[Model] Support Mamba (#6484)

This commit is contained in:
Tyler Michael Smith
2024-10-11 15:40:06 +00:00
committed by GitHub
parent df3dcdf49d
commit 7342a7d7f8
29 changed files with 1603 additions and 343 deletions
+44 -1
View File
@@ -271,7 +271,7 @@ class HasInnerState(Protocol):
"""
A flag that indicates this model has inner state.
Models that has inner state usually need access to the scheduler_config
for max_num_seqs ,etc... (Currently only used by Jamba)
for max_num_seqs, etc. True for e.g. both Mamba and Jamba.
"""
def __init__(self,
@@ -307,3 +307,46 @@ def has_inner_state(
return isinstance(model, _HasInnerStateType)
return isinstance(model, HasInnerState)
@runtime_checkable
class IsAttentionFree(Protocol):
"""The interface required for all models like Mamba that lack attention,
but do have state whose size is constant wrt the number of tokens."""
is_attention_free: ClassVar[Literal[True]] = True
"""
A flag that indicates this model has no attention.
Used for block manager and attention backend selection.
True for Mamba but not Jamba.
"""
def __init__(self) -> None:
...
@runtime_checkable
class _IsAttentionFreeType(Protocol):
is_attention_free: ClassVar[Literal[True]]
def __init__(self) -> None:
...
@overload
def is_attention_free(model: object) -> TypeIs[IsAttentionFree]:
...
@overload
def is_attention_free(model: Type[object]) -> TypeIs[Type[IsAttentionFree]]:
...
def is_attention_free(
model: Union[Type[object], object]
) -> Union[TypeIs[Type[IsAttentionFree]], TypeIs[IsAttentionFree]]:
if isinstance(model, type):
return isinstance(model, _IsAttentionFreeType)
return isinstance(model, IsAttentionFree)
+32 -229
View File
@@ -1,18 +1,16 @@
# coding=utf-8
"""Inference-only Jamba model."""
from dataclasses import dataclass
from typing import Dict, Iterable, List, Optional, Tuple
from typing import Iterable, List, Optional, Tuple
import torch
from torch import nn
from torch.nn.parameter import Parameter
from transformers import JambaConfig
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.attention.layer import Attention
from vllm.config import CacheConfig, LoRAConfig, SchedulerConfig
from vllm.distributed import (get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size)
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
@@ -29,7 +27,9 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
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.model_loader.weight_utils import (
composed_weight_loader, default_weight_loader, sharded_weight_loader)
from vllm.model_executor.models.mamba_cache import MambaCacheManager
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.model_executor.utils import set_weight_attrs
from vllm.sequence import IntermediateTensors
@@ -99,16 +99,6 @@ class JambaMambaMixer(nn.Module):
bias=True,
skip_bias_add=True)
def weight_loader(param: Parameter, loaded_weight: torch.Tensor):
tp_rank = get_tensor_model_parallel_rank()
tp_size = get_tensor_model_parallel_world_size()
param.data.copy_(
loaded_weight.data.split(loaded_weight.shape[0] // tp_size,
dim=0)[tp_rank])
def A_weight_loader(param: Parameter, loaded_weight: torch.Tensor):
weight_loader(param, -torch.exp(loaded_weight.float()))
tp_size = get_tensor_model_parallel_world_size()
self.A = nn.Parameter(
torch.empty(
@@ -118,8 +108,10 @@ class JambaMambaMixer(nn.Module):
))
self.D = nn.Parameter(torch.ones(self.intermediate_size // tp_size))
set_weight_attrs(self.D, {"weight_loader": weight_loader})
set_weight_attrs(self.A, {"weight_loader": A_weight_loader})
set_weight_attrs(self.D, {"weight_loader": sharded_weight_loader(0)})
a_weight_loader = composed_weight_loader(
sharded_weight_loader(0), lambda x: -torch.exp(x.float()))
set_weight_attrs(self.A, {"weight_loader": a_weight_loader})
self.out_proj = RowParallelLinear(
self.intermediate_size,
@@ -571,10 +563,8 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA):
if not lora_config else lora_config.lora_vocab_padding_size,
)
