Sync from v0.13
This commit is contained in:
@@ -1,4 +1,6 @@
|
||||
# coding=utf-8
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
# Adapted from
|
||||
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/opt/modeling_opt.py
|
||||
# Copyright 2023 The vLLM team.
|
||||
@@ -17,32 +19,46 @@
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""Inference-only OPT model compatible with HuggingFace weights."""
|
||||
from typing import Iterable, List, Optional, Tuple
|
||||
|
||||
from collections.abc import Iterable
|
||||
from itertools import islice
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import OPTConfig
|
||||
|
||||
from vllm.attention import Attention, AttentionMetadata
|
||||
from vllm.distributed import get_tensor_model_parallel_world_size
|
||||
from vllm.attention.layer 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_world_size
|
||||
from vllm.model_executor.layers.activation import get_act_fn
|
||||
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear,
|
||||
)
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig)
|
||||
from vllm.model_executor.layers.sampler import Sampler
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
VocabParallelEmbedding)
|
||||
ParallelLMHead,
|
||||
VocabParallelEmbedding,
|
||||
)
|
||||
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
||||
from vllm.model_executor.sampling_metadata import SamplingMetadata
|
||||
from vllm.sequence import SamplerOutput
|
||||
from vllm.sequence import IntermediateTensors
|
||||
|
||||
from .interfaces import SupportsLoRA, SupportsPP
|
||||
from .utils import (
|
||||
AutoWeightsLoader,
|
||||
WeightsMapper,
|
||||
is_pp_missing_parameter,
|
||||
make_empty_intermediate_tensors_factory,
|
||||
make_layers,
|
||||
maybe_prefix,
|
||||
)
|
||||
|
||||
|
||||
class OPTLearnedPositionalEmbedding(nn.Embedding):
|
||||
|
||||
def __init__(self, num_embeddings: int, embedding_dim: int):
|
||||
# OPT is set up so that if padding_idx is specified then offset the
|
||||
# embedding ids by 2 and adjust num_embeddings appropriately. Other
|
||||
@@ -55,18 +71,18 @@ class OPTLearnedPositionalEmbedding(nn.Embedding):
|
||||
|
||||
|
||||
class OPTAttention(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
embed_dim: int,
|
||||
num_heads: int,
|
||||
bias: bool = True,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.embed_dim = embed_dim
|
||||
tensor_model_parallel_world_size = (
|
||||
get_tensor_model_parallel_world_size())
|
||||
tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
|
||||
total_num_heads = num_heads
|
||||
assert num_heads % tensor_model_parallel_world_size == 0
|
||||
self.num_heads = total_num_heads // tensor_model_parallel_world_size
|
||||
@@ -79,36 +95,42 @@ class OPTAttention(nn.Module):
|
||||
total_num_heads,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.qkv_proj",
|
||||
)
|
||||
self.out_proj = RowParallelLinear(
|
||||
embed_dim,
|
||||
embed_dim,
|
||||
bias=bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.out_proj",
|
||||
)
|
||||
self.attn = Attention(
|
||||
self.num_heads,
|
||||
self.head_dim,
|
||||
scale=self.scaling,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.attn",
|
||||
)
|
||||
self.attn = Attention(self.num_heads,
|
||||
self.head_dim,
|
||||
scale=self.scaling)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
kv_cache: torch.Tensor,
|
||||
attn_metadata: AttentionMetadata,
|
||||
) -> torch.Tensor:
|
||||
qkv, _ = self.qkv_proj(hidden_states)
|
||||
q, k, v = qkv.chunk(chunks=3, dim=-1)
|
||||
attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
|
||||
attn_output = self.attn(q, k, v)
|
||||
output, _ = self.out_proj(attn_output)
|
||||
return output
|
||||
|
||||
|
||||
class OPTDecoderLayer(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: OPTConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
@@ -117,45 +139,44 @@ class OPTDecoderLayer(nn.Module):
|
