[Model] Add LongCat-Flash (#23991)
Signed-off-by: yangxurui <yangxurui@meituan.com> Co-authored-by: yangxurui <yangxurui@meituan.com>
This commit is contained in:
@@ -0,0 +1,712 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
# Apache License, Version 2.0:
|
||||
# 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.
|
||||
#
|
||||
# MIT License:
|
||||
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
||||
# of this software and associated documentation files (the "Software"), to deal
|
||||
# in the Software without restriction, including without limitation the rights
|
||||
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
||||
# copies of the Software, and to permit persons to whom the Software is
|
||||
# furnished to do so, subject to the following conditions:
|
||||
#
|
||||
# The above copyright notice and this permission notice shall be included in all
|
||||
# copies or substantial portions of the Software.
|
||||
#
|
||||
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
||||
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
||||
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
||||
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
||||
# SOFTWARE.
|
||||
"""Inference-only Flash model compatible with HuggingFace weights."""
|
||||
import typing
|
||||
from collections.abc import Callable, Iterable
|
||||
from typing import Optional, Union
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import CacheConfig, VllmConfig
|
||||
from vllm.distributed import get_pp_group
|
||||
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,
|
||||
ReplicatedLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
from vllm.model_executor.layers.quantization.utils.int8_utils import (
|
||||
block_dequant)
|
||||
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.deepseek_v2 import DeepseekV2MLAAttention
|
||||
from vllm.sequence import IntermediateTensors
|
||||
|
||||
from .interfaces import SupportsLoRA, SupportsPP
|
||||
from .utils import (PPMissingLayer, is_pp_missing_parameter,
|
||||
make_empty_intermediate_tensors_factory, make_layers,
|
||||
maybe_prefix)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class FlashConfig(PretrainedConfig):
|
||||
"""Flash model configuration."""
|
||||
model_type = "longcat_flash"
|
||||
keys_to_ignore_at_inference = ["past_key_values"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size=131072,
|
||||
hidden_size=4096,
|
||||
intermediate_size=8192,
|
||||
num_layers=28,
|
||||
num_hidden_layers=None,
|
||||
num_attention_heads=96,
|
||||
num_key_value_heads=128,
|
||||
ep_size=1,
|
||||
kv_lora_rank=512,
|
||||
q_lora_rank=1536,
|
||||
qk_rope_head_dim=64,
|
||||
v_head_dim=128,
|
||||
qk_nope_head_dim=128,
|
||||
num_experts_per_tok=None,
|
||||
norm_topk_prob=False,
|
||||
max_position_embeddings=8192,
|
||||
initializer_range=0.02,
|
||||
rms_norm_eps=1e-05,
|
||||
use_cache=True,
|
||||
pad_token_id=None,
|
||||
bos_token_id=100000,
|
||||
eos_token_id=100001,
|
||||
pretraining_tp=1,
|
||||
tie_word_embeddings=False,
|
||||
rope_theta=1000000.0,
|
||||
rope_scaling=None,
|
||||
attention_bias=False,
|
||||
attention_dropout=0.0,
|
||||
mla_scale_q_lora=False,
|
||||
mla_scale_kv_lora=False,
|
||||
torch_dtype="bfloat16",
|
||||
params_dtype="bfloat16",
|
||||
router_dtype="float32",
|
||||
router_bias=False,
|
||||
topk_method=None,
|
||||
routed_scaling_factor=None,
|
||||
zero_expert_num=0,
|
||||
zero_expert_type=None,
|
||||
nextn_use_scmoe=False,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__(
|
||||
pad_token_id=pad_token_id,
|
||||
bos_token_id=bos_token_id,
|
||||
eos_token_id=eos_token_id,
|
||||
tie_word_embeddings=tie_word_embeddings,
|
||||
torch_dtype=torch_dtype,
|
||||
params_dtype=params_dtype,
|
||||
router_dtype=router_dtype,
|
||||
topk_method=topk_method,
|
||||
router_bias=router_bias,
|
||||
nextn_use_scmoe=nextn_use_scmoe,
|
||||
**kwargs,
|
||||
)
|
||||
self.vocab_size = vocab_size
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = (num_hidden_layers if num_hidden_layers
|
||||
is not None else num_layers)
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.ep_size = ep_size
|
||||
self.kv_lora_rank = kv_lora_rank
|
||||
self.q_lora_rank = q_lora_rank
|
||||
self.qk_rope_head_dim = qk_rope_head_dim
|
||||
self.v_head_dim = v_head_dim
|
||||
self.qk_nope_head_dim = qk_nope_head_dim
|
||||
self.num_experts_per_tok = num_experts_per_tok
|
||||
self.norm_topk_prob = norm_topk_prob
|
||||
# for backward compatibility
|
||||
if num_key_value_heads is None:
|
||||
num_key_value_heads = num_attention_heads
|
||||
|
||||
self.num_key_value_heads = num_key_value_heads
|
||||
self.initializer_range = initializer_range
|
||||
self.rms_norm_eps = rms_norm_eps
|
||||
self.pretraining_tp = pretraining_tp
|
||||
self.use_cache = use_cache
|
||||
self.rope_theta = rope_theta
|
||||
self.rope_scaling = rope_scaling
|
||||
self.attention_bias = attention_bias
|
||||
self.attention_dropout = attention_dropout
|
||||
self.mla_scale_q_lora = mla_scale_q_lora
|
||||
self.mla_scale_kv_lora = mla_scale_kv_lora
|
||||
self.zero_expert_num = zero_expert_num
|
||||
self.zero_expert_type = zero_expert_type
|
||||
self.routed_scaling_factor = routed_scaling_factor
|
||||
self.hidden_act = "silu"
|
||||
self.intermediate_size = self.ffn_hidden_size if hasattr(
|
||||
self, "ffn_hidden_size") else self.intermediate_size
|
||||
if hasattr(self, "moe_intermediate_size"):
|
||||
self.moe_intermediate_size = self.moe_intermediate_size
|
||||
elif hasattr(self, "expert_ffn_hidden_size"):
|
||||
self.moe_intermediate_size = self.expert_ffn_hidden_size
|
||||
else:
|
||||
self.moe_intermediate_size = self.intermediate_size
|
||||
|
||||
|
||||
class FlashMLP(nn.Module):
|
||||
"""Flash MLP layer."""
