EngineX-Ascend/enginex-ascend-910-vllm
11
0
Fork 1
Code Issues Pull Requests Actions Projects Releases Wiki Activity

init v0.23.0

Browse Source 
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
This commit is contained in:
Sun Ruoxi
2026-08-27 15:11:51 +08:00
parent b582a8e7d1
commit 7f8a1b1f7a
2849 changed files with 712887 additions and 22001 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
vllm_ascend/ops/triton/kda/__init__.py Normal file
View File

328
vllm_ascend/ops/triton/kda/chunk_delta_h.py Normal file
View File

@@ -0,0 +1,328 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
#
# This file contains code copied from the flash-linear-attention project.
# The original source code was licensed under the MIT license and included
# the following copyright notice:
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
# ruff: noqa: E501
# mypy: ignore-errors
import torch
from vllm.triton_utils import tl, triton
from .utils import FLA_CHUNK_SIZE, prepare_chunk_indices, prepare_chunk_offsets
# KDA-specific h-state backend. It uses [V, K] state layout and vector gate
# semantics, so it is exposed through a distinct entry point.
@triton.heuristics(
{
"USE_G": lambda args: args["g"] is not None,
"USE_GK": lambda args: args["gk"] is not None,
"USE_INITIAL_STATE": lambda args: args["h0"] is not None,
"STORE_FINAL_STATE": lambda args: args["ht"] is not None,
"SAVE_NEW_VALUE": lambda args: args["v_new"] is not None,
"IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
}
)
@triton.autotune(
configs=[triton.Config({"BV": 64}, num_warps=4, num_stages=3)],
key=["H", "K", "V", "BT"],
)
@triton.jit(do_not_specialize=["T"])
def chunk_gated_delta_rule_fwd_kernel_h_blockdim64_kda(
k,
v,
w,
v_new,
g,
gk,
h,
h0,
ht,
cu_seqlens,
chunk_offsets,
T,
H: tl.constexpr,
Hg: tl.constexpr,
K: tl.constexpr,
V: tl.constexpr,
BT: tl.constexpr,
BV: tl.constexpr,
USE_G: tl.constexpr,
USE_GK: tl.constexpr,
USE_INITIAL_STATE: tl.constexpr,
STORE_FINAL_STATE: tl.constexpr,
SAVE_NEW_VALUE: tl.constexpr,
IS_VARLEN: tl.constexpr,
):
i_v, i_nh = tl.program_id(0), tl.program_id(1)
i_n, i_h = i_nh // H, i_nh % H
if IS_VARLEN:
bos, eos = (
tl.load(cu_seqlens + i_n).to(tl.int32),
tl.load(cu_seqlens + i_n + 1).to(tl.int32),
)
T = eos - bos
NT = tl.cdiv(T, BT)
boh = tl.load(chunk_offsets + i_n).to(tl.int32)
else:
bos, eos = i_n * T, i_n * T + T
NT = tl.cdiv(T, BT)
boh = i_n * NT
# [BV, BK]
b_h1 = tl.zeros([BV, 64], dtype=tl.float32)
if K > 64:
b_h2 = tl.zeros([BV, 64], dtype=tl.float32)
if K > 128:
b_h3 = tl.zeros([BV, 64], dtype=tl.float32)
if K > 192:
b_h4 = tl.zeros([BV, 64], dtype=tl.float32)
# calculate offset
h += ((boh * H + i_h) * V * K).to(tl.int64)
v += ((bos * H + i_h) * V).to(tl.int64)
k += ((bos * Hg + i_h // (H // Hg)) * K).to(tl.int64)
w += ((bos * H + i_h) * K).to(tl.int64)
if SAVE_NEW_VALUE:
v_new += ((bos * H + i_h) * V).to(tl.int64)
stride_v = H * V
stride_h = H * V * K
stride_k = Hg * K
stride_w = H * K
if USE_INITIAL_STATE:
h0 = h0 + i_nh * V * K
if STORE_FINAL_STATE:
ht = ht + i_nh * V * K
# load initial state
if USE_INITIAL_STATE:
p_h0_1 = tl.make_block_ptr(h0, (V, K), (K, 1), (i_v * BV, 0), (BV, 64), (1, 0))
b_h1 += tl.load(p_h0_1, boundary_check=(0, 1)).to(tl.float32)
if K > 64:
p_h0_2 = tl.make_block_ptr(h0, (V, K), (K, 1), (i_v * BV, 64), (BV, 64), (1, 0))
b_h2 += tl.load(p_h0_2, boundary_check=(0, 1)).to(tl.float32)
if K > 128:
p_h0_3 = tl.make_block_ptr(h0, (V, K), (K, 1), (i_v * BV, 128), (BV, 64), (1, 0))
b_h3 += tl.load(p_h0_3, boundary_check=(0, 1)).to(tl.float32)
if K > 192:
p_h0_4 = tl.make_block_ptr(h0, (V, K), (K, 1), (i_v * BV, 192), (BV, 64), (1, 0))
b_h4 += tl.load(p_h0_4, boundary_check=(0, 1)).to(tl.float32)
# main recurrence
for i_t in range(NT):
p_h1 = tl.make_block_ptr(
h + i_t.to(tl.int64) * stride_h,
(V, K),
(K, 1),
(i_v * BV, 0),
(BV, 64),
(1, 0),
)
tl.store(p_h1, b_h1.to(p_h1.dtype.element_ty), boundary_check=(0, 1))
if K > 64:
p_h2 = tl.make_block_ptr(
h + i_t.to(tl.int64) * stride_h,
(V, K),
(K, 1),
(i_v * BV, 64),
(BV, 64),
(1, 0),
)
tl.store(p_h2, b_h2.to(p_h2.dtype.element_ty), boundary_check=(0, 1))
if K > 128:
p_h3 = tl.make_block_ptr(
h + i_t.to(tl.int64) * stride_h,
(V, K),
(K, 1),
(i_v * BV, 128),
(BV, 64),
(1, 0),
)
tl.store(p_h3, b_h3.to(p_h3.dtype.element_ty), boundary_check=(0, 1))
if K > 192:
p_h4 = tl.make_block_ptr(
h + i_t.to(tl.int64) * stride_h,
(V, K),
(K, 1),
(i_v * BV, 192),
(BV, 64),
(1, 0),
)
tl.store(p_h4, b_h4.to(p_h4.dtype.element_ty), boundary_check=(0, 1))
m_t = (i_t.to(tl.int64) * BT + tl.arange(0, BT)) < T
p_w = tl.make_block_ptr(w, (T, K), (stride_w, 1), (i_t * BT, 0), (BT, 64), (1, 0))
b_w = tl.load(p_w, boundary_check=(0, 1)).to(tl.float32)
b_w = tl.where(m_t[:, None], b_w, 0.0)
b_v = tl.dot(b_w, tl.trans(b_h1), input_precision="ieee")
if K > 64:
p_w = tl.make_block_ptr(w, (T, K), (stride_w, 1), (i_t * BT, 64), (BT, 64), (1, 0))
b_w = tl.load(p_w, boundary_check=(0, 1)).to(tl.float32)
b_w = tl.where(m_t[:, None], b_w, 0.0)
b_v += tl.dot(b_w, tl.trans(b_h2), input_precision="ieee")
if K > 128:
p_w = tl.make_block_ptr(w, (T, K), (stride_w, 1), (i_t * BT, 128), (BT, 64), (1, 0))