# Used to track and store by the Mamba cache between steps.
self.mamba_cache: Tuple[torch.Tensor, torch.Tensor] = tuple()
# Maps between the request id and a dict that maps between the seq_id
# and its index inside the self.mamba_cache
self.mamba_cache_indices_mapping: Dict[str, Dict[int, int]] = {}
self.mamba_cache: Optional[MambaCacheManager] = None
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
config.vocab_size)
self.sampler = Sampler()
@@ -586,203 +576,36 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA):
attn_metadata: AttentionMetadata,
intermediate_tensors: Optional[IntermediateTensors] = None,
**kwargs):
if not self.mamba_cache:
self._prepare_mamba_cache()
if self.mamba_cache is None:
max_batch_size = (_get_graph_batch_size(
self.scheduler_config.max_num_seqs) if self.scheduler_config
else max(_BATCH_SIZES_TO_CAPTURE) + 2)
if "seqlen_agnostic_capture_inputs" not in kwargs:
# We get here only on Prefill/Eager mode runs
request_ids_to_seq_ids = kwargs["request_ids_to_seq_ids"]
finished_requests_ids = kwargs["finished_requests_ids"]
mamba_cache = self._release_finished_and_prepare_mamba_cache(
finished_requests_ids, request_ids_to_seq_ids)
else:
# CUDA graph capturing runs
mamba_cache = kwargs["seqlen_agnostic_capture_inputs"]
layers_type = self.config.layers_block_type
num_mamba_layers = sum(
[layer_type == "mamba" for layer_type in layers_type])
self.mamba_cache = MambaCacheManager(
self.lm_head.weight.dtype, num_mamba_layers, max_batch_size,
*self._get_mamba_cache_shape())
mamba_cache_tensors = self.mamba_cache.current_run_tensors(
input_ids, attn_metadata, **kwargs)
hidden_states = self.model(input_ids, positions, kv_caches,
attn_metadata, mamba_cache[0],
mamba_cache[1])
attn_metadata, mamba_cache_tensors[0],
mamba_cache_tensors[1])
return hidden_states
def _swap_mamba_cache(self, from_index: int, to_index: int):
assert len(self.mamba_cache) > 0
for cache_t in self.mamba_cache:
cache_t[:, [to_index,from_index]] = \
cache_t[:, [from_index,to_index]]
def _copy_mamba_cache(self, from_index: int, to_index: int):
assert len(self.mamba_cache) > 0
for cache_t in self.mamba_cache:
cache_t[:, to_index].copy_(cache_t[:, from_index],
non_blocking=True)
def _move_out_if_already_occupied(self, index: int,
all_occupied_indices: List[int]):
if index in all_occupied_indices:
first_free_index = self._first_free_index_in_mamba_cache()
# In case occupied, move the occupied to a new empty block
self._move_cache_index_and_mappings(from_index=index,
to_index=first_free_index)
def _assign_seq_id_to_mamba_cache_in_specific_dest(self, cur_rid: str,
seq_id: int,
destination_index: int):
"""
Assign (req_id,seq_id) pair to a `destination_index` index, if
already occupied, move the occupying index to a free index.
"""
all_occupied_indices = self._get_all_occupied_indices()
if cur_rid not in self.mamba_cache_indices_mapping:
self._move_out_if_already_occupied(
index=destination_index,
all_occupied_indices=all_occupied_indices)
self.mamba_cache_indices_mapping[cur_rid] = {
seq_id: destination_index
}
elif seq_id not in (seq_ids2indices :=
self.mamba_cache_indices_mapping[cur_rid]):
# parallel sampling , where n > 1, assume prefill have
# already happened now we only need to copy the already
# existing cache into the siblings seq_ids caches
self._move_out_if_already_occupied(
index=destination_index,
all_occupied_indices=all_occupied_indices)
index_exists = list(seq_ids2indices.values())[0]
# case of decoding n>1, copy prefill cache to decoding indices
self._copy_mamba_cache(from_index=index_exists,
to_index=destination_index)
self.mamba_cache_indices_mapping[cur_rid][
seq_id] = destination_index
else:
# already exists
cache_index_already_exists = self.mamba_cache_indices_mapping[
cur_rid][seq_id]
if cache_index_already_exists != destination_index:
# In case the seq id already exists but not in
# the right destination, swap it with what's occupying it
self._swap_pair_indices_and_mappings(
from_index=cache_index_already_exists,
to_index=destination_index)
def _prepare_current_run_mamba_cache(
self, request_ids_to_seq_ids: Dict[str, list[int]],
finished_requests_ids: List[str]
) -> Tuple[torch.Tensor, torch.Tensor]:
running_indices = []
request_ids_to_seq_ids_flatten = [
(req_id, seq_id)
for req_id, seq_ids in request_ids_to_seq_ids.items()
for seq_id in seq_ids
]
batch_size = len(request_ids_to_seq_ids_flatten)
for dest_index, (request_id,
seq_id) in enumerate(request_ids_to_seq_ids_flatten):
if request_id in finished_requests_ids:
# Do not allocate cache index for requests that run
# and finish right after
continue
self._assign_seq_id_to_mamba_cache_in_specific_dest(
request_id, seq_id, dest_index)
running_indices.append(dest_index)
self._clean_up_first_bs_blocks(batch_size, running_indices)
conv_state = self.mamba_cache[0][:, :batch_size]
temporal_state = self.mamba_cache[1][:, :batch_size]
return (conv_state, temporal_state)
def _get_all_occupied_indices(self):