|
||||
|
||||
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: torch.Tensor) -> torch.Tensor:
|
||||
if x.numel() == 0:
|
||||
return x
|
||||
|
||||
gate_up, _ = self.gate_up_proj(x)
|
||||
x = self.act_fn(gate_up)
|
||||
x, _ = self.down_proj(x)
|
||||
return x
|
||||
|
||||
|
||||
class LongcatRouter(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
config,
|
||||
zero_expert_num=0,
|
||||
rounter_params_dtype=torch.bfloat16,
|
||||
prefix: str = ""):
|
||||
super().__init__()
|
||||
self.n_routed_experts = config.n_routed_experts if hasattr(
|
||||
config, "n_routed_experts") else config.num_experts[0]
|
||||
self.n_routed_experts = self.n_routed_experts + zero_expert_num
|
||||
self.classifier = ReplicatedLinear(
|
||||
config.hidden_size,
|
||||
self.n_routed_experts,
|
||||
bias=config.router_bias,
|
||||
params_dtype=rounter_params_dtype,
|
||||
quant_config=None,
|
||||
prefix=f"{prefix}.classifier",
|
||||
)
|
||||
self.e_score_correction_bias = nn.Parameter(
|
||||
torch.zeros((self.n_routed_experts), dtype=rounter_params_dtype))
|
||||
|
||||
def forward(self, hidden_states):
|
||||
logits, _ = self.classifier(hidden_states)
|
||||
return logits
|
||||
|
||||
|
||||
class LongcatMoe(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: FlashConfig,
|
||||
num_experts: int,
|
||||
top_k: int,
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
params_dtype: Optional[torch.dtype] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
enable_eplb: bool = False,
|
||||
):
|
||||
super().__init__()
|
||||
self.hidden_size = hidden_size
|
||||
self.zero_expert_num = config.zero_expert_num
|
||||
self.zero_expert_type = config.zero_expert_type
|
||||
self.routed_scaling_factor = config.routed_scaling_factor
|
||||
self.enable_eplb = enable_eplb
|
||||
# Gate always runs at half / full precision for now.
|
||||
self.rounter_params_dtype = params_dtype
|
||||
if config.router_dtype == "float32":
|
||||
self.rounter_params_dtype = torch.float32
|
||||
|
||||
self.router = LongcatRouter(
|
||||
config=config,
|
||||
zero_expert_num=self.zero_expert_num,
|
||||
rounter_params_dtype=self.rounter_params_dtype,
|
||||
prefix=f"{prefix}.gate")
|
||||
|
||||
self.experts = FusedMoE(
|
||||
num_experts=num_experts,
|
||||
top_k=top_k,
|
||||
hidden_size=hidden_size,
|
||||
intermediate_size=intermediate_size,
|
||||
reduce_results=True,
|
||||
params_dtype=params_dtype,
|
||||
e_score_correction_bias=self.router.e_score_correction_bias,
|
||||
renormalize=False,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.experts",
|
||||
zero_expert_num=self.zero_expert_num,
|
||||
zero_expert_type=self.zero_expert_type,
|
||||
enable_eplb=self.enable_eplb,
|
||||
routed_scaling_factor=config.routed_scaling_factor,
|
||||
)
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
num_tokens, hidden_dim = hidden_states.shape
|
||||
hidden_states = hidden_states.view(-1, hidden_dim)
|
||||
|
||||
router_logits = self.router(hidden_states.to(
|
||||
self.rounter_params_dtype))
|
||||
final_hidden_states = self.experts(hidden_states=hidden_states,
|
||||
router_logits=router_logits)
|
||||
|
||||
return final_hidden_states.view(num_tokens, hidden_dim)
|
||||
|
||||
|
||||
class FlashDecoderLayer(nn.Module):
|
||||
"""Flash decoder layer with dual attention and MLP structure."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: FlashConfig,
|
||||
cache_config: Optional[CacheConfig] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = "",
|
||||
enable_eplb: bool = False,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.layer_idx = int(prefix.split(sep='.')[-1])
|
||||
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)
