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

177
vllm/model_executor/models/gemma2.py
View File

@@ -31,29 +31,35 @@ from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import GeluAndMul
from vllm.model_executor.layers.layernorm import GemmaRMSNorm
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 (
VocabParallelEmbedding)
from vllm.model_executor.layers.vocab_parallel_embedding import 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, extract_layer_index,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
AutoWeightsLoader,
extract_layer_index,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
logger = init_logger(__name__)
class Gemma2MLP(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -64,18 +70,17 @@ class Gemma2MLP(nn.Module):
) -> None:
super().__init__()
self.gate_up_proj = MergedColumnParallelLinear(
hidden_size, [intermediate_size] * 2,
bias=False,
quant_config=quant_config)
self.down_proj = RowParallelLinear(intermediate_size,
hidden_size,
bias=False,
quant_config=quant_config)
hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config
)
self.down_proj = RowParallelLinear(
intermediate_size, hidden_size, bias=False, quant_config=quant_config
)
if not (hidden_act == hidden_activation == "gelu_pytorch_tanh"):
raise ValueError(
"Gemma2 uses `gelu_pytorch_tanh` as the hidden activation "
"function. Please set `hidden_act` and `hidden_activation` to "
"`gelu_pytorch_tanh`.")
"`gelu_pytorch_tanh`."
)
self.act_fn = GeluAndMul(approximate="tanh")
def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -86,19 +91,20 @@ class Gemma2MLP(nn.Module):
class Gemma2Attention(nn.Module):
def __init__(self,
config: Gemma2Config,
hidden_size: int,
num_heads: int,
num_kv_heads: int,
head_dim: int,
max_position_embeddings: int,
rope_theta: float,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
attn_logits_soft_cap: Optional[float] = None,
prefix: str = "") -> None:
def __init__(
self,
config: Gemma2Config,
hidden_size: int,
num_heads: int,
num_kv_heads: int,
head_dim: int,
max_position_embeddings: int,
rope_theta: float,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
attn_logits_soft_cap: Optional[float] = None,
prefix: str = "",
) -> None:
super().__init__()
self.config = config
self.hidden_size = hidden_size
@@ -148,15 +154,17 @@ class Gemma2Attention(nn.Module):
is_sliding = config.layer_types[layer_idx] == "sliding_attention"
sliding_window = config.sliding_window if is_sliding else None
self.attn = Attention(self.num_heads,
self.head_dim,
self.scaling,
num_kv_heads=self.num_kv_heads,
cache_config=cache_config,
quant_config=quant_config,
logits_soft_cap=attn_logits_soft_cap,
per_layer_sliding_window=sliding_window,
prefix=f"{prefix}.attn")
self.attn = Attention(
self.num_heads,
self.head_dim,
self.scaling,
num_kv_heads=self.num_kv_heads,
cache_config=cache_config,
quant_config=quant_config,
logits_soft_cap=attn_logits_soft_cap,
per_layer_sliding_window=sliding_window,
prefix=f"{prefix}.attn",
)
def forward(
self,
@@ -172,7 +180,6 @@ class Gemma2Attention(nn.Module):
class Gemma2DecoderLayer(nn.Module):
def __init__(
self,
config: Gemma2Config,
@@ -203,14 +210,16 @@ class Gemma2DecoderLayer(nn.Module):
hidden_activation=config.hidden_activation,
quant_config=quant_config,
)
self.input_layernorm = GemmaRMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.post_attention_layernorm = GemmaRMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.pre_feedforward_layernorm = GemmaRMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.post_feedforward_layernorm = GemmaRMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.input_layernorm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.post_attention_layernorm = GemmaRMSNorm(
config.hidden_size, eps=config.rms_norm_eps
)
self.pre_feedforward_layernorm = GemmaRMSNorm(
config.hidden_size, eps=config.rms_norm_eps
)
self.post_feedforward_layernorm = GemmaRMSNorm(
config.hidden_size, eps=config.rms_norm_eps
)
def forward(
self,
@@ -222,8 +231,7 @@ class Gemma2DecoderLayer(nn.Module):
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
else:
hidden_states, residual = self.input_layernorm(
hidden_states, residual)
hidden_states, residual = self.input_layernorm(hidden_states, residual)
hidden_states = self.self_attn(
positions=positions,
hidden_states=hidden_states,
@@ -231,7 +239,8 @@ class Gemma2DecoderLayer(nn.Module):
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states, residual = self.pre_feedforward_layernorm(
hidden_states, residual)
hidden_states, residual
)
hidden_states = self.mlp(hidden_states)
hidden_states = self.post_feedforward_layernorm(hidden_states)
return hidden_states, residual
@@ -239,7 +248,6 @@ class Gemma2DecoderLayer(nn.Module):
@support_torch_compile
class Gemma2Model(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
@@ -255,8 +263,10 @@ class Gemma2Model(nn.Module):
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: Gemma2DecoderLayer(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.layers")
config, cache_config, quant_config, prefix=prefix
),
prefix=f"{prefix}.layers",
)
self.norm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
# Normalize the embedding by sqrt(hidden_size)
@@ -264,12 +274,10 @@ class Gemma2Model(nn.Module):
# data type such as bfloat16, not float32.
# See https://github.com/huggingface/transformers/pull/29402
normalizer = self.config.hidden_size**0.5
self.register_buffer("normalizer",
torch.tensor(normalizer),
persistent=False)
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size))
self.register_buffer("normalizer", torch.tensor(normalizer), persistent=False)
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)
@@ -299,15 +307,13 @@ class Gemma2Model(nn.Module):
residual,
)
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, residual)
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"),
@@ -319,17 +325,17 @@ class Gemma2Model(nn.Module):
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
for name, loaded_weight in weights:
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 scales for compressed-tensors quantization
param = params_dict[scale_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader = getattr(param, "weight_loader", default_weight_loader)
loaded_weight = loaded_weight[0]
weight_loader(param, loaded_weight)
loaded_params.add(scale_name)
continue
for (param_name, shard_name, shard_id) in stacked_params_mapping:
for param_name, shard_name, shard_id in stacked_params_mapping:
if shard_name not in name:
continue
name = name.replace(shard_name, param_name)
@@ -353,8 +359,7 @@ class Gemma2Model(nn.Module):
if is_pp_missing_parameter(name, self):
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)
@@ -384,12 +389,15 @@ class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
# currently all existing Gemma models have `tie_word_embeddings` enabled
assert config.tie_word_embeddings
self.quant_config = quant_config
self.model = Gemma2Model(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.model = Gemma2Model(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
self.logits_processor = LogitsProcessor(
config.vocab_size, soft_cap=config.final_logit_softcapping)
config.vocab_size, soft_cap=config.final_logit_softcapping
)
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)
@@ -401,8 +409,9 @@ class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
hidden_states = self.model(input_ids, positions, intermediate_tensors,
inputs_embeds)
hidden_states = self.model(
input_ids, positions, intermediate_tensors, inputs_embeds
)
return hidden_states
def compute_logits(
@@ -412,11 +421,9 @@ class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
logits = self.logits_processor(self.model.embed_tokens, 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,
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)