b_w = tl.load(p_w, boundary_check=(0, 1)).to(tl.float32)
b_w = tl.where(m_t[:, None], b_w, 0.0)
b_v += tl.dot(b_w, tl.trans(b_h3), input_precision="ieee")
if K > 192:
p_w = tl.make_block_ptr(w, (T, K), (stride_w, 1), (i_t * BT, 192), (BT, 64), (1, 0))
b_w = tl.load(p_w, boundary_check=(0, 1)).to(tl.float32)
b_w = tl.where(m_t[:, None], b_w, 0.0)
b_v += tl.dot(b_w, tl.trans(b_h4), input_precision="ieee")
p_v = tl.make_block_ptr(v, (T, V), (stride_v, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0))
b_v = tl.load(p_v, boundary_check=(0, 1)) - b_v
b_v = tl.where(m_t[:, None], b_v, 0.0)
if SAVE_NEW_VALUE:
p_v = tl.make_block_ptr(v_new, (T, V), (stride_v, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0))
tl.store(p_v, b_v.to(p_v.dtype.element_ty), boundary_check=(0, 1))
last_idx = min((i_t.to(tl.int64) + 1) * BT, T) - 1
if USE_G:
b_g_last = tl.load(g + bos * H + last_idx * H + i_h)
p_g = tl.make_block_ptr(g + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,))
b_g = tl.load(p_g, boundary_check=(0,))
b_v = b_v * tl.where(m_t, tl.exp(b_g_last - b_g), 0)[:, None]
b_g_last = tl.exp(b_g_last)
b_h1 *= b_g_last
if K > 64:
b_h2 *= b_g_last
if K > 128:
b_h3 *= b_g_last
if K > 192:
b_h4 *= b_g_last
if USE_GK:
o_k1 = tl.arange(0, 64)
b_gk_last1 = tl.load(
gk + (bos + last_idx) * H * K + i_h * K + o_k1,
mask=(o_k1 < K),
other=0.0,
)
b_h1 *= tl.exp2(b_gk_last1)[None, :]
if K > 64:
o_k2 = 64 + o_k1
b_gk_last2 = tl.load(
gk + (bos + last_idx) * H * K + i_h * K + o_k2,
mask=(o_k2 < K),
other=0.0,
)
b_h2 *= tl.exp2(b_gk_last2)[None, :]
if K > 128:
o_k3 = 128 + o_k1
b_gk_last3 = tl.load(
gk + (bos + last_idx) * H * K + i_h * K + o_k3,
mask=(o_k3 < K),
other=0.0,
)
b_h3 *= tl.exp2(b_gk_last3)[None, :]
if K > 192:
o_k4 = 192 + o_k1
b_gk_last4 = tl.load(
gk + (bos + last_idx) * H * K + i_h * K + o_k4,
mask=(o_k4 < K),
other=0.0,
)
b_h4 *= tl.exp2(b_gk_last4)[None, :]
b_v = b_v.to(k.dtype.element_ty)
p_k = tl.make_block_ptr(k, (K, T), (1, stride_k), (0, i_t * BT), (64, BT), (0, 1))
b_k = tl.load(p_k, boundary_check=(0, 1))
b_k = tl.where(m_t[None, :], b_k, 0.0)
b_h1 += tl.trans(tl.dot(b_k, b_v))
if K > 64:
p_k = tl.make_block_ptr(k, (K, T), (1, stride_k), (64, i_t * BT), (64, BT), (0, 1))
b_k = tl.load(p_k, boundary_check=(0, 1))
b_k = tl.where(m_t[None, :], b_k, 0.0)
b_h2 += tl.trans(tl.dot(b_k, b_v))
if K > 128:
p_k = tl.make_block_ptr(k, (K, T), (1, stride_k), (128, i_t * BT), (64, BT), (0, 1))
b_k = tl.load(p_k, boundary_check=(0, 1))
b_k = tl.where(m_t[None, :], b_k, 0.0)
b_h3 += tl.trans(tl.dot(b_k, b_v))
if K > 192:
p_k = tl.make_block_ptr(k, (K, T), (1, stride_k), (192, i_t * BT), (64, BT), (0, 1))
b_k = tl.load(p_k, boundary_check=(0, 1))
b_k = tl.where(m_t[None, :], b_k, 0.0)
b_h4 += tl.trans(tl.dot(b_k, b_v))
# epilogue
if STORE_FINAL_STATE:
p_ht = tl.make_block_ptr(ht, (V, K), (K, 1), (i_v * BV, 0), (BV, 64), (1, 0))
tl.store(p_ht, b_h1.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
if K > 64:
p_ht = tl.make_block_ptr(ht, (V, K), (K, 1), (i_v * BV, 64), (BV, 64), (1, 0))
tl.store(p_ht, b_h2.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
if K > 128:
p_ht = tl.make_block_ptr(ht, (V, K), (K, 1), (i_v * BV, 128), (BV, 64), (1, 0))
tl.store(p_ht, b_h3.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
if K > 192:
p_ht = tl.make_block_ptr(ht, (V, K), (K, 1), (i_v * BV, 192), (BV, 64), (1, 0))
tl.store(p_ht, b_h4.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
def chunk_gated_delta_rule_fwd_h_kda(
k: torch.Tensor,
w: torch.Tensor,
u: torch.Tensor,
g: torch.Tensor | None = None,
gk: torch.Tensor | None = None,
initial_state: torch.Tensor | None = None,
output_final_state: bool = False,
chunk_size: int = FLA_CHUNK_SIZE,
save_new_value: bool = True,
cu_seqlens: torch.Tensor | None = None,
chunk_indices: torch.Tensor | None = None,
chunk_offsets: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
# This kernel is slightly different from fla to support Q/K with different head numbers.
# In fla, Q/K always have the same head number, so Hg is always equal to H.
B, T, Hg, K, V = *k.shape, u.shape[-1]
H = u.shape[-2]
BT = chunk_size
if chunk_indices is None and cu_seqlens is not None:
chunk_indices = prepare_chunk_indices(cu_seqlens, chunk_size)
# N: the actual number of sequences in the batch with either equal or variable lengths
if cu_seqlens is None:
N, NT, chunk_offsets = B, triton.cdiv(T, BT), None
else:
N, NT = len(cu_seqlens) - 1, len(chunk_indices)
if chunk_offsets is None:
chunk_offsets = prepare_chunk_offsets(cu_seqlens, BT)
assert K <= 256, "current kernel does not support head dimension larger than 256."
h = k.new_empty(B, NT, H, V, K)
final_state = k.new_empty(N, H, V, K, dtype=torch.float32) if output_final_state else None
v_new = torch.empty_like(u) if save_new_value else None
def grid(meta):
return (triton.cdiv(V, meta["BV"]), N * H)
chunk_gated_delta_rule_fwd_kernel_h_blockdim64_kda[grid](
k=k,
v=u,
w=w,
v_new=v_new,
g=g,
gk=gk,
h=h,
h0=initial_state,
ht=final_state,
cu_seqlens=cu_seqlens,
chunk_offsets=chunk_offsets,
T=T,
H=H,
Hg=Hg,
K=K,
V=V,
BT=BT,
)
return h, v_new, final_state