|
||||
if rope_scaling is not None and getattr(
|
||||
config, "original_max_position_embeddings", None):
|
||||
rope_scaling["original_max_position_embeddings"] = (
|
||||
config.original_max_position_embeddings)
|
||||
|
||||
# Dual attention structure
|
||||
self.self_attn = nn.ModuleList([
|
||||
DeepseekV2MLAAttention(
|
||||
config=config,
|
||||
hidden_size=self.hidden_size,
|
||||
num_heads=config.num_attention_heads,
|
||||
qk_nope_head_dim=config.qk_nope_head_dim,
|
||||
qk_rope_head_dim=config.qk_rope_head_dim,
|
||||
v_head_dim=config.v_head_dim,
|
||||
q_lora_rank=(config.q_lora_rank if hasattr(
|
||||
config, "q_lora_rank") else None),
|
||||
kv_lora_rank=config.kv_lora_rank,
|
||||
rope_theta=rope_theta,
|
||||
rope_scaling=rope_scaling,
|
||||
max_position_embeddings=max_position_embeddings,
|
||||
cache_config=cache_config,
|
||||
quant_config=None if "self_attn" in getattr(
|
||||
config, "disable_quant_module", []) else quant_config,
|
||||
prefix=f"{prefix}.self_attn.{i}",
|
||||
) for i in range(2)
|
||||
])
|
||||
self.input_layernorm = nn.ModuleList([
|
||||
RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
for i in range(2)
|
||||
])
|
||||
self.post_attention_layernorm = nn.ModuleList([
|
||||
RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
for i in range(2)
|
||||
])
|
||||
|
||||
# Dual MLP structure
|
||||
self.mlps = nn.ModuleList([
|
||||
FlashMLP(
|
||||
hidden_size=self.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
hidden_act=config.hidden_act,
|
||||
quant_config=None if "mlps" in getattr(
|
||||
config, "disable_quant_module", []) else quant_config,
|
||||
prefix=f"{prefix}.mlps.{i}",
|
||||
) for i in range(2)
|
||||
])
|
||||
|
||||
self.mlp = LongcatMoe(
|
||||
config=config,
|
||||
num_experts=config.n_routed_experts if hasattr(
|
||||
config, "n_routed_experts") else
|
||||
config.num_experts[self.layer_idx],
|
||||
top_k=config.moe_topk
|
||||
if hasattr(config, "moe_topk") else config.num_experts_per_tok,
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.moe_intermediate_size,
|
||||
quant_config=quant_config,
|
||||
prefix=(f"{prefix}.mlp"),
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
positions: torch.Tensor,
|
||||
hidden_states: torch.Tensor,
|
||||
residual: Optional[torch.Tensor],
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
|
||||
if residual is None:
|
||||
residual = hidden_states
|
||||
hidden_states = self.input_layernorm[0](hidden_states)
|
||||
else:
|
||||
hidden_states, residual = self.input_layernorm[0](hidden_states,
|
||||
residual)
|
||||
|
||||
hidden_states = self.self_attn[0](
|
||||
positions=positions,
|
||||
hidden_states=hidden_states,
|
||||
)
|
||||
|
||||
hidden_states, residual = self.post_attention_layernorm[0](
|
||||
hidden_states, residual)
|
||||
|
||||
# moe
|
||||
hidden_states_copy = hidden_states.clone()
|
||||
moe_hidden_states = self.mlp(hidden_states_copy)
|
||||
|
||||
# first mlp
|
||||
hidden_states = self.mlps[0](hidden_states)
|
||||
|
||||
hidden_states, residual = self.input_layernorm[1](hidden_states,
|
||||
residual)
|
||||
|
||||
# second_attn
|
||||
hidden_states = self.self_attn[1](
|
||||
positions=positions,
|
||||
hidden_states=hidden_states,
|
||||
)
|
||||
hidden_states, residual = self.post_attention_layernorm[1](
|
||||
hidden_states, residual)
|
||||
|
||||
# second_mlp
|
||||
hidden_states = self.mlps[1](hidden_states)
|
||||
|
||||
hidden_states = hidden_states + moe_hidden_states
|
||||
|
||||
return hidden_states, residual
|
||||
|
||||
|
||||
@support_torch_compile
|
||||
class FlashModel(nn.Module):
|
||||
"""Flash model."""
