xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-14 18:15:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Kernel] Accelerate solve_tril with TMA (#26746)

Browse Source 
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
This commit is contained in:
Jiangyun Zhu 2025-10-20 13:39:02 +08:00 committed by GitHub
parent b63f2143f8
commit 9fce7bee74
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 414 additions and 303 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

43
vllm/model_executor/layers/fla/ops/op.py
View File

@ -11,29 +11,50 @@ import os
from vllm.triton_utils import tl, tldevice, triton from vllm.triton_utils import tl, tldevice, triton
from .utils import is_gather_supported
if os.environ.get("FLA_USE_FAST_OPS", "0") == "1": if os.environ.get("FLA_USE_FAST_OPS", "0") == "1":
div = tldevice.fast_dividef
exp = tldevice.fast_expf exp = tldevice.fast_expf
log = tldevice.fast_logf log = tldevice.fast_logf
log2 = tldevice.fast_log2f log2 = tldevice.fast_log2f
else: else:
@triton.jit
def div_normal(x, y):
return x / y
div = div_normal
exp = tl.exp exp = tl.exp
log = tl.log log = tl.log
log2 = tl.log2 log2 = tl.log2
if not hasattr(tl, "gather"): if not is_gather_supported:
@triton.jit @triton.jit
def gather(src, index, axis, _builder=None): def gather(src, index, axis, _builder=None):
# This is a fallback implementation when tl.gather is not supported """
# In order to pass triton compiler, there is no actual gather operation Gather operation that works when tl.gather is not supported.
return src This is a fallback implementation that returns None.
Just to make triton compiler happy.
"""
return None
else: else:
gather = tl.gather gather = tl.gather
if hasattr(triton.language, "_experimental_make_tensor_descriptor"):
# For Triton 3.3.x
make_tensor_descriptor = triton.language._experimental_make_tensor_descriptor
elif hasattr(triton.language, "make_tensor_descriptor"):
# For Triton 3.4.x and later
make_tensor_descriptor = triton.language.make_tensor_descriptor
else:
"""
Fallback implementation when TMA is not supported.
Returns None to indicate TMA descriptors are unavailable.
Just make triton compiler happy.
"""
@triton.jit
def make_tensor_descriptor(
base,
shape,
strides,
block_shape,
_builder=None,
):
return None

669
vllm/model_executor/layers/fla/ops/solve_tril.py
View File

@ -8,12 +8,21 @@
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang # Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
# ruff: noqa: E501 # ruff: noqa: E501
import os
import torch import torch
from vllm.triton_utils import tl, triton from vllm.triton_utils import tl, triton
from .index import prepare_chunk_indices from .index import prepare_chunk_indices
from .utils import input_guard from .op import make_tensor_descriptor
from .utils import input_guard, is_amd, is_tma_supported
FLA_TRIL_PRECISION = os.environ.get("FLA_TRIL_PRECISION", "ieee")
ALLOWED_TRIL_PRECISIONS = ["ieee", "tf32"] if is_amd else ["ieee", "tf32", "tf32x3"]
assert FLA_TRIL_PRECISION in ALLOWED_TRIL_PRECISIONS, (
f"FLA_TRIL_PRECISION must be one of {ALLOWED_TRIL_PRECISIONS}, but got {FLA_TRIL_PRECISION}"
)
@triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None}) @triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None})
@ -28,13 +37,15 @@ from .utils import input_guard
@triton.jit(do_not_specialize=["T"]) @triton.jit(do_not_specialize=["T"])
def solve_tril_16x16_kernel( def solve_tril_16x16_kernel(
A, A,
Ad, Ai,
cu_seqlens, cu_seqlens,
chunk_indices, chunk_indices,
T, T,
H: tl.constexpr, H: tl.constexpr,
BT: tl.constexpr, BT: tl.constexpr,
USE_TMA: tl.constexpr,
IS_VARLEN: tl.constexpr, IS_VARLEN: tl.constexpr,
DOT_PRECISION: tl.constexpr,
): ):
i_t, i_bh = tl.program_id(0), tl.program_id(1) i_t, i_bh = tl.program_id(0), tl.program_id(1)
i_b, i_h = i_bh // H, i_bh % H i_b, i_h = i_bh // H, i_bh % H
@ -50,30 +61,43 @@ def solve_tril_16x16_kernel(
T = eos - bos T = eos - bos
else: else:
bos, eos = i_b * T, i_b * T + T 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 A = A + (bos * H + i_h) * BT
Ad = Ad + (bos * H + i_h) * 16 Ai = Ai + (bos * H + i_h) * 16
offset = (i_t * 16) % BT offset = (i_t * 16) % BT
p_A = tl.make_block_ptr( if not USE_TMA:
A, (T, BT), (H * BT, 1), (i_t * 16, offset), (16, 16), (1, 0) p_A = tl.make_block_ptr(
) A, (T, BT), (H * BT, 1), (i_t * 16, offset), (16, 16), (1, 0)
p_Ai = tl.make_block_ptr(Ad, (T, 16), (H * 16, 1), (i_t * 16, 0), (16, 16), (1, 0)) )
b_A = tl.load(p_A, boundary_check=(0, 1)).to(tl.float32) # [16, 16]
b_A = -tl.where(tl.arange(0, 16)[:, None] > tl.arange(0, 16)[None, :], b_A, 0) b_A = tl.load(p_A, boundary_check=(0, 1)).to(tl.float32)
else:
desc = make_tensor_descriptor(A, [T, BT], [H * BT, 1], [16, 16])
desc_o = make_tensor_descriptor(Ai, [T, 16], [H * 16, 1], [16, 16])
b_A = desc.load([i_t * 16, offset]).to(tl.float32)
b_A = -tl.where(m_A, b_A, 0)
o_i = tl.arange(0, 16) for i in range(2, min(16, T - i_t * 16)):
for i in range(1, min(16, T - i_t * 16)): # [16]
b_a = -tl.load(A + (i_t * 16 + i) * H * BT + o_i + offset) 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 = b_a + tl.sum(b_a[:, None] * b_A, 0)
mask = o_i == i b_A = tl.where((o_i == i)[:, None], b_a, b_A)
b_A = tl.where(mask[:, None], b_a, b_A) b_A += m_I
b_A += o_i[:, None] == o_i[None, :] if not USE_TMA:
tl.store( p_Ai = tl.make_block_ptr(
p_Ai, Ai, (T, 16), (H * 16, 1), (i_t * 16, 0), (16, 16), (1, 0)
b_A.to(p_Ai.dtype.element_ty, fp_downcast_rounding="rtne"), )
boundary_check=(0, 1), tl.store(
) p_Ai,
b_A.to(p_Ai.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
else:
desc_o.store([i_t * 16, 0], b_A.to(desc_o.dtype, fp_downcast_rounding="rtne"))
@triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None}) @triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None})
@ -88,14 +112,15 @@ def solve_tril_16x16_kernel(
@triton.jit(do_not_specialize=["T"]) @triton.jit(do_not_specialize=["T"])
def merge_16x16_to_32x32_inverse_kernel( def merge_16x16_to_32x32_inverse_kernel(
A, A,
Ad,
Ai, Ai,
cu_seqlens, cu_seqlens,
chunk_indices, chunk_indices,
T, T,
H: tl.constexpr, H: tl.constexpr,
BT: tl.constexpr, BT: tl.constexpr,
USE_TMA: tl.constexpr,
IS_VARLEN: tl.constexpr, IS_VARLEN: tl.constexpr,
DOT_PRECISION: tl.constexpr,
): ):
i_t, i_bh = tl.program_id(0), tl.program_id(1) i_t, i_bh = tl.program_id(0), tl.program_id(1)
