Skip to content

GitLab

Unverified Commit 9d1c4747 authored by Jee Jee Li's avatar Jee Jee Li Committed by GitHub
Browse files

[LoRA][1/N]Remove LoRA extra vocab (#28382) 


Signed-off-by: default avatarJee Jee Li <pandaleefree@gmail.com>
parent 8c32c6e4
Hide whitespace changes
Inline Side-by-side
Showing with 53 additions and 205 deletions
vllm/model_executor/models/molmo.py
View file @ 9d1c4747
......@@ -1404,10 +1404,9 @@ class MolmoForCausalLM(
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
multimodal_config = vllm_config.model_config.multimodal_config
lora_config = vllm_config.lora_config
self.config = config
self.multimodal_config = multimodal_config
self.lora_config = lora_config
vision_config = VisionBackboneConfig()
self.vision_backbone = MolmoVisionBackbone(config, vision_config, quant_config)
......
vllm/model_executor/models/nemotron.py
View file @ 9d1c4747
......@@ -45,7 +45,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
......@@ -319,24 +318,18 @@ class NemotronModel(nn.Module):
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
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
self.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,
)
else:
self.embed_tokens = PPMissingLayer()
......@@ -467,29 +460,20 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
super().__init__()
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_config
assert isinstance(config, NemotronConfig)
self.config = config
self.lora_config = lora_config
self.quant_config = quant_config
self.model = NemotronModel(
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.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"),
)
......@@ -498,7 +482,7 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
logit_scale = getattr(config, "logit_scale", 1.0)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size, logit_scale
config.vocab_size, scale=logit_scale
)
else:
self.lm_head = PPMissingLayer()
......
vllm/model_executor/models/nemotron_h.py
View file @ 9d1c4747
......@@ -50,7 +50,6 @@ from vllm.model_executor.layers.mamba.mamba_utils import (
)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
......@@ -513,21 +512,14 @@ class NemotronHModel(nn.Module):
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
parallel_config = vllm_config.parallel_config
lora_config = vllm_config.lora_config
self.config = config
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
self.vocab_size = config.vocab_size
self.embed_tokens = VocabParallelEmbedding(
self.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
)
self.has_moe = "E" in config.hybrid_override_pattern
......@@ -768,7 +760,7 @@ class NemotronHForCausalLM(
config = vllm_config.model_config.hf_config
self.vllm_config = vllm_config
self.model_config = vllm_config.model_config
lora_config = vllm_config.lora_config
scheduler_config = vllm_config.scheduler_config
self.quant_config = vllm_config.quant_config
......@@ -779,24 +771,14 @@ class NemotronHForCausalLM(
self.model = NemotronHModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
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.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,
prefix=maybe_prefix(prefix, "lm_head"),
)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size
)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors
......
vllm/model_executor/models/nemotron_nas.py
View file @ 9d1c4747
......@@ -41,7 +41,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
......@@ -250,25 +249,19 @@ class DeciModel(nn.Module):
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
self.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:
......@@ -437,29 +430,17 @@ class DeciLMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, HasNoOps):
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.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"),
)
......@@ -468,7 +449,7 @@ class DeciLMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, HasNoOps):
logit_scale = getattr(config, "logit_scale", 1.0)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size, logit_scale
config.vocab_size, scale=logit_scale
)
else:
self.lm_head = PPMissingLayer()
......
vllm/model_executor/models/olmo.py
View file @ 9d1c4747
......@@ -368,11 +368,9 @@ class OlmoForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
if config.tie_word_embeddings:
self.lm_head = self.model.embed_tokens
else:
self.unpadded_vocab_size = config.vocab_size
self.lm_head = ParallelLMHead(
self.unpadded_vocab_size,
config.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "lm_head"),
)
......
vllm/model_executor/models/olmo2.py
View file @ 9d1c4747
......@@ -408,11 +408,9 @@ class Olmo2ForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
if config.tie_word_embeddings:
self.lm_head = self.model.embed_tokens
else:
self.unpadded_vocab_size = config.vocab_size
self.lm_head = ParallelLMHead(
config.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
quant_config=vllm_config.quant_config,
prefix=maybe_prefix(prefix, "lm_head"),
)
......