274
vllm_ascend/ops/triton/kda/cumsum.py Normal file
View File

@@ -0,0 +1,274 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
#
# This file contains code copied from the flash-linear-attention project.
# The original source code was licensed under the MIT license and included
# the following copyright notice:
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
# ruff: noqa: E501
# mypy: ignore-errors
import torch
from vllm.triton_utils import tl, triton
from .utils import prepare_chunk_indices
@triton.heuristics(
{"HAS_SCALE": lambda args: args["scale"] is not None, "IS_VARLEN": lambda args: args["cu_seqlens"] is not None}
)
@triton.jit(do_not_specialize=["T"])
def chunk_local_cumsum_scalar_kernel(
s,
o,
scale,
cu_seqlens,
chunk_indices,
T,
H: tl.constexpr,
BLOCK_T: tl.constexpr,
REVERSE: tl.constexpr,
HAS_SCALE: tl.constexpr,
IS_VARLEN: tl.constexpr,
HEAD_FIRST: tl.constexpr,
CHUNK_SIZE: tl.constexpr = 64,
):
i_block, i_b = tl.program_id(0), tl.program_id(1)
N_CHUNKS: tl.constexpr = BLOCK_T // CHUNK_SIZE
if IS_VARLEN:
i_s, i_block = (
tl.load(chunk_indices + i_block * 2).to(tl.int32),
tl.load(chunk_indices + i_block * 2 + 1).to(tl.int32),
)
bos, eos = tl.load(cu_seqlens + i_s).to(tl.int32), tl.load(cu_seqlens + i_s + 1).to(tl.int32)
T = eos - bos
else:
bos, eos = i_b * T, i_b * T + T
if HEAD_FIRST:
ptr_s = tl.make_block_ptr(s + bos * H, (H, T), (T, 1), (0, i_block * BLOCK_T), (H, BLOCK_T), (1, 0))
ptr_o = tl.make_block_ptr(o + bos * H, (H, T), (T, 1), (0, i_block * BLOCK_T), (H, BLOCK_T), (1, 0))
b_s = tl.load(ptr_s, boundary_check=(0,)).to(tl.float32)
b_s = tl.reshape(b_s, (H, N_CHUNKS, CHUNK_SIZE))
b_s = tl.trans(b_s, (2, 0, 1))
b_o = tl.cumsum(b_s, axis=0, reverse=REVERSE)
if HAS_SCALE:
b_o *= scale
b_o = tl.trans(b_o, (2, 0, 1))
b_o = tl.reshape(b_o, (H, BLOCK_T))
else:
ptr_s = tl.make_block_ptr(s + bos * H, (T, H), (H, 1), (i_block * BLOCK_T, 0), (BLOCK_T, H), (1, 0))
ptr_o = tl.make_block_ptr(o + bos * H, (T, H), (H, 1), (i_block * BLOCK_T, 0), (BLOCK_T, H), (1, 0))
b_s = tl.load(ptr_s, boundary_check=(0,)).to(tl.float32)
b_s = tl.reshape(b_s, (N_CHUNKS, CHUNK_SIZE, H))
b_s = tl.trans(b_s, (1, 0, 2))
b_o = tl.cumsum(b_s, axis=0, reverse=REVERSE)
if HAS_SCALE:
b_o *= scale
b_o = tl.trans(b_o, (1, 0, 2))
b_o = tl.reshape(b_o, (BLOCK_T, H))
tl.store(ptr_o, b_o.to(s.dtype.element_ty), boundary_check=(0,))
return
def chunk_local_cumsum_scalar(
g,
chunk_size,
reverse: bool = False,
scale: float = None,
cu_seqlens: torch.Tensor | None = None,
block_indices: torch.Tensor | None = None,
head_first: bool = False,
output_dtype: torch.Tensor | None = torch.float,
):
if head_first:
B, H, T = g.shape
else:
B, T, H = g.shape
assert chunk_size == 2 ** (chunk_size.bit_length() - 1), "chunk_size must be a power of 2"
OPTIM_BLOCK_SIZE = triton.next_power_of_2((2**18) // (H * chunk_size))
if cu_seqlens is not None and block_indices is None:
block_indices = prepare_chunk_indices(cu_seqlens, chunk_size=OPTIM_BLOCK_SIZE)
num_blocks = len(block_indices) if cu_seqlens is not None else triton.cdiv(T, OPTIM_BLOCK_SIZE)
g_org, g = g, torch.empty_like(g, dtype=output_dtype or g.dtype)
grid = (num_blocks, B)
chunk_local_cumsum_scalar_kernel[grid](
s=g_org,
o=g,
scale=scale,
cu_seqlens=cu_seqlens,
chunk_indices=block_indices,
T=T,
H=H,
BLOCK_T=OPTIM_BLOCK_SIZE,
CHUNK_SIZE=chunk_size,
HEAD_FIRST=head_first,
REVERSE=reverse,
num_warps=8,
num_stages=3,
)
return g
@triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None})
@triton.autotune(
configs=[triton.Config({"BS": 32}, num_warps=8)],
key=["B", "H", "S", "BT", "IS_VARLEN", "REVERSE"],
)
@triton.jit(do_not_specialize=["T"])
def chunk_local_cumsum_vector_kernel(
s,
o,
cu_seqlens,
chunk_indices,
T,
B: tl.constexpr,
H: tl.constexpr,
S: tl.constexpr,
BT: tl.constexpr,
BS: tl.constexpr,
REVERSE: tl.constexpr,
IS_VARLEN: tl.constexpr,
HEAD_FIRST: tl.constexpr,
):
i_s, i_t, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
i_b, i_h = i_bh // H, i_bh % H
if IS_VARLEN:
i_n, i_t = (
tl.load(chunk_indices + i_t * 2).to(tl.int32),
tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32),
)
bos, eos = tl.load(cu_seqlens + i_n).to(tl.int32), tl.load(cu_seqlens + i_n + 1).to(tl.int32)
T = eos - bos
else:
bos, eos = i_b * T, i_b * T + T
o_i = tl.arange(0, BT)
if REVERSE:
m_s = tl.where(o_i[:, None] <= o_i[None, :], 1.0, 0.0)
else:
m_s = tl.where(o_i[:, None] >= o_i[None, :], 1.0, 0.0)
if HEAD_FIRST:
p_s = tl.make_block_ptr(
s + (bos * H + i_h * T) * S,
(T, S),
(S, 1),
(i_t * BT, i_s * BS),
(BT, BS),
(1, 0),
)
p_o = tl.make_block_ptr(
o + (bos * H + i_h * T) * S,
(T, S),
(S, 1),
(i_t * BT, i_s * BS),
(BT, BS),
(1, 0),
)
else:
p_s = tl.make_block_ptr(
s + (bos * H + i_h) * S,
(T, S),
(H * S, 1),
(i_t * BT, i_s * BS),
(BT, BS),
(1, 0),
)
p_o = tl.make_block_ptr(
o + (bos * H + i_h) * S,
(T, S),
(H * S, 1),
(i_t * BT, i_s * BS),
(BT, BS),
(1, 0),
)
b_s = tl.load(p_s, boundary_check=(0, 1)).to(tl.float32)
b_o = tl.dot(m_s, b_s, allow_tf32=False)
tl.store(p_o, b_o.to(p_o.dtype.element_ty), boundary_check=(0, 1))
def chunk_local_cumsum_vector(
g: torch.Tensor,
chunk_size: int,
reverse: bool = False,
cu_seqlens: torch.Tensor | None = None,
chunk_indices: torch.Tensor | None = None,
head_first: bool = False,
output_dtype: torch.dtype | None = torch.float,
) -> torch.Tensor:
if head_first:
B, H, T, S = g.shape
else:
B, T, H, S = g.shape
assert chunk_size == 2 ** (chunk_size.bit_length() - 1), "chunk_size must be a power of 2"
if chunk_indices is None and cu_seqlens is not None:
chunk_indices = prepare_chunk_indices(cu_seqlens, chunk_size)
BT = chunk_size
NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)
g_org, g = g, torch.empty_like(g, dtype=output_dtype or g.dtype)
def grid(meta):
return (triton.cdiv(meta["S"], meta["BS"]), NT, B * H)
chunk_local_cumsum_vector_kernel[grid](
g_org,
g,
cu_seqlens,
chunk_indices,
T=T,
B=B,
H=H,
S=S,
BT=BT,
HEAD_FIRST=head_first,
REVERSE=reverse,
)
return g
def chunk_local_cumsum(
g: torch.Tensor,
chunk_size: int,
reverse: bool = False,
scale: float = None,
cu_seqlens: torch.Tensor | None = None,
chunk_indices: torch.Tensor | None = None,
head_first: bool = False,
output_dtype: torch.dtype | None = torch.float,
**kwargs,
) -> torch.Tensor:
if cu_seqlens is not None:
assert g.shape[0] == 1, "Only batch size 1 is supported when cu_seqlens are provided"
if len(g.shape) == 3:
block_indices = kwargs.get("block_indices", chunk_indices)
return chunk_local_cumsum_scalar(
g=g,
chunk_size=chunk_size,
reverse=reverse,
scale=scale,
cu_seqlens=cu_seqlens,
block_indices=block_indices,
head_first=head_first,
output_dtype=output_dtype,
)
elif len(g.shape) == 4:
return chunk_local_cumsum_vector(
g=g,
chunk_size=chunk_size,
reverse=reverse,
cu_seqlens=cu_seqlens,
chunk_indices=chunk_indices,
head_first=head_first,
output_dtype=output_dtype,
)
else:
raise ValueError(
f"Unsupported input shape {g.shape}, "
f"which should be (B, T, H, D) if `head_first=False` "
f"or (B, H, T, D) otherwise"
)