||||
embed_dim=self.embed_dim,
|
||||
num_heads=config.num_attention_heads,
|
||||
bias=config.enable_bias,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.self_attn",
|
||||
)
|
||||
self.do_layer_norm_before = config.do_layer_norm_before
|
||||
|
||||
self.self_attn_layer_norm = nn.LayerNorm(
|
||||
self.embed_dim,
|
||||
elementwise_affine=config.layer_norm_elementwise_affine)
|
||||
self.embed_dim, elementwise_affine=config.layer_norm_elementwise_affine
|
||||
)
|
||||
self.fc1 = ColumnParallelLinear(
|
||||
self.embed_dim,
|
||||
config.ffn_dim,
|
||||
bias=config.enable_bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.fc1",
|
||||
)
|
||||
self.activation_fn = get_act_fn(config.activation_function,
|
||||
quant_config, config.ffn_dim)
|
||||
self.activation_fn = get_act_fn(config.activation_function)
|
||||
self.fc2 = RowParallelLinear(
|
||||
config.ffn_dim,
|
||||
self.embed_dim,
|
||||
bias=config.enable_bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.fc2",
|
||||
)
|
||||
self.final_layer_norm = nn.LayerNorm(
|
||||
self.embed_dim,
|
||||
elementwise_affine=config.layer_norm_elementwise_affine)
|
||||
self.embed_dim, elementwise_affine=config.layer_norm_elementwise_affine
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
kv_cache: torch.Tensor,
|
||||
attn_metadata: AttentionMetadata,
|
||||
) -> torch.Tensor:
|
||||
# Self Attention
|
||||
residual = hidden_states
|
||||
# 125m, 1.7B, ..., 175B applies layer norm BEFORE attention
|
||||
if self.do_layer_norm_before:
|
||||
hidden_states = self.self_attn_layer_norm(hidden_states)
|
||||
hidden_states = self.self_attn(hidden_states=hidden_states,
|
||||
kv_cache=kv_cache,
|
||||
attn_metadata=attn_metadata)
|
||||
hidden_states = self.self_attn(hidden_states=hidden_states)
|
||||
hidden_states = residual + hidden_states
|
||||
# 350m applies layer norm AFTER attention
|
||||
if not self.do_layer_norm_before:
|
||||
@@ -177,15 +198,15 @@ class OPTDecoderLayer(nn.Module):
|
||||
|
||||
|
||||
class OPTDecoder(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: OPTConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
cache_config: CacheConfig | None = None,
|
||||
quant_config: QuantizationConfig | None = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.padding_idx = config.pad_token_id
|
||||
self.max_target_positions = config.max_position_embeddings
|
||||
self.vocab_size = config.vocab_size
|
||||
|
||||
@@ -195,22 +216,29 @@ class OPTDecoder(nn.Module):
|
||||
)
|
||||
# Positional embeddings are replicated (not sharded).
|
||||
self.embed_positions = OPTLearnedPositionalEmbedding(
|
||||
config.max_position_embeddings, config.hidden_size)
|
||||
config.max_position_embeddings, config.hidden_size
|
||||
)
|
||||
|
||||
# Project out & in will be replicated if they exist.
|
||||
if config.word_embed_proj_dim != config.hidden_size:
|
||||
self.project_out = ReplicatedLinear(config.hidden_size,
|
||||
config.word_embed_proj_dim,
|
||||
bias=False,
|
||||
quant_config=quant_config)
|
||||
self.project_out = ReplicatedLinear(
|
||||
config.hidden_size,
|
||||
config.word_embed_proj_dim,
|
||||
bias=False,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.project_out",
|
||||
)
|
||||
else:
|
||||
self.project_out = None
|
||||
|
||||
if config.word_embed_proj_dim != config.hidden_size:
|
||||
self.project_in = ReplicatedLinear(config.word_embed_proj_dim,
|
||||
config.hidden_size,
|
||||
bias=False,
|
||||
quant_config=quant_config)
|
||||
self.project_in = ReplicatedLinear(
|
||||
config.word_embed_proj_dim,
|
||||
config.hidden_size,
|
||||
bias=False,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.project_in",
|
||||
)
|
||||
else:
|
||||
self.project_in = None
|
||||
|
||||
@@ -221,32 +249,45 @@ class OPTDecoder(nn.Module):
|
||||
if config.do_layer_norm_before and not config._remove_final_layer_norm:
|
||||
self.final_layer_norm = nn.LayerNorm(
|