i_b, i_h = i_bh // H, i_bh % H i_b, i_h = i_bh // H, i_bh % H
@ -112,50 +137,92 @@ def merge_16x16_to_32x32_inverse_kernel(
else: else:
bos, eos = i_b * T, i_b * T + T bos, eos = i_b * T, i_b * T + T
A += (bos * H + i_h) * 32 o_i = tl.arange(0, 16)
Ad += (bos * H + i_h) * 16 m_A = o_i[:, None] > o_i[None, :]
Ai += (bos * H + i_h) * 32 m_I = o_i[:, None] == o_i[None, :]
A += (bos * H + i_h) * BT
Ai += (bos * H + i_h) * BT
p_A_21 = tl.make_block_ptr( if not USE_TMA:
A, (T, 32), (H * 32, 1), (i_t * 32 + 16, 0), (16, 16), (1, 0) p_A_11 = tl.make_block_ptr(
) A, (T, BT), (H * BT, 1), (i_t * BT, 0), (16, 16), (1, 0)
p_Ad_11 = tl.make_block_ptr( )
Ad, (T, 16), (H * 16, 1), (i_t * 32, 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_Ad_22 = tl.make_block_ptr( )
Ad, (T, 16), (H * 16, 1), (i_t * 32 + 16, 0), (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)
p_Ai_11 = tl.make_block_ptr( else:
Ai, (T, 32), (H * 32, 1), (i_t * 32, 0), (16, 16), (1, 0) desc = make_tensor_descriptor(A, [T, BT], [H * BT, 1], [16, 16])
) desc_o = make_tensor_descriptor(Ai, [T, BT], [H * BT, 1], [16, 16])
p_Ai_22 = tl.make_block_ptr( b_Ai_11 = desc.load([i_t * BT + 0, 0]).to(tl.float32)
Ai, (T, 32), (H * 32, 1), (i_t * 32 + 16, 16), (16, 16), (1, 0) b_Ai_22 = desc.load([i_t * BT + 16, 16]).to(tl.float32)
)
p_Ai_21 = tl.make_block_ptr( # [16, 16]
Ai, (T, 32), (H * 32, 1), (i_t * 32 + 16, 0), (16, 16), (1, 0) 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
if not USE_TMA:
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)
else:
b_A_21 = desc.load([i_t * BT + 16, 0]).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,
) )
A_21 = tl.load(p_A_21, boundary_check=(0, 1)).to(tl.float32) if not USE_TMA:
Ai_11 = tl.load(p_Ad_11, boundary_check=(0, 1)).to(tl.float32) p_Ai_11 = tl.make_block_ptr(
Ai_22 = tl.load(p_Ad_22, boundary_check=(0, 1)).to(tl.float32) Ai, (T, BT), (H * BT, 1), (i_t * BT, 0), (16, 16), (1, 0)
Ai_21 = -tl.dot( )
tl.dot(Ai_22, A_21, input_precision="ieee"), Ai_11, input_precision="ieee" p_Ai_21 = tl.make_block_ptr(
) Ai, (T, BT), (H * BT, 1), (i_t * BT + 16, 0), (16, 16), (1, 0)
tl.store( )
p_Ai_11, p_Ai_22 = tl.make_block_ptr(
Ai_11.to(p_Ai_11.dtype.element_ty, fp_downcast_rounding="rtne"), Ai, (T, BT), (H * BT, 1), (i_t * BT + 16, 16), (16, 16), (1, 0)
boundary_check=(0, 1), )
) tl.store(
tl.store( p_Ai_11,
p_Ai_22, b_Ai_11.to(p_Ai_11.dtype.element_ty, fp_downcast_rounding="rtne"),
Ai_22.to(p_Ai_22.dtype.element_ty, fp_downcast_rounding="rtne"), boundary_check=(0, 1),
boundary_check=(0, 1), )
) tl.store(
tl.store( p_Ai_22,
p_Ai_21, b_Ai_22.to(p_Ai_22.dtype.element_ty, fp_downcast_rounding="rtne"),
Ai_21.to(p_Ai_21.dtype.element_ty, fp_downcast_rounding="rtne"), boundary_check=(0, 1),
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),
)
else:
desc_o.store(
[i_t * BT + 0, 0], b_Ai_11.to(desc_o.dtype, fp_downcast_rounding="rtne")
)
desc_o.store(
[i_t * BT + 16, 0], b_Ai_21.to(desc_o.dtype, fp_downcast_rounding="rtne")
)
desc_o.store(
[i_t * BT + 16, 16], b_Ai_22.to(desc_o.dtype, fp_downcast_rounding="rtne")
)
@triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None}) @triton.heuristics({"IS_VARLEN": lambda args: args["cu_seqlens"] is not None})