158
vllm_ascend/ops/triton/kda/fused_recurrent_kda.py Normal file
View File

@@ -0,0 +1,158 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
#
# This file contains code copied from the flash-linear-attention project.
# The original source code was licensed under the MIT license and included
# the following copyright notice:
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
# ruff: noqa: E501
# mypy: ignore-errors
from vllm.triton_utils import tl, triton
@triton.heuristics(
{
"USE_INITIAL_STATE": lambda args: args["h0"] is not None,
"IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
"IS_CONTINUOUS_BATCHING": lambda args: args["ssm_state_indices"] is not None,
"IS_SPEC_DECODING": lambda args: args["num_accepted_tokens"] is not None,
}
)
@triton.jit(do_not_specialize=["N", "T"])
def fused_recurrent_gated_delta_rule_fwd_kernel(
q,
k,
v,
g,
beta,
o,
h0,
ht,
cu_seqlens,
ssm_state_indices,
num_accepted_tokens,
scale,
N: tl.int64, # num of sequences
T: tl.int64, # num of tokens
B: tl.constexpr,
H: tl.constexpr,
HV: tl.constexpr,
K: tl.constexpr,
V: tl.constexpr,
BK: tl.constexpr,
BV: tl.constexpr,
stride_init_state_token: tl.constexpr,
stride_final_state_token: tl.constexpr,
stride_indices_seq: tl.constexpr,
stride_indices_tok: tl.constexpr,
USE_INITIAL_STATE: tl.constexpr, # whether to use initial state
INPLACE_FINAL_STATE: tl.constexpr, # whether to store final state inplace
IS_BETA_HEADWISE: tl.constexpr, # whether beta is headwise vector or scalar,
USE_QK_L2NORM_IN_KERNEL: tl.constexpr,
IS_VARLEN: tl.constexpr,
IS_CONTINUOUS_BATCHING: tl.constexpr,
IS_SPEC_DECODING: tl.constexpr,
IS_KDA: tl.constexpr,
):
i_k, i_v, i_nh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
i_n, i_hv = i_nh // HV, i_nh % HV
i_h = i_hv // (HV // H)
if IS_VARLEN:
bos, eos = (
tl.load(cu_seqlens + i_n).to(tl.int64),
tl.load(cu_seqlens + i_n + 1).to(tl.int64),
)
all = T
T = eos - bos
else:
bos, eos = i_n * T, i_n * T + T
all = B * T
if T == 0:
return
o_k = i_k * BK + tl.arange(0, BK)
o_v = i_v * BV + tl.arange(0, BV)
p_q = q + (bos * H + i_h) * K + o_k
p_k = k + (bos * H + i_h) * K + o_k
p_v = v + (bos * HV + i_hv) * V + o_v
if IS_BETA_HEADWISE:
p_beta = beta + (bos * HV + i_hv) * V + o_v
else:
p_beta = beta + bos * HV + i_hv
if not IS_KDA:
p_g = g + bos * HV + i_hv
else:
p_gk = g + (bos * HV + i_hv) * K + o_k
p_o = o + ((i_k * all + bos) * HV + i_hv) * V + o_v
mask_k = o_k < K
mask_v = o_v < V
mask_h = mask_v[:, None] & mask_k[None, :]
b_h = tl.zeros([BV, BK], dtype=tl.float32)
if USE_INITIAL_STATE:
if IS_CONTINUOUS_BATCHING:
if IS_SPEC_DECODING:
i_t = tl.load(num_accepted_tokens + i_n).to(tl.int64) - 1
else:
i_t = 0
state_idx = tl.load(ssm_state_indices + i_n * stride_indices_seq + i_t).to(tl.int64)
if state_idx <= 0:
return
p_h0 = h0 + state_idx * stride_init_state_token
else:
p_h0 = h0 + bos * HV * V * K
p_h0 = p_h0 + i_hv * V * K + o_v[:, None] * K + o_k[None, :]
b_h += tl.load(p_h0, mask=mask_h, other=0).to(tl.float32)
for i_t in range(0, T):
b_q = tl.load(p_q, mask=mask_k, other=0).to(tl.float32)
b_k = tl.load(p_k, mask=mask_k, other=0).to(tl.float32)
b_v = tl.load(p_v, mask=mask_v, other=0).to(tl.float32)
if USE_QK_L2NORM_IN_KERNEL:
b_q = b_q / tl.sqrt(tl.sum(b_q * b_q) + 1e-6)
b_k = b_k / tl.sqrt(tl.sum(b_k * b_k) + 1e-6)
b_q = b_q * scale
if not IS_KDA:
b_g = tl.load(p_g).to(tl.float32)
b_h *= tl.exp(b_g)
else:
b_gk = tl.load(p_gk).to(tl.float32)
b_h *= tl.exp(b_gk[None, :])
b_v -= tl.sum(b_h * b_k[None, :], 1)
if IS_BETA_HEADWISE:
b_beta = tl.load(p_beta, mask=mask_v, other=0).to(tl.float32)
else:
b_beta = tl.load(p_beta).to(tl.float32)
b_v *= b_beta
b_h += b_v[:, None] * b_k[None, :]
b_o = tl.sum(b_h * b_q[None, :], 1)
tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=mask_v)
if INPLACE_FINAL_STATE:
final_state_idx = tl.load(ssm_state_indices + i_n * stride_indices_seq + i_t).to(tl.int64)
if final_state_idx > 0:
p_ht = ht + final_state_idx * stride_final_state_token
p_ht = p_ht + i_hv * V * K + o_v[:, None] * K + o_k[None, :]
tl.store(p_ht, b_h.to(p_ht.dtype.element_ty), mask=mask_h)
else:
p_ht = ht + (bos + i_t) * stride_final_state_token
p_ht = p_ht + i_hv * V * K + o_v[:, None] * K + o_k[None, :]
tl.store(p_ht, b_h.to(p_ht.dtype.element_ty), mask=mask_h)
p_q += H * K
p_k += H * K
p_o += HV * V
p_v += HV * V
if not IS_KDA:
p_g += HV
else:
p_gk += HV * K
p_beta += HV * (V if IS_BETA_HEADWISE else 1)

1331
vllm_ascend/ops/triton/kda/kda.py Normal file
View File

File diff suppressed because it is too large Load Diff

100
vllm_ascend/ops/triton/kda/l2norm.py Normal file
View File

@@ -0,0 +1,100 @@
# Adapt from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fla/ops/l2norm.py
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
#
# This file contains code copied from the flash-linear-attention project.
# The original source code was licensed under the MIT license and included
# the following copyright notice:
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
# mypy: ignore-errors
import torch
from vllm.triton_utils import tl, triton
from vllm_ascend.ops.triton.triton_utils import get_vectorcore_num
@triton.jit(do_not_specialize=["eps", "M", "NUM_CHUNKS"])
def l2norm_fwd_persistent_kernel(X, Y, eps, M, N: tl.constexpr, MBLOCK: tl.constexpr, NUM_CHUNKS):
# One program per vector core; each program loops over NUM_CHUNKS blocks
# of MBLOCK rows so the grid stays resident on the device.
base_row = tl.program_id(0) * (NUM_CHUNKS * MBLOCK)
rindex = tl.arange(0, N)[None, :]
for chunk in range(NUM_CHUNKS):
row_idx = base_row + chunk * MBLOCK + tl.arange(0, MBLOCK)[:, None]
xmask = row_idx < M
xs = tl.load(X + (rindex + N * row_idx), mask=xmask, other=0.0).to(tl.float32)
square = xs * xs
square_sum = tl.sum(square, 1)[:, None]
rsqrt = tl.rsqrt(square_sum + eps)
tl.store(Y + (rindex + N * row_idx), xs * rsqrt, xmask)
@triton.jit
def l2norm_fwd_tiled_kernel(X, Y, eps, M, N: tl.constexpr, BD: tl.constexpr, MBLOCK: tl.constexpr):
# One program per MBLOCK-row tile; columns are padded to BD and masked.
xoffset = tl.program_id(0) * MBLOCK
row_idx = xoffset + tl.arange(0, MBLOCK)[:, None]
xmask = row_idx < M
rindex = tl.arange(0, BD)[None, :]
cmask = rindex < N
mask = xmask & cmask
xs = tl.load(X + (rindex + N * row_idx), mask, other=0.0).to(tl.float32)
square = tl.broadcast_to(xs * xs, [MBLOCK, BD])
square_sum = tl.sum(tl.where(xmask, square, 0), 1)[:, None]
rsqrt = tl.rsqrt(square_sum + eps)
tl.store(Y + (rindex + N * row_idx), xs * rsqrt, mask)
def l2norm_fwd(
x: torch.Tensor,
eps: float = 1e-6,
output_dtype: torch.dtype | None = None,
use_tiled_kernel: bool = False,
):
x_shape_og = x.shape
x = x.reshape(-1, x.shape[-1])
# allocate output
if output_dtype is None:
y = torch.empty_like(x)
else:
y = torch.empty_like(x, dtype=output_dtype)
assert y.stride(-1) == 1
T, D = x.shape[0], x.shape[-1]
# Less than 64KB per feature: enqueue fused kernel
MAX_FUSED_SIZE = 65536 // x.element_size()
BD = min(MAX_FUSED_SIZE, triton.next_power_of_2(D))
if D > BD:
raise RuntimeError("This layer doesn't support feature dim >= 64KB.")
if use_tiled_kernel:
MBLOCK = 32
l2norm_fwd_tiled_kernel[(triton.cdiv(T, MBLOCK),)](
x,
y,
eps,
T,
D,
BD,
MBLOCK,
)
else:
MBLOCK = 69
num_core = get_vectorcore_num()
main_bs = triton.cdiv(T, num_core)
num_sub_blocks = triton.cdiv(main_bs, MBLOCK)
l2norm_fwd_persistent_kernel[(num_core,)](
X=x,
Y=y,
eps=eps,
M=T,
N=D,
MBLOCK=MBLOCK,
NUM_CHUNKS=num_sub_blocks,
)
return y.view(x_shape_og)

