153 lines
6.1 KiB
Python
153 lines
6.1 KiB
Python
# Copyright 2023-2024 SGLang Team
|
|
# 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 Qwen3Next MTP Speculative Decoding."""
|
|
|
|
import copy
|
|
import logging
|
|
from contextlib import ExitStack
|
|
from typing import Iterable, Optional, Tuple
|
|
|
|
import torch
|
|
from torch import nn
|
|
from transformers import PretrainedConfig
|
|
|
|
from sglang.srt.distributed import get_pp_group
|
|
from sglang.srt.environ import envs
|
|
from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
|
|
from sglang.srt.layers.layernorm import GemmaRMSNorm
|
|
from sglang.srt.layers.logits_processor import LogitsProcessor
|
|
from sglang.srt.layers.quantization.base_config import QuantizationConfig
|
|
from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
|
|
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
|
|
from sglang.srt.models.qwen3_next import Qwen3NextForCausalLM, Qwen3NextModel
|
|
from sglang.srt.runtime_context import get_model, get_parallel, get_spec
|
|
from sglang.srt.utils import add_prefix, is_npu
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
class Qwen3NextForCausalLMMTP(Qwen3NextForCausalLM):
|
|
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
nn.Module.__init__(self)
|
|
# Deep-copy so MTP mutations below don't leak into the target's config.
|
|
config = copy.deepcopy(config)
|
|
self.config = config
|
|
self.tp_size = get_parallel().tp_size
|
|
if is_npu() and get_spec().speculative_draft_model_quantization is None:
|
|
quant_config = None
|
|
self.quant_config = quant_config
|
|
# if not set, model load will be broken in Qwen3NextForCausalLM load_weights()
|
|
self.pp_group = get_pp_group()
|
|
|
|
# currently based on the provided ckpt, we:
|
|
# (1) do not use_dedicated_mtp_embeddings provided in ckpt since not provided and directly use the target model embeddings
|
|
# (2) hardcode bias=False since not provided
|
|
self.fc = nn.Linear(2 * config.hidden_size, config.hidden_size, bias=False)
|
|
RMSNorm_cls = GemmaRMSNorm
|
|
self.pre_fc_norm_embedding = RMSNorm_cls(
|
|
config.hidden_size, config.rms_norm_eps
|
|
)
|
|
self.pre_fc_norm_hidden = RMSNorm_cls(config.hidden_size, config.rms_norm_eps)
|
|
mtp_config = copy.deepcopy(config)
|
|
mtp_config.num_hidden_layers = 1
|
|
mtp_config.full_attention_interval = 1
|
|
self.model = Qwen3NextModel(
|
|
mtp_config,
|
|
quant_config,
|
|
prefix=add_prefix("model", prefix),
|
|
is_nextn=True,
|
|
)
|
|
self.lm_head = ParallelLMHead(
|
|
config.vocab_size,
|
|
config.hidden_size,
|
|
quant_config=quant_config,
|
|
prefix=add_prefix("model.shared_head.head", prefix),
|
|
use_attn_tp_group=get_parallel().config.enable_dp_lm_head,
|
|
)
|
|
self.logits_processor = LogitsProcessor(config)
|
|
# Mirror Qwen3NextForCausalLM.__init__'s shared-expert fusion setup so
|
|
# the inherited load_weights() can find the attribute on the MTP path.
|
|
# We compute it from the actual MTP MoE layer (1 layer with is_nextn=True),
|
|
# not hardcode it — when the layer's MoE pre-fuses the shared expert,
|
|
# load_weights must remap mlp.shared_expert.* into the fused slot.
|
|
self.num_fused_shared_experts = self._get_num_fused_shared_experts()
|
|
if self.num_fused_shared_experts > 1:
|
|
raise ValueError(
|
|
"Qwen3-Next MTP shared expert fusion currently supports exactly one "
|
|
"shared expert because checkpoint weight remapping maps it into "
|
|
"a single fused MoE expert slot."
|
|
)
|
|
self.enable_shared_expert_fusion = self.num_fused_shared_experts > 0
|
|
|
|
@torch.no_grad()
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
forward_batch: ForwardBatch,
|
|
input_embeds: Optional[torch.Tensor] = None,
|
|
**kwargs,
|
|
):
|
|
exit_stack = ExitStack()
|
|
if (
|
|
is_npu()
|
|
and self.quant_config is None
|
|
and get_model().quantization is not None
|
|
):
|
|
# ascend mtp unquant
|
|
exit_stack.enter_context(envs.SGLANG_DEEPEP_BF16_DISPATCH.override(True))
|
|
exit_stack.enter_context(
|
|
envs.DEEP_NORMAL_MODE_USE_INT8_QUANT.override(False)
|
|
)
|
|
|
|
try:
|
|
if input_embeds is None:
|
|
input_embeds = self.model.embed_tokens(input_ids)
|
|
|
|
hidden_states = forward_batch.spec_info.hidden_states
|
|
# Some idle batch has 0 batch size. GemmaRMSNorm.forward would fail due to bs=0.
|
|
if not forward_batch.forward_mode.is_idle():
|
|
input_embeds = self.pre_fc_norm_embedding(input_embeds)
|
|
hidden_states = self.pre_fc_norm_hidden(hidden_states)
|
|
hidden_states = self.fc(torch.cat((input_embeds, hidden_states), dim=-1))
|
|
|
|
with get_global_expert_distribution_recorder().disable_this_region():
|
|
hidden_states = self.model(
|
|
input_ids,
|
|
positions,
|
|
forward_batch,
|
|
hidden_states,
|
|
)
|
|
finally:
|
|
exit_stack.close()
|
|
|
|
return self.logits_processor(
|
|
input_ids, hidden_states, self.lm_head, forward_batch
|
|
)
|
|
|
|
def load_weights(
|
|
self, weights: Iterable[Tuple[str, torch.Tensor]], is_mtp: bool = False
|
|
):
|
|
super().load_weights(weights, is_mtp=True)
|
|
|
|
|
|
EntryClass = [Qwen3NextForCausalLMMTP]
|