EngineX-Cambricon/enginex-mlu370-vllm
10
0
Fork 1
Code Issues Pull Requests Actions Projects Releases Wiki Activity

add qwen3

Browse Source
This commit is contained in:
Chranos
2026-02-04 17:22:39 +08:00
parent d1c0f68ab4
commit 8511fe8530
1932 changed files with 300426 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
vllm-v0.6.2/vllm/model_executor/layers/mamba/__init__.py Normal file
View File

217
vllm-v0.6.2/vllm/model_executor/layers/mamba/mamba_mixer.py Normal file
View File

@@ -0,0 +1,217 @@
import torch
from torch import nn
from torch.nn.parameter import Parameter
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.distributed.parallel_state import (
get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
from vllm.model_executor.custom_op import CustomOp
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
MergedColumnParallelLinear,
RowParallelLinear)
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.models.mamba_cache import MambaCacheParams
from vllm.model_executor.utils import set_weight_attrs
# Adapted from transformers.models.mamba.modeling_mamba.MambaMixer
@CustomOp.register("mamba_mixer")
class MambaMixer(CustomOp):
"""
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,
hidden_size: int,
ssm_state_size: int,
conv_kernel_size: int,
intermediate_size: int,
time_step_rank: int,
use_conv_bias: bool,
use_bias: bool,
use_rms_norm: bool,
rms_norm_eps: float = 1e-5,
activation="silu"):
super().__init__()
self.time_step_rank = time_step_rank
self.ssm_state_size = ssm_state_size
self.use_rms_norm = use_rms_norm
self.activation = activation
self.conv1d = ColumnParallelLinear(
input_size=conv_kernel_size,
output_size=intermediate_size,
bias=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(hidden_size,
[intermediate_size] * 2,
bias=use_bias)
# selective projection used to make dt, B and C input dependent
self.x_proj = RowParallelLinear(
intermediate_size,
time_step_rank + 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(time_step_rank,
intermediate_size,
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(
intermediate_size // tp_size,
ssm_state_size,
dtype=torch.float32,
))
self.D = nn.Parameter(torch.ones(intermediate_size // tp_size))
set_weight_attrs(self.D, {"weight_loader": weight_loader})
set_weight_attrs(self.A, {"weight_loader": A_weight_loader})
self.out_proj = RowParallelLinear(
intermediate_size,
hidden_size,
bias=use_bias,
input_is_parallel=True,
)
self.dt_layernorm = RMSNorm(time_step_rank,
eps=rms_norm_eps) if use_rms_norm else None
self.b_layernorm = RMSNorm(ssm_state_size,
eps=rms_norm_eps) if use_rms_norm else None
self.c_layernorm = RMSNorm(ssm_state_size,
eps=rms_norm_eps) if use_rms_norm else None
def forward_native(self, hidden_states: torch.Tensor,
attn_metadata: AttentionMetadata,
conv_state: torch.Tensor, ssm_state: torch.Tensor):
pass
def forward_cuda(self, hidden_states: torch.Tensor,
attn_metadata: AttentionMetadata,
mamba_cache_params: MambaCacheParams):
# 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=mamba_cache_params.conv_state,
has_initial_state=attn_metadata.context_lens_tensor > 0,
cache_indices=mamba_cache_params.state_indices_tensor,
query_start_loc=attn_metadata.query_start_loc)
else:
hidden_states = causal_conv1d_update(
hidden_states.transpose(0, 1),
mamba_cache_params.conv_state,
conv_weights,
self.conv1d.bias,
self.activation,
conv_state_indices=mamba_cache_params.state_indices_tensor)
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,
)
if self.use_rms_norm:
assert self.dt_layernorm is not None
assert self.b_layernorm is not None
assert self.c_layernorm is not None
time_step = self.dt_layernorm(time_step.contiguous())
B = self.b_layernorm(B.contiguous())
C = self.c_layernorm(C.contiguous())
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,
mamba_cache_params.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,
cache_indices=mamba_cache_params.state_indices_tensor,
has_initial_state=attn_metadata.context_lens_tensor > 0,
query_start_loc=attn_metadata.query_start_loc)
else:
scan_outputs = selective_state_update(
mamba_cache_params.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,
state_batch_indices=mamba_cache_params.state_indices_tensor)
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

0
vllm-v0.6.2/vllm/model_executor/layers/mamba/ops/__init__.py Normal file
View File

102
vllm-v0.6.2/vllm/model_executor/layers/mamba/ops/causal_conv1d.py Normal file
View File