402
vllm_ascend/ops/triton/kda/solve_tril.py Normal file
View File

@@ -0,0 +1,402 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
#
# This file contains code copied from the flash-linear-attention project.
# The original source code was licensed under the MIT license and included
# the following copyright notice:
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
# ruff: noqa: E501
# mypy: ignore-errors
import torch
from vllm.triton_utils import tl, triton
from .utils import input_guard, prepare_chunk_indices
@triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None})
@triton.autotune(
configs=[
triton.Config({}, num_warps=num_warps, num_stages=num_stages)
for num_warps in [1, 2, 4, 8]
for num_stages in [2, 3, 4, 5]
],
key=["BT"],
)
@triton.jit(do_not_specialize=["T"])
def solve_tril_16x16_kernel_kda(
A,
Ai,
cu_seqlens,
chunk_indices,
T,
H: tl.constexpr,
BT: tl.constexpr,
IS_VARLEN: tl.constexpr,
DOT_PRECISION: tl.constexpr,
):
i_t, i_bh = tl.program_id(0), tl.program_id(1)
i_b, i_h = i_bh // H, i_bh % H
if IS_VARLEN:
i_n, i_t = (
tl.load(chunk_indices + i_t * 2).to(tl.int32),
tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32),
)
bos, eos = (
tl.load(cu_seqlens + i_n).to(tl.int32),
tl.load(cu_seqlens + i_n + 1).to(tl.int32),
)
T = eos - bos
else:
bos, eos = i_b * T, i_b * T + T
o_i = tl.arange(0, 16)
m_A = o_i[:, None] > o_i[None, :]
m_I = o_i[:, None] == o_i[None, :]
A = A + (bos * H + i_h) * BT
Ai = Ai + (bos * H + i_h) * 16
offset = (i_t * 16) % BT
p_A = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * 16, offset), (16, 16), (1, 0))
b_A = tl.load(p_A, boundary_check=(0, 1)).to(tl.float32)
b_A = -tl.where(m_A, b_A, 0)
for i in range(2, min(16, T - i_t * 16)):
b_a = -tl.load(A + (i_t * 16 + i) * H * BT + o_i + offset)
b_a = b_a + tl.sum(b_a[:, None] * b_A, 0)
b_A = tl.where((o_i == i)[:, None], b_a, b_A)
b_A += m_I
p_Ai = tl.make_block_ptr(Ai, (T, 16), (H * 16, 1), (i_t * 16, 0), (16, 16), (1, 0))
tl.store(
p_Ai,
b_A.to(p_Ai.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
@triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None})
@triton.autotune(
configs=[
triton.Config({}, num_warps=num_warps, num_stages=num_stages)
for num_warps in [1, 2, 4, 8]
for num_stages in [2, 3, 4, 5]
],
key=["H", "BT", "IS_VARLEN"],
)
@triton.jit(do_not_specialize=["T"])
def merge_16x16_to_32x32_inverse_kernel_kda(
A,
Ai,
cu_seqlens,
chunk_indices,
T,
H: tl.constexpr,
BT: tl.constexpr,
IS_VARLEN: tl.constexpr,
DOT_PRECISION: tl.constexpr,
):
i_t, i_bh = tl.program_id(0), tl.program_id(1)
i_b, i_h = i_bh // H, i_bh % H
if IS_VARLEN:
i_n, i_t = (
tl.load(chunk_indices + i_t * 2).to(tl.int32),
tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32),
)
bos, eos = (
tl.load(cu_seqlens + i_n).to(tl.int32),
tl.load(cu_seqlens + i_n + 1).to(tl.int32),
)
T = eos - bos
else:
bos, eos = i_b * T, i_b * T + T
o_i = tl.arange(0, 16)
m_A = o_i[:, None] > o_i[None, :]
m_I = o_i[:, None] == o_i[None, :]
A += (bos * H + i_h) * BT
Ai += (bos * H + i_h) * BT
p_A_11 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT, 0), (16, 16), (1, 0))
p_A_22 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 16, 16), (16, 16), (1, 0))
b_Ai_11 = tl.load(p_A_11, boundary_check=(0, 1)).to(tl.float32)
b_Ai_22 = tl.load(p_A_22, boundary_check=(0, 1)).to(tl.float32)
b_Ai_11 = -tl.where(m_A, b_Ai_11, 0)
b_Ai_22 = -tl.where(m_A, b_Ai_22, 0)
for i in range(2, min(16, T - i_t * BT)):
b_a_11 = -tl.load(A + (i_t * BT + i) * H * BT + o_i)
b_a_11 += tl.sum(b_a_11[:, None] * b_Ai_11, 0)
b_Ai_11 = tl.where((o_i == i)[:, None], b_a_11, b_Ai_11)
for i in range(16 + 2, min(32, T - i_t * BT)):
b_a_22 = -tl.load(A + (i_t * BT + i) * H * BT + o_i + 16)
b_a_22 += tl.sum(b_a_22[:, None] * b_Ai_22, 0)
b_Ai_22 = tl.where((o_i == i - 16)[:, None], b_a_22, b_Ai_22)
b_Ai_11 += m_I
b_Ai_22 += m_I
p_A_21 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 16, 0), (16, 16), (1, 0))
b_A_21 = tl.load(p_A_21, boundary_check=(0, 1)).to(tl.float32)
b_Ai_21 = -tl.dot(
tl.dot(b_Ai_22, b_A_21, input_precision=DOT_PRECISION),
b_Ai_11,
input_precision=DOT_PRECISION,
)
p_Ai_11 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT, 0), (16, 16), (1, 0))
p_Ai_21 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 16, 0), (16, 16), (1, 0))
p_Ai_22 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 16, 16), (16, 16), (1, 0))
tl.store(
p_Ai_11,
b_Ai_11.to(p_Ai_11.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_22,
b_Ai_22.to(p_Ai_22.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_21,
b_Ai_21.to(p_Ai_21.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
@triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None})
@triton.autotune(
configs=[triton.Config({}, num_warps=8, num_stages=5)],
key=["H", "BT", "IS_VARLEN"],
)
@triton.jit(do_not_specialize=["T"])
def merge_16x16_to_64x64_inverse_kernel_kda(
A,
Ai,
cu_seqlens,
chunk_indices,
T,
H: tl.constexpr,
BT: tl.constexpr,
IS_VARLEN: tl.constexpr,
DOT_PRECISION: tl.constexpr,
):
i_t, i_bh = tl.program_id(0), tl.program_id(1)
i_b, i_h = i_bh // H, i_bh % H
if IS_VARLEN:
i_n, i_t = (
tl.load(chunk_indices + i_t * 2).to(tl.int32),
tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32),
)
bos, eos = (
tl.load(cu_seqlens + i_n).to(tl.int32),
tl.load(cu_seqlens + i_n + 1).to(tl.int32),
)
T = eos - bos
else:
bos, eos = i_b * T, i_b * T + T
o_i = tl.arange(0, 16)
m_A = o_i[:, None] > o_i[None, :]
m_I = o_i[:, None] == o_i[None, :]
A += (bos * H + i_h) * BT
Ai += (bos * H + i_h) * BT
p_A_11 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT, 0), (16, 16), (1, 0))
p_A_22 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 16, 16), (16, 16), (1, 0))
p_A_33 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 32, 32), (16, 16), (1, 0))
p_A_44 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 48, 48), (16, 16), (1, 0))
b_Ai_11 = tl.load(p_A_11, boundary_check=(0, 1)).to(tl.float32)
b_Ai_22 = tl.load(p_A_22, boundary_check=(0, 1)).to(tl.float32)
b_Ai_33 = tl.load(p_A_33, boundary_check=(0, 1)).to(tl.float32)