@ -170,14 +237,15 @@ def merge_16x16_to_32x32_inverse_kernel(
@triton.jit(do_not_specialize=["T"]) @triton.jit(do_not_specialize=["T"])
def merge_16x16_to_64x64_inverse_kernel( def merge_16x16_to_64x64_inverse_kernel(
A, A,
Ad,
Ai, Ai,
cu_seqlens, cu_seqlens,
chunk_indices, chunk_indices,
T, T,
H: tl.constexpr, H: tl.constexpr,
BT: tl.constexpr, BT: tl.constexpr,
USE_TMA: tl.constexpr,
IS_VARLEN: tl.constexpr, IS_VARLEN: tl.constexpr,
DOT_PRECISION: tl.constexpr,
): ):
i_t, i_bh = tl.program_id(0), tl.program_id(1) i_t, i_bh = tl.program_id(0), tl.program_id(1)
i_b, i_h = i_bh // H, i_bh % H i_b, i_h = i_bh // H, i_bh % H
@ -194,213 +262,245 @@ def merge_16x16_to_64x64_inverse_kernel(
else: else:
bos, eos = i_b * T, i_b * T + T bos, eos = i_b * T, i_b * T + T
A += (bos * H + i_h) * 64 o_i = tl.arange(0, 16)
Ad += (bos * H + i_h) * 16 m_A = o_i[:, None] > o_i[None, :]
Ai += (bos * H + i_h) * 64 m_I = o_i[:, None] == o_i[None, :]
A += (bos * H + i_h) * BT
Ai += (bos * H + i_h) * BT
p_A_21 = tl.make_block_ptr( if not USE_TMA:
A, (T, 64), (H * 64, 1), (i_t * 64 + 16, 0), (16, 16), (1, 0) 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)
else:
desc = make_tensor_descriptor(A, [T, BT], [H * BT, 1], [16, 16])
desc_o = make_tensor_descriptor(Ai, [T, BT], [H * BT, 1], [16, 16])
b_Ai_11 = desc.load([i_t * BT + 0, 0]).to(tl.float32)
b_Ai_22 = desc.load([i_t * BT + 16, 16]).to(tl.float32)
b_Ai_33 = desc.load([i_t * BT + 32, 32]).to(tl.float32)
b_Ai_44 = desc.load([i_t * BT + 48, 48]).to(tl.float32)
# [16, 16]
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
if not USE_TMA:
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)
else:
b_A_21 = desc.load([i_t * BT + 16, 0]).to(tl.float32)
b_A_31 = desc.load([i_t * BT + 32, 0]).to(tl.float32)
b_A_32 = desc.load([i_t * BT + 32, 16]).to(tl.float32)
b_A_41 = desc.load([i_t * BT + 48, 0]).to(tl.float32)
b_A_42 = desc.load([i_t * BT + 48, 16]).to(tl.float32)
b_A_43 = desc.load([i_t * BT + 48, 32]).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_A_32 = tl.make_block_ptr( b_Ai_32 = -tl.dot(
A, (T, 64), (H * 64, 1), (i_t * 64 + 32, 16), (16, 16), (1, 0) tl.dot(b_Ai_33, b_A_32, input_precision=DOT_PRECISION),
b_Ai_22,
input_precision=DOT_PRECISION,
) )
p_A_31 = tl.make_block_ptr( b_Ai_43 = -tl.dot(
A, (T, 64), (H * 64, 1), (i_t * 64 + 32, 0), (16, 16), (1, 0) tl.dot(b_Ai_44, b_A_43, input_precision=DOT_PRECISION),
) b_Ai_33,
p_A_43 = tl.make_block_ptr( input_precision=DOT_PRECISION,
A, (T, 64), (H * 64, 1), (i_t * 64 + 48, 32), (16, 16), (1, 0)
)
p_A_42 = tl.make_block_ptr(
A, (T, 64), (H * 64, 1), (i_t * 64 + 48, 16), (16, 16), (1, 0)
)
p_A_41 = tl.make_block_ptr(
A, (T, 64), (H * 64, 1), (i_t * 64 + 48, 0), (16, 16), (1, 0)
)
p_Ad_11 = tl.make_block_ptr(
Ad, (T, 16), (H * 16, 1), (i_t * 64, 0), (16, 16), (1, 0)
)
p_Ad_22 = tl.make_block_ptr(
Ad, (T, 16), (H * 16, 1), (i_t * 64 + 16, 0), (16, 16), (1, 0)
)
p_Ad_33 = tl.make_block_ptr(
Ad, (T, 16), (H * 16, 1), (i_t * 64 + 32, 0), (16, 16), (1, 0)
)
p_Ad_44 = tl.make_block_ptr(
Ad, (T, 16), (H * 16, 1), (i_t * 64 + 48, 0), (16, 16), (1, 0)
