biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

136
vllm/model_executor/models/exaone4.py
View File

@@ -35,27 +35,38 @@ from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear,
)
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)
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader, maybe_remap_kv_scale_name)
default_weight_loader,
maybe_remap_kv_scale_name,
)
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
extract_layer_index,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
class Exaone4GatedMLP(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -81,8 +92,9 @@ class Exaone4GatedMLP(nn.Module):
prefix=f"{prefix}.down_proj",
)
if hidden_act != "silu":
raise ValueError(f"Unsupported activation: {hidden_act}. "
"Only silu is supported for now.")
raise ValueError(
f"Unsupported activation: {hidden_act}. Only silu is supported for now."
)
self.act_fn = SiluAndMul()
def forward(self, x):
@@ -93,7 +105,6 @@ class Exaone4GatedMLP(nn.Module):
class Exaone4Attention(nn.Module):
def __init__(
self,
config: Exaone4Config,
@@ -208,7 +219,6 @@ class Exaone4Attention(nn.Module):
class Exaone4DecoderLayer(nn.Module):
def __init__(
self,
config: Exaone4Config,
@@ -221,22 +231,25 @@ class Exaone4DecoderLayer(nn.Module):
rope_theta = getattr(config, "rope_theta", 1000000)
rope_scaling = getattr(config, "rope_scaling", None)
if rope_scaling is not None and getattr(
config, "original_max_position_embeddings", None):
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)
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)
config, "bias", False
)
self.self_attn = Exaone4Attention(
config=config,
hidden_size=self.hidden_size,
num_heads=config.num_attention_heads,
num_kv_heads=getattr(config, "num_key_value_heads",
config.num_attention_heads),
num_kv_heads=getattr(
config, "num_key_value_heads", config.num_attention_heads
),
rope_theta=rope_theta,
rope_scaling=rope_scaling,
max_position_embeddings=max_position_embeddings,
@@ -253,10 +266,12 @@ class Exaone4DecoderLayer(nn.Module):
bias=getattr(config, "mlp_bias", False),
prefix=f"{prefix}.mlp",
)
self.post_attention_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.post_feedforward_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.post_attention_layernorm = RMSNorm(
config.hidden_size, eps=config.rms_norm_eps
)
self.post_feedforward_layernorm = RMSNorm(
config.hidden_size, eps=config.rms_norm_eps
)
def forward(
self,
@@ -290,7 +305,6 @@ class Exaone4DecoderLayer(nn.Module):
@support_torch_compile
class Exaone4Model(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -301,11 +315,15 @@ class Exaone4Model(nn.Module):
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)
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
if get_pp_group().is_first_rank or (config.tie_word_embeddings
and get_pp_group().is_last_rank):
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,
@@ -329,9 +347,9 @@ class Exaone4Model(nn.Module):
else:
self.norm = PPMissingLayer()
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size))
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)
@@ -362,16 +380,14 @@ class Exaone4Model(nn.Module):
)
if not get_pp_group().is_last_rank:
return IntermediateTensors({
"hidden_states": hidden_states,
"residual": residual
})
return IntermediateTensors(
{"hidden_states": hidden_states, "residual": residual}
)
hidden_states = self.norm(hidden_states)
return hidden_states
def load_weights(self, weights: Iterable[tuple[str,
torch.Tensor]]) -> set[str]:
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"),
@@ -385,19 +401,19 @@ class Exaone4Model(nn.Module):
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):
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))):
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 = 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
@@ -430,8 +446,7 @@ class Exaone4Model(nn.Module):
continue
param = params_dict[name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader = getattr(param, "weight_loader", default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params
@@ -482,7 +497,8 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
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,
if not lora_config
else lora_config.lora_vocab_padding_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "lm_head"),
)
@@ -490,14 +506,15 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
self.lm_head.weight = self.model.embed_tokens.weight
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(
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)
self.model.make_empty_intermediate_tensors
)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.get_input_embeddings(input_ids)
@@ -509,8 +526,9 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
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)
model_output = self.model(
input_ids, positions, intermediate_tensors, inputs_embeds
)
return model_output
def compute_logits(
@@ -520,14 +538,12 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
logits = self.logits_processor(self.lm_head, hidden_states)
return logits
def load_weights(self, weights: Iterable[tuple[str,
torch.Tensor]]) -> set[str]:
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
loader = AutoWeightsLoader(
self,
# With tie_word_embeddings, we can skip lm_head.weight
# The weight might appear unnecessarily in the files if the model is
# processed with quantization, LoRA, fine-tuning, etc.
skip_prefixes=(["lm_head."]
if self.config.tie_word_embeddings else None),
skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
)
return loader.load_weights(weights)