return [
cache_idx
for seq_ids2indices in self.mamba_cache_indices_mapping.values()
for cache_idx in seq_ids2indices.values()
]
def _clean_up_first_bs_blocks(self, batch_size: int,
indices_for_current_run: List[int]):
# move out all of the occupied but currently not running blocks
# outside of the first n blocks
destination_indices = range(batch_size)
max_possible_batch_size = self.mamba_cache[0].shape[1]
for destination_index in destination_indices:
if destination_index in self._get_all_occupied_indices() and \
destination_index not in indices_for_current_run:
# move not running indices outside of the batch
all_other_indices = list(
range(batch_size, max_possible_batch_size))
first_avail_index = self._first_free_index_in_mamba_cache(
all_other_indices)
self._swap_indices(from_index=destination_index,
to_index=first_avail_index)
def _move_cache_index_and_mappings(self, from_index: int, to_index: int):
self._copy_mamba_cache(from_index=from_index, to_index=to_index)
self._update_mapping_index(from_index=from_index, to_index=to_index)
def _swap_pair_indices_and_mappings(self, from_index: int, to_index: int):
self._swap_mamba_cache(from_index=from_index, to_index=to_index)
self._swap_mapping_index(from_index=from_index, to_index=to_index)
def _swap_mapping_index(self, from_index: int, to_index: int):
for seq_ids2index in self.mamba_cache_indices_mapping.values():
for seq_id, index in seq_ids2index.items():
if from_index == index:
seq_ids2index.update({seq_id: to_index})
elif to_index == index:
seq_ids2index.update({seq_id: from_index})
def _update_mapping_index(self, from_index: int, to_index: int):
for seq_ids2index in self.mamba_cache_indices_mapping.values():
for seq_id, index in seq_ids2index.items():
if from_index == index:
seq_ids2index.update({seq_id: to_index})
return
def _release_finished_and_prepare_mamba_cache(
self, finished_requests_ids,
request_ids_to_seq_ids) -> Tuple[torch.Tensor, torch.Tensor]:
self._release_mamba_cache(finished_requests_ids)
return self._prepare_current_run_mamba_cache(request_ids_to_seq_ids,
finished_requests_ids)
def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs):
"""
Copy the relevant Mamba cache into the CUDA graph input buffer
that was provided during the capture runs
(JambaForCausalLM.mamba_gc_cache_buffer).
"""
self._release_finished_and_prepare_mamba_cache(
kwargs["finished_requests_ids"], kwargs["request_ids_to_seq_ids"])
return self.mamba_cache.copy_inputs_before_cuda_graphs(
input_buffers, **kwargs)
def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
"""
Provide the CUDA graph capture runs with a buffer in adjusted size.
The buffer is used to maintain the Mamba Cache during the CUDA graph
replay runs.
"""
return tuple(buffer[:, :batch_size] for buffer in self.mamba_cache)
def _release_mamba_cache(self, finished_seq_groups_req_ids: List[str]):
for req_id in finished_seq_groups_req_ids:
if req_id in self.mamba_cache_indices_mapping:
self.mamba_cache_indices_mapping.pop(req_id)
def _first_free_index_in_mamba_cache(
self, indices_range: Optional[List[int]] = None) -> int:
assert self.mamba_cache is not None
if indices_range is None:
max_possible_batch_size = self.mamba_cache[0].shape[1]
indices_range = list(range(max_possible_batch_size))
all_occupied_indices = self._get_all_occupied_indices()
for i in indices_range:
if i not in all_occupied_indices:
return i
raise Exception("Couldn't find a free spot in the mamba cache! This"
"should never happen")
return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size)
def _get_mamba_cache_shape(
self
) -> Tuple[Optional[Tuple[int, int]], Optional[Tuple[int, int]]]:
self) -> Tuple[Tuple[int, int], Tuple[int, int]]:
world_size = get_tensor_model_parallel_world_size()
hidden_size = self.config.hidden_size
conv_state_shape = (
@@ -790,31 +613,11 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA):
self.config.mamba_d_conv - 1,
)
temporal_state_shape = (
self.config.mamba_expand * self.config.hidden_size // world_size,
self.config.mamba_expand * hidden_size // world_size,
self.config.mamba_d_state,
)
return conv_state_shape, temporal_state_shape
def _prepare_mamba_cache(self):
dtype = self.lm_head.weight.dtype
layers_type = self.config.layers_block_type
mamba_layers = sum(
[layer_type == "mamba" for layer_type in layers_type])
max_batch_size = (_get_graph_batch_size(
self.scheduler_config.max_num_seqs) if self.scheduler_config else
max(_BATCH_SIZES_TO_CAPTURE) + 2)
conv_state_shape, temporal_state_shape = self._get_mamba_cache_shape()
assert conv_state_shape is not None and temporal_state_shape is not None
self.mamba_cache = (torch.empty(size=(mamba_layers, max_batch_size) +
conv_state_shape,
dtype=dtype,
device="cuda"),
torch.empty(size=(mamba_layers, max_batch_size) +
temporal_state_shape,
dtype=dtype,
device="cuda"))
def compute_logits(
self,
hidden_states: torch.Tensor,
+499
View File
@@ -0,0 +1,499 @@
# coding=utf-8
"""PyTorch MAMBA model."""
