Files
enginex-biren-vllm/vllm_br/model_executor/models/qwen3_moe.py
2026-03-10 13:31:25 +08:00

301 lines
13 KiB
Python

################################################################################
# Copyright(c)2020-2025 Shanghai Biren Technology Co., Ltd. All rights reserved.
# 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.
#
################################################################################
from collections.abc import Iterable
from typing import Optional, Union
import torch
import torch_br
from fastcore.basics import patch_to
import vllm
from vllm.distributed import get_pp_group, tensor_model_parallel_all_reduce
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.models.qwen3_moe import Qwen3MoeModel
from vllm.model_executor.models.utils import is_pp_missing_parameter
from vllm.sequence import IntermediateTensors
from vllm_br.v1.attention.backends.attention_v1 import (
SUPAFlashAttentionMetadata)
logger = init_logger(__name__)
@patch_to(vllm.model_executor.models.qwen3_moe.Qwen3MoeSparseMoeBlock)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
# NOTE: hidden_states can have either 1D or 2D shape.
orig_shape = hidden_states.shape
if len(hidden_states.shape) == 3:
hidden_states = hidden_states.squeeze(0)
final_hidden_states = self.experts(hidden_states=hidden_states,
router_logits=(self.gate.weight, None,
None))
if hasattr(final_hidden_states, 'all_reduced'):
# NOTE: this flag indicates that the final_hidden_states has been reduced in fused_moe
delattr(final_hidden_states, 'all_reduced')
elif self.tp_size > 1:
final_hidden_states = tensor_model_parallel_all_reduce(
final_hidden_states)
return final_hidden_states.view(orig_shape)
@patch_to(vllm.model_executor.models.qwen3_moe.Qwen3MoeAttention)
def forward(
self,
positions: torch.Tensor,
hidden_states: torch.Tensor,
) -> torch.Tensor:
forward_context: ForwardContext = get_forward_context()
attn_metadata: SUPAFlashAttentionMetadata = forward_context.attn_metadata
if attn_metadata is None:
## for dummy run
return hidden_states
seq_len = hidden_states.shape[-2]
decode_seql = 512
if isinstance(attn_metadata, dict):
attn_metadata = attn_metadata[self.attn.layer_name]
kv_cache = self.attn.kv_cache[forward_context.virtual_engine]
if kv_cache is not None:
if seq_len <= decode_seql:
if hasattr(self.qkv_proj, "qweight"):
qkv_weight = self.qkv_proj.qweight.data
qkv_scales = self.qkv_proj.scales.data
elif hasattr(self.qkv_proj, "weight_packed"):
qkv_weight = self.qkv_proj.weight_packed.data
qkv_scales = self.qkv_proj.weight_scale.data
else:
qkv_weight = self.qkv_proj.weight
qkv_scales = None
if isinstance(self.rotary_emb, MRotaryEmbedding):
assert len(
self.rotary_emb.mrope_section
) == 3 and self.rotary_emb.mrope_section[
1] == self.rotary_emb.mrope_section[
2], "current only support mrope_section width and height are equal!"
q, k, v = torch_br.br_qwen3_vl_prefix_attn_infer(
hidden_states,
qkv_weight, [self.q_size, self.kv_size, self.kv_size],
self.head_dim,
self.q_norm.variance_epsilon,
self.q_norm.weight,
self.k_norm.weight,
self.rotary_emb.cos_sin_cache,
kv_cache,
positions,
attn_metadata.slot_mapping,
self.rotary_emb.mrope_section[1],
bias=self.qkv_proj.bias,
scales=qkv_scales)
else:
q, k, v = torch_br.br_qwen3_prefix_attn_infer(
hidden_states,
qkv_weight, [self.q_size, self.kv_size, self.kv_size],
self.head_dim,
self.q_norm.variance_epsilon,
self.q_norm.weight,
self.k_norm.weight,
self.rotary_emb.sin_cache,
self.rotary_emb.cos_cache,
kv_cache,
positions,
attn_metadata.slot_mapping,
bias=self.qkv_proj.bias,
scales=qkv_scales)
else:
qkv, _ = self.qkv_proj(hidden_states)
if isinstance(self.rotary_emb, MRotaryEmbedding):
assert len(
self.rotary_emb.mrope_section
) == 3 and self.rotary_emb.mrope_section[
1] == self.rotary_emb.mrope_section[
2], "current only support mrope_section width and height are equal!"