||||
config.hidden_size,
|
||||
elementwise_affine=config.layer_norm_elementwise_affine)
|
||||
elementwise_affine=config.layer_norm_elementwise_affine,
|
||||
)
|
||||
else:
|
||||
self.final_layer_norm = None
|
||||
|
||||
self.layers = nn.ModuleList([
|
||||
OPTDecoderLayer(config, quant_config)
|
||||
for _ in range(config.num_hidden_layers)
|
||||
])
|
||||
self.start_layer, self.end_layer, self.layers = make_layers(
|
||||
config.num_hidden_layers,
|
||||
lambda prefix: OPTDecoderLayer(
|
||||
config, cache_config, quant_config, prefix=prefix
|
||||
),
|
||||
prefix=f"{prefix}.layers",
|
||||
)
|
||||
|
||||
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.embed_tokens(input_ids)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
kv_caches: List[torch.Tensor],
|
||||
attn_metadata: AttentionMetadata,
|
||||
) -> torch.Tensor:
|
||||
inputs_embeds = self.embed_tokens(input_ids)
|
||||
pos_embeds = self.embed_positions(positions)
|
||||
if self.project_in is not None:
|
||||
inputs_embeds, _ = self.project_in(inputs_embeds)
|
||||
hidden_states = inputs_embeds + pos_embeds
|
||||
intermediate_tensors: IntermediateTensors | None,
|
||||
inputs_embeds: torch.Tensor | None = None,
|
||||
) -> torch.Tensor | IntermediateTensors:
|
||||
if get_pp_group().is_first_rank:
|
||||
if inputs_embeds is None:
|
||||
inputs_embeds = self.embed_input_ids(input_ids)
|
||||
pos_embeds = self.embed_positions(positions)
|
||||
if self.project_in is not None:
|
||||
inputs_embeds, _ = self.project_in(inputs_embeds)
|
||||
hidden_states = inputs_embeds + pos_embeds
|
||||
else:
|
||||
assert intermediate_tensors is not None
|
||||
hidden_states = intermediate_tensors["hidden_states"]
|
||||
|
||||
for i in range(len(self.layers)):
|
||||
layer = self.layers[i]
|
||||
hidden_states = layer(hidden_states, kv_caches[i], attn_metadata)
|
||||
for layer in islice(self.layers, self.start_layer, self.end_layer):
|
||||
hidden_states = layer(hidden_states)
|
||||
|
||||
if not get_pp_group().is_last_rank:
|
||||
return IntermediateTensors({"hidden_states": hidden_states})
|
||||
if self.final_layer_norm is not None:
|
||||
hidden_states = self.final_layer_norm(hidden_states)
|
||||
if self.project_out is not None:
|
||||
@@ -254,67 +295,37 @@ class OPTDecoder(nn.Module):
|
||||
return hidden_states
|
||||
|
||||
|
||||
@support_torch_compile
|
||||
class OPTModel(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: OPTConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
):
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
self.decoder = OPTDecoder(config, quant_config)
|
||||
|
||||
config = vllm_config.model_config.hf_config
|
||||
cache_config = vllm_config.cache_config
|
||||
quant_config = vllm_config.quant_config
|
||||
|
||||
self.decoder = OPTDecoder(
|
||||
config, cache_config, quant_config, prefix=f"{prefix}.decoder"
|
||||
)
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states"], config.hidden_size
|
||||
)
|
||||
|
||||
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.decoder.embed_input_ids(input_ids)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
kv_caches: List[torch.Tensor],
|
||||
attn_metadata: AttentionMetadata,
|
||||
) -> torch.Tensor:
|
||||
return self.decoder(input_ids, positions, kv_caches, attn_metadata)
|
||||
intermediate_tensors: IntermediateTensors | None,
|
||||
inputs_embeds: torch.Tensor | None = None,
|
||||
) -> torch.Tensor | IntermediateTensors:
|
||||
return self.decoder(
|
||||
input_ids, positions, intermediate_tensors, inputs_embeds=inputs_embeds
|
||||
)
|
||||
|
||||
|
||||
class OPTForCausalLM(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.quant_config = quant_config
|
||||
self.model = OPTModel(config, quant_config)
|
||||
self.lm_head_weight = self.model.decoder.embed_tokens.weight
|
||||
self.logits_processor = LogitsProcessor(config.vocab_size)
|
||||
self.sampler = Sampler()
