837 lines
31 KiB
Python
837 lines
31 KiB
Python
#
|
|
# Copyright (c) 2025 Baidu, Inc. All Rights Reserved.
|
|
# Adapted from vllm/model_executor/models/qwen3_moe.py
|
|
# Copyright 2023 The vLLM team.
|
|
#
|
|
# This file is a part of the vllm-kunlun project.
|
|
#
|
|
# 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 Qwen3MoE model compatible with HuggingFace weights."""
|
|
import os
|
|
from collections.abc import Iterable
|
|
from typing import Any, Optional, Union, Tuple, Set
|
|
|
|
import torch
|
|
import os
|
|
from torch import nn
|
|
from transformers import PretrainedConfig
|
|
|
|
from vllm_kunlun.ops.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.logger import init_logger
|
|
from vllm_kunlun.ops.activation import SiluAndMul
|
|
from vllm_kunlun.ops.fused_moe.layer import FusedMoE
|
|
from vllm.model_executor.layers.layernorm import RMSNorm
|
|
from vllm.model_executor.layers.linear import (
|
|
MergedColumnParallelLinear,
|
|
QKVParallelLinear,
|
|
RowParallelLinear,
|
|
)
|
|
from vllm_kunlun.ops.linear import ReplicatedLinear
|
|
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_kunlun.ops.vocab_parallel_embedding import 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 IntermediateTensors
|
|
|
|
from vllm.model_executor.models.interfaces import SupportsPP
|
|
from vllm.model_executor.models.utils import (
|
|
AutoWeightsLoader,
|
|
extract_layer_index,
|
|
is_pp_missing_parameter,
|
|
make_empty_intermediate_tensors_factory,
|
|
make_layers,
|
|
maybe_prefix,
|
|
)
|
|
from vllm_kunlun.ops.rotary_embedding import Split_Norm_Rope
|
|
|
|
logger = init_logger(__name__)
|
|
|
|
|
|
class Qwen3MoeMLP(nn.Module):
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
intermediate_size: int,
|
|
hidden_act: str,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
reduce_results: bool = True,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.gate_up_proj = MergedColumnParallelLinear(
|
|
hidden_size,
|
|
[intermediate_size] * 2,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.gate_up_proj",
|
|
)
|
|
self.down_proj = RowParallelLinear(
|
|
intermediate_size,
|
|
hidden_size,
|
|
bias=False,
|
|
quant_config=quant_config,
|
|
reduce_results=reduce_results,
|
|
prefix=f"{prefix}.down_proj",
|
|
)
|
|
if hidden_act != "silu":
|
|
raise ValueError(
|
|
f"Unsupported activation: {hidden_act}. "
|
|
"Only silu is supported for now."
|
|
)
|
|
self.act_fn = SiluAndMul()
|
|
|
|
def forward(self, x):
|
|
gate_up, _ = self.gate_up_proj(x)
|
|
x = self.act_fn(gate_up)
|
|
x, _ = self.down_proj(x)
|
|
return x
|
|
|
|
|
|
class Qwen3MoeSparseMoeBlock(nn.Module):
|
|
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
self.tp_size = get_tensor_model_parallel_world_size()
|
|
|
|
if self.tp_size > config.num_experts:
|
|
raise ValueError(
|
|
f"Tensor parallel size {self.tp_size} is greater than "
|
|
f"the number of experts {config.num_experts}."
|
|
)
|
|
|
|
self.experts = FusedMoE(
|
|
num_experts=config.num_experts,
|
|
top_k=config.num_experts_per_tok,
|
|
hidden_size=config.hidden_size,
|
|
intermediate_size=config.moe_intermediate_size,
|
|
reduce_results=False,
|
|
renormalize=config.norm_topk_prob,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.experts",
|
|
)
|
|
self.quant_config = quant_config
|
|
self.gate = ReplicatedLinear(
|
|
config.hidden_size,
|
|
config.num_experts,
|
|
bias=False,
|
|
quant_config=None,
|
|
prefix=f"{prefix}.gate",
|
|
)
