EngineX-Mthreads/enginex-mthreads-vllm
4
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Sync from v0.13

Browse Source
This commit is contained in:
xiezhongtao
2026-01-19 10:38:50 +08:00
parent b2ef04d792
commit 5aef6c175a
3714 changed files with 854317 additions and 89342 deletions
Download Patch File Download Diff File Expand all files Collapse all files

315
vllm/model_executor/models/starcoder2.py
View File

@@ -1,4 +1,6 @@
# coding=utf-8
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# Copyright 2024 BigCode and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
@@ -17,36 +19,56 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch Starcoder2 model."""
from typing import Iterable, List, Optional, Tuple
"""PyTorch Starcoder2 model."""
from collections.abc import Iterable
from itertools import islice
import torch
from torch import nn
from transformers import Starcoder2Config
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,
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.base_config import (
QuantizationConfig)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE, 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
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import (
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,
)
class Starcoder2Attention(nn.Module):
def __init__(self,
config: Starcoder2Config,
quant_config: Optional[QuantizationConfig] = None):
def __init__(
self,
config: Starcoder2Config,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
):
super().__init__()
self.config = config
@@ -69,10 +91,8 @@ class Starcoder2Attention(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.rope_theta = config.rope_theta
self.max_position_embeddings = config.max_position_embeddings
self.use_bias = config.use_bias
self.sliding_window = config.sliding_window
self.qkv_proj = QKVParallelLinear(
self.hidden_size,
@@ -81,18 +101,19 @@ class Starcoder2Attention(nn.Module):
self.total_num_kv_heads,
bias=self.use_bias,
quant_config=quant_config,
prefix=f"{prefix}.qkv_proj",
)
self.o_proj = RowParallelLinear(
self.total_num_heads * self.head_dim,
self.hidden_size,
bias=self.use_bias,
quant_config=quant_config,
prefix=f"{prefix}.o_proj",
)
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=self.max_position_embeddings,
base=int(self.rope_theta),
rope_parameters=config.rope_parameters,
is_neox_style=True,
)
self.attn = Attention(
@@ -100,44 +121,47 @@ class Starcoder2Attention(nn.Module):
self.head_dim,
self.scaling,
num_kv_heads=self.num_kv_heads,
sliding_window=self.sliding_window,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn",
)
def forward(
self,
positions: torch.Tensor,
hidden_states: torch.Tensor,
kv_cache: torch.Tensor,
attn_metadata: AttentionMetadata,
) -> torch.Tensor:
qkv, _ = self.qkv_proj(hidden_states)
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
q, k = self.rotary_emb(positions, q, k)
attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
attn_output = self.attn(q, k, v)
output, _ = self.o_proj(attn_output)
return output
class Starcoder2MLP(nn.Module):
def __init__(self,
config: Starcoder2Config,
quant_config: Optional[QuantizationConfig] = None):
def __init__(
self,
config: Starcoder2Config,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
):
super().__init__()
self.c_fc = ColumnParallelLinear(
config.hidden_size,
config.intermediate_size,
bias=config.use_bias,
quant_config=quant_config,
prefix=f"{prefix}.c_fc",
)
self.c_proj = RowParallelLinear(
config.intermediate_size,
config.hidden_size,
bias=config.use_bias,
quant_config=quant_config,
prefix=f"{prefix}.c_proj",
)
self.act = get_act_fn(config.hidden_act, quant_config,
config.intermediate_size)
self.act = get_act_fn(config.hidden_act)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
hidden_states, _ = self.c_fc(hidden_states)
@@ -147,25 +171,33 @@ class Starcoder2MLP(nn.Module):
class Starcoder2DecoderLayer(nn.Module):
def __init__(self,
config: Starcoder2Config,
quant_config: Optional[QuantizationConfig] = None):
def __init__(
self,
config: Starcoder2Config,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
):
super().__init__()
self.hidden_size = config.hidden_size
self.self_attn = Starcoder2Attention(config, quant_config=quant_config)
self.mlp = Starcoder2MLP(config, quant_config=quant_config)
self.input_layernorm = nn.LayerNorm(config.hidden_size,
eps=config.norm_epsilon)
self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
eps=config.norm_epsilon)
self.self_attn = Starcoder2Attention(
config,
cache_config,
quant_config=quant_config,
prefix=f"{prefix}.self_attn",
)
self.mlp = Starcoder2MLP(
config, quant_config=quant_config, prefix=f"{prefix}.mlp"
