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

167
vllm/model_executor/models/jais.py
View File

@@ -32,48 +32,57 @@ from torch import nn
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.model_executor.layers.linear import (ColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear)
from vllm.distributed import (
get_pp_group,
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
)
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 vllm.transformers_utils.configs import JAISConfig
from .interfaces import SupportsPP
from .utils import (is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
class SwiGLUActivation(nn.Module):
def forward(self, x1: torch.Tensor, x2: torch.Tensor) -> torch.Tensor:
return x1 * nn.functional.silu(x2)
def _get_alibi_slopes(n):
def get_slopes_power_of_2(n):
start = 2**(-(2**-(math.log2(n) - 3)))
start = 2 ** (-(2 ** -(math.log2(n) - 3)))
ratio = start
return [start * ratio**i for i in range(n)]
if math.log2(n).is_integer():
return get_slopes_power_of_2(n)
else:
closest_power_of_2 = 2**math.floor(math.log2(n))
return (get_slopes_power_of_2(closest_power_of_2) + _get_alibi_slopes(
2 * closest_power_of_2)[0::2][:n - closest_power_of_2])
closest_power_of_2 = 2 ** math.floor(math.log2(n))
return (
get_slopes_power_of_2(closest_power_of_2)
+ _get_alibi_slopes(2 * closest_power_of_2)[0::2][: n - closest_power_of_2]
)
class JAISAttention(nn.Module):
def __init__(
self,
config: JAISConfig,
@@ -84,8 +93,7 @@ class JAISAttention(nn.Module):
super().__init__()
self.hidden_size = config.hidden_size
total_num_heads = config.num_attention_heads
tensor_model_parallel_world_size = (
get_tensor_model_parallel_world_size())
tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
assert total_num_heads % tensor_model_parallel_world_size == 0
self.num_heads = total_num_heads // tensor_model_parallel_world_size
self.head_dim = self.hidden_size // total_num_heads
@@ -113,13 +121,15 @@ class JAISAttention(nn.Module):
head_end = (tp_rank + 1) * self.num_heads
alibi_slopes = _get_alibi_slopes(total_num_heads)
alibi_slopes = alibi_slopes[head_start:head_end]
self.attn = Attention(self.num_heads,
self.head_dim,
scale=self.scale,
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,
scale=self.scale,
alibi_slopes=alibi_slopes,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn",
)
def forward(
self,
@@ -133,7 +143,6 @@ class JAISAttention(nn.Module):
class JAISMLP(nn.Module):
def __init__(
self,
intermediate_size: int,
@@ -149,12 +158,16 @@ class JAISMLP(nn.Module):
bias=True,
quant_config=quant_config,
)
self.c_fc2 = (ColumnParallelLinear(
hidden_size,
intermediate_size,
bias=True,
quant_config=quant_config,
) if self.swiglu else None)
self.c_fc2 = (
ColumnParallelLinear(
hidden_size,
intermediate_size,
bias=True,
quant_config=quant_config,
)
if self.swiglu
else None
)
self.c_proj = RowParallelLinear(
intermediate_size,
hidden_size,
@@ -168,14 +181,16 @@ class JAISMLP(nn.Module):
if self.swiglu:
hidden_states2, _ = self.c_fc2(hidden_states)
hidden_states, _ = self.c_fc(hidden_states)
hidden_states = (self.act(hidden_states, hidden_states2)
if self.swiglu else self.act(hidden_states))
hidden_states = (
self.act(hidden_states, hidden_states2)
if self.swiglu
else self.act(hidden_states)
)
hidden_states, _ = self.c_proj(hidden_states)
return hidden_states
class JAISBlock(nn.Module):
def __init__(
self,
config: JAISConfig,
@@ -185,14 +200,12 @@ class JAISBlock(nn.Module):
):
super().__init__()
hidden_size = config.hidden_size
inner_dim = (config.n_inner if config.n_inner is not None else 4 *
hidden_size)
inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
self.attn = JAISAttention(config,
cache_config,
quant_config,
prefix=f"{prefix}.attn")
self.attn = JAISAttention(
