quant_config=quant_config,

543 - prefix=f"{prefix}.vision_tower",

543 + prefix=maybe_prefix(prefix, "vision_tower"),

544 )

545 self.multi_modal_projector = AriaProjector(

546 config, prefix=maybe_prefix(prefix, "multi_modal_projector")

vllm/model_executor/models/blip2.py

+1 -1

@@ -561,7 +561,7 @@ class Blip2ForConditionalGeneration(

561 config.qformer_config,

562 cache_config=cache_config,

563 quant_config=quant_config,

564 - prefix=f"{prefix}.qformer",

564 + prefix=maybe_prefix(prefix, "qformer"),

565 )

566 self.language_projection = nn.Linear(

567 config.qformer_config.hidden_size,

vllm/model_executor/models/clip.py

+1 -1

@@ -777,7 +777,7 @@ class CLIPVisionModel(nn.Module):

777 quant_config=quant_config,

778 num_hidden_layers_override=num_hidden_layers_override,

779 require_post_norm=require_post_norm,

780 - prefix=f"{prefix}.vision_model",

780 + prefix=maybe_prefix(prefix, "vision_model"),

781 )

782

783 def forward(

vllm/model_executor/models/cohere_asr.py

+3 -2

@@ -64,7 +64,7 @@ from .interfaces import (

64 SupportsMultiModal,

65 SupportsTranscription,

66 )

67 -from .utils import AutoWeightsLoader, WeightsMapper, make_layers

67 +from .utils import AutoWeightsLoader, WeightsMapper, make_layers, maybe_prefix

68

69 logger = init_logger(__name__)

70

@@ -1717,7 +1717,8 @@ class CohereASRModel(nn.Module):

1717 self.encoder = ConformerEncoder(vllm_config=vllm_config)

1718

1719 self.decoder = CohereASRDecoder(

1720 - vllm_config=vllm_config, prefix=f"{prefix}.decoder"

1720 + vllm_config=vllm_config,

1721 + prefix=maybe_prefix(prefix, "decoder"),

1722 )

1723

1724 if self.encoder.d_model != self.decoder.hidden_size:

vllm/model_executor/models/deepseek_eagle.py

+3 -1

@@ -198,7 +198,9 @@ class EagleDeepseekV3ForCausalLM(DeepseekV3ForCausalLM):

198 vllm_config.parallel_config

199 )

200 self.model = DeepseekV2Model(

201 - vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num

201 + vllm_config=vllm_config,

202 + prefix=maybe_prefix(prefix, "model"),

203 + start_layer_id=target_layer_num,

204 )

205

206 self.lm_head = ParallelLMHead(

vllm/model_executor/models/deepseek_eagle3.py

+3 -1

@@ -318,7 +318,9 @@ class Eagle3DeepseekV2ForCausalLM(DeepseekV2ForCausalLM):

318 self.config.target_layer_count = target_layer_num

319

320 self.model = DeepseekV2Eagle3Model(

321 - vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num

321 + vllm_config=vllm_config,

322 + prefix=maybe_prefix(prefix, "model"),

323 + start_layer_id=target_layer_num,

324 )

325

326 logit_scale = getattr(self.config, "logit_scale", 1.0)

vllm/model_executor/models/granite4_vision.py

+1 -1

@@ -149,7 +149,7 @@ class WindowQFormerDownsampler(nn.Module):

149 qformer_config,

150 quant_config=quant_config,

151 cache_config=cache_config,

152 - prefix=f"{prefix}.qformer",

152 + prefix=maybe_prefix(prefix, "qformer"),

153 )

154

155 self.image_side = (

vllm/model_executor/models/granite_speech.py

+2 -2

@@ -620,7 +620,7 @@ class GraniteSpeechForConditionalGeneration(

620 self.encoder = GraniteSpeechCTCEncoder(

621 config=config.encoder_config,

622 quant_config=quant_config,

623 - prefix=f"{prefix}.encoder",

623 + prefix=maybe_prefix(prefix, "encoder"),

624 )

625

626 # Blip2 QFormer

@@ -628,7 +628,7 @@ class GraniteSpeechForConditionalGeneration(

628 config=config,

629 quant_config=quant_config,

630 cache_config=cache_config,

631 - prefix=f"{prefix}.projector",

631 + prefix=maybe_prefix(prefix, "projector"),

632 )