from dataclasses import dataclass
from typing import Iterable, List, Optional, Tuple
import torch
from torch import nn
from transformers import MambaConfig
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.config import CacheConfig, LoRAConfig, SchedulerConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
MergedColumnParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
causal_conv1d_fn, causal_conv1d_update)
from vllm.model_executor.layers.mamba.ops.mamba_ssm import (
selective_scan_fn, selective_state_update)
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig)
from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import (
composed_weight_loader, default_weight_loader, sharded_weight_loader)
from vllm.model_executor.models.interfaces import (HasInnerState,
IsAttentionFree)
from vllm.model_executor.models.mamba_cache import MambaCacheManager
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.model_executor.utils import set_weight_attrs
from vllm.sequence import IntermediateTensors
from vllm.worker.model_runner import (_BATCH_SIZES_TO_CAPTURE,
_get_graph_batch_size)
KVCache = Tuple[torch.Tensor, torch.Tensor]
@dataclass
class MambaCacheParams:
is_prompt: bool = False
conv_state: torch.Tensor = torch.Tensor()
ssm_state: torch.Tensor = torch.Tensor()
# Adapted from transformers.models.mamba.modeling_mamba.MambaMixer
class MambaMixer(nn.Module):
"""
Compute ∆, A, B, C, and D the state space parameters and compute
the `contextualized_states`. A, D are input independent
(see Mamba paper [1] Section 3.5.2 "Interpretation of A"
for why A isn't selective) ∆, B, C are input-dependent
(this is a key difference between Mamba and the linear time
invariant S4, and is why Mamba is called
**selective** state spaces)
"""
def __init__(self, config: MambaConfig, layer_idx):
super().__init__()
self.config = config
self.layer_idx = layer_idx
self.hidden_size = config.hidden_size
self.ssm_state_size = config.state_size
self.conv_kernel_size = config.conv_kernel
self.intermediate_size = config.intermediate_size
self.time_step_rank = int(config.time_step_rank)
self.conv1d = ColumnParallelLinear(
input_size=self.conv_kernel_size,
output_size=self.intermediate_size,
bias=config.use_conv_bias,
)
# unsqueeze to fit conv1d weights shape into the linear weights shape.
# Can't do this in `weight_loader` since it already exists in
# `ColumnParallelLinear` and `set_weight_attrs`
# doesn't allow to override it
self.conv1d.weight.data = self.conv1d.weight.data.unsqueeze(1)
self.in_proj = MergedColumnParallelLinear(self.hidden_size,
[self.intermediate_size] * 2,
bias=config.use_bias)
# selective projection used to make dt, B and C input dependent
self.x_proj = RowParallelLinear(
self.intermediate_size,
self.time_step_rank + self.ssm_state_size * 2,
bias=False,
)
