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:
@@ -11,25 +11,39 @@ from transformers import DbrxConfig
|
||||
|
||||
from vllm.attention import Attention
|
||||
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.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.linear import (QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.linear import (
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
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 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 DbrxRouter(nn.Module):
|
||||
@@ -60,7 +74,6 @@ class DbrxRouter(nn.Module):
|
||||
|
||||
|
||||
class DbrxExperts(FusedMoE):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: DbrxConfig,
|
||||
@@ -82,12 +95,16 @@ class DbrxExperts(FusedMoE):
|
||||
)
|
||||
self.config = config
|
||||
self.d_model = config.d_model
|
||||
self.intermediate_size = (self.config.ffn_config.ffn_hidden_size //
|
||||
self.tp_size)
|
||||
self.intermediate_size = self.config.ffn_config.ffn_hidden_size // self.tp_size
|
||||
|
||||
# Define custom weight loader for dbrx model
|
||||
def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
|
||||
weight_name: str, param_name: str):
|
||||
def weight_loader(
|
||||
self,
|
||||
param: nn.Parameter,
|
||||
loaded_weight: torch.Tensor,
|
||||
weight_name: str,
|
||||
param_name: str,
|
||||
):
|
||||
tp_rank = get_tensor_model_parallel_rank()
|
||||
param_data = param.data
|
||||
shard_size = self.intermediate_size
|
||||
@@ -111,8 +128,9 @@ class DbrxExperts(FusedMoE):
|
||||
loaded_weight,
|
||||
[-1, self.intermediate_size * self.tp_size, self.d_model],
|
||||
)
|
||||
param_data[:, shard_size:2 *
|
||||
shard_size, :] = loaded_weight[:, shard, :]
|
||||
param_data[:, shard_size : 2 * shard_size, :] = loaded_weight[
|
||||
:, shard, :
|
||||
]
|
||||
elif param_name.endswith("weight_scale"):
|
||||
param_data[:, 1] = loaded_weight
|
||||
else:
|
||||
@@ -151,10 +169,12 @@ class DbrxMoE(nn.Module):
|
||||
|
||||
self.router = DbrxRouter(config, self.params_dtype)
|
||||
|
||||
self.experts = DbrxExperts(config=config,
|
||||
quant_config=quant_config,
|
||||
params_dtype=self.params_dtype,
|
||||
prefix=f"{prefix}.experts")
|
||||
self.experts = DbrxExperts(
|
||||
config=config,
|
||||
quant_config=quant_config,
|
||||
params_dtype=self.params_dtype,
|
||||
prefix=f"{prefix}.experts",
|
||||
)
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
orig_shape = hidden_states.shape
|
||||
@@ -166,7 +186,6 @@ class DbrxMoE(nn.Module):
|
||||
|
||||
|
||||
class DbrxAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: DbrxConfig,
|
||||
@@ -222,13 +241,15 @@ class DbrxAttention(nn.Module):
|
||||
self.q_size = self.num_heads * self.head_dim
|
||||
self.kv_size = self.num_kv_heads * self.head_dim
|
||||
self.scaling = self.head_dim**-0.5
|
||||
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,
|
||||
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,
|
||||
prefix=f"{prefix}.attn",
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
@@ -246,7 +267,6 @@ class DbrxAttention(nn.Module):
|
||||
|
||||
|
||||
class DbrxFusedNormAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: DbrxConfig,
|
||||
@@ -256,10 +276,9 @@ class DbrxFusedNormAttention(nn.Module):
|
||||
):
|
||||
super().__init__()
|
||||
self.d_model = config.d_model
|
||||
self.attn = DbrxAttention(config,
|
||||
cache_config,
|
||||
quant_config,
|
||||
prefix=f"{prefix}.attn")
|
||||
self.attn = DbrxAttention(
|
||||
config, cache_config, quant_config, prefix=f"{prefix}.attn"
|
||||
)
|
||||
self.norm_1 = nn.LayerNorm(self.d_model)
|
||||
self.norm_2 = nn.LayerNorm(self.d_model)
|
||||
|
||||
@@ -281,7 +300,6 @@ class DbrxFusedNormAttention(nn.Module):
|
||||
|
||||
|
||||
class DbrxBlock(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: DbrxConfig,
|
||||
@@ -291,10 +309,8 @@ class DbrxBlock(nn.Module):
|
||||
):
|
||||
super().__init__()
|