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
config = FlashConfig(**vllm_config.model_config.hf_config.__dict__)
|
||||
cache_config = vllm_config.cache_config
|
||||
quant_config = vllm_config.quant_config
|
||||
self.config = config
|
||||
|
||||
self.padding_idx = getattr(config, "pad_token_id", None)
|
||||
self.vocab_size = config.vocab_size
|
||||
|
||||
if get_pp_group().is_first_rank:
|
||||
self.embed_tokens = VocabParallelEmbedding(
|
||||
config.vocab_size,
|
||||
config.hidden_size,
|
||||
prefix=maybe_prefix(prefix, "embed_tokens"),
|
||||
)
|
||||
else:
|
||||
self.embed_tokens = PPMissingLayer()
|
||||
self.start_layer, self.end_layer, self.layers = make_layers(
|
||||
config.num_hidden_layers,
|
||||
lambda prefix: FlashDecoderLayer(
|
||||
config,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
prefix=prefix,
|
||||
),
|
||||
prefix=f"{prefix}.layers")
|
||||
if get_pp_group().is_last_rank:
|
||||
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
else:
|
||||
self.norm = PPMissingLayer()
|
||||
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
|
||||
|
||||
|
||||
class LongcatFlashForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
|
||||
"""Flash model for causal language modeling."""
|
||||
|
||||
packed_modules_mapping = {
|
||||
"qkv_proj": [
|
||||
"q_proj",
|
||||
"k_proj",
|
||||
"v_proj",
|
||||
],
|
||||
"gate_up_proj": [
|
||||
"gate_proj",
|
||||
"up_proj",
|
||||
],
|
||||
}
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
config = FlashConfig(**vllm_config.model_config.hf_config.__dict__)
|
||||
quant_config = vllm_config.quant_config
|
||||
lora_config = vllm_config.lora_config
|
||||
|
||||
self.config = config
|
||||
config.intermediate_size = config.ffn_hidden_size if hasattr(
|
||||
config, "ffn_hidden_size") else config.intermediate_size
|
||||
self.lora_config = lora_config
|
||||
self.quant_config = quant_config
|
||||
|
||||
self.model = FlashModel(vllm_config=vllm_config,
|
||||
prefix=maybe_prefix(prefix, "model"))
|
||||
|
||||
if get_pp_group().is_last_rank:
|
||||
self.lm_head = ParallelLMHead(config.vocab_size,
|
||||
config.hidden_size,
|
||||
quant_config=quant_config,
|
||||
prefix=maybe_prefix(
|
||||
prefix, "lm_head"))
|
||||
else:
|
||||
self.lm_head = PPMissingLayer()
|
||||
|
||||
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,
|
||||
) -> 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,
|
||||
) -> Optional[torch.Tensor]:
|
||||
logits = self.logits_processor(self.lm_head, hidden_states)
|
||||
return logits
|
||||
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return 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.n_routed_experts if hasattr(
|
||||
self.config, "n_routed_experts") else
|
||||
self.config.num_experts[0],
|
||||
)
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
|
||||
stacked_params_mapping = [
|
||||
("fused_qkv_a_proj", "q_a_proj", 0),
|
||||
("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1),
|
||||
(".gate_up_proj", ".gate_proj", 0),
|
||||
(".gate_up_proj", ".up_proj", 1),
|
||||
]
|
||||
|
||||
expert_params_mapping = self.get_expert_mapping()
|
||||
loaded_params: set[str] = set()
|
||||
|
||||
params_dict = dict(self.named_parameters())
|
||||
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" in name and "mlps" not 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 mtp
|
||||
if ".mtp." in name:
|
||||
continue
|
||||
if is_pp_missing_parameter(name, self):
|
||||
continue
|
||||
param = params_dict[name]
|
||||
weight_loader = param.weight_loader
|
||||
weight_loader(param, loaded_weight, shard_id)
|
||||
break
|
||||
else:
|
||||
is_expert_weight = False
|
||||
for mapping in expert_params_mapping:
|
||||
param_name, weight_name, expert_id, shard_id = mapping
|
||||
if weight_name not in name:
|
||||
continue
|
||||
is_expert_weight = True
|
||||
name_mapped = name.replace(weight_name, param_name)
|
||||
# Skip mtp
|
||||
if ".mtp." in name_mapped:
|
||||
continue
|
||||
if (name_mapped.endswith(".bias")
|
||||
or name_mapped.endswith("_bias")
|
||||
) and name not in params_dict:
|
||||
continue
|
||||
if is_pp_missing_parameter(name, self):
|
||||
continue
|
||||
param = params_dict[name_mapped]
|
||||
weight_loader = param.weight_loader
|
||||
weight_loader = typing.cast(Callable[..., bool],
|
||||
param.weight_loader)
|
||||
success = weight_loader(param,
|
||||
loaded_weight,
|
||||
name_mapped,
|
||||
shard_id=shard_id,
|
||||
expert_id=expert_id,
|
||||
return_success=True)
|
||||
if success:
|
||||
name = name_mapped
|
||||
break
|
||||
else:
|
||||
if is_expert_weight:
|
||||
# We've checked that this is an expert weight
|
||||
# However it's not mapped locally to this rank
|
||||
# So we simply skip it
|
||||
continue
|
||||
# Skip loading extra bias for GPTQ models.
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
# Skip loading kv_scale from ckpts towards new design.
|
||||
if name.endswith(".kv_scale") and name not in params_dict:
|
||||
continue
|
||||
# Skip mtp
|
||||
if ".mtp." in name:
|
||||
continue
|
||||
if name is None:
|
||||
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)
|
||||
for layer_id in range(self.config.num_hidden_layers):
|
||||
for i in range(2):
|
||||
if isinstance(self.model.layers[layer_id], PPMissingLayer):
|
||||
continue
|
||||
self_attn = self.model.layers[layer_id].self_attn[i]
|
||||
if hasattr(self.quant_config, "weight_block_size"
|
||||
) and self_attn.kv_b_proj.weight.dtype in (
|
||||
torch.float8_e4m3fn,
|
||||
torch.float8_e4m3fnuz,
|
||||
):
|
||||
weight_block_size = self.quant_config.weight_block_size
|
||||
if weight_block_size is not None:
|
||||
assert hasattr(self_attn.kv_b_proj, "weight_scale_inv")
|
||||
dtype = torch.get_default_dtype()
|
||||
w = block_dequant(self_attn.kv_b_proj.weight,
|
||||
self_attn.kv_b_proj.weight_scale_inv,
|
||||
weight_block_size).to(dtype)
|
||||
else:
|
||||
w = self_attn.kv_b_proj.weight
|
||||
|
||||
w_kc, w_vc = w.unflatten(
|
||||
0,
|
||||
(-1,
|
||||
self_attn.qk_nope_head_dim + self_attn.v_head_dim)).split(
|
||||
[self_attn.qk_nope_head_dim, self_attn.v_head_dim],
|
||||
dim=1)
|
||||
self_attn.w_kc = w_kc.transpose(1, 2).contiguous().transpose(
|
||||
1, 2)
|
||||
self_attn.w_vc = w_vc.contiguous().transpose(1, 2)
|
||||
if self.config.mla_scale_q_lora:
|
||||
self_attn.q_a_layernorm.weight.data *= (
|
||||
self.config.hidden_size / self.config.q_lora_rank)**0.5
|
||||
if self.config.mla_scale_kv_lora:
|
||||
self_attn.kv_a_layernorm.weight.data *= (
|
||||
self.config.hidden_size /
|
||||
self.config.kv_lora_rank)**0.5
|
||||
return loaded_params
|
||||
@@ -0,0 +1,352 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
# Adapted from
|
||||
# https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/model_executor/models/deepseek_mtp.py
|
||||
from collections.abc import Iterable
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.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.quantization.utils.int8_utils import (
|
||||
block_dequant)
|
||||
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.longcat_flash import FlashConfig
|
||||
from vllm.sequence import IntermediateTensors
|
||||
|
||||
from .deepseek_v2 import DeepseekV2DecoderLayer
|
||||
from .interfaces import SupportsPP
|
||||
from .utils import maybe_prefix
|
||||
|
||||
|
||||
class LongCatMultiTokenPredictorLayer(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
config: PretrainedConfig,
|
||||
prefix: str,
|
||||
vllm_config: VllmConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
self.eh_proj = ReplicatedLinear(2 * config.hidden_size,
|
||||
config.hidden_size,
|
||||
bias=False,
|
||||
quant_config=quant_config,
|
||||
prefix="eh_proj")
|
||||
self.mtp_block = DeepseekV2DecoderLayer(vllm_config, prefix)
|
||||
self.final_layernorm = RMSNorm(config.hidden_size,
|
||||
eps=config.rms_norm_eps)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
previous_hidden_states: torch.Tensor,
|
||||
inputs_embeds: Optional[torch.Tensor] = None,
|
||||
spec_step_index: int = 0,
|
||||
) -> torch.Tensor:
|
||||
assert inputs_embeds is not None
|
||||
inputs_embeds = self.enorm(inputs_embeds)
|
||||
previous_hidden_states = self.hnorm(previous_hidden_states)
|
||||
|
||||
hidden_states, _ = self.eh_proj(
|
||||
torch.cat([inputs_embeds, previous_hidden_states], dim=-1))
|
||||
|
||||
hidden_states, residual = self.mtp_block(positions=positions,
|
||||
hidden_states=hidden_states,
|
||||
residual=None)
|
||||
hidden_states, _ = self.final_layernorm(hidden_states, residual)
|
||||
return hidden_states
|
||||
|
||||
|
||||
class LongCatMultiTokenPredictor(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
*,
|
||||
vllm_config: VllmConfig,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = ""):
|
||||
super().__init__()
|
||||
config = FlashConfig(**vllm_config.model_config.hf_config.__dict__)