# time step projection (discretization) -
# In the forward we need to apply dt_proj without the bias,
# as the bias is added in the selective scan kernel.
self.dt_proj = ColumnParallelLinear(self.time_step_rank,
self.intermediate_size,
bias=True,
skip_bias_add=True)
tp_size = get_tensor_model_parallel_world_size()
self.A = nn.Parameter(
torch.empty(
self.intermediate_size // tp_size,
self.ssm_state_size,
dtype=torch.float32,
))
self.D = nn.Parameter(torch.ones(self.intermediate_size // tp_size))
set_weight_attrs(self.D, {"weight_loader": sharded_weight_loader(0)})
a_weight_loader = composed_weight_loader(
sharded_weight_loader(0), lambda x: -torch.exp(x.float()))
set_weight_attrs(self.A, {"weight_loader": a_weight_loader})
self.out_proj = RowParallelLinear(
self.intermediate_size,
self.hidden_size,
bias=config.use_bias,
input_is_parallel=True,
)
self.activation = config.hidden_act
def forward(self, hidden_states: torch.Tensor,
attn_metadata: AttentionMetadata, conv_state: torch.Tensor,
ssm_state: torch.Tensor):
# 1. Gated MLP's linear projection
projected_states = self.in_proj(hidden_states)[0].transpose(-2, -1)
hidden_states, gate = projected_states.chunk(2, dim=-2)
# 2. Convolution sequence transformation
conv_weights = self.conv1d.weight.view(self.conv1d.weight.size(0),
self.conv1d.weight.size(2))
if attn_metadata.query_start_loc is not None \
and attn_metadata.context_lens_tensor is not None:
# |---------- N-1 iteration --------|
# |---------------- N iteration ---------------------|
# |- tokenA -|......................|-- newTokens ---|
# |---------- context_len ----------|
# |-------------------- seq_len ---------------------|
# |-- query_len ---|
hidden_states = causal_conv1d_fn(
hidden_states,
conv_weights,
self.conv1d.bias,
activation=self.activation,
conv_states=conv_state,
has_initial_state=attn_metadata.context_lens_tensor > 0,
query_start_loc=attn_metadata.query_start_loc)
else:
hidden_states = causal_conv1d_update(
hidden_states.transpose(0, 1),
conv_state,
conv_weights,
self.conv1d.bias,
self.activation,
)
hidden_states = hidden_states.transpose(0, 1)
# 3. State Space Model sequence transformation
# 3.a. input varying initialization of time_step, B and C
ssm_parameters = self.x_proj(hidden_states.transpose(-2, -1))[0]
time_step, B, C = torch.split(
ssm_parameters,
[self.time_step_rank, self.ssm_state_size, self.ssm_state_size],
dim=-1,
)
# Note that Jamba normalizes B, C, and time_step here but Mamba doesn't.
discrete_time_step = self.dt_proj(time_step)[0].transpose(-2, -1)
# 3.c perform the recurrence y ← SSM(A, B, C)(x)
time_proj_bias = (self.dt_proj.bias.float() if hasattr(
self.dt_proj, "bias") else None)
if attn_metadata.query_start_loc is not None \
and attn_metadata.context_lens_tensor is not None:
scan_outputs = selective_scan_fn(
hidden_states,
ssm_state,
discrete_time_step,
self.A,
B.transpose(-2, -1),
C.transpose(-2, -1),
self.D.float(),
gate,
time_proj_bias,
delta_softplus=True,
has_initial_state=attn_metadata.context_lens_tensor > 0,
query_start_loc=attn_metadata.query_start_loc)
else:
scan_outputs = selective_state_update(
ssm_state,
hidden_states.transpose(0, 1),
discrete_time_step.transpose(0, 1),
self.A,
B,
C,
self.D,
gate.transpose(0, 1),
time_proj_bias,
dt_softplus=True,
)
scan_outputs = scan_outputs.transpose(0, 1)
# 4. Final linear projection
contextualized_states = self.out_proj(scan_outputs.transpose(-2,
-1))[0]
return contextualized_states
class MambaMLP(nn.Module):
def __init__(
self,
config: MambaConfig,
quant_config: Optional[QuantizationConfig] = None,
) -> None:
super().__init__()
hidden_size = config.hidden_size
intermediate_size = config.intermediate_size
hidden_act = config.hidden_act
self.gate_up_proj = MergedColumnParallelLinear(
hidden_size, [intermediate_size] * 2,
bias=False,
quant_config=quant_config)
self.down_proj = RowParallelLinear(intermediate_size,
hidden_size,
bias=False,
quant_config=quant_config)
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 MambaDecoderLayer(nn.Module):
def __init__(self,
config: MambaConfig,
layer_idx: int,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None) -> None:
super().__init__()
self.layer_idx = layer_idx
self.config = config
self.mixer = MambaMixer(config, layer_idx)
self.feed_forward = MambaMLP(config, quant_config=quant_config)
self.norm = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
self.pre_ff_layernorm = RMSNorm(config.hidden_size,
eps=config.layer_norm_epsilon)
def forward(
self,
hidden_states: torch.Tensor,
attn_metadata: AttentionMetadata,
residual: Optional[torch.Tensor],
conv_state: torch.Tensor,
ssm_state: 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)
hidden_states = self.mixer(hidden_states, attn_metadata, conv_state,
ssm_state)
# Fully Connected
hidden_states, residual = self.pre_ff_layernorm(
hidden_states, residual)
hidden_states = self.feed_forward(hidden_states)
return hidden_states, residual
class MambaModel(nn.Module):
def __init__(
self,
config: MambaConfig,
quant_config: Optional[QuantizationConfig] = None,
cache_config: Optional[CacheConfig] = None,