@@ -0,0 +1,102 @@
# Copyright (c) 2024, Tri Dao.
# Adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/causal_conv1d/causal_conv1d_interface.py
from typing import Optional
import torch
from vllm import _custom_ops as ops
from vllm.attention.backends.utils import PAD_SLOT_ID
def causal_conv1d_fn(x: torch.Tensor,
weight: torch.Tensor,
bias: Optional[torch.Tensor] = None,
query_start_loc: Optional[torch.Tensor] = None,
cache_indices: Optional[torch.Tensor] = None,
has_initial_state: Optional[torch.Tensor] = None,
conv_states: Optional[torch.Tensor] = None,
activation: Optional[str] = "silu",
pad_slot_id: int = PAD_SLOT_ID):
"""
x: (batch, dim, seqlen) or (dim,cu_seq_len) for varlen
sequences are concatenated from left to right for varlen
weight: (dim, width)
bias: (dim,)
query_start_loc: (batch + 1) int32
The cumulative sequence lengths of the sequences in
the batch, used to index into sequence. prepended by 0.
for example: query_start_loc = torch.Tensor([0,10,16,17]),
x.shape=(dim,17)
cache_indices: (batch) int32
indicates the corresponding state index,
like so: conv_state = conv_states[cache_indices[batch_id]]
has_initial_state: (batch) bool
indicates whether should the kernel take the current state as initial
state for the calculations
conv_states: (...,dim,width - 1) itype
updated inplace if provided
activation: either None or "silu" or "swish"
pad_slot_id: int
if cache_indices is passed, lets the kernel identify padded
entries that will not be processed,
for example: cache_indices = [pad_slot_id, 1, 20, pad_slot_id]
in this case, the kernel will not process entries at
indices 0 and 3
out: (batch, dim, seqlen)
"""
if activation not in [None, "silu", "swish"]:
raise NotImplementedError("activation must be None, silu, or swish")
if x.stride(-1) != 1:
x = x.contiguous()
bias = bias.contiguous() if bias is not None else None
ops.causal_conv1d_fwd(x, weight, bias, conv_states, query_start_loc,
cache_indices, has_initial_state, activation
in ["silu", "swish"], pad_slot_id)
return x
def causal_conv1d_update(x: torch.Tensor,
conv_state: torch.Tensor,
weight: torch.Tensor,
bias: Optional[torch.Tensor] = None,
activation: Optional[str] = None,
cache_seqlens: Optional[torch.Tensor] = None,
conv_state_indices: Optional[torch.Tensor] = None,
pad_slot_id: int = PAD_SLOT_ID):
"""
x: (batch, dim) or (batch, dim, seqlen)
conv_state: (batch, dim, state_len), where state_len >= width - 1
weight: (dim, width)
bias: (dim,)
cache_seqlens: (batch,), dtype int32.
If not None, the conv_state is treated as a circular buffer.
The conv_state will be updated by copying x to the conv_state
starting at the index
@cache_seqlens % state_len.
conv_state_indices: (batch,), dtype int32
If not None, the conv_state is a larger tensor along the batch dim,
and we are selecting the batch coords specified by conv_state_indices.
Useful for a continuous batching scenario.
pad_slot_id: int
if cache_indices is passed, lets the kernel identify padded
entries that will not be processed,
for example: cache_indices = [pad_slot_id, 1 ,20 ,pad_slot_id]
in this case, the kernel will not process entries at
indices 0 and 3
out: (batch, dim) or (batch, dim, seqlen)
"""
if activation not in [None, "silu", "swish"]:
raise NotImplementedError("activation must be None, silu, or swish")
activation_val = activation in ["silu", "swish"]
unsqueeze = x.dim() == 2
if unsqueeze:
x = x.unsqueeze(-1)
ops.causal_conv1d_update(x, conv_state, weight, bias, activation_val,
cache_seqlens, conv_state_indices, pad_slot_id)
if unsqueeze:
x = x.squeeze(-1)
return x

411
vllm-v0.6.2/vllm/model_executor/layers/mamba/ops/mamba_ssm.py Normal file
View File