|
||||
vllm_config.model_config.hf_config.intermediate_size \
|
||||
= config.intermediate_size
|
||||
self.mtp_start_layer_idx = config.num_hidden_layers * 2
|
||||
self.num_mtp_layers = 1
|
||||
self.layers = torch.nn.ModuleDict({
|
||||
str(idx):
|
||||
LongCatMultiTokenPredictorLayer(
|
||||
config,
|
||||
prefix=f"{prefix}.layers.{idx}",
|
||||
vllm_config=vllm_config,
|
||||
quant_config=quant_config,
|
||||
)
|
||||
for idx in range(self.mtp_start_layer_idx,
|
||||
self.mtp_start_layer_idx + self.num_mtp_layers)
|
||||
})
|
||||
self.embed_tokens = VocabParallelEmbedding(
|
||||
config.vocab_size,
|
||||
config.hidden_size,
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
previous_hidden_states: torch.Tensor,
|
||||
inputs_embeds: Optional[torch.Tensor] = None,
|
||||
spec_step_idx: int = 0,
|
||||
) -> torch.Tensor:
|
||||
if inputs_embeds is None:
|
||||
inputs_embeds = self.embed_tokens(input_ids)
|
||||
current_step_idx = (spec_step_idx % self.num_mtp_layers)
|
||||
return self.layers[str(self.mtp_start_layer_idx + current_step_idx)](
|
||||
input_ids,
|
||||
positions,
|
||||
previous_hidden_states,
|
||||
inputs_embeds,
|
||||
current_step_idx,
|
||||
)
|
||||
|
||||
|
||||
class LongCatFlashMTP(nn.Module, SupportsPP):
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
||||
super().__init__()
|
||||
# LongCat MTP without MoE layers
|
||||
vllm_config.model_config.hf_config.n_routed_experts = None
|
||||
self.config = FlashConfig(
|
||||
**vllm_config.model_config.hf_config.__dict__)
|
||||
self.quant_config = None if "mtp" in getattr(
|
||||
self.config, "disable_quant_module",
|
||||
[]) else vllm_config.quant_config
|
||||
|
||||
self.model = LongCatMultiTokenPredictor(vllm_config=vllm_config,
|
||||
quant_config=self.quant_config,
|
||||
prefix=maybe_prefix(
|
||||
prefix, "model"))
|
||||
self.lm_head = ParallelLMHead(
|
||||
self.config.vocab_size,
|
||||
self.config.hidden_size,
|
||||
quant_config=self.quant_config,
|
||||
)
|
||||
self.logits_processor = LogitsProcessor(self.config.vocab_size)
|
||||
|
||||
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:
|
||||
hidden_states = self.model(input_ids, positions, hidden_states,
|
||||
inputs_embeds, spec_step_idx)
|
||||
return hidden_states
|
||||
|
||||
def compute_logits(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
spec_step_idx: int = 0,
|
||||
) -> Optional[torch.Tensor]:
|
||||
|
||||
logits = self.logits_processor(self.lm_head, hidden_states)
|
||||
return logits
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
stacked_params_mapping = [
|
||||
("gate_up_proj", "gate_proj", 0),
|
||||
("gate_up_proj", "up_proj", 1),
|
||||
("fused_qkv_a_proj", "q_a_proj", 0),
|
||||
("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1),
|
||||
]
|
||||
|
||||
new_to_old_names_mapping = {
|
||||
"model.mtp.embed_tokens.weight":
|
||||
"model.layers.0.embed_tokens.weight",
|
||||
"model.mtp.layers.0.eh_proj.weight": "eh_proj.weight",
|
||||
"model.mtp.layers.0.eh_proj.weight_scale_inv":
|
||||
"eh_proj.weight_scale_inv",
|
||||
"model.mtp.layers.0.enorm.m.weight": "enorm.weight",
|
||||
"model.mtp.layers.0.hnorm.m.weight": "hnorm.weight",
|
||||
"model.mtp.layers.0.input_layernorm.weight":
|
||||
"model.layers.0.input_layernorm.weight",
|
||||
"model.mtp.layers.0.post_attention_layernorm.weight":
|
||||
"model.layers.0.post_attention_layernorm.weight",
|
||||
"model.mtp.layers.0.self_attn.kv_a_layernorm.weight":
|
||||
"model.layers.0.self_attn.kv_a_layernorm.weight",
|
||||
"model.mtp.layers.0.self_attn.kv_a_proj_with_mqa.weight":
|
||||
"model.layers.0.self_attn.kv_a_proj_with_mqa.weight",
|
||||
"model.mtp.layers.0.self_attn.kv_a_proj_with_mqa.weight_scale_inv":
|
||||
"model.layers.0.self_attn.kv_a_proj_with_mqa.weight_scale_inv",
|
||||
"model.mtp.layers.0.self_attn.kv_b_proj.weight":
|
||||
"model.layers.0.self_attn.kv_b_proj.weight",
|
||||
"model.mtp.layers.0.self_attn.kv_b_proj.weight_scale_inv":
|
||||
"model.layers.0.self_attn.kv_b_proj.weight_scale_inv",
|
||||
"model.mtp.layers.0.self_attn.o_proj.weight":
|
||||
"model.layers.0.self_attn.o_proj.weight",
|
||||
"model.mtp.layers.0.self_attn.o_proj.weight_scale_inv":
|
||||
"model.layers.0.self_attn.o_proj.weight_scale_inv",
|
||||
