mirror of
https://git.datalinker.icu/vllm-project/vllm.git
synced 2025-12-20 01:15:01 +08:00
292 lines
12 KiB
Python
292 lines
12 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
"""PyTorch MAMBA2 model."""
|
|
from collections.abc import Iterable
|
|
from typing import Optional
|
|
|
|
import torch
|
|
from torch import nn
|
|
from transformers import MambaConfig
|
|
|
|
from vllm.compilation.decorators import support_torch_compile
|
|
from vllm.config import CacheConfig, ModelConfig, VllmConfig
|
|
from vllm.distributed.parallel_state import get_pp_group
|
|
from vllm.model_executor.layers.layernorm import RMSNorm
|
|
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
|
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
|
|
from vllm.model_executor.layers.mamba.mamba_utils import (
|
|
MambaStateDtypeCalculator, MambaStateShapeCalculator)
|
|
from vllm.model_executor.layers.quantization.base_config import (
|
|
QuantizationConfig)
|
|
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
|
DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
|
|
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
|
from vllm.model_executor.models.interfaces import (HasInnerState,
|
|
IsAttentionFree)
|
|
from vllm.sequence import IntermediateTensors
|
|
|
|
from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
|
|
make_empty_intermediate_tensors_factory, make_layers,
|
|
maybe_prefix)
|
|
|
|
KVCache = tuple[torch.Tensor, torch.Tensor]
|
|
|
|
|
|
class Mamba2DecoderLayer(nn.Module):
|
|
|
|
def __init__(self,
|
|
config: MambaConfig,
|
|
model_config: Optional[ModelConfig] = None,
|
|
cache_config: Optional[CacheConfig] = None,
|
|
quant_config: Optional[QuantizationConfig] = None,
|
|
prefix: str = "") -> None:
|
|
super().__init__()
|
|
self.config = config
|
|
self.mixer = MambaMixer2(hidden_size=config.hidden_size,
|
|
ssm_state_size=config.state_size,
|
|
conv_kernel_size=config.conv_kernel,
|
|
intermediate_size=getattr(
|
|
config, "intermediate_size",
|
|
config.expand * config.hidden_size),
|
|
use_conv_bias=config.use_conv_bias,
|
|
use_bias=config.use_bias,
|
|
n_groups=config.n_groups,
|
|
num_heads=config.num_heads,
|
|
head_dim=config.head_dim,
|
|
rms_norm_eps=config.layer_norm_epsilon,
|
|
activation=config.hidden_act,
|
|
model_config=model_config,
|
|
cache_config=cache_config,
|
|
quant_config=quant_config,
|
|
prefix=f"{prefix}.mixer")
|
|
|
|
self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
residual: Optional[torch.Tensor],
|
|
**kwargs,
|
|
):
|
|
if residual is None:
|
|
residual = hidden_states
|
|
hidden_states = self.norm(hidden_states)
|
|
else:
|
|
hidden_states, residual = self.norm(hidden_states, residual)
|
|
|
|
output = torch.empty_like(hidden_states)
|
|
self.mixer(hidden_states, output)
|
|
return output, residual
|
|
|
|
|
|
@support_torch_compile
|
|
class Mamba2Model(nn.Module):
|
|
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
|
super().__init__()
|
|
|
|
config = vllm_config.model_config.hf_config
|
|
model_config = vllm_config.model_config
|
|
cache_config = vllm_config.cache_config
|
|
quant_config = vllm_config.quant_config
|
|
lora_config = vllm_config.lora_config
|
|
is_lora_enabled = bool(lora_config)
|
|
assert not is_lora_enabled
|
|
|
|
self.config = config
|
|
lora_vocab = ((lora_config.lora_extra_vocab_size *
|
|
(lora_config.max_loras or 1)) if lora_config else 0)
|
|
self.vocab_size = config.vocab_size + lora_vocab
|
|
self.org_vocab_size = config.vocab_size
|
|
|
|
self.embeddings = VocabParallelEmbedding(
|
|
self.vocab_size,
|
|
config.hidden_size,
|
|
org_num_embeddings=config.vocab_size,
|
|
)
|
|
|
|
self.start_layer, self.end_layer, self.layers = make_layers(
|
|
config.num_hidden_layers,
|
|
lambda prefix: Mamba2DecoderLayer(config,
|
|
model_config=model_config,
|
|
cache_config=cache_config,
|
|
quant_config=quant_config,
|
|
prefix=prefix),
|
|
prefix=f"{prefix}.layers")
|
|
|
|
self.norm_f = RMSNorm(config.hidden_size,
|
|
eps=config.layer_norm_epsilon)
|
|
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.embeddings(input_ids)
|
|
|
|
def forward(
|
|
self,
|
|
input_ids: torch.Tensor,
|
|
positions: torch.Tensor,
|
|
intermediate_tensors: Optional[IntermediateTensors] = None,
|
|
inputs_embeds: Optional[torch.Tensor] = None,
|
|
) -> torch.Tensor:
|
|
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, layer in enumerate(self.layers):
|
|
hidden_states, residual = layer(positions=positions,
|
|
hidden_states=hidden_states,
|
|
residual=residual)
|
|
|
|
if not get_pp_group().is_last_rank:
|
|
return IntermediateTensors({
|
|
"hidden_states": hidden_states,
|
|
"residual": residual
|
|
})
|
|
|
|
hidden_states, _ = self.norm_f(hidden_states, residual)
|
|
|
|
return hidden_states
|
|
|
|
def load_weights(self, weights: Iterable[tuple[str,
|
|
torch.Tensor]]) -> set[str]:
|
|
params_dict = dict(self.named_parameters())
|
|
loaded_params: set[str] = set()
|
|
for name, loaded_weight in weights:
|
|
if "A_log" in name:
|
|
name = name.replace("A_log", "A")