) )
A_21 = tl.load(p_A_21, boundary_check=(0, 1)).to(tl.float32) b_Ai_31 = -tl.dot(
A_32 = tl.load(p_A_32, boundary_check=(0, 1)).to(tl.float32) b_Ai_33,
A_31 = tl.load(p_A_31, boundary_check=(0, 1)).to(tl.float32) tl.dot(b_A_31, b_Ai_11, input_precision=DOT_PRECISION)
A_43 = tl.load(p_A_43, boundary_check=(0, 1)).to(tl.float32) + tl.dot(b_A_32, b_Ai_21, input_precision=DOT_PRECISION),
A_42 = tl.load(p_A_42, boundary_check=(0, 1)).to(tl.float32) input_precision=DOT_PRECISION,
A_41 = tl.load(p_A_41, boundary_check=(0, 1)).to(tl.float32)
Ai_11 = tl.load(p_Ad_11, boundary_check=(0, 1)).to(tl.float32)
Ai_22 = tl.load(p_Ad_22, boundary_check=(0, 1)).to(tl.float32)
Ai_33 = tl.load(p_Ad_33, boundary_check=(0, 1)).to(tl.float32)
Ai_44 = tl.load(p_Ad_44, boundary_check=(0, 1)).to(tl.float32)
Ai_21 = -tl.dot(
tl.dot(Ai_22, A_21, input_precision="ieee"), Ai_11, input_precision="ieee"
) )
Ai_32 = -tl.dot( b_Ai_42 = -tl.dot(
tl.dot(Ai_33, A_32, input_precision="ieee"), Ai_22, input_precision="ieee" 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,
) )
Ai_43 = -tl.dot( b_Ai_41 = -tl.dot(
tl.dot(Ai_44, A_43, input_precision="ieee"), Ai_33, input_precision="ieee" 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,
) )
Ai_31 = -tl.dot( if not USE_TMA:
Ai_33, p_Ai_11 = tl.make_block_ptr(
tl.dot(A_31, Ai_11, input_precision="ieee") Ai, (T, BT), (H * BT, 1), (i_t * BT, 0), (16, 16), (1, 0)
+ tl.dot(A_32, Ai_21, input_precision="ieee"), )
input_precision="ieee", p_Ai_22 = tl.make_block_ptr(
) Ai, (T, BT), (H * BT, 1), (i_t * BT + 16, 16), (16, 16), (1, 0)
Ai_42 = -tl.dot( )
Ai_44, p_Ai_33 = tl.make_block_ptr(
tl.dot(A_42, Ai_22, input_precision="ieee") Ai, (T, BT), (H * BT, 1), (i_t * BT + 32, 32), (16, 16), (1, 0)
+ tl.dot(A_43, Ai_32, input_precision="ieee"), )
input_precision="ieee", p_Ai_44 = tl.make_block_ptr(
) Ai, (T, BT), (H * BT, 1), (i_t * BT + 48, 48), (16, 16), (1, 0)
Ai_41 = -tl.dot( )
Ai_44, p_Ai_21 = tl.make_block_ptr(
tl.dot(A_41, Ai_11, input_precision="ieee") Ai, (T, BT), (H * BT, 1), (i_t * BT + 16, 0), (16, 16), (1, 0)
+ tl.dot(A_42, Ai_21, input_precision="ieee") )
+ tl.dot(A_43, Ai_31, input_precision="ieee"), p_Ai_31 = tl.make_block_ptr(
input_precision="ieee", Ai, (T, BT), (H * BT, 1), (i_t * BT + 32, 0), (16, 16), (1, 0)
) )
p_Ai_32 = tl.make_block_ptr(
p_Ai_11 = tl.make_block_ptr( Ai, (T, BT), (H * BT, 1), (i_t * BT + 32, 16), (16, 16), (1, 0)
Ai, (T, 64), (H * 64, 1), (i_t * 64, 0), (16, 16), (1, 0) )
) p_Ai_41 = tl.make_block_ptr(
p_Ai_22 = tl.make_block_ptr( Ai, (T, BT), (H * BT, 1), (i_t * BT + 48, 0), (16, 16), (1, 0)
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 16, 16), (16, 16), (1, 0) )
) p_Ai_42 = tl.make_block_ptr(
p_Ai_33 = tl.make_block_ptr( Ai, (T, BT), (H * BT, 1), (i_t * BT + 48, 16), (16, 16), (1, 0)
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 32, 32), (16, 16), (1, 0) )
) p_Ai_43 = tl.make_block_ptr(
p_Ai_44 = tl.make_block_ptr( Ai, (T, BT), (H * BT, 1), (i_t * BT + 48, 32), (16, 16), (1, 0)
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 48, 48), (16, 16), (1, 0) )
) tl.store(
p_Ai_21 = tl.make_block_ptr( p_Ai_11,
