Support Pangu Pro MoE model (#1204)
### What this PR does / why we need it? Support Pangu Pro MoE model (https://arxiv.org/abs/2505.21411) ### Does this PR introduce _any_ user-facing change? Yes, new model supported ### How was this patch tested? Test locally --------- Signed-off-by: angazenn <zengyanjia@huawei.com> Co-authored-by: angazenn <zengyanjia@huawei.com>
This commit is contained in:
Angazenn
@@ -47,3 +47,7 @@ def register_model():
|
||||
ModelRegistry.register_model(
|
||||
"Qwen3MoeForCausalLM",
|
||||
"vllm_ascend.models.qwen3_moe:CustomQwen3MoeForCausalLM")
|
||||
|
||||
ModelRegistry.register_model(
|
||||
"PanguProMoEForCausalLM",
|
||||
"vllm_ascend.models.pangu_moe:PanguProMoEForCausalLM")
|
||||
639
vllm_ascend/models/pangu_moe.py
Normal file
639
vllm_ascend/models/pangu_moe.py
Normal file
@@ -0,0 +1,639 @@
|
||||
#
|
||||
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
||||
#
|
||||
# This file is a part of the vllm-ascend 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.
|
||||
|
||||
from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Union
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
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,
|
||||
tensor_model_parallel_all_reduce)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
|
||||
QKVParallelLinear,
|
||||
ReplicatedLinear,
|
||||
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.sampler import SamplerOutput, get_sampler
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
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 (
|
||||
extract_layer_index, is_pp_missing_parameter,
|
||||
make_empty_intermediate_tensors_factory, make_layers, maybe_prefix)
|
||||
from vllm.model_executor.sampling_metadata import SamplingMetadata
|
||||
from vllm.sequence import IntermediateTensors
|
||||
|
||||
from vllm_ascend.distributed.parallel_state import get_ep_group
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
_ROUTER_SCALE = None
|
||||
|
||||
|
||||
class PanguProMoEMLP(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 PanguProMoESparseMoeBlock(nn.Module):
|
||||
|
||||
@staticmethod
|
||||
def pangu_group8_topk(
|
||||
hidden_states: torch.Tensor,
|
||||
gating_output: torch.Tensor,
|
||||
topk: int,
|
||||
renormalize: bool,
|
||||
num_expert_group: int = 0,
|
||||
topk_group: int = 0,
|
||||
global_num_experts: int = 0,
|
||||
):
|
||||
ep_size = get_ep_group().world_size
|
||||
local_num_experts = global_num_experts // ep_size
|
||||
local_num_group = topk // ep_size
|
||||
router_scale = _ROUTER_SCALE.squeeze() # type: ignore
|
||||
scores = F.softmax(gating_output, dim=1)
|
||||
scores = scores[...,
|
||||
get_ep_group().rank_in_group *
|
||||
local_num_experts:(get_ep_group().rank_in_group + 1) *
|
||||
local_num_experts]
|
||||
|
||||
router_weights = router_scale[get_ep_group().rank_in_group *
|
||||
local_num_experts:
|
||||
(get_ep_group().rank_in_group + 1) *
|
||||
local_num_experts]
|
||||
topk_weights, topk_ids = torch.max(scores.view(scores.shape[0],
|
||||
local_num_group, -1),
|
||||
dim=-1)
|
||||
bias = torch.arange(0,
|
||||
local_num_experts,
|
||||
topk,
|
||||
device=scores.device,
|
||||
dtype=torch.int32).unsqueeze(0)
|
||||
topk_ids = topk_ids.to(torch.int32) + bias
|
||||
|
||||
flatten_topk_ids = topk_ids.view(-1)
|
||||
router_weights = router_weights.index_select(0, flatten_topk_ids).view(
|
||||
topk_ids.shape)
|
||||
topk_weights *= router_weights
|
||||
|
||||
return topk_weights, topk_ids
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
):
|
||||
super().__init__()
|
||||
self.tp_size = get_tensor_model_parallel_world_size()
|
||||
self.num_experts = config.num_experts
|
||||
|
||||
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.local_num_group = config.num_experts_per_tok // get_ep_group(
|
||||
).world_size
|
||||
self.num_experts_per_tok = config.num_experts_per_tok
|
||||
self.local_num_experts = config.num_experts // get_ep_group(
|
||||
).world_size
|
||||
self.router_scale = torch.nn.Parameter(
|
||||
torch.ones((1, self.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,
|
||||
quant_config=quant_config,
|
||||
custom_routing_function=PanguProMoESparseMoeBlock.
