[LoRA][2/2]Remove LoRA extra vocab (#28545)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li
2025-11-21 09:46:43 +08:00
committed by GitHub
parent df44df0143
commit 9875be6431
28 changed files with 133 additions and 528 deletions
+5 -29
View File
@@ -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,
)
@@ -261,29 +260,16 @@ class GraniteModel(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
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.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,
)
else:
@@ -420,28 +406,18 @@ class GraniteForCausalLM(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 = GraniteModel(
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"),
)
@@ -453,7 +429,7 @@ class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
logit_scale /= config.logits_scaling
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size, scale=logit_scale
config.vocab_size, scale=logit_scale
)
else:
self.lm_head = PPMissingLayer()
+5 -25
View File
@@ -47,7 +47,6 @@ 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,
)
@@ -368,24 +367,18 @@ class LlamaModel(nn.Module):
config = vllm_config.model_config.hf_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,
quant_config=quant_config,
)
else:
@@ -562,9 +555,7 @@ class LlamaForCausalLM(
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,
@@ -573,20 +564,9 @@ class LlamaForCausalLM(
)
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"),
)
@@ -595,7 +575,7 @@ class LlamaForCausalLM(
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()
+8 -24
View File
@@ -51,7 +51,6 @@ 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,
)
@@ -301,23 +300,18 @@ class MixtralModel(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
parallel_config = vllm_config.parallel_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.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,
)
self.enable_eplb = parallel_config.enable_eplb
@@ -508,34 +502,24 @@ class MixtralForCausalLM(nn.Module, SupportsLoRA, SupportsPP, MixtureOfExperts):
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 = MixtralModel(
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=quant_config,
prefix=maybe_prefix(prefix, "lm_head"),
)
if self.config.tie_word_embeddings:
self.lm_head.weight = self.model.embed_tokens.weight
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
)
+1 -1
View File
@@ -74,5 +74,5 @@ class TeleFLMForCausalLM(LlamaForCausalLM):
self.output_mult = self.config.output_mult / self.mup_scale_factor
logit_scale = self.output_mult
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, self.config.vocab_size, logit_scale
self.config.vocab_size, scale=logit_scale
)