@@ -0,0 +1,411 @@
# Copyright (c) 2024, Tri Dao, Albert Gu.
# Adapted from https://github.com/state-spaces/mamba/blob/main/mamba_ssm/ops/triton/selective_state_update.py
import torch
import triton
import triton.language as tl
from packaging import version
from vllm import _custom_ops as ops
from vllm.attention.backends.utils import PAD_SLOT_ID
TRITON3 = version.parse(triton.__version__) >= version.parse("3.0.0")
if TRITON3:
@triton.jit
def softplus(dt):
dt = tl.where(dt <= 20.0, tl.math.log(tl.math.exp(dt) + 1), dt)
return dt
else:
@triton.jit
def softplus(dt):
dt = tl.where(dt <= 20.0, tl.math.log1p(tl.exp(dt)), dt)
return dt
@triton.heuristics(
{"HAS_DT_BIAS": lambda args: args["dt_bias_ptr"] is not None})
@triton.heuristics({"HAS_D": lambda args: args["D_ptr"] is not None})
@triton.heuristics({"HAS_Z": lambda args: args["z_ptr"] is not None})
@triton.heuristics({
"HAS_STATE_BATCH_INDICES":
lambda args: args["state_batch_indices_ptr"] is not None
})
@triton.heuristics(
{"BLOCK_SIZE_DSTATE": lambda args: triton.next_power_of_2(args["dstate"])})
@triton.jit
def _selective_scan_update_kernel(
# Pointers to matrices
state_ptr,
x_ptr,
dt_ptr,
dt_bias_ptr,
A_ptr,
B_ptr,
C_ptr,
D_ptr,
z_ptr,
out_ptr,
state_batch_indices_ptr,
pad_slot_id,
# Matrix dimensions
batch,
nheads,
dim,
dstate,
nheads_ngroups_ratio,
# Strides
stride_state_batch,
stride_state_head,
stride_state_dim,
stride_state_dstate,
stride_x_batch,
stride_x_head,
stride_x_dim,
stride_dt_batch,
stride_dt_head,
stride_dt_dim,
stride_dt_bias_head,
stride_dt_bias_dim,
stride_A_head,
stride_A_dim,
stride_A_dstate,
stride_B_batch,
stride_B_group,
stride_B_dstate,
stride_C_batch,
stride_C_group,
stride_C_dstate,
stride_D_head,
stride_D_dim,
stride_z_batch,
stride_z_head,
stride_z_dim,
stride_out_batch,
stride_out_head,
stride_out_dim,
# Meta-parameters
DT_SOFTPLUS: tl.constexpr,
TIE_HDIM: tl.constexpr,
BLOCK_SIZE_M: tl.constexpr,
HAS_DT_BIAS: tl.constexpr,
HAS_D: tl.constexpr,
HAS_Z: tl.constexpr,
HAS_STATE_BATCH_INDICES: tl.constexpr,
BLOCK_SIZE_DSTATE: tl.constexpr,
):
pid_m = tl.program_id(axis=0)
pid_b = tl.program_id(axis=1)
pid_h = tl.program_id(axis=2)
# If HAS_STATE_BATCH_INDICES is true, then the ssm state's batch coordinate
# is taken from the state_batch_indices_ptr Otherwise, the state coordinate
# is the same as the batch id.
if HAS_STATE_BATCH_INDICES:
state_batch_indices_ptr += pid_b
state_batch_idx = tl.load(state_batch_indices_ptr)
state_ptr += (state_batch_idx * stride_state_batch +
pid_h * stride_state_head)
else:
state_ptr += pid_b * stride_state_batch + pid_h * stride_state_head
x_ptr += pid_b * stride_x_batch + pid_h * stride_x_head
dt_ptr += pid_b * stride_dt_batch + pid_h * stride_dt_head
if HAS_DT_BIAS:
dt_bias_ptr += pid_h * stride_dt_bias_head
A_ptr += pid_h * stride_A_head
B_ptr += pid_b * stride_B_batch + (pid_h //
nheads_ngroups_ratio) * stride_B_group
C_ptr += pid_b * stride_C_batch + (pid_h //
nheads_ngroups_ratio) * stride_C_group
if HAS_Z:
z_ptr += pid_b * stride_z_batch + pid_h * stride_z_head
out_ptr += pid_b * stride_out_batch + pid_h * stride_out_head
offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_n = tl.arange(0, BLOCK_SIZE_DSTATE)
state_ptrs = state_ptr + (offs_m[:, None] * stride_state_dim +
offs_n[None, :] * stride_state_dstate)
x_ptrs = x_ptr + offs_m * stride_x_dim
dt_ptrs = dt_ptr + offs_m * stride_dt_dim
if HAS_DT_BIAS:
dt_bias_ptrs = dt_bias_ptr + offs_m * stride_dt_bias_dim
if HAS_D:
D_ptr += pid_h * stride_D_head
A_ptrs = A_ptr + (offs_m[:, None] * stride_A_dim +
offs_n[None, :] * stride_A_dstate)
B_ptrs = B_ptr + offs_n * stride_B_dstate