633

634 self.make_empty_intermediate_tensors = (

vllm/model_executor/models/hunyuan_v1.py

+4 -1

@@ -930,7 +930,10 @@ class HunyuanV1ModelBase(

930 self.config = config

931 self.quant_config = quant_config

932

933 - self.model = HunYuanModel(vllm_config=vllm_config, prefix="model")

933 + self.model = HunYuanModel(

934 + vllm_config=vllm_config,

935 + prefix=maybe_prefix(prefix, "model"),

936 + )

937 if get_pp_group().is_last_rank:

938 self.lm_head = ParallelLMHead(

939 config.vocab_size,

vllm/model_executor/models/keye.py

+1 -1

@@ -729,7 +729,7 @@ class KeyeSiglipVisionModel(nn.Module):

729 self.vision_model = KeyeSiglipVisionTransformer(

730 config,

731 quant_config=quant_config,

732 - prefix=f"{prefix}.vision_model",

732 + prefix=maybe_prefix(prefix, "vision_model"),

733 )

734 self.quant_config = quant_config

735

vllm/model_executor/models/lfm2_siglip2.py

+2 -1

@@ -25,6 +25,7 @@ from vllm.model_executor.layers.linear import (

25 from vllm.model_executor.layers.quantization import QuantizationConfig

26 from vllm.model_executor.model_loader.weight_utils import default_weight_loader

27

28 +from .utils import maybe_prefix

29 from .vision import (

30 is_vit_use_data_parallel,

31 resolve_visual_encoder_outputs,

@@ -472,7 +473,7 @@ class Siglip2Model(torch.nn.Module):

473 quant_config=quant_config,

474 num_hidden_layers_override=num_hidden_layers_override,

475 require_post_norm=require_post_norm,

475 - prefix=f"{prefix}.vision_model",

476 + prefix=maybe_prefix(prefix, "vision_model"),

477 )

478

479 def forward(

vllm/model_executor/models/llama4_eagle.py

+1 -1

@@ -176,7 +176,7 @@ class EagleLlama4ForCausalLM(Llama4ForCausalLM):

176 )

177 self.model = LlamaModel(

178 vllm_config=vllm_config,

179 - prefix="model",

179 + prefix=maybe_prefix(prefix, "model"),

180 start_layer_id=target_layer_num,

181 quant_config=quant_config,

182 )

vllm/model_executor/models/llama_eagle.py

+3 -1

@@ -174,7 +174,9 @@ class EagleLlamaForCausalLM(LlamaForCausalLM):

174 vllm_config.parallel_config

175 )

176 self.model = LlamaModel(

177 - vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num

177 + vllm_config=vllm_config,

178 + prefix=maybe_prefix(prefix, "model"),

179 + start_layer_id=target_layer_num,

180 )

181

182 logit_scale = getattr(self.config, "logit_scale", 1.0)

vllm/model_executor/models/llama_eagle3.py

+3 -1

@@ -287,7 +287,9 @@ class Eagle3LlamaForCausalLM(LlamaForCausalLM):

287 # proper layer_types indexing in draft models

288 self.config.target_layer_count = target_layer_num

289 self.model = LlamaModel(

290 - vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num

290 + vllm_config=vllm_config,

291 + prefix=maybe_prefix(prefix, "model"),

292 + start_layer_id=target_layer_num,

293 )

294

295 logit_scale = getattr(self.config, "logit_scale", 1.0)

vllm/model_executor/models/mistral_eagle.py

+3 -1

@@ -128,7 +128,9 @@ class EagleMistralForCausalLM(MistralForCausalLM):

128 vllm_config.parallel_config

129 )

130 self.model = EagleMistralModel(

131 - vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num

131 + vllm_config=vllm_config,

132 + prefix=maybe_prefix(prefix, "model"),

133 + start_layer_id=target_layer_num,

134 )

135

136 logit_scale = getattr(self.config, "logit_scale", 1.0)

vllm/model_executor/models/molmo2.py

+1 -1

@@ -489,7 +489,7 @@ class Molmo2VisionTransformer(nn.Module):

489 self.transformer = Molmo2VisionBlockCollection(

490 config,

491 quant_config,

492 - prefix=f"{prefix}.transformer",

492 + prefix=maybe_prefix(prefix, "transformer"),

493 )

494

495 def add_pos_emb(self, x: torch.Tensor, patch_num: int) -> torch.Tensor:

vllm/model_executor/models/openpangu_vl.py

+1 -1

@@ -857,7 +857,7 @@ class OpenPanguVLForConditionalGeneration(

857 with self._mark_language_model(vllm_config):