|
|
|
|
# Skip loading extra bias for GPTQ models.
|
|
if name.endswith(".bias") and name not in params_dict:
|
|
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)
|
|
return loaded_params
|
|
|
|
|
|
class Mamba2ForCausalLM(nn.Module, HasInnerState, IsAttentionFree):
|
|
|
|
@classmethod
|
|
def get_mamba_state_dtype_from_config(
|
|
cls,
|
|
vllm_config: "VllmConfig",
|
|
) -> tuple[torch.dtype, torch.dtype]:
|
|
|
|
return MambaStateDtypeCalculator.mamba2_state_dtype(
|
|
vllm_config.model_config.dtype,
|
|
vllm_config.cache_config.mamba_cache_dtype,
|
|
vllm_config.cache_config.mamba_ssm_cache_dtype,
|
|
)
|
|
|
|
@classmethod
|
|
def get_mamba_state_shape_from_config(
|
|
cls,
|
|
vllm_config: "VllmConfig",
|
|
) -> tuple[tuple[int, int], tuple[int, int, int]]:
|
|
"""Calculate shapes for Mamba's convolutional and state caches.
|
|
|
|
Args:
|
|
vllm_config: vLLM config
|
|
|
|
Returns:
|
|
Tuple containing:
|
|
- conv_state_shape: Shape for convolutional state cache
|
|
- temporal_state_shape: Shape for state space model cache
|
|
"""
|
|
parallel_config = vllm_config.parallel_config
|
|
hf_config = vllm_config.model_config.hf_config
|
|
intermediate_size = hf_config.expand * hf_config.hidden_size
|
|
|
|
return MambaStateShapeCalculator.mamba2_state_shape(
|
|
intermediate_size=intermediate_size,
|
|
tp_world_size=parallel_config.tensor_parallel_size,
|
|
n_groups=hf_config.n_groups,
|
|
num_heads=hf_config.num_heads,
|
|
head_dim=hf_config.head_dim,
|
|
state_size=hf_config.state_size,
|
|
conv_kernel=hf_config.conv_kernel,
|
|
)
|
|
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
|
config = vllm_config.model_config.hf_config
|
|
cache_config = vllm_config.cache_config
|
|
lora_config = vllm_config.lora_config
|
|
scheduler_config = vllm_config.scheduler_config
|
|
assert not cache_config.enable_prefix_caching, \
|
|
"Mamba does not support prefix caching"
|
|
|
|
super().__init__()
|
|
self.config = config
|
|
self.vllm_config = vllm_config
|
|
self.scheduler_config = scheduler_config
|
|
self.model_config = vllm_config.model_config
|
|
self.backbone = Mamba2Model(vllm_config=vllm_config,
|
|
prefix=maybe_prefix(prefix, "backbone"))
|
|
self.unpadded_vocab_size = config.vocab_size
|
|
if lora_config:
|
|
self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
|
|
|
|
self.lm_head = ParallelLMHead(
|
|
self.unpadded_vocab_size,
|
|
config.hidden_size,
|
|
org_num_embeddings=config.vocab_size,
|
|
padding_size=DEFAULT_VOCAB_PADDING_SIZE
|
|
# We need bigger padding if using lora for kernel
|
|
# compatibility
|
|
if not lora_config else lora_config.lora_vocab_padding_size,
|
|
prefix=maybe_prefix(prefix, "lm_head"),
|
|
)
|
|
if config.tie_word_embeddings:
|
|
self.lm_head = self.lm_head.tie_weights(self.backbone.embeddings)
|
|
|
|
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
|
|
config.vocab_size)
|
|
|
|
self.make_empty_intermediate_tensors = (
|
|
self.backbone.make_empty_intermediate_tensors)
|
|
|
|
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
|
|
return self.backbone.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,
|
|
**kwargs):
|
|
|
|
hidden_states = self.backbone(input_ids, positions,
|
|
intermediate_tensors, inputs_embeds)
|
|
|
|
return hidden_states
|
|
|
|
def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs):
|
|
return self.mamba_cache.copy_inputs_before_cuda_graphs(
|
|
input_buffers, **kwargs)
|
|
|
|
def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
|
|
return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size)
|
|
|
|
def compute_logits(self, hidden_states: torch.Tensor) -> 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]:
|
|
loader = AutoWeightsLoader(self)
|
|
return loader.load_weights(weights)
|