|
|
|
|
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
|
# NOTE: hidden_states can have either 1D or 2D shape.
|
|
orig_shape = hidden_states.shape
|
|
hidden_dim = hidden_states.shape[-1]
|
|
hidden_states = hidden_states.view(-1, hidden_dim)
|
|
|
|
if self.quant_config is None:
|
|
kunlun_linear_weights = self.gate.get_weights()
|
|
final_hidden_states = self.experts(
|
|
hidden_states=hidden_states, linear_weights=kunlun_linear_weights
|
|
)
|
|
else:
|
|
kunlun_linear_weights = self.gate.get_weights()
|
|
router_logits, _ = self.gate(hidden_states)
|
|
final_hidden_states = self.experts(
|
|
hidden_states=hidden_states,
|
|
router_logits=router_logits,
|
|
linear_weights=kunlun_linear_weights,
|
|
)
|
|
|
|
if self.tp_size > 1:
|
|
final_hidden_states = (
|
|
self.experts.maybe_all_reduce_tensor_model_parallel( # noqa E501
|
|
final_hidden_states
|
|
)
|
|
)
|
|
|
|
return final_hidden_states.view(orig_shape)
|
|
|
|
|
|
class Qwen3MoeAttention(nn.Module):
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
num_heads: int,
|
|
num_kv_heads: int,
|
|
rope_theta: float = 10000,
|
|
rope_scaling: Optional[dict[str, Any]] = 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.rope_theta = rope_theta
|
|
self.max_position_embeddings = max_position_embeddings
|
|
if rope_scaling is not None:
|
|
scaling_factor = rope_scaling["factor"]
|
|
self.max_position_embeddings = int(
|
|
self.max_position_embeddings * scaling_factor
|
|
)
|
|
|
|
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=self.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",
|
|
)
|
|
|
|
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,
|
|
) -> torch.Tensor:
|
|
qkv, _ = self.qkv_proj(hidden_states)
|
|
if os.getenv("FUSED_QK_ROPE_OP") == "1":
|
|
# Rope fusion operators
|
|
q, k, v = Split_Norm_Rope(
|
|
qkv,
|
|
self.rotary_emb.cos_sin_cache,
|
|
self.q_norm.weight,
|
|
self.k_norm.weight,
|
|
positions,
|
|
self.max_position_embeddings,
|
|
self.num_heads,
|
|
self.num_kv_heads,
|
|
self.head_dim,
|
|
)
|
|
else:
|
|
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
|
|
# Add qk-norm
|
|
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)
|
|
output, _ = self.o_proj(attn_output)
|
|
return output
|
|
|
|
|
|
class Qwen3MoeDecoderLayer(nn.Module):
|
|
|
|
def __init__(
|
|
self,
|
|
config: PretrainedConfig,
|
|
cache_config: Optional[CacheConfig] = None,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "",
|
|
) -> None:
|
|
super().__init__()
|
|
self.hidden_size = config.hidden_size
|
|
rope_theta = getattr(config, "rope_theta", 10000)
|
|
rope_scaling = getattr(config, "rope_scaling", None)
|
|
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
|
|
self.self_attn = Qwen3MoeAttention(
|
|
hidden_size=self.hidden_size,
|
|
num_heads=config.num_attention_heads,
|
|
num_kv_heads=config.num_key_value_heads,
|
|
rope_theta=rope_theta,
|
|
rope_scaling=rope_scaling,
|
|
max_position_embeddings=max_position_embeddings,
|
|
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,
|
|
prefix=f"{prefix}.self_attn",
|
|
)
