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

150
vllm/model_executor/models/bloom.py
View File

@@ -18,6 +18,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only BLOOM model compatible with HuggingFace weights."""
import math
from collections.abc import Iterable
from itertools import islice
@@ -30,29 +31,40 @@ from transformers import BloomConfig
from vllm.attention import Attention
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_rank,
get_tensor_model_parallel_world_size)
from vllm.distributed import (
get_pp_group,
get_tensor_model_parallel_rank,
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.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, SupportsQuant
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,
)
def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
closest_power_of_2 = 2**math.floor(math.log2(total_num_heads))
closest_power_of_2 = 2 ** math.floor(math.log2(total_num_heads))
base = torch.tensor(
2**(-(2**-(math.log2(closest_power_of_2) - 3))),
2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))),
dtype=torch.float32,
)
powers = torch.arange(1, 1 + closest_power_of_2, dtype=torch.int32)
@@ -60,22 +72,20 @@ def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
if closest_power_of_2 != total_num_heads:
extra_base = torch.tensor(
2**(-(2**-(math.log2(2 * closest_power_of_2) - 3))),
2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))),
dtype=torch.float32,
)
num_remaining_heads = min(closest_power_of_2,
total_num_heads - closest_power_of_2)
extra_powers = torch.arange(start=1,
end=1 + 2 * num_remaining_heads,
step=2,
dtype=torch.int32)
slopes = torch.cat(
[slopes, torch.pow(extra_base, extra_powers)], dim=0)
num_remaining_heads = min(
closest_power_of_2, total_num_heads - closest_power_of_2
)
extra_powers = torch.arange(
start=1, end=1 + 2 * num_remaining_heads, step=2, dtype=torch.int32
)
slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
return slopes
class BloomAttention(nn.Module):
def __init__(
self,
config: BloomConfig,
@@ -115,13 +125,15 @@ class BloomAttention(nn.Module):
alibi_slopes = alibi_slopes[head_start:head_end].tolist()
scaling = self.head_dim**-0.5
self.attn = Attention(self.num_heads,
self.head_dim,
scaling,
alibi_slopes=alibi_slopes,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn")
self.attn = Attention(
self.num_heads,
self.head_dim,
scaling,
alibi_slopes=alibi_slopes,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn",
)
def forward(
self,
@@ -137,7 +149,6 @@ class BloomAttention(nn.Module):
class BloomMLP(nn.Module):
def __init__(
self,
config: BloomConfig,
@@ -165,7 +176,6 @@ class BloomMLP(nn.Module):
class BloomBlock(nn.Module):
def __init__(
self,
config: BloomConfig,
@@ -176,17 +186,17 @@ class BloomBlock(nn.Module):
super().__init__()
hidden_size = config.hidden_size
self.input_layernorm = nn.LayerNorm(hidden_size,
eps=config.layer_norm_epsilon)
self.self_attention = BloomAttention(config,
cache_config,
quant_config,
prefix=f"{prefix}.self_attention")
self.input_layernorm = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
self.self_attention = BloomAttention(
config, cache_config, quant_config, prefix=f"{prefix}.self_attention"
)
self.post_attention_layernorm = nn.LayerNorm(
hidden_size, eps=config.layer_norm_epsilon)
hidden_size, eps=config.layer_norm_epsilon
)
self.mlp = BloomMLP(config, quant_config)
self.apply_residual_connection_post_layernorm = (
config.apply_residual_connection_post_layernorm)
config.apply_residual_connection_post_layernorm
)
def forward(
self,
@@ -223,7 +233,6 @@ class BloomBlock(nn.Module):
@support_torch_compile
class BloomModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -240,20 +249,23 @@ class BloomModel(nn.Module):
self.embed_dim,
)
self.word_embeddings_layernorm = nn.LayerNorm(
self.embed_dim, eps=config.layer_norm_epsilon)
self.embed_dim, eps=config.layer_norm_epsilon
)
# Transformer blocks
self.start_layer, self.end_layer, self.h = make_layers(
config.num_hidden_layers,
lambda prefix: BloomBlock(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.h")
config, cache_config, quant_config, prefix=prefix
),
prefix=f"{prefix}.h",
)
# Final Layer Norm
self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(["hidden_states"],
config.hidden_size))
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.word_embeddings(input_ids)
@@ -281,8 +293,7 @@ class BloomModel(nn.Module):
hidden_states = self.ln_f(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(remove_duplicate=False))
loaded_params: set[str] = set()
for name, loaded_weight in weights:
@@ -300,14 +311,14 @@ class BloomModel(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)
@@ -315,27 +326,28 @@ class BloomModel(nn.Module):
class BloomForCausalLM(nn.Module, SupportsPP, SupportsQuant):
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.transformer = BloomModel(vllm_config=vllm_config,
prefix=maybe_prefix(
prefix, "transformer"))
self.transformer = BloomModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "transformer")
)
if self.config.tie_word_embeddings:
self.lm_head = self.transformer.word_embeddings
else:
self.lm_head = ParallelLMHead(self.config.vocab_size,
self.config.hidden_size,
prefix=maybe_prefix(
prefix, "lm_head"))
self.lm_head = ParallelLMHead(
self.config.vocab_size,
self.config.hidden_size,
prefix=maybe_prefix(prefix, "lm_head"),
)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.transformer.make_empty_intermediate_tensors)
self.transformer.make_empty_intermediate_tensors
)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.transformer.get_input_embeddings(input_ids)
@@ -347,8 +359,9 @@ class BloomForCausalLM(nn.Module, SupportsPP, SupportsQuant):
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
hidden_states = self.transformer(input_ids, positions,
intermediate_tensors, inputs_embeds)
hidden_states = self.transformer(
input_ids, positions, intermediate_tensors, inputs_embeds
)
return hidden_states
def compute_logits(
@@ -358,17 +371,16 @@ class BloomForCausalLM(nn.Module, SupportsPP, SupportsQuant):
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, skip_prefixes=["lm_head.weight"])
weights = _add_transformer_prefix(weights)
return loader.load_weights(weights)
def _add_transformer_prefix(
weights: Iterable[tuple[str, torch.Tensor]]
weights: Iterable[tuple[str, torch.Tensor]],
) -> Iterable[tuple[str, torch.Tensor]]:
for name, tensor in weights:
if not name.startswith('transformer.'):
name = 'transformer.' + name
if not name.startswith("transformer."):
name = "transformer." + name
yield name, tensor