lora_config: Optional[LoRAConfig] = None,
) -> None:
super().__init__()
self.config = config
self.padding_idx = config.pad_token_id
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,
)
decoder_layers = []
for i in range(config.num_hidden_layers):
decoder_layers.append(
MambaDecoderLayer(config,
layer_idx=i,
cache_config=cache_config,
quant_config=quant_config))
self.layers = nn.ModuleList(decoder_layers)
self.norm_f = RMSNorm(config.hidden_size,
eps=config.layer_norm_epsilon)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
kv_caches: List[torch.Tensor],
attn_metadata: AttentionMetadata,
conv_state: torch.Tensor,
ssm_state: torch.Tensor,
) -> torch.Tensor:
hidden_states = self.embeddings(input_ids)
residual = None
for i in range(len(self.layers)):
layer = self.layers[i]
current_ssm_state = ssm_state[i]
current_conv_state = conv_state[i]
hidden_states, residual = layer(
positions=positions,
hidden_states=hidden_states,
attn_metadata=attn_metadata,
residual=residual,
conv_state=current_conv_state,
ssm_state=current_ssm_state,
)
hidden_states, _ = self.norm_f(hidden_states, residual)
return hidden_states
class MambaForCausalLM(nn.Module, HasInnerState, IsAttentionFree):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
"k_proj",
"v_proj",
],
}
# LoRA specific attributes
supported_lora_modules = [
"qkv_proj",
"o_proj",
"embed_tokens",
"lm_head",
]
embedding_modules = {
"embeddings": "input_embeddings",
"lm_head": "output_embeddings",
}
embedding_padding_modules = ["lm_head"]
def __init__(
self,
config: MambaConfig,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
lora_config: Optional[LoRAConfig] = None,
scheduler_config: Optional[SchedulerConfig] = None,
) -> None:
assert not cache_config.enable_prefix_caching, \
"Mamba does not support prefix caching"
super().__init__()
self.config = config
self.scheduler_config = scheduler_config
self.backbone = MambaModel(config,
cache_config=cache_config,
quant_config=quant_config,
lora_config=lora_config)
self.unpadded_vocab_size = config.vocab_size
if lora_config:
self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
self.lm_head = self.backbone.embeddings
# Used to track and store by the Mamba cache between steps.
self.mamba_cache: Optional[MambaCacheManager] = None
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
config.vocab_size)
self.sampler = Sampler()
def forward(self,
input_ids: torch.Tensor,
positions: torch.Tensor,
kv_caches: List[KVCache],
attn_metadata: AttentionMetadata,
intermediate_tensors: Optional[IntermediateTensors] = None,
**kwargs):
if self.mamba_cache is None:
max_batch_size = (_get_graph_batch_size(
self.scheduler_config.max_num_seqs) if self.scheduler_config
else max(_BATCH_SIZES_TO_CAPTURE) + 2)
self.mamba_cache = MambaCacheManager(
self.lm_head.weight.dtype, self.config.num_hidden_layers,
max_batch_size, *self._get_mamba_cache_shape())
mamba_cache_tensors = self.mamba_cache.current_run_tensors(
input_ids, attn_metadata, **kwargs)
hidden_states = self.backbone(input_ids, positions, kv_caches,
attn_metadata, mamba_cache_tensors[0],
mamba_cache_tensors[1])
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 _get_mamba_cache_shape(
self) -> Tuple[Tuple[int, int], Tuple[int, int]]:
world_size = get_tensor_model_parallel_world_size()
conv_state_shape = (
self.config.intermediate_size // world_size,
self.config.conv_kernel - 1,
)
temporal_state_shape = (
self.config.intermediate_size // world_size,
self.config.state_size,
)
return conv_state_shape, temporal_state_shape
def compute_logits(self, hidden_states: torch.Tensor,
sampling_metadata: SamplingMetadata) -> 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_dict = dict(self.named_parameters())
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
if "A_log" in name:
name = name.replace("A_log", "A")
if ".self_attn." in name:
name = name.replace(".self_attn", "")
for param_name, weight_name, shard_id in stacked_params_mapping:
if weight_name not in name:
continue
name = name.replace(weight_name, param_name)
# 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
param = params_dict[name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
+222
View File
@@ -0,0 +1,222 @@
from typing import Dict, List, Optional
import torch
from vllm.attention.backends.abstract import AttentionMetadata
class MambaCacheManager:
def __init__(self, dtype, num_mamba_layers, max_batch_size,
conv_state_shape, temporal_state_shape):
conv_state = torch.empty(size=(num_mamba_layers, max_batch_size) +
conv_state_shape,
dtype=dtype,
device="cuda")
temporal_state = torch.empty(size=(num_mamba_layers, max_batch_size) +
temporal_state_shape,
dtype=dtype,
device="cuda")
self.mamba_cache = (conv_state, temporal_state)
# Maps between the request id and a dict that maps between the seq_id
# and its index inside the self.mamba_cache
self.mamba_cache_indices_mapping: Dict[str, Dict[int, int]] = {}
def current_run_tensors(self, input_ids: torch.Tensor,
attn_metadata: AttentionMetadata, **kwargs):
"""
Return the tensors for the current run's conv and ssm state.