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 16, 0), (16, 16), (1, 0) b_Ai_11.to(p_Ai_11.dtype.element_ty, fp_downcast_rounding="rtne"),
) boundary_check=(0, 1),
p_Ai_31 = tl.make_block_ptr( )
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 32, 0), (16, 16), (1, 0) tl.store(
) p_Ai_22,
p_Ai_32 = tl.make_block_ptr( b_Ai_22.to(p_Ai_22.dtype.element_ty, fp_downcast_rounding="rtne"),
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 32, 16), (16, 16), (1, 0) boundary_check=(0, 1),
) )
p_Ai_41 = tl.make_block_ptr( tl.store(
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 48, 0), (16, 16), (1, 0) p_Ai_33,
) b_Ai_33.to(p_Ai_33.dtype.element_ty, fp_downcast_rounding="rtne"),
p_Ai_42 = tl.make_block_ptr( boundary_check=(0, 1),
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 48, 16), (16, 16), (1, 0) )
) tl.store(
p_Ai_43 = tl.make_block_ptr( p_Ai_44,
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 48, 32), (16, 16), (1, 0) b_Ai_44.to(p_Ai_44.dtype.element_ty, fp_downcast_rounding="rtne"),
) boundary_check=(0, 1),
tl.store( )
p_Ai_11, tl.store(
Ai_11.to(p_Ai_11.dtype.element_ty, fp_downcast_rounding="rtne"), p_Ai_21,
boundary_check=(0, 1), b_Ai_21.to(p_Ai_21.dtype.element_ty, fp_downcast_rounding="rtne"),
) boundary_check=(0, 1),
tl.store( )
p_Ai_22, tl.store(
Ai_22.to(p_Ai_22.dtype.element_ty, fp_downcast_rounding="rtne"), p_Ai_31,
boundary_check=(0, 1), b_Ai_31.to(p_Ai_31.dtype.element_ty, fp_downcast_rounding="rtne"),
) boundary_check=(0, 1),
tl.store( )
p_Ai_33, tl.store(
Ai_33.to(p_Ai_33.dtype.element_ty, fp_downcast_rounding="rtne"), p_Ai_32,
boundary_check=(0, 1), b_Ai_32.to(p_Ai_32.dtype.element_ty, fp_downcast_rounding="rtne"),
) boundary_check=(0, 1),
tl.store( )
p_Ai_44, tl.store(
Ai_44.to(p_Ai_44.dtype.element_ty, fp_downcast_rounding="rtne"), p_Ai_41,
boundary_check=(0, 1), b_Ai_41.to(p_Ai_41.dtype.element_ty, fp_downcast_rounding="rtne"),
) boundary_check=(0, 1),
tl.store( )
p_Ai_21, tl.store(
Ai_21.to(p_Ai_21.dtype.element_ty, fp_downcast_rounding="rtne"), p_Ai_42,
boundary_check=(0, 1), b_Ai_42.to(p_Ai_42.dtype.element_ty, fp_downcast_rounding="rtne"),
) boundary_check=(0, 1),
tl.store( )
p_Ai_31, tl.store(
Ai_31.to(p_Ai_31.dtype.element_ty, fp_downcast_rounding="rtne"), p_Ai_43,
boundary_check=(0, 1), b_Ai_43.to(p_Ai_43.dtype.element_ty, fp_downcast_rounding="rtne"),
) boundary_check=(0, 1),
tl.store( )
p_Ai_32, else:
Ai_32.to(p_Ai_32.dtype.element_ty, fp_downcast_rounding="rtne"), desc_o.store(
boundary_check=(0, 1), [i_t * BT + 0, 0], b_Ai_11.to(desc_o.dtype, fp_downcast_rounding="rtne")
) )
tl.store( desc_o.store(
p_Ai_41, [i_t * BT + 16, 16], b_Ai_22.to(desc_o.dtype, fp_downcast_rounding="rtne")
Ai_41.to(p_Ai_41.dtype.element_ty, fp_downcast_rounding="rtne"), )
boundary_check=(0, 1), desc_o.store(
) [i_t * BT + 32, 32], b_Ai_33.to(desc_o.dtype, fp_downcast_rounding="rtne")
tl.store( )
p_Ai_42, desc_o.store(
Ai_42.to(p_Ai_42.dtype.element_ty, fp_downcast_rounding="rtne"), [i_t * BT + 48, 48], b_Ai_44.to(desc_o.dtype, fp_downcast_rounding="rtne")
boundary_check=(0, 1), )
) desc_o.store(
tl.store( [i_t * BT + 16, 0], b_Ai_21.to(desc_o.dtype, fp_downcast_rounding="rtne")
p_Ai_43, )