C_ptrs = C_ptr + offs_n * stride_C_dstate
if HAS_D:
D_ptrs = D_ptr + offs_m * stride_D_dim
if HAS_Z:
z_ptrs = z_ptr + offs_m * stride_z_dim
out_ptrs = out_ptr + offs_m * stride_out_dim
mask = (offs_m[:, None] < dim) & (offs_n[None, :] < dstate)
if HAS_STATE_BATCH_INDICES:
mask &= (state_batch_idx != pad_slot_id)
state = tl.load(state_ptrs, mask=mask, other=0.0)
x = tl.load(x_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
if not TIE_HDIM:
dt = tl.load(dt_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
if HAS_DT_BIAS:
dt += tl.load(dt_bias_ptrs, mask=offs_m < dim,
other=0.0).to(tl.float32)
if DT_SOFTPLUS:
dt = softplus(dt)
A = tl.load(A_ptrs,
mask=(offs_m[:, None] < dim) & (offs_n[None, :] < dstate),
other=0.0).to(tl.float32)
dA = tl.exp(A * dt[:, None])
else:
dt = tl.load(dt_ptr).to(tl.float32)
if HAS_DT_BIAS:
dt += tl.load(dt_bias_ptr).to(tl.float32)
if DT_SOFTPLUS:
dt = softplus(dt)
A = tl.load(A_ptr).to(tl.float32)
dA = tl.exp(A * dt) # scalar, not a matrix
B = tl.load(B_ptrs, mask=offs_n < dstate, other=0.0).to(tl.float32)
C = tl.load(C_ptrs, mask=offs_n < dstate, other=0.0).to(tl.float32)
if HAS_D:
D = tl.load(D_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
if HAS_Z:
z = tl.load(z_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
dB = B[None, :] * dt[:, None] if not TIE_HDIM else B * dt
state = state * dA + dB * x[:, None]
mask = (offs_m[:, None] < dim) & (offs_n[None, :] < dstate)
if HAS_STATE_BATCH_INDICES:
mask &= (state_batch_idx != pad_slot_id)
tl.store(state_ptrs, state, mask=mask)
out = tl.sum(state * C[None, :], axis=1)
if HAS_D:
out += x * D
if HAS_Z:
out *= z * tl.sigmoid(z)
tl.store(out_ptrs, out, mask=offs_m < dim)
def selective_state_update(state,
x,
dt,
A,
B,
C,
D=None,
z=None,
dt_bias=None,
dt_softplus=False,
state_batch_indices=None,
pad_slot_id=PAD_SLOT_ID):
"""
Argument:
state: (batch, dim, dstate) or (batch, nheads, dim, dstate)
x: (batch, dim) or (batch, nheads, dim)
dt: (batch, dim) or (batch, nheads, dim)
A: (dim, dstate) or (nheads, dim, dstate)
B: (batch, dstate) or (batch, ngroups, dstate)
C: (batch, dstate) or (batch, ngroups, dstate)
D: (dim,) or (nheads, dim)
z: (batch, dim) or (batch, nheads, dim)
dt_bias: (dim,) or (nheads, dim)
pad_slot_id: int
if cache_indices is passed, lets the kernel identify padded
entries that will not be processed,
for example: cache_indices = [pad_slot_id, 1, 20, pad_slot_id]
in this case, the kernel will not process entries at
indices 0 and 3
Return:
out: (batch, dim) or (batch, nheads, dim)
"""
has_heads = state.dim() > 3
if state.dim() == 3:
state = state.unsqueeze(1)
if x.dim() == 2:
x = x.unsqueeze(1)
if dt.dim() == 2:
dt = dt.unsqueeze(1)
if A.dim() == 2:
A = A.unsqueeze(0)
if B.dim() == 2:
B = B.unsqueeze(1)
if C.dim() == 2:
C = C.unsqueeze(1)
if D is not None and D.dim() == 1:
D = D.unsqueeze(0)
if z is not None and z.dim() == 2:
z = z.unsqueeze(1)
if dt_bias is not None and dt_bias.dim() == 1:
dt_bias = dt_bias.unsqueeze(0)
_, nheads, dim, dstate = state.shape
batch = x.shape[0]
assert x.shape == (batch, nheads, dim)
assert dt.shape == x.shape
assert A.shape == (nheads, dim, dstate)
ngroups = B.shape[1]
assert nheads % ngroups == 0, "nheads must be divisible by ngroups"
assert B.shape == (batch, ngroups, dstate)
assert C.shape == B.shape
if D is not None:
assert D.shape == (nheads, dim)
if z is not None:
assert z.shape == x.shape
if dt_bias is not None:
assert dt_bias.shape == (nheads, dim)
if state_batch_indices is not None:
assert state_batch_indices.shape == (batch, )
out = torch.empty_like(x)
grid = lambda META: (triton.cdiv(dim, META['BLOCK_SIZE_M']), batch, nheads)