q, k, v = torch_br.br_fused_rms_mrope_kvstore_infer(
qkv, [self.q_size, self.kv_size, self.kv_size],
self.head_dim, self.q_norm.variance_epsilon,
self.q_norm.weight, self.k_norm.weight,
self.rotary_emb.cos_sin_cache, kv_cache, positions,
attn_metadata.slot_mapping, attn_metadata.block_table,
attn_metadata.query_start_loc, attn_metadata.context_lens,
self.rotary_emb.mrope_section[1])
else:
q, k, v = torch_br.br_fused_rms_rope_kvstore_infer(
qkv, [self.q_size, self.kv_size, self.kv_size],
self.head_dim, self.q_norm.variance_epsilon,
self.q_norm.weight, self.k_norm.weight,
self.rotary_emb.sin_cache, self.rotary_emb.cos_cache,
kv_cache, positions, attn_metadata.slot_mapping,
attn_metadata.block_table, attn_metadata.query_start_loc,
attn_metadata.context_lens)
if hasattr(attn_metadata, 'do_cache'):
attn_metadata.do_cache = False
attn_output = self.attn(q, k, v)
output, _ = self.o_proj(attn_output)
return output
else:
return hidden_states
def model_forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
if get_pp_group().is_first_rank:
if inputs_embeds is not None:
hidden_states = inputs_embeds
else:
hidden_states = self.get_input_embeddings(input_ids)
residual = None
else:
assert intermediate_tensors is not None
hidden_states = intermediate_tensors["hidden_states"]
residual = intermediate_tensors["residual"]
if len(hidden_states.shape) == 2:
hidden_states = hidden_states.unsqueeze(0)
for i in range(self.start_layer, self.end_layer):
layer = self.layers[i]
hidden_states, residual = layer(positions, hidden_states, residual)
if not get_pp_group().is_last_rank:
return IntermediateTensors({
"hidden_states":
hidden_states.squeeze(0) if hidden_states is not None else None,
"residual":
residual.squeeze(0) if residual is not None else None
})
hidden_states, _ = self.norm(hidden_states, residual)
# NOTE: convert back to 2D
hidden_states = hidden_states.squeeze()
if hidden_states.dim() == 1:
hidden_states = hidden_states.unsqueeze(0)
return hidden_states
Qwen3MoeModel.forward = model_forward
def Qwen3MoeModel_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:
for (param_name, weight_name, shard_id) in stacked_params_mapping:
# Skip non-stacked layers and experts (experts handled below).
if weight_name not in name:
continue
# We have mlp.experts[0].gate_proj in the checkpoint.
# Since we handle the experts below in expert_params_mapping,
# we need to skip here BEFORE we update the name, otherwise
# name will be updated to mlp.experts[0].gate_up_proj, which
# will then be updated below in expert_params_mapping
# for mlp.experts[0].gate_gate_up_proj, which breaks load.
if "mlp.experts" in name:
continue
name = name.replace(weight_name, param_name)
# Skip loading extra bias for GPTQ models.
if ((name.endswith(".bias") or 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") or name.endswith("_bias"))
and name not in params_dict):
continue
# Skip layers on other devices.
if is_pp_missing_parameter(name, self):
continue
# Remapping the name of FP8 kv-scale.
if name.endswith("kv_scale"):
remapped_kv_scale_name = name.replace(
".kv_scale", ".attn.kv_scale")
if remapped_kv_scale_name not in params_dict:
logger.warning_once(
"Found kv scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv-scale is not loaded.", # noqa: E501
name,
remapped_kv_scale_name,
)
continue
else:
name = remapped_kv_scale_name
param = params_dict[name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
if name.find("norm.weight") != -1:
param.data = param.data.to(torch.float32)
loaded_params.add(name)
return loaded_params
Qwen3MoeModel.load_weights = Qwen3MoeModel_load_weights