"model.mtp.layers.0.self_attn.q_a_layernorm.weight":
|
||||
"model.layers.0.self_attn.q_a_layernorm.weight",
|
||||
"model.mtp.layers.0.self_attn.q_a_proj.weight":
|
||||
"model.layers.0.self_attn.q_a_proj.weight",
|
||||
"model.mtp.layers.0.self_attn.q_a_proj.weight_scale_inv":
|
||||
"model.layers.0.self_attn.q_a_proj.weight_scale_inv",
|
||||
"model.mtp.layers.0.self_attn.q_b_proj.weight":
|
||||
"model.layers.0.self_attn.q_b_proj.weight",
|
||||
"model.mtp.layers.0.self_attn.q_b_proj.weight_scale_inv":
|
||||
"model.layers.0.self_attn.q_b_proj.weight_scale_inv",
|
||||
"model.mtp.layers.0.transformer_layer.mlp.down_proj.weight":
|
||||
"model.layers.0.mlp.down_proj.weight",
|
||||
"model.mtp.layers.0.transformer_layer.mlp.down_proj.weight_scale_inv":
|
||||
"model.layers.0.mlp.down_proj.weight_scale_inv",
|
||||
"model.mtp.layers.0.transformer_layer.mlp.gate_proj.weight":
|
||||
"model.layers.0.mlp.gate_proj.weight",
|
||||
"model.mtp.layers.0.transformer_layer.mlp.gate_proj.weight_scale_inv":
|
||||
"model.layers.0.mlp.gate_proj.weight_scale_inv",
|
||||
"model.mtp.layers.0.transformer_layer.mlp.up_proj.weight":
|
||||
"model.layers.0.mlp.up_proj.weight",
|
||||
"model.mtp.layers.0.transformer_layer.mlp.up_proj.weight_scale_inv":
|
||||
"model.layers.0.mlp.up_proj.weight_scale_inv",
|
||||
"model.mtp.norm.weight": "final_layernorm.weight",
|
||||
}
|
||||
|
||||
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
|
||||
spec_layer = self.get_spec_layer_idx_from_weight_name(
|
||||
self.config, name)
|
||||
if spec_layer is None:
|
||||
continue
|
||||
name = self._rewrite_spec_layer_name(spec_layer, name,
|
||||
new_to_old_names_mapping)
|
||||
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) and name not in params_dict):
|
||||
continue
|
||||
name = name.replace(weight_name, param_name)
|
||||
|
||||
# QKV fusion is optional, fall back to normal
|
||||
# weight loading if it's not enabled
|
||||
if ((param_name == "fused_qkv_a_proj")
|
||||
and name not in params_dict):
|
||||
continue
|
||||
|
||||
# Skip loading extra bias for GPTQ models.
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
|
||||
param = params_dict[name]
|
||||
weight_loader = param.weight_loader
|
||||
weight_loader(param, loaded_weight, shard_id)
|
||||
break
|
||||
else:
|
||||
# Skip loading extra bias for GPTQ models.
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
|
||||
# According to DeepSeek-V3 Technical Report, MTP modules
|
||||
# shares embedding layer. We only load the first weights.
|
||||
if (spec_layer != self.model.mtp_start_layer_idx
|
||||
and ".layers" not in name):
|
||||
continue
|
||||
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader(param, loaded_weight)
|
||||
loaded_params.add(name)
|
||||
spec_layer_id = self.config.num_hidden_layers * 2
|
||||
self_attn = self.model.layers[str(spec_layer_id)].mtp_block.self_attn
|
||||
if hasattr(
|
||||
self.quant_config,
|
||||
"weight_block_size") and self_attn.kv_b_proj.weight.dtype in (
|
||||
torch.float8_e4m3fn,
|
||||
torch.float8_e4m3fnuz,
|
||||
):
|
||||
weight_block_size = self.quant_config.weight_block_size
|
||||
if weight_block_size is not None:
|
||||
dtype = torch.get_default_dtype()
|
||||
w = block_dequant(self_attn.kv_b_proj.weight,
|
||||
self_attn.kv_b_proj.weight_scale_inv,
|
||||
weight_block_size).to(dtype)
|
||||
else:
|
||||
w = self_attn.kv_b_proj.weight
|
||||
else:
|
||||
w = self_attn.kv_b_proj.weight
|
||||
w_kc, w_vc = w.unflatten(
|
||||
0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)).split(
|
||||
[self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)
|
||||
self_attn.w_kc = w_kc.transpose(1, 2).contiguous().transpose(1, 2)
|
||||
self_attn.w_vc = w_vc.contiguous().transpose(1, 2)
|
||||
if self.config.mla_scale_q_lora:
|
||||
self_attn.q_a_layernorm.weight.data *= (
|
||||
self.config.hidden_size / self.config.q_lora_rank)**0.5
|
||||
if self.config.mla_scale_kv_lora:
|
||||
self_attn.kv_a_layernorm.weight.data *= (
|
||||
self.config.hidden_size / self.config.kv_lora_rank)**0.5
|
||||
return loaded_params
|
||||
|
||||
def _rewrite_spec_layer_name(self, spec_layer: int, name: str,
|
||||
new_to_old_names_mapping: dict) -> str:
|
||||
"""
|
||||
Rewrite the weight name to match the format of the original model.