858 self.language_model = init_vllm_registered_model(

859 vllm_config=vllm_config,

860 - prefix=maybe_prefix("openpangu", "language_model"),

860 + prefix=maybe_prefix(prefix, "openpangu.language_model"),

861 architectures=["PanguEmbeddedForCausalLM"],

862 )

863

vllm/model_executor/models/ovis.py

+2 -2

@@ -96,7 +96,7 @@ class VisualTokenizer(torch.nn.Module):

96 self.backbone = self._init_backbone(

97 config=config,

98 quant_config=quant_config,

99 - prefix=f"{prefix}.backbone",

99 + prefix=maybe_prefix(prefix, "backbone"),

100 )

101 # reserved tokens for IMAGE_INDICATORS

102 head_dim = config.vocab_size - len(IMAGE_INDICATOR_IDS)

@@ -442,7 +442,7 @@ class Ovis(nn.Module, SupportsMultiModal, SupportsPP):

442 self.visual_tokenizer = VisualTokenizer(

443 config=config.visual_tokenizer_config,

444 quant_config=quant_config,

445 - prefix=f"{prefix}.visual_tokenizer",

445 + prefix=maybe_prefix(prefix, "visual_tokenizer"),

446 )

447 self.vte = VisualEmbedding(

448 self.config.visual_tokenizer_config.vocab_size, self.config.hidden_size

vllm/model_executor/models/ovis2_5.py

+1 -1

@@ -465,7 +465,7 @@ class Ovis2_5(nn.Module, SupportsMultiModal, SupportsPP):

465 config=config.vit_config,

466 visual_vocab_size=config.visual_vocab_size,

467 quant_config=quant_config,

468 - prefix=f"{prefix}.visual_tokenizer",

468 + prefix=maybe_prefix(prefix, "visual_tokenizer"),

469 )

470 self.vte = VisualEmbedding(config.visual_vocab_size, config.hidden_size)

471

vllm/model_executor/models/paddleocr_vl.py

+1 -1

@@ -895,7 +895,7 @@ class SiglipVisionModel(nn.Module):

895 self.vision_model = SiglipVisionTransformer(

896 config,

897 quant_config=quant_config,

898 - prefix=f"{prefix}.vision_model",

898 + prefix=maybe_prefix(prefix, "vision_model"),

899 )

900 self.quant_config = quant_config

901

vllm/model_executor/models/phi4mm.py

+1 -1

@@ -1034,7 +1034,7 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):

1034 self.vision_encoder = Phi4MMImageEncoder(

1035 config,

1036 quant_config,

1037 - prefix="model.vision_embed_tokens",

1037 + prefix=maybe_prefix(prefix, "model.vision_embed_tokens"),

1038 model_dir=config._name_or_path,

1039 )

1040

vllm/model_executor/models/qwen3_dflash.py

+1 -1

@@ -510,7 +510,7 @@ class DFlashQwen3ForCausalLM(Qwen3ForCausalLM):

510 self.config.target_layer_count = target_layer_num

511 self.model = DFlashQwen3Model(

512 vllm_config=vllm_config,

513 - prefix="model",

513 + prefix=maybe_prefix(prefix, "model"),

514 start_layer_id=target_layer_num,

515 )

516

vllm/model_executor/models/siglip.py

+1 -1

@@ -867,7 +867,7 @@ class SiglipVisionModel(nn.Module):

867 quant_config=quant_config,

868 num_hidden_layers_override=num_hidden_layers_override,

869 require_post_norm=require_post_norm,

870 - prefix=f"{prefix}.vision_model",

870 + prefix=maybe_prefix(prefix, "vision_model"),

871 use_head=use_head,

872 )

873

vllm/model_executor/models/siglip2navit.py

+2 -1

@@ -29,6 +29,7 @@ from vllm.model_executor.layers.rotary_embedding.common import (

29 from vllm.model_executor.model_loader.weight_utils import default_weight_loader

30 from vllm.platforms import current_platform

31

32 +from .utils import maybe_prefix

33 from .vision import is_vit_use_data_parallel

34

35

@@ -598,7 +599,7 @@ class Siglip2NavitModel(torch.nn.Module):

599 self.vision_model = Siglip2VisionTransformer(

600 config,

601 quant_config=quant_config,

601 - prefix=f"{prefix}.vision_model",

602 + prefix=maybe_prefix(prefix, "vision_model"),

603 )

604

605 def forward(

[Model] Fix missing `maybe_prefix` (#42280)