||||
self.norm_attn_norm = DbrxFusedNormAttention(
|
||||
config,
|
||||
cache_config,
|
||||
quant_config,
|
||||
prefix=f"{prefix}.norm_attn_norm")
|
||||
config, cache_config, quant_config, prefix=f"{prefix}.norm_attn_norm"
|
||||
)
|
||||
self.ffn = DbrxMoE(config, quant_config, prefix=f"{prefix}.ffn")
|
||||
|
||||
def forward(
|
||||
@@ -312,7 +328,6 @@ class DbrxBlock(nn.Module):
|
||||
|
||||
|
||||
class DbrxModel(nn.Module):
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
|
||||
@@ -327,19 +342,17 @@ class DbrxModel(nn.Module):
|
||||
)
|
||||
self.start_layer, self.end_layer, self.blocks = make_layers(
|
||||
config.n_layers,
|
||||
lambda prefix: DbrxBlock(
|
||||
config, cache_config, quant_config, prefix=prefix),
|
||||
lambda prefix: DbrxBlock(config, cache_config, quant_config, prefix=prefix),
|
||||
prefix=f"{prefix}.blocks",
|
||||
)
|
||||
self.norm_f = nn.LayerNorm(config.d_model, eps=1e-5)
|
||||
for module in self.modules():
|
||||
if hasattr(module, "bias") and isinstance(module.bias,
|
||||
nn.Parameter):
|
||||
if hasattr(module, "bias") and isinstance(module.bias, nn.Parameter):
|
||||
# Remove the bias term in Linear and LayerNorm.
|
||||
module.register_parameter("bias", None)
|
||||
self.make_empty_intermediate_tensors = (
|
||||
make_empty_intermediate_tensors_factory(["hidden_states"],
|
||||
config.d_model))
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states"], config.d_model
|
||||
)
|
||||
|
||||
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.wte(input_ids)
|
||||
@@ -366,24 +379,27 @@ class DbrxModel(nn.Module):
|
||||
hidden_states = self.norm_f(hidden_states)
|
||||
return hidden_states
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
expert_params_mapping = [(
|
||||
"w13" if weight_name in ["w1", "v1"] else "w2",
|
||||
f"mlp.{weight_name}",
|
||||
) for weight_name in ["w1", "v1", "w2"]]
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
expert_params_mapping = [
|
||||
(
|
||||
"w13" if weight_name in ["w1", "v1"] else "w2",
|
||||
f"mlp.{weight_name}",
|
||||
)
|
||||
for weight_name in ["w1", "v1", "w2"]
|
||||
]
|
||||
params_dict = dict(self.named_parameters(remove_duplicate=False))
|
||||
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 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
|
||||
@@ -409,28 +425,25 @@ class DbrxModel(nn.Module):
|
||||
if name is None:
|
||||
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
|
||||
|
||||
|
||||
class DbrxForCausalLM(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
|
||||
if config.tie_word_embeddings:
|
||||
raise ValueError(
|
||||
"tie_word_embeddings is not supported for Dbrx models.")
|
||||
raise ValueError("tie_word_embeddings is not supported for Dbrx models.")
|
||||
self.quant_config = quant_config
|
||||
self.unpadded_vocab_size = config.vocab_size
|
||||
self.transformer = DbrxModel(vllm_config=vllm_config,
|
||||
prefix=maybe_prefix(
|
||||
prefix, "transformer"))
|
||||
self.transformer = DbrxModel(
|
||||
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "transformer")
|
||||
)
|
||||
self.lm_head = ParallelLMHead(
|
||||
config.vocab_size,
|
||||
config.d_model,
|
||||
@@ -439,10 +452,12 @@ class DbrxForCausalLM(nn.Module, SupportsPP):
|
||||
quant_config=quant_config,
|
||||
prefix=maybe_prefix(prefix, "lm_head"),
|
||||
)
|
||||
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
|
||||
config.vocab_size)
|
||||
self.logits_processor = LogitsProcessor(
|
||||
self.unpadded_vocab_size, 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)
|
||||
@@ -454,8 +469,9 @@ class DbrxForCausalLM(nn.Module, SupportsPP):
|
||||
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(
|
||||
@@ -465,7 +481,6 @@ class DbrxForCausalLM(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]:
|
||||
loader = AutoWeightsLoader(self)
|
||||
return loader.load_weights(weights)
|
||||
|
||||
Reference in New Issue
Block a user