)
self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.norm_epsilon)
self.post_attention_layernorm = nn.LayerNorm(
config.hidden_size, eps=config.norm_epsilon
)
def forward(
self,
positions: torch.Tensor,
hidden_states: torch.Tensor,
kv_cache: torch.Tensor,
attn_metadata: AttentionMetadata,
) -> torch.Tensor:
# Self Attention
residual = hidden_states
@@ -173,8 +205,6 @@ class Starcoder2DecoderLayer(nn.Module):
hidden_states = self.self_attn(
positions=positions,
hidden_states=hidden_states,
kv_cache=kv_cache,
attn_metadata=attn_metadata,
)
hidden_states = residual + hidden_states
@@ -187,92 +217,62 @@ class Starcoder2DecoderLayer(nn.Module):
return hidden_states
@support_torch_compile
class Starcoder2Model(nn.Module):
def __init__(self,
config: Starcoder2Config,
quant_config: Optional[QuantizationConfig] = None):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
self.config = config
self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size
# TODO: consider padding_idx (currently removed)
self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
config.hidden_size)
self.layers = nn.ModuleList([
Starcoder2DecoderLayer(config, quant_config=quant_config)
for _ in range(config.num_hidden_layers)
])
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=f"{prefix}.embed_tokens",
)
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: Starcoder2DecoderLayer(
config, cache_config, quant_config=quant_config, prefix=prefix
),
prefix=f"{prefix}.layers",
)
self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_epsilon)
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.embed_tokens(input_ids)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
kv_caches: List[torch.Tensor],
attn_metadata: AttentionMetadata,
) -> torch.Tensor:
hidden_states = self.embed_tokens(input_ids)
for i in range(len(self.layers)):
layer = self.layers[i]
hidden_states = layer(positions, hidden_states, kv_caches[i],
attn_metadata)
intermediate_tensors: IntermediateTensors | None,
inputs_embeds: torch.Tensor | None = None,
) -> torch.Tensor | IntermediateTensors:
if get_pp_group().is_first_rank:
if inputs_embeds is not None:
hidden_states = inputs_embeds
else:
hidden_states = self.embed_input_ids(input_ids)
else:
assert intermediate_tensors is not None
hidden_states = intermediate_tensors["hidden_states"]
for layer in islice(self.layers, self.start_layer, self.end_layer):
hidden_states = layer(positions, hidden_states)
if not get_pp_group().is_last_rank:
return IntermediateTensors({"hidden_states": hidden_states})
hidden_states = self.norm(hidden_states)
return hidden_states
class Starcoder2ForCausalLM(nn.Module):
def __init__(self,
config: Starcoder2Config,
quant_config: Optional[QuantizationConfig] = None):
super().__init__()
self.config = config
self.model = Starcoder2Model(config, quant_config=quant_config)
self.vocab_size = config.vocab_size
self.unpadded_vocab_size = config.vocab_size
if config.tie_word_embeddings:
self.lm_head_weight = self.model.embed_tokens.weight
else:
self.unpadded_vocab_size = config.vocab_size
self.lm_head = ParallelLMHead(
self.unpadded_vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
padding_size=DEFAULT_VOCAB_PADDING_SIZE,
)
self.lm_head_weight = self.lm_head.weight
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
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: Optional[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"),
@@ -281,22 +281,85 @@ class Starcoder2ForCausalLM(nn.Module):
]
params_dict = dict(self.named_parameters(remove_duplicate=False))
loaded_params: set[str] = set()
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
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)
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
weight_loader = param.weight_loader
weight_loader(param, loaded_weight, shard_id)
break
else:
if self.config.tie_word_embeddings and "lm_head.weight" in name:
name = maybe_remap_kv_scale_name(name, params_dict)
if name is None:
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 Starcoder2ForCausalLM(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.model = Starcoder2Model(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
self.vocab_size = config.vocab_size
if config.tie_word_embeddings:
self.lm_head = self.model.embed_tokens
else:
self.lm_head = ParallelLMHead(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=f"{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,
# Models trained using ColossalAI may include these tensors in
# the checkpoint. Skip them.
skip_prefixes=(
["lm_head.weight"] if self.config.tie_word_embeddings else None
),
)
return loader.load_weights(weights)