364 lines
13 KiB
Python
364 lines
13 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
# Copyright 2024 The Qwen team.
|
|
# Copyright 2023 The vLLM team.
|
|
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
|
|
#
|
|
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
|
|
# and OPT implementations in this library. It has been modified from its
|
|
# original forms to accommodate minor architectural differences compared
|
|
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# 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.
|
|
"""Inference-only Qwen3 model compatible with HuggingFace weights."""
|
|
|
|
from collections.abc import Iterable
|
|
from typing import Any, Optional, Tuple
|
|
|
|
import torch
|
|
from torch import nn
|
|
from transformers import Qwen2Config as Qwen3Config
|
|
|
|
from vllm.attention import Attention, AttentionMetadata
|
|
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.logger import init_logger
|
|
from vllm.model_executor.layers.layernorm import RMSNorm
|
|
from vllm.model_executor.layers.linear import QKVParallelLinear, 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 ParallelLMHead
|
|
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
|
from vllm.sequence import IntermediateTensors
|
|
|
|
from .interfaces import SupportsLoRA, SupportsPP
|
|
from .qwen2 import Qwen2MLP as Qwen3MLP
|
|
from .qwen2 import Qwen2Model
|
|
from .utils import AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter, maybe_prefix
|
|
|
|
logger = init_logger(__name__)
|
|
|
|
|
|
class Qwen3Attention(nn.Module):
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
num_heads: int,
|
|
num_kv_heads: int,
|
|
rope_theta: float = 10000,
|
|
rope_scaling: Optional[dict] = None,
|
|
max_position_embeddings: int = 8192,
|
|
head_dim: Optional[int] = None,
|
|
rms_norm_eps: float = 1e-06,
|
|
qkv_bias: bool = False,
|
|
cache_config: Optional[CacheConfig] = None,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.hidden_size = hidden_size
|
|
tp_size = get_tensor_model_parallel_world_size()
|
|
self.total_num_heads = num_heads
|
|
assert self.total_num_heads % tp_size == 0
|
|
self.num_heads = self.total_num_heads // tp_size
|
|
self.total_num_kv_heads = num_kv_heads
|
|
if self.total_num_kv_heads >= tp_size:
|
|
# Number of KV heads is greater than TP size, so we partition
|
|
# the KV heads across multiple tensor parallel GPUs.
|
|
assert self.total_num_kv_heads % tp_size == 0
|
|
else:
|
|
# Number of KV heads is less than TP size, so we replicate
|
|
# the KV heads across multiple tensor parallel GPUs.
|
|
assert tp_size % self.total_num_kv_heads == 0
|
|
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
|
|
self.head_dim = head_dim or hidden_size // self.total_num_heads
|
|
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.qkv_proj = QKVParallelLinear(
|
|
hidden_size,
|
|
self.head_dim,
|
|
self.total_num_heads,
|
|
self.total_num_kv_heads,
|
|
bias=qkv_bias,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.qkv_proj",
|
|
)
|
|
self.o_proj = RowParallelLinear(
|
|
self.total_num_heads * self.head_dim,
|
|
hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.o_proj",
|
|
)
|
|
|
|
self.rotary_emb = get_rope(
|
|
self.head_dim,
|
|
rotary_dim=self.head_dim,
|
|
max_position=max_position_embeddings,
|
|
base=rope_theta,
|
|
rope_scaling=rope_scaling,
|
|
)
|
|
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",
|
|
)
|
|
# Qwen3 specific: QK normalization
|
|
self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
|
|
self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
|
|
|
|
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)
|
|
|
|
# Qwen3 specific: Apply QK normalization
|
|
q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim, self.head_dim)
|
|
q_by_head = self.q_norm(q_by_head)
|
|
q = q_by_head.view(q.shape)
|
|
|
|
k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim, self.head_dim)
|
|
k_by_head = self.k_norm(k_by_head)
|
|
k = k_by_head.view(k.shape)
|
|
|
|
q, k = self.rotary_emb(positions, q, k)
|
|
attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
|
|
output, _ = self.o_proj(attn_output)
|
|
return output
|
|
|
|
|
|
class Qwen3DecoderLayer(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: Qwen3Config,
|
|
cache_config: Optional[CacheConfig] = None,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.hidden_size = config.hidden_size
|
|
|
|
# Set default rope_theta if not present
|
|
rope_theta = getattr(config, "rope_theta", 10000)
|
|
rope_scaling = getattr(config, "rope_scaling", None)
|
|
|
|
self.self_attn = Qwen3Attention(
|
|
hidden_size=self.hidden_size,
|
|
num_heads=config.num_attention_heads,
|
|
max_position_embeddings=config.max_position_embeddings,
|
|
num_kv_heads=config.num_key_value_heads,
|
|
rms_norm_eps=config.rms_norm_eps,
|
|
qkv_bias=getattr(config, "attention_bias", False),
|
|
head_dim=getattr(config, "head_dim", None),
|
|
cache_config=cache_config,
|
|
quant_config=quant_config,
|
|
rope_theta=rope_theta,
|
|
rope_scaling=rope_scaling,
|
|
prefix=f"{prefix}.self_attn",
|
|
)
|
|
self.mlp = Qwen3MLP(
|
|
hidden_size=self.hidden_size,
|
|
intermediate_size=config.intermediate_size,
|
|
hidden_act=config.hidden_act,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.mlp",
|
|
)
|
|
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
self.post_attention_layernorm = RMSNorm(
|
|
config.hidden_size, eps=config.rms_norm_eps
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
positions: torch.Tensor,
|
|
hidden_states: torch.Tensor,
|
|
kv_cache: torch.Tensor,
|
|
attn_metadata: AttentionMetadata,
|
|
residual: Optional[torch.Tensor],
|
|
) -> Tuple[torch.Tensor, torch.Tensor]:
|
|
# Self Attention
|
|
if residual is None:
|
|
residual = hidden_states
|
|
hidden_states = self.input_layernorm(hidden_states)
|
|
else:
|
|
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
|
hidden_states = self.self_attn(
|
|
positions=positions,
|
|
hidden_states=hidden_states,
|
|
kv_cache=kv_cache,
|
|
attn_metadata=attn_metadata,
|
|
)
|
|
|
|
# Fully Connected
|
|
hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
|
|
hidden_states = self.mlp(hidden_states)
|
|
return hidden_states, residual
|
|
|
|
|
|
@support_torch_compile(
|
|
dynamic_arg_dims={
|
|
"input_ids": 0,
|
|
"positions": 0,
|
|
"intermediate_tensors": 0,
|
|
"inputs_embeds": 0,
|
|
}
|
|
)
|
|
class Qwen3Model(Qwen2Model):
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
|
# Initialize with Qwen3DecoderLayer
|
|
config = vllm_config.model_config.hf_config
|
|
cache_config = vllm_config.cache_config
|
|
quant_config = vllm_config.quant_config
|
|
|
|
# Call nn.Module.__init__ directly to avoid Qwen2Model's __init__
|
|
nn.Module.__init__(self)
|
|
self.config = config
|
|
self.padding_idx = config.pad_token_id
|
|
self.vocab_size = config.vocab_size
|
|
|
|
from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
|
|
if get_pp_group().is_first_rank or (config.tie_word_embeddings
|
|
and get_pp_group().is_last_rank):
|
|
self.embed_tokens = VocabParallelEmbedding(
|
|
config.vocab_size,
|
|
config.hidden_size,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.embed_tokens",
|
|
)
|
|
else:
|
|
self.embed_tokens = PPMissingLayer()
|
|
|
|
# Use Qwen3DecoderLayer instead of Qwen2DecoderLayer
|
|
from .utils import make_layers
|
|
self.start_layer, self.end_layer, self.layers = make_layers(
|
|
config.num_hidden_layers,
|
|
lambda prefix: Qwen3DecoderLayer(
|
|
config=config,
|
|
cache_config=cache_config,
|
|
quant_config=quant_config,
|
|
prefix=prefix,
|
|
),
|
|
prefix=f"{prefix}.layers",
|
|
)
|
|
if get_pp_group().is_last_rank:
|
|
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
else:
|
|
self.norm = PPMissingLayer()
|
|
|
|
from .utils import make_empty_intermediate_tensors_factory
|
|
self.make_empty_intermediate_tensors = (
|
|
make_empty_intermediate_tensors_factory(
|
|
["hidden_states", "residual"], config.hidden_size
|
|
)
|
|
)
|
|
|
|
|
|
class Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
|
packed_modules_mapping = {
|
|
"qkv_proj": [
|
|
"q_proj",
|
|
"k_proj",
|
|
"v_proj",
|
|
],
|
|
"gate_up_proj": [
|
|
"gate_proj",
|
|
"up_proj",
|
|
],
|
|
}
|
|
|
|
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 = Qwen3Model(
|
|
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
|
|
)
|
|
|
|
if get_pp_group().is_last_rank:
|
|
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=maybe_prefix(prefix, "lm_head"),
|
|
)
|
|
else:
|
|
self.lm_head = PPMissingLayer()
|
|
|
|
self.logits_processor = LogitsProcessor(config.vocab_size)
|
|
|
|
from .utils import make_empty_intermediate_tensors_factory
|
|
self.make_empty_intermediate_tensors = (
|
|
make_empty_intermediate_tensors_factory(
|
|
["hidden_states", "residual"], config.hidden_size
|
|
)
|
|
)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
kv_caches: list,
|
|
attn_metadata: AttentionMetadata,
|
|
intermediate_tensors: Optional[IntermediateTensors] = None,
|
|
inputs_embeds: Optional[torch.Tensor] = None,
|
|
) -> torch.Tensor:
|
|
hidden_states = self.model(
|
|
input_ids, positions, kv_caches, attn_metadata, intermediate_tensors, inputs_embeds
|
|
)
|
|
return hidden_states
|
|
|
|
def compute_logits(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
sampling_metadata,
|
|
) -> torch.Tensor:
|
|
logits = self.logits_processor(self.lm_head, hidden_states, sampling_metadata)
|
|
return logits
|
|
|
|
def sample(
|
|
self,
|
|
logits: torch.Tensor,
|
|
sampling_metadata,
|
|
):
|
|
from vllm.model_executor.layers.sampler import get_sampler
|
|
sampler = get_sampler()
|
|
next_tokens = sampler(logits, sampling_metadata)
|
|
return next_tokens
|
|
|
|
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]) -> set:
|
|
loader = AutoWeightsLoader(
|
|
self,
|
|
skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
|
|
)
|
|
return loader.load_weights(weights)
|