z_strides = ((z.stride(0), z.stride(1), z.stride(2)) if z is not None else
(0, 0, 0))
# We don't want autotune since it will overwrite the state
# We instead tune by hand.
BLOCK_SIZE_M, num_warps = ((32, 4) if dstate <= 16 else
((16, 4) if dstate <= 32 else
((8, 4) if dstate <= 64 else
((4, 4) if dstate <= 128 else ((4, 8))))))
tie_hdim = A.stride(-1) == 0 and A.stride(-2) == 0 and dt.stride(
-1) == 0 and dt_bias.stride(-1) == 0
with torch.cuda.device(x.device.index):
_selective_scan_update_kernel[grid](
state,
x,
dt,
dt_bias,
A,
B,
C,
D,
z,
out,
state_batch_indices,
pad_slot_id,
batch,
nheads,
dim,
dstate,
nheads // ngroups,
state.stride(0),
state.stride(1),
state.stride(2),
state.stride(3),
x.stride(0),
x.stride(1),
x.stride(2),
dt.stride(0),
dt.stride(1),
dt.stride(2),
*(dt_bias.stride(0),
dt_bias.stride(1)) if dt_bias is not None else 0,
A.stride(0),
A.stride(1),
A.stride(2),
B.stride(0),
B.stride(1),
B.stride(2),
C.stride(0),
C.stride(1),
C.stride(2),
*(D.stride(0), D.stride(1)) if D is not None else 0,
z_strides[0],
z_strides[1],
z_strides[2],
out.stride(0),
out.stride(1),
out.stride(2),
dt_softplus,
tie_hdim,
BLOCK_SIZE_M,
num_warps=num_warps,
)
if not has_heads:
out = out.squeeze(1)
return out
def selective_scan_fn(u,
ssm_states,
delta,
A,
B,
C,
D=None,
z=None,
delta_bias=None,
delta_softplus=False,
query_start_loc=None,
cache_indices=None,
has_initial_state=None,
pad_slot_id=PAD_SLOT_ID) -> torch.Tensor:
"""
u: (dim, total_length) for varlen or (batch, dim, seqlen)
applies changes in place.
ssm_states: (batch, dim, dstate) or (batch, nheads, dim, dstate)
applies changes in place.
delta: (dim, total_length) for varlen or (batch, dim, seqlen)
A: (dim, dstate)
B: (ngroups, dstate, total_length) for varlen or
(batch,ngroups,dstate,seqlen)
C: (ngroups, dstate, total_length) for varlen or
(batch,ngroups,dstate,seqlen)
D: (dim,)
z: (dim, total_length) for varlen or (batch, dim, seqlen)
dt_bias: (dim,) or (dim)
query_start_loc: (batch + 1) int32
The cumulative sequence lengths of the sequences in
the batch, used to index into sequence. prepended with 0.
for example: query_start_loc = torch.Tensor([0,10,16,17]),
x.shape=(dim,17)
cache_indices: (batch) int32
A tensor with each cell is a correspondent
input and output ssm_state index
has_initial_state: (batch) bool
A tensor populated with ones and zeros,
indicate if the ssm_state at the corresponding index should be
used as initial state. Not providing argument assumes
there's no initial state
pad_slot_id: int
if cache_indices is passed, lets the kernel identify padding entries
that will not be processed,
for example: cache_indices = [pad_slot_id, 1 ,20 ,pad_slot_id]
in this case, the kernel will not process entries at indices 0 and 3
returns
output: (dim, total_length) for varlen or (batch, dim, seqlen)
supports inplace replacement
"""
if u.stride(-1) != 1:
u = u.contiguous()
if delta.stride(-1) != 1:
delta = delta.contiguous()
if D is not None:
D = D.contiguous()
if B.stride(-1) != 1:
B = B.contiguous()
if C.stride(-1) != 1:
C = C.contiguous()
if z is not None and z.stride(-1) != 1:
z = z.contiguous()
if B.dim() == 3 and query_start_loc is None:
B = B.unsqueeze(1)
if B.dim() == 2 and query_start_loc is not None:
B = B.unsqueeze(0)
if C.dim() == 3 and query_start_loc is None:
C = C.unsqueeze(1)
if C.dim() == 2 and query_start_loc is not None:
C = C.unsqueeze(0)
ops.selective_scan_fwd(u, delta, A, B, C, D, z, delta_bias, delta_softplus,
query_start_loc, cache_indices, has_initial_state,
ssm_states, pad_slot_id)
if z is None:
return delta # output written inplace to delta
else:
return z # output written inplace to z