|
||||
Add .mtp_block for modules in transformer layer block for spec layer
|
||||
and rename shared layer weights to be top level.
|
||||
"""
|
||||
if name in new_to_old_names_mapping:
|
||||
name = new_to_old_names_mapping[name]
|
||||
spec_layer_weight_names = [
|
||||
"embed_tokens", "enorm", "hnorm", "eh_proj", "shared_head"
|
||||
]
|
||||
if name.startswith("enorm") or name.startswith(
|
||||
"hnorm") or name.startswith("eh_proj") or name.startswith(
|
||||
"final_layernorm"):
|
||||
name = "model.layers." + str(spec_layer) + "." + name
|
||||
shared_weight_names = ["embed_tokens"]
|
||||
spec_layer_weight = False
|
||||
shared_weight = False
|
||||
for weight_name in spec_layer_weight_names:
|
||||
if weight_name in name:
|
||||
spec_layer_weight = True
|
||||
if weight_name in shared_weight_names:
|
||||
shared_weight = True
|
||||
break
|
||||
if not spec_layer_weight:
|
||||
# treat rest weights as weights for transformer layer block
|
||||
name = name.replace("model.layers.0.",
|
||||
f"model.layers.{spec_layer}.mtp_block.")
|
||||
elif shared_weight:
|
||||
# treat shared weights as top level weights
|
||||
name = name.replace("model.layers.0.", "model.")
|
||||
return name
|
||||
|
||||
def get_spec_layer_idx_from_weight_name(self, config: PretrainedConfig,
|
||||
weight_name: str) -> Optional[int]:
|
||||
if "model.mtp" in weight_name:
|
||||
return config.num_hidden_layers * 2
|
||||
return None
|
||||
@@ -109,6 +109,7 @@ _TEXT_GENERATION_MODELS = {
|
||||
"Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"), # noqa: E501
|
||||
# For decapoda-research/llama-*
|
||||
"LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
|
||||
"LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
|
||||
"MambaForCausalLM": ("mamba", "MambaForCausalLM"),
|
||||
"FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
|
||||
"FalconH1ForCausalLM":("falcon_h1", "FalconH1ForCausalLM"),
|
||||
@@ -287,6 +288,7 @@ _SPECULATIVE_DECODING_MODELS = {
|
||||
"EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
|
||||
"DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
|
||||
"ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
|
||||
"LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
|
||||
"Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
|
||||
"MedusaModel": ("medusa", "Medusa"),
|
||||
"Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
|
||||
|
||||
@@ -691,14 +691,14 @@ def maybe_prefix(prefix: str, name: str) -> str:
|
||||
return name if not prefix else f"{prefix}.{name}"
|
||||
|
||||
|
||||
def extract_layer_index(layer_name: str) -> int:
|
||||
def extract_layer_index(layer_name: str, num_attn_module: int = 1) -> int:
|
||||
"""
|
||||
Extract the layer index from the module name.
|
||||
Examples:
|
||||
- "encoder.layers.0" -> 0
|
||||
- "encoder.layers.1.self_attn" -> 1
|
||||
- "2.self_attn" -> 2
|
||||
- "model.encoder.layers.0.sub.1" -> ValueError
|
||||
- "model.encoder.layers.0.sub.1" -> ValueError if num_attn_module == 1
|
||||
"""
|
||||
subnames = layer_name.split(".")
|
||||
int_vals: list[int] = []
|
||||
@@ -707,9 +707,17 @@ def extract_layer_index(layer_name: str) -> int:
|
||||
int_vals.append(int(subname))
|
||||
except ValueError:
|
||||
continue
|
||||
assert len(int_vals) == 1, (f"layer name {layer_name} should"
|
||||
" only contain one integer")
|
||||
return int_vals[0]
|
||||
if num_attn_module == 1 or "attn" not in layer_name:
|
||||
assert len(int_vals) == 1, (f"layer name {layer_name} should"
|
||||
" only contain one integer")
|
||||
|
||||
return int_vals[0]
|
||||
else:
|
||||
assert len(int_vals) <= 2, (f"layer name {layer_name} should"
|
||||
" contain most two integers")
|
||||
layer_index = int_vals[0] * num_attn_module + int_vals[1] if len(
|
||||
int_vals) == 2 else int_vals[0]
|
||||
return layer_index
|
||||
|
||||
|
||||
def cast_overflow_tensors(
|
||||
|
||||
Reference in New Issue
Block a user