|
|
|
|
# `mlp_only_layers` in the config.
|
|
layer_idx = extract_layer_index(prefix)
|
|
mlp_only_layers = (
|
|
[] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
|
|
)
|
|
if (layer_idx not in mlp_only_layers) and (
|
|
config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
|
|
):
|
|
self.mlp = Qwen3MoeSparseMoeBlock(
|
|
config=config, quant_config=quant_config, prefix=f"{prefix}.mlp"
|
|
)
|
|
else:
|
|
self.mlp = Qwen3MoeMLP(
|
|
hidden_size=config.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,
|
|
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,
|
|
)
|
|
|
|
# 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
|
|
class Qwen3MoeModel(nn.Module):
|
|
|
|
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.padding_idx = config.pad_token_id
|
|
self.vocab_size = config.vocab_size
|
|
self.config = config
|
|
self.embed_tokens = VocabParallelEmbedding(
|
|
config.vocab_size, config.hidden_size, prefix=f"{prefix}.embed_tokens"
|
|
)
|
|
self.start_layer, self.end_layer, self.layers = make_layers(
|
|
config.num_hidden_layers,
|
|
lambda prefix: Qwen3MoeDecoderLayer(
|
|
config=config,
|
|
cache_config=cache_config,
|
|
quant_config=quant_config,
|
|
prefix=prefix,
|
|
),
|
|
prefix=f"{prefix}.layers",
|
|
)
|
|
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
|
["hidden_states", "residual"], config.hidden_size
|
|
)
|
|
|
|
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,
|
|
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"]
|
|
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, "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()
|
|
weights_to_quantize = {}
|
|
|
|
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)
|
|
loaded_params.add(name)
|
|
break
|
|
else:
|
|
for mapping in expert_params_mapping:
|
|
param_name, weight_name, expert_id, shard_id = mapping
|
|
if weight_name not in name:
|
|
continue
|
|
# Map to the parameter name in the model
|
|
name_mapped = name.replace(weight_name, param_name)
|
|
|
|
# Layer/PP skip judgment
|
|
if is_pp_missing_parameter(name_mapped, self):
|
|
continue
|
|
if (
|
|
name_mapped.endswith(".bias") or name_mapped.endswith("_bias")
|
|
) and name_mapped not in params_dict:
|
|
continue
|
|
|
|
# Get the param and target module
|
|
param = params_dict.get(name_mapped, None)
|
|
if param is None:
|
|
continue
|
|
|
|
# === Only when the target MoE layer has int8 weights and scales, and the name matches, the "streaming quantization" is performed ===
|
|
if self._should_stream_quantize(name_mapped):
|
|
# Note: Pass the mapped name_mapped instead of the original name
|
|
self._stream_quantize_moe_weight(
|
|
name_mapped,
|
|
param,
|
|
loaded_weight,
|
|
expert_id=expert_id,
|
|
shard_id=shard_id,
|
|
)
|
|
loaded_params.add(name_mapped)
|
|
else:
|
|
# Fallback: Normal weight loading (non-quantized)
|
|
weight_loader = getattr(
|
|
param, "weight_loader", default_weight_loader
|
|
)
|
|
weight_loader(
|
|
param,
|
|
loaded_weight,
|
|
name_mapped,
|
|
shard_id=shard_id,
|
|
expert_id=expert_id,
|
|
)
|
|
loaded_params.add(name_mapped)
|
|
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 "
|
|
f"(e.g. {name}), but not found the expected "
|
|
f"name in the model "
|
|
f"(e.g. {remapped_kv_scale_name}). "
|
|
"kv-scale is not loaded."
|
|
)
|
|
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)
|
|
loaded_params.add(name)
|
|
# loaded_params.add(name)
|
|
return loaded_params
|
|
|
|
def _is_moe_weight(self, name: str) -> bool:
|
|
"""Check if the weight is MoE weight"""
|
|
return name.endswith("w13_weight") or name.endswith("w2_weight")
|
|
|
|
def _is_expert_complete(self, cache_key):
|
|
cache = self._moe_weight_cache.get(cache_key)
|
|
if cache is None:
|
|
return False
|
|
w13_ok = (0 in cache["w13_shards"]) and (1 in cache["w13_shards"])
|
|
w2_ok = cache["w2_weight"] is not None
|
|
return w13_ok and w2_ok
|
|
|
|
@torch.no_grad()
|
|
def _stream_quantize_moe_weight(
|
|
self,
|
|
param_name: str,
|
|
param: nn.Parameter,
|
|
loaded_weight: torch.Tensor,
|
|
*,
|
|
expert_id,
|
|
shard_id,
|
|
):
|
|
|
|
rank = os.environ.get("RANK", "0")
|
|
|
|
# Ensure expert_id is an integer
|
|
try:
|
|
expert_id = int(expert_id)
|
|
except (ValueError, TypeError):
|
|
if isinstance(expert_id, str):
|
|
expert_id = int(expert_id)
|
|
|
|
# Process shard_id
|
|
if isinstance(shard_id, str):
|
|
if shard_id in ("gate", "w1"):
|
|
shard_id = 0
|
|
elif shard_id in ("up", "w3"):
|
|
shard_id = 1
|
|
elif shard_id == "w2":
|
|
shard_id = 0
|
|
else:
|
|
try:
|
|
shard_id = int(shard_id)
|
|
except ValueError:
|
|
shard_id = 0
|
|
else:
|
|
shard_id = int(shard_id)
|
|
|
|
# Initialize cache
|
|
if not hasattr(self, "_moe_weight_cache"):
|
|
self._moe_weight_cache = {}
|
|
self._expert_batch_count = 0 # Batch counter
|
|
|
|
module_path = ".".join(param_name.split(".")[:-1])
|
|
cache_key = (module_path, expert_id)
|
|
|
|
cache = self._moe_weight_cache.get(cache_key)
|
|
if cache is None:
|
|
cache = {
|
|
"w13_shards": {},
|
|
"w2_weight": None,
|
|
"target_module": self.get_submodule(module_path),
|
|
"done": False,
|
|
}
|
|
self._moe_weight_cache[cache_key] = cache
|
|
|
|
if cache.get("done", False):
|
|
return
|
|
|
|
# Cache weights (keep original precision)
|
|
if "w13_weight" in param_name:
|
|
cache["w13_shards"][shard_id] = loaded_weight.clone()
|
|
elif "w2_weight" in param_name:
|
|
cache["w2_weight"] = loaded_weight.clone()
|
|
|
|
# Check if complete
|
|
if self._is_expert_complete(cache_key):
|
|
# Quantize this expert
|
|
self._quantize_expert_weights(cache_key)
|
|
cache["done"] = True
|
|
self._moe_weight_cache.pop(cache_key, None)
|
|
|
|
# Force synchronization every 4 experts
|
|
self._expert_batch_count += 1
|
|
if self._expert_batch_count % 4 == 0:
|
|
torch.cuda.synchronize() # Force synchronization
|
|
# print(f"[Rank {rank}] Completed batch of {self._expert_batch_count} experts")
|
|
|
|
def _quantize_expert_weights(self, cache_key):
|
|
"""Quantize the complete weights of an expert (supports TP sharding)"""
|
|
module_path, expert_id = cache_key
|
|
cache = self._moe_weight_cache[cache_key]
|
|
target_module = cache["target_module"]
|
|
|
|
# Get TP config
|
|
from vllm.distributed import (
|
|
get_tensor_model_parallel_rank,
|
|
get_tensor_model_parallel_world_size,
|
|
)
|
|
|
|
tp_rank = get_tensor_model_parallel_rank()
|
|
tp_size = get_tensor_model_parallel_world_size()
|
|
|
|
# Get actual shapes
|
|
E, twoN, H = target_module.w13_weight.shape
|
|
_, H2, N = target_module.w2_weight.shape
|
|
|
|
qmax = 127.0
|
|
|
|
# Process w13_weight: concatenate gate and up
|
|
gate_weight = cache["w13_shards"][0] # [768, 2048]
|
|
up_weight = cache["w13_shards"][1] # [768, 2048]
|
|
|
|
# TP sharding
|
|
if tp_size > 1:
|
|
# Calculate shard for each TP rank
|
|
gate_per_rank = gate_weight.shape[0] // tp_size
|
|
up_per_rank = up_weight.shape[0] // tp_size
|
|
|
|
gate_start = tp_rank * gate_per_rank
|
|
gate_end = (tp_rank + 1) * gate_per_rank
|
|
up_start = tp_rank * up_per_rank
|
|
up_end = (tp_rank + 1) * up_per_rank
|
|
|
|
gate_weight = gate_weight[gate_start:gate_end, :] # [192, 2048]
|
|
up_weight = up_weight[up_start:up_end, :] # [192, 2048]
|
|
|
|
w13_complete = torch.cat([gate_weight, up_weight], dim=0) # [384, 2048]
|
|
|
|
# Quantize w13_weight
|
|
w13_f = w13_complete.float()
|
|
w13_abs_max = torch.amax(torch.abs(w13_f), dim=-1) # [384]
|
|
w13_scale_2d = torch.clamp(w13_abs_max, min=1e-6) / qmax # [384]
|
|
w13_scale_3d = w13_scale_2d.unsqueeze(-1) # [384, 1]
|
|
w13_q = torch.round(w13_f / w13_scale_3d).clamp_(-128, 127).to(torch.int8)
|
|
|
|
# Write w13_weight
|
|
target_module.w13_weight.data[expert_id, :, :].copy_(
|
|
w13_q.to(target_module.w13_weight.device)
|
|
)
|
|
|
|
# Update w13_scale - pre-multiply 127
|
|
s = getattr(target_module, "w13_weight_scale")