b_Ai_44 = tl.load(p_A_44, boundary_check=(0, 1)).to(tl.float32)
b_Ai_11 = -tl.where(m_A, b_Ai_11, 0)
b_Ai_22 = -tl.where(m_A, b_Ai_22, 0)
b_Ai_33 = -tl.where(m_A, b_Ai_33, 0)
b_Ai_44 = -tl.where(m_A, b_Ai_44, 0)
for i in range(2, min(16, T - i_t * BT)):
b_a_11 = -tl.load(A + (i_t * BT + i) * H * BT + o_i)
b_a_11 += tl.sum(b_a_11[:, None] * b_Ai_11, 0)
b_Ai_11 = tl.where((o_i == i)[:, None], b_a_11, b_Ai_11)
for i in range(16 + 2, min(32, T - i_t * BT)):
b_a_22 = -tl.load(A + (i_t * BT + i) * H * BT + o_i + 16)
b_a_22 += tl.sum(b_a_22[:, None] * b_Ai_22, 0)
b_Ai_22 = tl.where((o_i == i - 16)[:, None], b_a_22, b_Ai_22)
for i in range(32 + 2, min(48, T - i_t * BT)):
b_a_33 = -tl.load(A + (i_t * BT + i) * H * BT + o_i + 32)
b_a_33 += tl.sum(b_a_33[:, None] * b_Ai_33, 0)
b_Ai_33 = tl.where((o_i == i - 32)[:, None], b_a_33, b_Ai_33)
for i in range(48 + 2, min(64, T - i_t * BT)):
b_a_44 = -tl.load(A + (i_t * BT + i) * H * BT + o_i + 48)
b_a_44 += tl.sum(b_a_44[:, None] * b_Ai_44, 0)
b_Ai_44 = tl.where((o_i == i - 48)[:, None], b_a_44, b_Ai_44)
b_Ai_11 += m_I
b_Ai_22 += m_I
b_Ai_33 += m_I
b_Ai_44 += m_I
p_A_21 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 16, 0), (16, 16), (1, 0))
p_A_31 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 32, 0), (16, 16), (1, 0))
p_A_32 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 32, 16), (16, 16), (1, 0))
p_A_41 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 48, 0), (16, 16), (1, 0))
p_A_42 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 48, 16), (16, 16), (1, 0))
p_A_43 = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * BT + 48, 32), (16, 16), (1, 0))
b_A_21 = tl.load(p_A_21, boundary_check=(0, 1)).to(tl.float32)
b_A_31 = tl.load(p_A_31, boundary_check=(0, 1)).to(tl.float32)
b_A_32 = tl.load(p_A_32, boundary_check=(0, 1)).to(tl.float32)
b_A_41 = tl.load(p_A_41, boundary_check=(0, 1)).to(tl.float32)
b_A_42 = tl.load(p_A_42, boundary_check=(0, 1)).to(tl.float32)
b_A_43 = tl.load(p_A_43, boundary_check=(0, 1)).to(tl.float32)
b_Ai_21 = -tl.dot(
tl.dot(b_Ai_22, b_A_21, input_precision=DOT_PRECISION),
b_Ai_11,
input_precision=DOT_PRECISION,
)
b_Ai_32 = -tl.dot(
tl.dot(b_Ai_33, b_A_32, input_precision=DOT_PRECISION),
b_Ai_22,
input_precision=DOT_PRECISION,
)
b_Ai_43 = -tl.dot(
tl.dot(b_Ai_44, b_A_43, input_precision=DOT_PRECISION),
b_Ai_33,
input_precision=DOT_PRECISION,
)
b_Ai_31 = -tl.dot(
b_Ai_33,
tl.dot(b_A_31, b_Ai_11, input_precision=DOT_PRECISION) + tl.dot(b_A_32, b_Ai_21, input_precision=DOT_PRECISION),
input_precision=DOT_PRECISION,
)
b_Ai_42 = -tl.dot(
b_Ai_44,
tl.dot(b_A_42, b_Ai_22, input_precision=DOT_PRECISION) + tl.dot(b_A_43, b_Ai_32, input_precision=DOT_PRECISION),
input_precision=DOT_PRECISION,
)
b_Ai_41 = -tl.dot(
b_Ai_44,
tl.dot(b_A_41, b_Ai_11, input_precision=DOT_PRECISION)
+ tl.dot(b_A_42, b_Ai_21, input_precision=DOT_PRECISION)
+ tl.dot(b_A_43, b_Ai_31, input_precision=DOT_PRECISION),
input_precision=DOT_PRECISION,
)
p_Ai_11 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT, 0), (16, 16), (1, 0))
p_Ai_22 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 16, 16), (16, 16), (1, 0))
p_Ai_33 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 32, 32), (16, 16), (1, 0))
p_Ai_44 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 48, 48), (16, 16), (1, 0))
p_Ai_21 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 16, 0), (16, 16), (1, 0))
p_Ai_31 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 32, 0), (16, 16), (1, 0))
p_Ai_32 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 32, 16), (16, 16), (1, 0))
p_Ai_41 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 48, 0), (16, 16), (1, 0))
p_Ai_42 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 48, 16), (16, 16), (1, 0))
p_Ai_43 = tl.make_block_ptr(Ai, (T, BT), (H * BT, 1), (i_t * BT + 48, 32), (16, 16), (1, 0))
tl.store(
p_Ai_11,
b_Ai_11.to(p_Ai_11.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_22,
b_Ai_22.to(p_Ai_22.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_33,
b_Ai_33.to(p_Ai_33.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_44,
b_Ai_44.to(p_Ai_44.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_21,
b_Ai_21.to(p_Ai_21.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_31,
b_Ai_31.to(p_Ai_31.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_32,
b_Ai_32.to(p_Ai_32.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_41,
b_Ai_41.to(p_Ai_41.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_42,
b_Ai_42.to(p_Ai_42.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_43,
b_Ai_43.to(p_Ai_43.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
@input_guard
def solve_tril_kda(
A: torch.Tensor,
cu_seqlens: torch.Tensor | None = None,
chunk_indices: torch.Tensor | None = None,
output_dtype: torch.dtype = torch.float,
) -> torch.Tensor:
"""
Compute the inverse of the matrix I + A
A should be strictly lower triangular, i.e., A.triu() == 0.
Args:
A (torch.Tensor):
[B, T, H, BT], where BT should only be 16, 32, or 64.
cu_seqlens (torch.Tensor):
The cumulative sequence lengths of the input tensor. Default: `None`.
chunk_indices (torch.Tensor):
Pre-computed chunk indices. Default: `None`.
output_dtype (torch.dtype):
The dtype of the output tensor. Default: `torch.float`.
If `None`, the output dtype will be the same as the input dtype.
Returns:
(I + A)^-1 with the same shape as A
"""
assert A.shape[-1] in [16, 32, 64]
output_dtype = A.dtype if output_dtype is None else output_dtype
B, T, H, BT = A.shape
if chunk_indices is None and cu_seqlens is not None:
chunk_indices = prepare_chunk_indices(cu_seqlens, BT)
NT = len(chunk_indices) if cu_seqlens is not None else triton.cdiv(T, BT)
Ai = torch.zeros_like(A, dtype=output_dtype)
if BT == 16:
merge_fn = solve_tril_16x16_kernel_kda
elif BT == 32:
merge_fn = merge_16x16_to_32x32_inverse_kernel_kda
elif BT == 64:
merge_fn = merge_16x16_to_64x64_inverse_kernel_kda
merge_fn[NT, B * H](
A=A,
Ai=Ai,
cu_seqlens=cu_seqlens,
chunk_indices=chunk_indices,
T=T,
H=H,
BT=BT,
DOT_PRECISION="ieee",
)
return Ai