vllm/model_executor/models/ouro.py
View file @ 9d1c4747
......@@ -462,10 +462,8 @@ class OuroForCausalLM(nn.Module, SupportsLoRA):
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.quant_config = quant_config
self.model = OuroModel(
......
vllm/model_executor/models/phi.py
View file @ 9d1c4747
......@@ -323,11 +323,10 @@ class PhiForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
super().__init__()
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_config
self.config = config
# lm_head use bias, cannot share word embeddings
assert not config.tie_word_embeddings
self.lora_config = lora_config
self.quant_config = quant_config
......
vllm/model_executor/models/phi3v.py
View file @ 9d1c4747
......@@ -591,7 +591,6 @@ class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsQuant)
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
quant_config=self.quant_config,
prefix=maybe_prefix(prefix, "model.embed_tokens"),
)
......
vllm/model_executor/models/phi4mm.py
View file @ 9d1c4747
......@@ -21,7 +21,6 @@ from vllm.distributed import get_pp_group
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,
)
from vllm.model_executor.models.llama import LlamaModel
......@@ -1023,12 +1022,10 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
multimodal_config = vllm_config.model_config.multimodal_config
assert multimodal_config, "multimodal_config is required"
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_config
self.config = config
self.multimodal_config = multimodal_config
self.quant_config = quant_config
self.lora_config = lora_config
# Tensor/Pipeline parallel not supported for now.
assert get_pp_group().world_size == 1, "pipeline parallel is not supported"
......@@ -1055,23 +1052,16 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
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.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
padding_size=DEFAULT_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
)
self.logits_processor = LogitsProcessor(config.vocab_size, scale=logit_scale)
def _parse_and_validate_audio_input(
self, **kwargs: object
......
vllm/model_executor/models/phimoe.py
View file @ 9d1c4747
......@@ -45,7 +45,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
......@@ -458,22 +457,15 @@ class PhiMoEModel(nn.Module):
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
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
self.vocab_size = config.vocab_size
self.config = config
self.quant_config = quant_config
self.embed_tokens = VocabParallelEmbedding(
self.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
)
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
......@@ -634,35 +626,23 @@ class PhiMoEForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
lora_config = vllm_config.lora_config
self.config = config
self.lora_config = lora_config
self.quant_config = vllm_config.quant_config
self.model = PhiMoEModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
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.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=None,
bias=True,
prefix=maybe_prefix(prefix, "lm_head"),
)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size
)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors
......
vllm/model_executor/models/plamo2.py
View file @ 9d1c4747
......@@ -46,7 +46,6 @@ from vllm.model_executor.layers.mamba.ops.ssd_combined import (
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
......@@ -751,12 +750,10 @@ class Plamo2Model(torch.nn.Module):
self.config = config
self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size
self.org_vocab_size = config.vocab_size
self.embed_tokens = VocabParallelEmbedding(
self.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
prefix=f"{prefix}.embed_tokens",
)
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
......@@ -827,20 +824,16 @@ class Plamo2ForCausalLM(torch.nn.Module, HasInnerState, SupportsPP, IsHybrid):
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
self.vocab_size = self.config.vocab_size
self.unpadded_vocab_size = self.config.vocab_size
num_embeddings = ((self.vocab_size + 15) // 16) * 16
self.lm_head = ParallelLMHead(
num_embeddings,
self.vocab_size,
self.config.hidden_size,
org_num_embeddings=self.config.vocab_size,
padding_size=DEFAULT_VOCAB_PADDING_SIZE,
prefix=f"{prefix}.lm_head",
)
if self.config.tie_word_embeddings:
self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, self.config.vocab_size
config.vocab_size, self.config.vocab_size
)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors
......
vllm/model_executor/models/qwen2.py
View file @ 9d1c4747
......@@ -477,10 +477,8 @@ class Qwen2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
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.quant_config = quant_config
self.model = Qwen2Model(
......
vllm/model_executor/models/qwen2_rm.py
View file @ 9d1c4747
......@@ -43,10 +43,8 @@ class Qwen2RewardBaseModel(nn.Module, SupportsLoRA, SupportsPP):
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.quant_config = quant_config
self.model = Qwen2Model(
......