|
||||
pangu_group8_topk,
|
||||
prefix=f"{prefix}.experts",
|
||||
)
|
||||
|
||||
self.gate = ReplicatedLinear(
|
||||
config.hidden_size,
|
||||
config.num_experts,
|
||||
bias=False,
|
||||
quant_config=None,
|
||||
prefix=f"{prefix}.gate",
|
||||
)
|
||||
|
||||
if config.shared_expert_intermediate_size > 0:
|
||||
self.shared_expert = PanguProMoEMLP(
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.shared_expert_intermediate_size,
|
||||
hidden_act=config.hidden_act,
|
||||
quant_config=quant_config,
|
||||
reduce_results=False,
|
||||
prefix=f"{prefix}.shared_expert",
|
||||
)
|
||||
else:
|
||||
self.shared_expert = None # type: ignore
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
attn_metadata: Optional[AttentionMetadata] = None) -> torch.Tensor:
|
||||
# NOTE: hidden_states can have either 1D or 2D shape.
|
||||
num_tokens, hidden_dim = hidden_states.shape
|
||||
hidden_states = hidden_states.view(-1, hidden_dim)
|
||||
shared_output = None
|
||||
if self.shared_expert is not None:
|
||||
shared_output = self.shared_expert(hidden_states)
|
||||
# router_logits: (num_tokens, n_experts)
|
||||
router_logits, _ = self.gate(hidden_states)
|
||||
global _ROUTER_SCALE
|
||||
_ROUTER_SCALE = self.router_scale
|
||||
final_hidden_states = self.experts(hidden_states=hidden_states,
|
||||
router_logits=router_logits)
|
||||
if shared_output is not None:
|
||||
final_hidden_states = final_hidden_states + shared_output
|
||||
if self.tp_size > 1:
|
||||
final_hidden_states = tensor_model_parallel_all_reduce(
|
||||
final_hidden_states)
|
||||
|
||||
return final_hidden_states.view(num_tokens, hidden_dim)
|
||||
|
||||
|
||||
class PanguProMoEAttention(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,
|
||||
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 = 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
|
||||
|
||||
self.qkv_proj = QKVParallelLinear(
|
||||
hidden_size,
|
||||
self.head_dim,
|
||||
self.total_num_heads,
|
||||
self.total_num_kv_heads,
|
||||
bias=True,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.qkv_proj",
|
||||
)
|
||||
|
||||
self.o_proj = RowParallelLinear(
|
||||
self.total_num_heads * self.head_dim,
|
||||
hidden_size,
|
||||
bias=True,
|
||||
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",
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
positions: torch.Tensor,
|
||||
hidden_states: torch.Tensor,
|
||||
kv_cache: Optional[torch.Tensor] = None,
|
||||
attn_metadata: Optional[AttentionMetadata] = None,
|
||||
) -> torch.Tensor:
|
||||
qkv, _ = self.qkv_proj(hidden_states)
|
||||
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
|
||||
q, k = self.rotary_emb(positions, q, k)
|
||||
attn_output = self.attn(q, k, v)
|
||||
output, _ = self.o_proj(attn_output)
|
||||
return output
|
||||
|
||||
|
||||
class PanguProMoEDecoderLayer(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 = PanguProMoEAttention(
|
||||
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,
|
||||
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): ### ???