97
vllm_ascend/ops/triton/kda/utils.py Normal file
View File

@@ -0,0 +1,97 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
#
# This file contains code copied from the flash-linear-attention project.
# The original source code was licensed under the MIT license and included
# the following copyright notice:
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
# ruff: noqa: E501
# mypy: ignore-errors
import contextlib
import functools
from collections.abc import Callable
from typing import Any
import torch
from vllm.triton_utils import triton
# Default chunk size used across FLA Triton kernels.
FLA_CHUNK_SIZE = 64
def tensor_cache(fn: Callable[..., torch.Tensor]) -> Callable[..., torch.Tensor]:
"""Cache recent tensor-derived metadata keyed by tensor object identity."""
cache_entries: tuple[tuple | None, dict | None, Any] = []
cache_size = 8
@functools.wraps(fn)
def wrapper(*args: Any, **kwargs: Any) -> Any:
nonlocal cache_entries
for i, entry in enumerate(cache_entries):
last_args, last_kwargs, last_result = entry
if (
len(args) == len(last_args)
and len(kwargs) == len(last_kwargs)
and all(a is b for a, b in zip(args, last_args))
and all(k in last_kwargs and v is last_kwargs[k] for k, v in kwargs.items())
):
cache_entries = cache_entries[:i] + cache_entries[i + 1 :] + [(args, kwargs, last_result)]
return last_result
result = fn(*args, **kwargs)
if len(cache_entries) >= cache_size:
cache_entries = cache_entries[1:]
cache_entries.append((args, kwargs, result))
return result
return wrapper
@tensor_cache
def prepare_lens(cu_seqlens: torch.LongTensor) -> torch.LongTensor:
return cu_seqlens[1:] - cu_seqlens[:-1]
@tensor_cache
def prepare_chunk_indices(cu_seqlens: torch.LongTensor, chunk_size: int) -> torch.LongTensor:
indices = torch.cat([torch.arange(n) for n in triton.cdiv(prepare_lens(cu_seqlens), chunk_size).tolist()])
return torch.stack([indices.eq(0).cumsum(0) - 1, indices], 1).to(cu_seqlens)
@tensor_cache
def prepare_chunk_offsets(cu_seqlens: torch.LongTensor, chunk_size: int) -> torch.LongTensor:
return torch.cat([cu_seqlens.new_tensor([0]), triton.cdiv(prepare_lens(cu_seqlens), chunk_size)]).cumsum(-1)
def input_guard(fn: Callable[..., torch.Tensor]) -> Callable[..., torch.Tensor]:
"""
A decorator to make sure all input tensors are contiguous and set the device based on input tensors.
"""
@functools.wraps(fn)
def wrapper(*args, **kwargs):
contiguous_args = (i if not isinstance(i, torch.Tensor) else i.contiguous() for i in args)
contiguous_kwargs = {k: (v if not isinstance(v, torch.Tensor) else v.contiguous()) for k, v in kwargs.items()}
tensor = None
for arg in args:
if isinstance(arg, torch.Tensor):
tensor = arg
break
if tensor is None:
for value in kwargs.values():
if isinstance(value, torch.Tensor):
tensor = value
break
if tensor is not None:
ctx = torch.npu.device(tensor.device.index)
else:
ctx = contextlib.nullcontext()
with ctx:
return fn(*contiguous_args, **contiguous_kwargs)
return wrapper