"""
if "seqlen_agnostic_capture_inputs" not in kwargs:
# We get here only on Prefill/Eager mode runs
request_ids_to_seq_ids = kwargs["request_ids_to_seq_ids"]
finished_requests_ids = kwargs["finished_requests_ids"]
self._release_finished_requests(finished_requests_ids)
mamba_cache_tensors = self._prepare_current_run_mamba_cache(
request_ids_to_seq_ids, finished_requests_ids)
else:
# CUDA graph capturing runs
mamba_cache_tensors = kwargs["seqlen_agnostic_capture_inputs"]
return mamba_cache_tensors
def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs):
"""
Copy the relevant Mamba cache into the CUDA graph input buffer
that was provided during the capture runs
(JambaForCausalLM.mamba_gc_cache_buffer).
"""
assert all(
key in kwargs
for key in ["request_ids_to_seq_ids", "finished_requests_ids"])
finished_requests_ids = kwargs["finished_requests_ids"]
request_ids_to_seq_ids = kwargs["request_ids_to_seq_ids"]
self._release_finished_requests(finished_requests_ids)
self._prepare_current_run_mamba_cache(request_ids_to_seq_ids,
finished_requests_ids)
def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
"""
Provide the CUDA graph capture runs with a buffer in adjusted size.
The buffer is used to maintain the Mamba Cache during the CUDA graph
replay runs.
"""
return tuple(buffer[:, :batch_size] for buffer in self.mamba_cache)
def _swap_mamba_cache(self, from_index: int, to_index: int):
assert len(self.mamba_cache) > 0
for cache_t in self.mamba_cache:
cache_t[:, [to_index,from_index]] = \
cache_t[:, [from_index,to_index]]
def _copy_mamba_cache(self, from_index: int, to_index: int):
assert len(self.mamba_cache) > 0
for cache_t in self.mamba_cache:
cache_t[:, to_index].copy_(cache_t[:, from_index],
non_blocking=True)
def _move_out_if_already_occupied(self, index: int,
all_occupied_indices: List[int]):
if index in all_occupied_indices:
first_free_index = self._first_free_index_in_mamba_cache()
# In case occupied, move the occupied to a new empty block
self._move_cache_index_and_mappings(from_index=index,
to_index=first_free_index)
def _assign_seq_id_to_mamba_cache_in_specific_dest(self, cur_rid: str,
seq_id: int,
destination_index: int):
"""
Assign (req_id,seq_id) pair to a `destination_index` index, if
already occupied, move the occupying index to a free index.
"""
all_occupied_indices = self._get_all_occupied_indices()
if cur_rid not in self.mamba_cache_indices_mapping:
self._move_out_if_already_occupied(
index=destination_index,
all_occupied_indices=all_occupied_indices)
self.mamba_cache_indices_mapping[cur_rid] = {
seq_id: destination_index
}
elif seq_id not in (seq_ids2indices :=
self.mamba_cache_indices_mapping[cur_rid]):
# parallel sampling , where n > 1, assume prefill have
# already happened now we only need to copy the already
# existing cache into the siblings seq_ids caches
self._move_out_if_already_occupied(
index=destination_index,
all_occupied_indices=all_occupied_indices)
index_exists = list(seq_ids2indices.values())[0]
# case of decoding n>1, copy prefill cache to decoding indices
self._copy_mamba_cache(from_index=index_exists,
to_index=destination_index)
self.mamba_cache_indices_mapping[cur_rid][
seq_id] = destination_index
else:
# already exists
cache_index_already_exists = self.mamba_cache_indices_mapping[
cur_rid][seq_id]
if cache_index_already_exists != destination_index:
# In case the seq id already exists but not in
# the right destination, swap it with what's occupying it
self._swap_pair_indices_and_mappings(
from_index=cache_index_already_exists,
to_index=destination_index)
def _prepare_current_run_mamba_cache(
self, request_ids_to_seq_ids: Dict[str, list[int]],
finished_requests_ids: List[str]):
running_indices = []
request_ids_to_seq_ids_flatten = [
(req_id, seq_id)
for req_id, seq_ids in request_ids_to_seq_ids.items()
for seq_id in seq_ids
]
batch_size = len(request_ids_to_seq_ids_flatten)
for dest_index, (request_id,
seq_id) in enumerate(request_ids_to_seq_ids_flatten):
if request_id in finished_requests_ids:
# Do not allocate cache index for requests that run
# and finish right after
continue
self._assign_seq_id_to_mamba_cache_in_specific_dest(
request_id, seq_id, dest_index)
running_indices.append(dest_index)
self._clean_up_first_bs_blocks(batch_size, running_indices)
conv_state = self.mamba_cache[0][:, :batch_size]
temporal_state = self.mamba_cache[1][:, :batch_size]
return (conv_state, temporal_state)
def _get_all_occupied_indices(self):
return [
cache_idx
for seq_ids2indices in self.mamba_cache_indices_mapping.values()
for cache_idx in seq_ids2indices.values()
]
def _clean_up_first_bs_blocks(self, batch_size: int,
indices_for_current_run: List[int]):
# move out all of the occupied but currently not running blocks
# outside of the first n blocks
destination_indices = range(batch_size)
max_possible_batch_size = self.mamba_cache[0].shape[1]
for destination_index in destination_indices:
if destination_index in self._get_all_occupied_indices() and \
destination_index not in indices_for_current_run:
# move not running indices outside of the batch
all_other_indices = list(
range(batch_size, max_possible_batch_size))
first_avail_index = self._first_free_index_in_mamba_cache(
all_other_indices)
self._swap_indices(from_index=destination_index,
to_index=first_avail_index)
def _move_cache_index_and_mappings(self, from_index: int, to_index: int):
self._copy_mamba_cache(from_index=from_index, to_index=to_index)
self._update_mapping_index(from_index=from_index, to_index=to_index)
def _swap_pair_indices_and_mappings(self, from_index: int, to_index: int):
self._swap_mamba_cache(from_index=from_index, to_index=to_index)
self._swap_mapping_index(from_index=from_index, to_index=to_index)
def _swap_mapping_index(self, from_index: int, to_index: int):
for seq_ids2index in self.mamba_cache_indices_mapping.values():
for seq_id, index in seq_ids2index.items():
if from_index == index:
seq_ids2index.update({seq_id: to_index})
elif to_index == index:
seq_ids2index.update({seq_id: from_index})
def _update_mapping_index(self, from_index: int, to_index: int):
for seq_ids2index in self.mamba_cache_indices_mapping.values():
for seq_id, index in seq_ids2index.items():
if from_index == index:
seq_ids2index.update({seq_id: to_index})
return
def _release_finished_requests(self,
finished_seq_groups_req_ids: List[str]):
for req_id in finished_seq_groups_req_ids:
if req_id in self.mamba_cache_indices_mapping:
self.mamba_cache_indices_mapping.pop(req_id)
def _first_free_index_in_mamba_cache(
self, indices_range: Optional[List[int]] = None) -> int:
assert self.mamba_cache is not None
if indices_range is None:
max_possible_batch_size = self.mamba_cache[0].shape[1]
indices_range = list(range(max_possible_batch_size))
all_occupied_indices = self._get_all_occupied_indices()
for i in indices_range:
if i not in all_occupied_indices:
return i
raise Exception("Couldn't find a free spot in the mamba cache! This"
"should never happen")
+15 -1
View File
@@ -14,7 +14,8 @@ import torch.nn as nn
from vllm.logger import init_logger
from vllm.utils import is_hip
from .interfaces import supports_multimodal, supports_pp
from .interfaces import (has_inner_state, is_attention_free,
supports_multimodal, supports_pp)
from .interfaces_base import is_embedding_model, is_text_generation_model
logger = init_logger(__name__)
@@ -52,6 +53,7 @@ _TEXT_GENERATION_MODELS = {
"LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
# For decapoda-research/llama-*
"LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
"MambaForCausalLM": ("mamba", "MambaForCausalLM"),
"MistralForCausalLM": ("llama", "LlamaForCausalLM"),
"MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
"QuantMixtralForCausalLM": ("mixtral_quant", "MixtralForCausalLM"),
@@ -157,6 +159,8 @@ class _ModelInfo:
is_embedding_model: bool
supports_multimodal: bool
supports_pp: bool
has_inner_state: bool
is_attention_free: bool
@staticmethod
def from_model_cls(model: Type[nn.Module]) -> "_ModelInfo":
@@ -165,6 +169,8 @@ class _ModelInfo:
is_embedding_model=is_embedding_model(model),
supports_multimodal=supports_multimodal(model),
supports_pp=supports_pp(model),
has_inner_state=has_inner_state(model),
is_attention_free=is_attention_free(model),
)
@@ -380,6 +386,14 @@ class _ModelRegistry:
) -> bool:
return self.inspect_model_cls(architectures).supports_pp
def model_has_inner_state(self, architectures: Union[str,
List[str]]) -> bool:
return self.inspect_model_cls(architectures).has_inner_state
def is_attention_free_model(self, architectures: Union[str,
List[str]]) -> bool:
return self.inspect_model_cls(architectures).is_attention_free
ModelRegistry = _ModelRegistry({
model_arch: _LazyRegisteredModel(