vllm/model_executor/models/qwen3.py
View file @ 9d1c4747
......@@ -272,10 +272,8 @@ class Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
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.quant_config = quant_config
self.model = Qwen3Model(
......
vllm/model_executor/models/qwen3_next.py
View file @ 9d1c4747
......@@ -59,7 +59,6 @@ from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
......@@ -967,22 +966,17 @@ class Qwen3NextModel(nn.Module):
config: Qwen3NextConfig = vllm_config.model_config.hf_config
parallel_config = vllm_config.parallel_config
lora_config = vllm_config.lora_config
eplb_config = parallel_config.eplb_config
self.num_redundant_experts = eplb_config.num_redundant_experts
self.config = config
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.vocab_size = config.vocab_size
self.embed_tokens = VocabParallelEmbedding(
self.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
)
def get_layer(prefix: str):
......@@ -1196,7 +1190,7 @@ class Qwen3NextForCausalLM(
self.vllm_config = vllm_config
self.model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
lora_config = vllm_config.lora_config
scheduler_config = vllm_config.scheduler_config
assert not cache_config.enable_prefix_caching, (
"Qwen3Next currently does not support prefix caching"
......@@ -1209,23 +1203,13 @@ class Qwen3NextForCausalLM(
self.model = Qwen3NextModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
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.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,
prefix=maybe_prefix(prefix, "lm_head"),
)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size
)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors
)
......
vllm/model_executor/models/qwen3_next_mtp.py
View file @ 9d1c4747
......@@ -15,7 +15,6 @@ from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.linear import ColumnParallelLinear
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
......@@ -48,17 +47,12 @@ class Qwen3NextMultiTokenPredictor(nn.Module):
model_config = vllm_config.model_config
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_config
config: Qwen3NextConfig = model_config.hf_config
self.config = config
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
self.vocab_size = config.vocab_size
self.mtp_start_layer_idx = config.num_hidden_layers
self.num_mtp_layers = getattr(config, "num_nextn_predict_layers", 1)
......@@ -66,7 +60,6 @@ class Qwen3NextMultiTokenPredictor(nn.Module):
self.embed_tokens = VocabParallelEmbedding(
self.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
)
self.fc = ColumnParallelLinear(
......@@ -252,17 +245,13 @@ class Qwen3NextMTP(nn.Module, SupportsPP, QwenNextMixtureOfExperts):
self.model = Qwen3NextMultiTokenPredictor(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "mtp")
)
self.unpadded_vocab_size = config.vocab_size
self.lm_head = ParallelLMHead(
self.unpadded_vocab_size,
config.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
padding_size=DEFAULT_VOCAB_PADDING_SIZE,
prefix=maybe_prefix(prefix, "lm_head"),
)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size
)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors
)
......
vllm/model_executor/models/qwen3_vl.py
View file @ 9d1c4747
......@@ -1136,10 +1136,8 @@ class Qwen3LLMForCausalLM(Qwen3ForCausalLM):
super(Qwen3ForCausalLM, self).__init__()
config = vllm_config.model_config.hf_config.text_config
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_config
self.config = config
self.lora_config = lora_config
self.quant_config = quant_config
self.model = Qwen3LLMModel(vllm_config=vllm_config, prefix=prefix)
......
vllm/model_executor/models/seed_oss.py
View file @ 9d1c4747
......@@ -440,10 +440,8 @@ class SeedOssForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
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.quant_config = quant_config
self.model = SeedOssModel(
......
vllm/model_executor/models/solar.py
View file @ 9d1c4747
......@@ -46,7 +46,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
......@@ -277,24 +276,18 @@ class SolarModel(nn.Module):
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
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
self.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,
)
else:
self.embed_tokens = PPMissingLayer()
......@@ -455,9 +448,9 @@ class SolarForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
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.quant_config = quant_config
self.model = SolarModel(
......@@ -465,18 +458,9 @@ class SolarForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
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.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"),
)
......@@ -485,7 +469,7 @@ class SolarForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
logit_scale = getattr(config, "logit_scale", 1.0)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size, logit_scale
config.vocab_size, scale=logit_scale
)
else:
self.lm_head = PPMissingLayer()
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment