[Feature] Merge branch 'Qwen3-Next' into main && Support Qwen-next (#222)
Signed-off-by: xyDong0223 <dongxinyu03@baidu.com> Co-authored-by: xyDong0223 <dongxinyu03@baidu.com>
This commit is contained in:
chanzhennan
@@ -3,95 +3,113 @@ from vllm import ModelRegistry
|
||||
|
||||
def register_model():
|
||||
# from .demo_model import DemoModel # noqa: F401
|
||||
from .qwen2_vl import Qwen2VLForConditionalGeneration #noqa: F401
|
||||
from .qwen2_5_vl import Qwen2_5_VLForConditionalGeneration #noqa: F401
|
||||
from .qwen3_moe import Qwen3MoeForCausalLM #noqa: F401
|
||||
from .qwen3_vl import Qwen3VLForConditionalGeneration
|
||||
from .qwen3_vl_moe import Qwen3VLMoeForConditionalGeneration
|
||||
from .qwen3_omni_moe_thinker import Qwen3OmniMoeThinkerForConditionalGeneration
|
||||
from .qwen2_5_vl import Qwen2_5_VLForConditionalGeneration # noqa: F401
|
||||
from .qwen2_vl import Qwen2VLForConditionalGeneration # noqa: F401
|
||||
from .qwen3_moe import Qwen3MoeForCausalLM # noqa: F401
|
||||
from .qwen3_omni_moe_thinker import ( # noqa: F401
|
||||
Qwen3OmniMoeThinkerForConditionalGeneration,
|
||||
)
|
||||
from .qwen3_vl import Qwen3VLForConditionalGeneration # noqa: F401
|
||||
from .qwen3_vl_moe import Qwen3VLMoeForConditionalGeneration # noqa: F401
|
||||
|
||||
# from .llama4 import Llama4ForCausalLM #noqa: F401
|
||||
# from .mllama4 import Llama4ForConditionalGeneration #noqa: F401
|
||||
# from .deepseek_v2 import KunlunDeepseekV2MoE
|
||||
|
||||
# ModelRegistry.register_model(
|
||||
# "DemoModel",
|
||||
# "vllm_kunlun.model_executor.models.demo_model:DemoModel")
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"Qwen2VLForConditionalGeneration",
|
||||
"vllm_kunlun.models.qwen2_vl:Qwen2VLForConditionalGeneration")
|
||||
"vllm_kunlun.models.qwen2_vl:Qwen2VLForConditionalGeneration",
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"Qwen2_5_VLForConditionalGeneration",
|
||||
"vllm_kunlun.models.qwen2_5_vl:Qwen2_5_VLForConditionalGeneration")
|
||||
"vllm_kunlun.models.qwen2_5_vl:Qwen2_5_VLForConditionalGeneration",
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"Qwen3ForCausalLM",
|
||||
"vllm_kunlun.models.qwen3:Qwen3ForCausalLM")
|
||||
"Qwen3ForCausalLM", "vllm_kunlun.models.qwen3:Qwen3ForCausalLM"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"Qwen3MoeForCausalLM",
|
||||
"vllm_kunlun.models.qwen3_moe:Qwen3MoeForCausalLM")
|
||||
"Qwen3MoeForCausalLM", "vllm_kunlun.models.qwen3_moe:Qwen3MoeForCausalLM"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"Qwen3NextForCausalLM",
|
||||
"vllm_kunlun.models.qwen3_next:Qwen3NextForCausalLM")
|
||||
"Qwen3NextForCausalLM", "vllm_kunlun.models.qwen3_next:Qwen3NextForCausalLM"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"GptOssForCausalLM",
|
||||
"vllm_kunlun.models.gpt_oss:GptOssForCausalLM")
|
||||
"Qwen3NextMTP", "vllm_kunlun.models.qwen3_next_mtp:Qwen3NextMTP"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"InternLM2ForCausalLM",
|
||||
"vllm_kunlun.models.internlm2:InternLM2ForCausalLM")
|
||||
|
||||
"GlmForCausalLM", "vllm_kunlun.models.glm:GlmForCausalLM"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"InternVLChatModel",
|
||||
"vllm_kunlun.models.internvl:InternVLChatModel")
|
||||
"GptOssForCausalLM", "vllm_kunlun.models.gpt_oss:GptOssForCausalLM"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"InternLM2ForCausalLM", "vllm_kunlun.models.internlm2:InternLM2ForCausalLM"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"InternVLChatModel", "vllm_kunlun.models.internvl:InternVLChatModel"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"InternS1ForConditionalGeneration",
|
||||
"vllm_kunlun.models.interns1:InternS1ForConditionalGeneration")
|
||||
|
||||
"vllm_kunlun.models.interns1:InternS1ForConditionalGeneration",
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"Qwen3VLForConditionalGeneration",
|
||||
"vllm_kunlun.models.qwen3_vl:Qwen3VLForConditionalGeneration")
|
||||
|
||||
"vllm_kunlun.models.qwen3_vl:Qwen3VLForConditionalGeneration",
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"Qwen3VLMoeForConditionalGeneration",
|
||||
"vllm_kunlun.models.qwen3_vl_moe:Qwen3VLMoeForConditionalGeneration")
|
||||
"vllm_kunlun.models.qwen3_vl_moe:Qwen3VLMoeForConditionalGeneration",
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"Qwen3OmniMoeForConditionalGeneration",
|
||||
"vllm_kunlun.models.qwen3_omni_moe_thinker:Qwen3OmniMoeThinkerForConditionalGeneration")
|
||||
"vllm_kunlun.models.qwen3_omni_moe_thinker:Qwen3OmniMoeThinkerForConditionalGeneration",
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"SeedOssForCausalLM",
|
||||
"vllm_kunlun.models.seed_oss:SeedOssForCausalLM")
|
||||
"SeedOssForCausalLM", "vllm_kunlun.models.seed_oss:SeedOssForCausalLM"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"MiMoV2FlashForCausalLM",
|
||||
"vllm_kunlun.models.mimo_v2_flash:MiMoV2FlashForCausalLM")
|
||||
"vllm_kunlun.models.mimo_v2_flash:MiMoV2FlashForCausalLM",
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"GptOssForCausalLM",
|
||||
"vllm_kunlun.models.gpt_oss:GptOssForCausalLM")
|
||||
"GptOssForCausalLM", "vllm_kunlun.models.gpt_oss:GptOssForCausalLM"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"DeepseekV3ForCausalLM",
|
||||
"vllm_kunlun.models.deepseek_v2:DeepseekV3ForCausalLM")
|
||||
"DeepseekV3ForCausalLM", "vllm_kunlun.models.deepseek_v2:DeepseekV3ForCausalLM"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"DeepseekV32ForCausalLM",
|
||||
"vllm_kunlun.models.deepseek_v2:DeepseekV3ForCausalLM")
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"DeepSeekMTPModel",
|
||||
"vllm_kunlun.models.deepseek_mtp:DeepSeekMTP")
|
||||
"DeepseekV32ForCausalLM", "vllm_kunlun.models.deepseek_v2:DeepseekV3ForCausalLM"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"GlmMoeDsaForCausalLM",
|
||||
"vllm_kunlun.models.deepseek_v2:GlmMoeDsaForCausalLM")
|
||||
"DeepSeekMTPModel", "vllm_kunlun.models.deepseek_mtp:DeepSeekMTP"
|
||||
)
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"GlmMoeDsaForCausalLM", "vllm_kunlun.models.deepseek_v2:GlmMoeDsaForCausalLM"
|
||||
)
|
||||
|
||||
|
||||
def register_quant_method():
|
||||
"""to do"""
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
303
vllm_kunlun/models/qwen3_next_mtp.py
Normal file
303
vllm_kunlun/models/qwen3_next_mtp.py
Normal file
@@ -0,0 +1,303 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Inference-only Qwen3Next MTP model."""
|
||||
|
||||
from collections.abc import Iterable
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.distributed.parallel_state import get_pp_group
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.linear import ColumnParallelLinear
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
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.models.interfaces import SupportsPP
|
||||
from vllm.model_executor.models.utils import (
|
||||
AutoWeightsLoader,
|
||||
is_pp_missing_parameter,
|
||||
make_empty_intermediate_tensors_factory,
|
||||
maybe_prefix,
|
||||
)
|
||||
from vllm.sequence import IntermediateTensors
|
||||
from vllm.transformers_utils.configs import Qwen3NextConfig
|
||||
|
||||
from .qwen3_next import Qwen3NextDecoderLayer, Qwen3NextRMSNorm
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
KVCache = tuple[torch.Tensor, torch.Tensor]
|
||||
|
||||
|
||||
@support_torch_compile
|
||||
class Qwen3NextMultiTokenPredictor(nn.Module):
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
|
||||
model_config = vllm_config.model_config
|
||||
quant_config = vllm_config.quant_config
|
||||
lora_config = vllm_config.lora_config
|
||||
config: Qwen3NextConfig = model_config.hf_config
|
||||
|
||||
self.config = config
|
||||
lora_vocab = (
|
||||
(lora_config.lora_extra_vocab_size * (lora_config.max_loras or 1))
|
||||
if lora_config
|
||||
else 0
|
||||
)
|
||||
self.vocab_size = config.vocab_size + lora_vocab
|
||||
self.org_vocab_size = config.vocab_size
|
||||
|
||||
self.mtp_start_layer_idx = config.num_hidden_layers
|
||||
self.num_mtp_layers = getattr(config, "num_nextn_predict_layers", 1)
|
||||
|
||||
self.embed_tokens = VocabParallelEmbedding(
|
||||
self.vocab_size,
|
||||
config.hidden_size,
|
||||
org_num_embeddings=config.vocab_size,
|
||||
)
|
||||
|
||||
self.fc = ColumnParallelLinear(
|
||||
self.config.hidden_size * 2,
|
||||
self.config.hidden_size,
|
||||
gather_output=True,
|
||||
bias=False,
|
||||
return_bias=False,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.fc",
|
||||
)
|
||||
|
||||
self.layers = torch.nn.ModuleList(
|
||||
Qwen3NextDecoderLayer(
|
||||
vllm_config,
|
||||
layer_type="full_attention",
|
||||
prefix=f"{prefix}.layers.{idx}",
|
||||
)
|
||||
for idx in range(self.num_mtp_layers)
|
||||
)
|
||||
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], config.hidden_size
|
||||
)
|
||||
|
||||
self.norm = Qwen3NextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
self.pre_fc_norm_hidden = Qwen3NextRMSNorm(
|
||||
config.hidden_size, eps=config.rms_norm_eps
|
||||
)
|
||||
self.pre_fc_norm_embedding = Qwen3NextRMSNorm(
|
||||
config.hidden_size, eps=config.rms_norm_eps
|
||||
)
|
||||
|
||||
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.embed_tokens(input_ids)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
hidden_states: torch.Tensor,
|
||||
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||||
inputs_embeds: Optional[torch.Tensor] = None,
|
||||
spec_step_idx: int = 0,
|
||||
) -> torch.Tensor:
|
||||
if get_pp_group().is_first_rank:
|
||||
if inputs_embeds is None:
|
||||
inputs_embeds = self.get_input_embeddings(input_ids)
|
||||
assert hidden_states.shape[-1] == inputs_embeds.shape[-1]
|
||||
inputs_embeds = self.pre_fc_norm_embedding(inputs_embeds)
|
||||
hidden_states = self.pre_fc_norm_hidden(hidden_states)
|
||||
hidden_states = torch.cat([inputs_embeds, hidden_states], dim=-1)
|
||||
hidden_states = self.fc(hidden_states)
|
||||
residual = None
|
||||
else:
|
||||
assert intermediate_tensors is not None
|
||||
hidden_states = intermediate_tensors["hidden_states"]
|
||||
residual = intermediate_tensors["residual"]
|
||||
|
||||
current_step_idx = spec_step_idx % self.num_mtp_layers
|
||||
hidden_states, residual = self.layers[current_step_idx](
|
||||
positions=positions,
|
||||
hidden_states=hidden_states,
|
||||
residual=residual,
|
||||
)
|
||||
|
||||
if not get_pp_group().is_last_rank:
|
||||
return IntermediateTensors(
|
||||
{"hidden_states": hidden_states, "residual": residual}
|
||||
)
|
||||
|
||||
hidden_states, _ = self.norm(hidden_states, residual)
|
||||
return hidden_states
|
||||
|
||||
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"),
|
||||
("qkv_proj", "k_proj", "k"),
|
||||
("qkv_proj", "v_proj", "v"),
|
||||
("gate_up_proj", "gate_proj", 0),
|
||||
("gate_up_proj", "up_proj", 1),
|
||||
]
|
||||
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
ckpt_up_proj_name="up_proj",
|
||||
num_experts=self.config.num_experts,
|
||||
)
|
||||
|
||||
params_dict = dict(self.named_parameters())
|
||||
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:
|
||||
if weight_name not in name:
|
||||
continue
|
||||
|
||||
if "mlp.experts" 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
|
||||
# Skip layers on other devices.
|
||||
if is_pp_missing_parameter(name, self):
|
||||
continue
|
||||
if name not in params_dict:
|
||||
continue
|
||||
param = params_dict[name]
|
||||
weight_loader = param.weight_loader
|
||||
weight_loader(param, loaded_weight, shard_id)
|
||||
break
|
||||
else:
|
||||
for mapping in expert_params_mapping:
|
||||
param_name, weight_name, expert_id, shard_id = mapping
|
||||
if weight_name not in name:
|
||||
continue
|
||||
name = name.replace(weight_name, param_name)
|
||||
# Skip layers on other devices.
|
||||
if is_pp_missing_parameter(name, self):
|
||||
continue
|
||||
# Skip loading extra bias for GPTQ models.
|
||||
if (
|
||||
name.endswith(".bias") or name.endswith("_bias")
|
||||
) and name not in params_dict:
|
||||
continue
|
||||
param = params_dict[name]
|
||||
weight_loader = param.weight_loader
|
||||
weight_loader(
|
||||
param,
|
||||
loaded_weight,
|
||||
name,
|
||||
shard_id=shard_id,
|
||||
expert_id=expert_id,
|
||||
)
|
||||
break
|
||||
else:
|
||||
# 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(param, loaded_weight)
|
||||
loaded_params.add(name)
|
||||
return loaded_params
|
||||
|
||||
|
||||
@support_torch_compile
|
||||
class Qwen3NextMTP(nn.Module, SupportsPP):
|
||||
packed_modules_mapping = {
|
||||
"qkv_proj": [
|
||||
"q_proj",
|
||||
"k_proj",
|
||||
"v_proj",
|
||||
],
|
||||
"gate_up_proj": ["up_proj", "down_proj"],
|
||||
}
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
config = vllm_config.model_config.hf_config
|
||||
self.vllm_config = vllm_config
|
||||
cache_config = vllm_config.cache_config
|
||||
assert (
|
||||
not cache_config.enable_prefix_caching
|
||||
), "Qwen3NextMTP currently does not support prefix caching"
|
||||
|
||||
self.quant_config = vllm_config.quant_config
|
||||
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.model = Qwen3NextMultiTokenPredictor(
|
||||
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "mtp")
|
||||
)
|
||||
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,
|
||||
prefix=maybe_prefix(prefix, "lm_head"),
|
||||
)
|
||||
self.logits_processor = LogitsProcessor(
|
||||
self.unpadded_vocab_size, config.vocab_size
|
||||
)
|
||||
self.make_empty_intermediate_tensors = (
|
||||
self.model.make_empty_intermediate_tensors
|
||||
)
|
||||
|
||||
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
||||
return self.model.get_input_embeddings(input_ids)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
hidden_states: torch.Tensor,
|
||||
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||||
inputs_embeds: Optional[torch.Tensor] = None,
|
||||
**kwargs: object,
|
||||
):
|
||||
hidden_states = self.model(
|
||||
input_ids, positions, hidden_states, intermediate_tensors, inputs_embeds
|
||||
)
|
||||
return hidden_states
|
||||
|
||||
def compute_logits(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
spec_step_idx: int = 0,
|
||||
) -> Optional[torch.Tensor]:
|
||||
return self.logits_processor(self.lm_head, hidden_states)
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
shared_weight_names = ["embed_tokens", "lm_head"]
|
||||
|
||||
def remap_weight_names(weights):
|
||||
for name, weight in weights:
|
||||
if name.startswith("mtp."):
|
||||
name = name.replace("mtp.", "model.")
|
||||
elif not any(key in name for key in shared_weight_names):
|
||||
continue
|
||||
yield name, weight
|
||||
|
||||
loader = AutoWeightsLoader(self)
|
||||
return loader.load_weights(remap_weight_names(weights))
|
||||
Reference in New Issue
Block a user