|
||||
self.mlp = PanguProMoESparseMoeBlock(
|
||||
config=config,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.mlp",
|
||||
)
|
||||
else:
|
||||
self.mlp = PanguProMoEMLP(
|
||||
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],
|
||||
kv_cache: Optional[torch.Tensor] = None,
|
||||
attn_metadata: Optional[AttentionMetadata] = None,
|
||||
) -> 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
|
||||
class PanguProMoEModel(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.embed_tokens = VocabParallelEmbedding(
|
||||
config.vocab_size,
|
||||
config.hidden_size,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.embed_tokens")
|
||||
|
||||
self.start_layer, self.end_layer, self.layers = make_layers(
|
||||
config.num_hidden_layers,
|
||||
lambda prefix: PanguProMoEDecoderLayer(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,
|
||||
kv_caches: Optional[List[torch.Tensor]] = None,
|
||||
attn_metadata: Optional[AttentionMetadata] = None,
|
||||
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,
|
||||
kv_caches[i -
|
||||
self.start_layer] if kv_caches is not None else None,
|
||||
attn_metadata)
|
||||
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
|
||||
|
||||
|
||||
class PanguProMoEForCausalLM(nn.Module, SupportsPP):
|
||||
|
||||
fall_back_to_pt_during_load = False
|
||||
|
||||
packed_modules_mapping = {
|
||||
"qkv_proj": ["q_proj", "k_proj", "v_proj"],
|
||||
"gate_up_proj": ["gate_proj", "up_proj"],
|
||||
"experts":
|
||||
["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_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 = PanguProMoEModel(vllm_config=vllm_config,
|
||||
prefix=maybe_prefix(prefix, "model"))
|
||||
self.lm_head = ParallelLMHead(
|
||||
config.vocab_size,
|
||||
config.hidden_size,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.lm_head",
|
||||
)
|
||||
if self.config.tie_word_embeddings:
|
||||
self.lm_head.weight = self.model.embed_tokens.weight
|
||||
self.logits_processor = LogitsProcessor(config.vocab_size)
|
||||
self.sampler = get_sampler()
|
||||
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,
|
||||
kv_caches: Optional[List[torch.Tensor]] = None,
|
||||
attn_metadata: Optional[AttentionMetadata] = None,
|
||||
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||||
inputs_embeds: Optional[torch.Tensor] = None,
|
||||
) -> Union[torch.Tensor, IntermediateTensors]:
|
||||
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: SamplingMetadata,
|
||||
) -> Optional[torch.Tensor]:
|
||||
logits = self.logits_processor(self.lm_head, hidden_states,
|
||||
sampling_metadata)
|
||||
return logits
|
||||
|
||||
def sample(
|
||||
self,
|
||||
logits: Optional[torch.Tensor],
|
||||
sampling_metadata: SamplingMetadata,
|
||||
) -> Optional[SamplerOutput]:
|
||||
next_tokens = self.sampler(logits, sampling_metadata)
|
||||
return next_tokens
|
||||
|
||||
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
|
||||
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)
|
||||
|
||||
# expert_params_mapping = []
|
||||
|
||||
params_dict = dict(self.named_parameters()) # from model
|
||||
loaded_params: Set[str] = set()
|
||||
for name, loaded_weight in weights:
|
||||
# =======================================================
|
||||
# BF: add this to load with less layers
|
||||
if 'layers' in name:
|
||||
layer_idx = int(name.split('layers.')[-1].split('.')[0])
|
||||
if layer_idx >= self.model.end_layer:
|
||||
continue
|
||||
|
||||
if "rotary_emb.inv_freq" in name:
|
||||
continue
|
||||
|
||||
if "module" in name:
|
||||
continue
|
||||
|
||||
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
|
||||
# breakpoint()
|
||||
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
|
||||
# breakpoint()
|
||||
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 "
|
||||
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)
|
||||
return loaded_params
|
||||
@@ -43,6 +43,7 @@ def forward_oot(
|
||||
activation: str = "silu",
|
||||
) -> torch.Tensor:
|
||||
topk_weights, topk_ids = select_experts(
|
||||
global_num_experts=global_num_experts,
|
||||
hidden_states=x,
|
||||
router_logits=router_logits,
|
||||
top_k=top_k,
|
||||
|
||||
@@ -759,6 +759,7 @@ def select_experts(
|
||||
custom_routing_function: Optional[Callable] = None,
|
||||
scoring_func: str = "softmax",
|
||||
e_score_correction_bias: Optional[torch.Tensor] = None,
|
||||
global_num_experts: Optional[torch.Tensor] = None
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Select top-k experts based on router logits.
|
||||
@@ -826,7 +827,8 @@ def select_experts(
|
||||
hidden_states=hidden_states,
|
||||
gating_output=router_logits,
|
||||
topk=top_k,
|
||||
renormalize=renormalize)
|
||||
renormalize=renormalize,
|
||||
global_num_experts=global_num_experts)
|
||||
# Required by npu_moe_init_routing
|
||||
topk_ids = topk_ids.to(torch.int32)
|
||||
return topk_weights, topk_ids
|
||||
|
||||
Reference in New Issue
Block a user