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

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 07:06:22 -07:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115205 additions and 94107 deletions
+66 -57
View File
@@ -18,6 +18,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only GPT-NeoX model compatible with HuggingFace weights."""
from collections.abc import Iterable
from itertools import islice
from typing import Optional, Union
@@ -31,25 +32,32 @@ from vllm.compilation.decorators import support_torch_compile
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 get_act_fn
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
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 (
ParallelLMHead, VocabParallelEmbedding)
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsPP
from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
AutoWeightsLoader,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
class GPTNeoXAttention(nn.Module):
def __init__(
self,
config: GPTNeoXConfig,
@@ -63,11 +71,9 @@ class GPTNeoXAttention(nn.Module):
self.head_size = self.hidden_size // self.total_num_heads
self.bias = getattr(config, "attention_bias", True)
tensor_model_parallel_world_size = (
get_tensor_model_parallel_world_size())
tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
assert self.total_num_heads % tensor_model_parallel_world_size == 0
self.num_heads = (self.total_num_heads //
tensor_model_parallel_world_size)
self.num_heads = self.total_num_heads // tensor_model_parallel_world_size
self.query_key_value = QKVParallelLinear(
config.hidden_size,
@@ -86,20 +92,21 @@ class GPTNeoXAttention(nn.Module):
rotary_dim = int(self.head_size * config.rotary_pct)
assert rotary_dim % 2 == 0
rope_theta = getattr(config, "rope_theta", 10000)
max_position_embeddings = getattr(config, "max_position_embeddings",
8192)
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
self.rotary_emb = get_rope(
self.head_size,
rotary_dim=rotary_dim,
max_position=max_position_embeddings,
base=rope_theta,
)
self.attn = Attention(self.num_heads,
self.head_size,
scaling,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn")
self.attn = Attention(
self.num_heads,
self.head_size,
scaling,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn",
)
def forward(
self,
@@ -115,7 +122,6 @@ class GPTNeoXAttention(nn.Module):
class GPTNeoXMLP(nn.Module):
def __init__(
self,
config: GPTNeoXConfig,
@@ -142,7 +148,6 @@ class GPTNeoXMLP(nn.Module):
class GPTNeoXLayer(nn.Module):
def __init__(
self,
config: GPTNeoXConfig,
@@ -152,14 +157,15 @@ class GPTNeoXLayer(nn.Module):
):
super().__init__()
self.use_parallel_residual = config.use_parallel_residual
self.input_layernorm = nn.LayerNorm(config.hidden_size,
eps=config.layer_norm_eps)
self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
eps=config.layer_norm_eps)
self.attention = GPTNeoXAttention(config,
cache_config,
quant_config,
prefix=f"{prefix}.attention")
self.input_layernorm = nn.LayerNorm(
config.hidden_size, eps=config.layer_norm_eps
)
self.post_attention_layernorm = nn.LayerNorm(
config.hidden_size, eps=config.layer_norm_eps
)
self.attention = GPTNeoXAttention(
config, cache_config, quant_config, prefix=f"{prefix}.attention"
)
self.mlp = GPTNeoXMLP(config, quant_config)
def forward(
@@ -192,7 +198,6 @@ class GPTNeoXLayer(nn.Module):
@support_torch_compile
class GPTNeoXModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -209,14 +214,16 @@ class GPTNeoXModel(nn.Module):
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: GPTNeoXLayer(
config, cache_config, quant_config, prefix=prefix),
config, cache_config, quant_config, prefix=prefix
),
prefix=f"{prefix}.layers",
)
self.final_layer_norm = nn.LayerNorm(config.hidden_size,
eps=config.layer_norm_eps)
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(["hidden_states"],
config.hidden_size))
self.final_layer_norm = nn.LayerNorm(
config.hidden_size, eps=config.layer_norm_eps
)
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
["hidden_states"], config.hidden_size
)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.embed_in(input_ids)
@@ -242,16 +249,17 @@ class GPTNeoXModel(nn.Module):
hidden_states = self.final_layer_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]:
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
for name, loaded_weight in weights:
if ("attention.bias" in name or "attention.masked_bias" in name
or "rotary_emb.inv_freq" in name):
if (
"attention.bias" in name
or "attention.masked_bias" in name
or "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 OpenRLHF may include
# these tensors in the checkpoint. Skip them.
continue
@@ -269,29 +277,29 @@ class GPTNeoXModel(nn.Module):
if output_dim is not None:
loaded_weight_shape = loaded_weight.shape
loaded_weight = loaded_weight.view(
loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
loaded_weight_shape[output_dim + 1:])
loaded_weight = loaded_weight.transpose(
output_dim, output_dim + 1)
loaded_weight_shape[:output_dim]
+ (num_heads, 3, -1)
+ loaded_weight_shape[output_dim + 1 :]
)
loaded_weight = loaded_weight.transpose(output_dim, output_dim + 1)
loaded_weight = loaded_weight.reshape(loaded_weight_shape)
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
class GPTNeoXForCausalLM(nn.Module, SupportsPP):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
self.config = config
self.quant_config = quant_config
self.gpt_neox = GPTNeoXModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "gpt_neox"))
self.gpt_neox = GPTNeoXModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "gpt_neox")
)
self.embed_out = ParallelLMHead(
config.vocab_size,
config.hidden_size,
@@ -302,7 +310,8 @@ class GPTNeoXForCausalLM(nn.Module, SupportsPP):
self.embed_out.weight = self.gpt_neox.embed_in.weight
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.gpt_neox.make_empty_intermediate_tensors)
self.gpt_neox.make_empty_intermediate_tensors
)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.gpt_neox.get_input_embeddings(input_ids)
@@ -314,8 +323,9 @@ class GPTNeoXForCausalLM(nn.Module, SupportsPP):
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
hidden_states = self.gpt_neox(input_ids, positions,
intermediate_tensors, inputs_embeds)
hidden_states = self.gpt_neox(
input_ids, positions, intermediate_tensors, inputs_embeds
)
return hidden_states
def compute_logits(
@@ -325,7 +335,6 @@ class GPTNeoXForCausalLM(nn.Module, SupportsPP):
logits = self.logits_processor(self.embed_out, 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)
return loader.load_weights(weights)