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
kv_caches: List[torch.Tensor],
|
||||
attn_metadata: AttentionMetadata,
|
||||
) -> torch.Tensor:
|
||||
hidden_states = self.model(input_ids, positions, kv_caches,
|
||||
attn_metadata)
|
||||
return hidden_states
|
||||
|
||||
def compute_logits(self, hidden_states: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata) -> torch.Tensor:
|
||||
logits = self.logits_processor(self.lm_head_weight, hidden_states,
|
||||
sampling_metadata)
|
||||
return logits
|
||||
|
||||
def sample(
|
||||
self,
|
||||
logits: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata,
|
||||
) -> Optional[SamplerOutput]:
|
||||
next_tokens = self.sampler(logits, sampling_metadata)
|
||||
return next_tokens
|
||||
|
||||
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
|
||||
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"),
|
||||
@@ -322,19 +333,17 @@ class OPTForCausalLM(nn.Module):
|
||||
("qkv_proj", "v_proj", "v"),
|
||||
]
|
||||
params_dict = dict(self.named_parameters(remove_duplicate=False))
|
||||
loaded_params: set[str] = set()
|
||||
for name, loaded_weight in weights:
|
||||
if "lm_head.weight" in name:
|
||||
continue
|
||||
if name.startswith("decoder."):
|
||||
name = "model." + name
|
||||
|
||||
for (param_name, weight_name, shard_id) in stacked_params_mapping:
|
||||
for param_name, weight_name, shard_id in stacked_params_mapping:
|
||||
if weight_name not in name:
|
||||
continue
|
||||
name = name.replace(weight_name, param_name)
|
||||
# Skip loading extra bias for GPTQ models.
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
if is_pp_missing_parameter(name, self):
|
||||
continue
|
||||
param = params_dict[name]
|
||||
weight_loader = param.weight_loader
|
||||
weight_loader(param, loaded_weight, shard_id)
|
||||
@@ -343,7 +352,75 @@ class OPTForCausalLM(nn.Module):
|
||||
# Skip loading extra bias for GPTQ models.
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
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)
|
||||
return loaded_params
|
||||
|
||||
|
||||
class OPTForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
|
||||
packed_modules_mapping = {
|
||||
"qkv_proj": ["q_proj", "k_proj", "v_proj"],
|
||||
}
|
||||
|
||||
hf_to_vllm_mapper = WeightsMapper(
|
||||
orig_to_new_prefix={
|
||||
"decoder.": "model.decoder.",
|
||||
}
|
||||
)
|
||||
|
||||
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.model = OPTModel(
|
||||
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
|
||||
)
|
||||
if self.config.tie_word_embeddings:
|
||||
self.lm_head = self.model.decoder.embed_tokens
|
||||
else:
|
||||
self.lm_head = ParallelLMHead(
|
||||
config.vocab_size,
|
||||
config.word_embed_proj_dim,
|
||||
prefix=maybe_prefix(prefix, "lm_head"),
|
||||
)
|
||||
self.logits_processor = LogitsProcessor(config.vocab_size)
|
||||
self.make_empty_intermediate_tensors = (
|
||||
self.model.make_empty_intermediate_tensors
|
||||
)
|
||||
|
||||
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.model.embed_input_ids(input_ids)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
intermediate_tensors: IntermediateTensors | None = None,
|
||||
inputs_embeds: torch.Tensor | None = None,
|
||||
) -> torch.Tensor | IntermediateTensors:
|
||||
hidden_states = self.model(
|
||||
input_ids, positions, intermediate_tensors, inputs_embeds
|
||||
)
|
||||
return hidden_states
|
||||
|
||||
def compute_logits(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
) -> torch.Tensor | None:
|
||||
logits = self.logits_processor(self.lm_head, hidden_states)
|
||||
return logits
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
loader = AutoWeightsLoader(
|
||||
self,
|
||||
skip_prefixes=(
|
||||
["lm_head.weight"] if self.config.tie_word_embeddings else None
|
||||
),
|
||||
)
|
||||
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
|
||||
|
||||
Reference in New Issue
Block a user