|
|
s.data[expert_id, :].copy_((w13_scale_2d * 127.0).to(s.device))
|
|
|
|
# Process w2_weight
|
|
w2_weight = cache["w2_weight"] # [2048, 768]
|
|
|
|
# TP sharding for w2 weight
|
|
if tp_size > 1:
|
|
w2_per_rank = w2_weight.shape[1] // tp_size
|
|
w2_start = tp_rank * w2_per_rank
|
|
w2_end = (tp_rank + 1) * w2_per_rank
|
|
w2_weight = w2_weight[:, w2_start:w2_end] # [2048, 192]
|
|
|
|
w2_f = w2_weight.float() # [2048, 192]
|
|
w2_abs_max = torch.amax(torch.abs(w2_f), dim=-1) # [2048]
|
|
w2_scale_2d = torch.clamp(w2_abs_max, min=1e-6) / qmax # [2048]
|
|
w2_scale_3d = w2_scale_2d.unsqueeze(-1) # [2048, 1]
|
|
w2_q = torch.round(w2_f / w2_scale_3d).clamp_(-128, 127).to(torch.int8)
|
|
|
|
# Write w2_weight
|
|
w2_param = getattr(target_module, "w2_weight")
|
|
w2_param.data[expert_id, :, :].copy_(w2_q.to(w2_param.device))
|
|
|
|
# Update w2_scale - pre-multiply 127
|
|
w2_s = getattr(target_module, "w2_weight_scale")
|
|
w2_s.data[expert_id, :].copy_((w2_scale_2d * 127.0).to(w2_s.device))
|
|
|
|
# Clear cache
|
|
cache["w13_shards"].clear()
|
|
cache["w2_weight"] = None
|
|
|
|
def _is_int8_moe_target_module(self, module_path: str) -> bool:
|
|
"""Check if a module_path is a FusedMoE target using INT8(W8A8).
|
|
Determine by the actual existing parameters and dtype, not relying on quant_config names.
|
|
"""
|
|
try:
|
|
mod = self.get_submodule(module_path)
|
|
except Exception:
|
|
return False
|
|
# Need to have both int8 weights and float32 scales, and dimensions come from CompressedTensorsW8A8 path
|
|
if not (
|
|
hasattr(mod, "w13_weight")
|
|
and hasattr(mod, "w2_weight")
|
|
and hasattr(mod, "w13_weight_scale")
|
|
and hasattr(mod, "w2_weight_scale")
|
|
):
|
|
return False
|
|
try:
|
|
return (
|
|
mod.w13_weight.dtype == torch.int8
|
|
and mod.w2_weight.dtype == torch.int8
|
|
and mod.w13_weight_scale.dtype == torch.float32
|
|
and mod.w2_weight_scale.dtype == torch.float32
|
|
)
|
|
except Exception:
|
|
return False
|
|
|
|
def _should_stream_quantize(self, param_name: str) -> bool:
|
|
"""Only when (1) the parameter name corresponds to the MoE weights we defined; and
|
|
(2) the MoE layer is indeed the INT8 path (exists int8 weights + scales)
|
|
Stream quantization is enabled; otherwise, it falls back to the default loading.
|
|
"""
|
|
# First, determine if it is the MoE weight name we want to process (w13_weight / w2_weight)
|
|
if not self._is_moe_weight(param_name):
|
|
return False
|
|
# Then, check if the module containing this param is the INT8 path
|
|
module_path = ".".join(param_name.split(".")[:-1])
|
|
return self._is_int8_moe_target_module(module_path)
|
|
|
|
|
|
class Qwen3MoeForCausalLM(nn.Module, SupportsPP):
|
|
packed_modules_mapping = {
|
|
"qkv_proj": [
|
|
"q_proj",
|
|
"k_proj",
|
|
"v_proj",
|
|
],
|
|
"gate_up_proj": [
|
|
"gate_proj",
|
|
"up_proj",
|
|
],
|
|
}
|
|
|
|
fall_back_to_pt_during_load = False
|
|
|
|
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 = Qwen3MoeModel(
|
|
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
|
|
)
|
|
self.lm_head = ParallelLMHead(
|
|
config.vocab_size, config.hidden_size, quant_config=quant_config
|
|
)
|
|
if self.config.tie_word_embeddings:
|
|
self.lm_head.weight = self.model.embed_tokens.weight
|
|
self.logits_processor = LogitsProcessor(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,
|
|
intermediate_tensors: Optional[IntermediateTensors] = None,
|
|
inputs_embeds: Optional[torch.Tensor] = None,
|
|
kv_caches: list[torch.Tensor] = None,
|
|
) -> Union[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,
|
|
sampling_metadata: SamplingMetadata,
|
|
) -> Optional[torch.Tensor]:
|
|
logits = self.logits_processor(self.lm_head, hidden_states, sampling_metadata)
|
|
return logits
|
|
|
|
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
|
loader = AutoWeightsLoader(self)
|
|
return loader.load_weights(weights)
|