config, cache_config, quant_config, prefix=f"{prefix}.attn"
)
self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
self.mlp = JAISMLP(inner_dim, config, quant_config)
@@ -202,7 +215,9 @@ class JAISBlock(nn.Module):
) -> torch.Tensor:
residual = hidden_states
hidden_states = self.ln_1(hidden_states)
attn_output = self.attn(hidden_states=hidden_states, )
attn_output = self.attn(
hidden_states=hidden_states,
)
# residual connection
hidden_states = attn_output + residual
@@ -216,7 +231,6 @@ class JAISBlock(nn.Module):
@support_torch_compile
class JAISModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -230,9 +244,11 @@ class JAISModel(nn.Module):
assert not config.reorder_and_upcast_attn
self.embed_dim = config.hidden_size
self.wte = VocabParallelEmbedding(config.vocab_size, self.embed_dim)
self.wpe = (nn.Embedding(config.max_position_embeddings,
self.embed_dim)
if config.position_embedding_type != "alibi" else None)
self.wpe = (
nn.Embedding(config.max_position_embeddings, self.embed_dim)
if config.position_embedding_type != "alibi"
else None
)
if hasattr(config, "embeddings_scale"):
self.embeddings_scale = config.embeddings_scale
else:
@@ -240,17 +256,19 @@ class JAISModel(nn.Module):
self.start_layer, self.end_layer, self.h = make_layers(
config.num_hidden_layers,
lambda prefix: JAISBlock(config=config,
cache_config=cache_config,
quant_config=quant_config,
prefix=prefix),
lambda prefix: JAISBlock(
config=config,
cache_config=cache_config,
quant_config=quant_config,
prefix=prefix,
),
prefix=f"{prefix}.h",
)
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.n_embd))
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
["hidden_states"], config.n_embd
)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.wte(input_ids)
@@ -270,8 +288,9 @@ class JAISModel(nn.Module):
hidden_states = inputs_embeds + position_embeds
else:
hidden_states = inputs_embeds
hidden_states *= torch.tensor(float(self.embeddings_scale),
dtype=hidden_states.dtype)
hidden_states *= torch.tensor(
float(self.embeddings_scale), dtype=hidden_states.dtype
)
else:
assert intermediate_tensors is not None
hidden_states = intermediate_tensors["hidden_states"]
@@ -287,32 +306,33 @@ class JAISModel(nn.Module):
class JAISLMHeadModel(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.transformer = JAISModel(vllm_config=vllm_config,
prefix=maybe_prefix(
prefix, "transformer"))
self.transformer = JAISModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "transformer")
)
if self.config.tie_word_embeddings:
self.lm_head = self.transformer.wte
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"),
)
if hasattr(config, "width_scale"):
self.output_logits_scale = config.width_scale
else:
self.output_logits_scale = (config.mup_output_alpha *
config.mup_width_scale)
self.logits_processor = LogitsProcessor(vocab_size=config.vocab_size,
scale=self.output_logits_scale)
self.output_logits_scale = config.mup_output_alpha * config.mup_width_scale
self.logits_processor = LogitsProcessor(
vocab_size=config.vocab_size, scale=self.output_logits_scale
)
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)
@@ -324,8 +344,9 @@ class JAISLMHeadModel(nn.Module, SupportsPP):
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[IntermediateTensors, torch.Tensor]:
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(
@@ -335,8 +356,7 @@ class JAISLMHeadModel(nn.Module, 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]:
params_dict = dict(self.named_parameters(remove_duplicate=False))
loaded_params: set[str] = set()
for name, loaded_weight in weights:
@@ -366,8 +386,7 @@ class JAISLMHeadModel(nn.Module, SupportsPP):
if not name.endswith(".weight"):
continue
loaded_weight = loaded_weight.t()
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