Ai_43.to(p_Ai_43.dtype.element_ty, fp_downcast_rounding="rtne"), desc_o.store(
boundary_check=(0, 1), [i_t * BT + 32, 0], b_Ai_31.to(desc_o.dtype, fp_downcast_rounding="rtne")
) )
desc_o.store(
fill_zeros = tl.zeros((16, 16), dtype=tl.float32) [i_t * BT + 32, 16], b_Ai_32.to(desc_o.dtype, fp_downcast_rounding="rtne")
p_Ai_12 = tl.make_block_ptr( )
Ai, (T, 64), (H * 64, 1), (i_t * 64, 16), (16, 16), (1, 0) desc_o.store(
) [i_t * BT + 48, 0], b_Ai_41.to(desc_o.dtype, fp_downcast_rounding="rtne")
p_Ai_13 = tl.make_block_ptr( )
Ai, (T, 64), (H * 64, 1), (i_t * 64, 32), (16, 16), (1, 0) desc_o.store(
) [i_t * BT + 48, 16], b_Ai_42.to(desc_o.dtype, fp_downcast_rounding="rtne")
p_Ai_14 = tl.make_block_ptr( )
Ai, (T, 64), (H * 64, 1), (i_t * 64, 48), (16, 16), (1, 0) desc_o.store(
) [i_t * BT + 48, 32], b_Ai_43.to(desc_o.dtype, fp_downcast_rounding="rtne")
p_Ai_23 = tl.make_block_ptr( )
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 16, 32), (16, 16), (1, 0)
)
p_Ai_24 = tl.make_block_ptr(
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 16, 48), (16, 16), (1, 0)
)
p_Ai_34 = tl.make_block_ptr(
Ai, (T, 64), (H * 64, 1), (i_t * 64 + 32, 48), (16, 16), (1, 0)
)
tl.store(
p_Ai_12,
fill_zeros.to(p_Ai_12.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_13,
fill_zeros.to(p_Ai_13.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_14,
fill_zeros.to(p_Ai_14.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_23,
fill_zeros.to(p_Ai_23.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_24,
fill_zeros.to(p_Ai_24.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
tl.store(
p_Ai_34,
fill_zeros.to(p_Ai_34.dtype.element_ty, fp_downcast_rounding="rtne"),
boundary_check=(0, 1),
)
@input_guard @input_guard
@ -410,62 +510,47 @@ def solve_tril(
output_dtype: torch.dtype = torch.float, output_dtype: torch.dtype = torch.float,
) -> torch.Tensor: ) -> torch.Tensor:
""" """
Compute the inverse of the lower triangular matrix Compute the inverse of the matrix I + A
A should be strictly lower triangular, i.e., A.triu() == 0. A should be strictly lower triangular, i.e., A.triu() == 0.
Args: Args:
A (torch.Tensor): A (torch.Tensor):
[B, T, H, K] [B, T, H, BT], where BT should only be 16, 32, or 64.
cu_seqlens (torch.Tensor): cu_seqlens (torch.Tensor):
The cumulative sequence lengths of the input tensor. The cumulative sequence lengths of the input tensor. Default: `None`.
Default: None.
output_dtype (torch.dtype): output_dtype (torch.dtype):
The dtype of the output tensor. Default: `torch.float` The dtype of the output tensor. Default: `torch.float`.
If `None`, the output dtype will be the same as the input dtype.
Returns: Returns:
(I + A)^-1 with the same shape as A (I + A)^-1 with the same shape as A
""" """
assert A.shape[-1] in [16, 32, 64] 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 B, T, H, BT = A.shape
Ad = torch.empty(
B, T, H, 16, device=A.device, dtype=torch.float if BT != 16 else output_dtype
)
chunk_indices = (
prepare_chunk_indices(cu_seqlens, 16) if cu_seqlens is not None else None
)
NT = len(chunk_indices) if cu_seqlens is not None else triton.cdiv(T, 16)
solve_tril_16x16_kernel[NT, B * H](
A=A,
Ad=Ad,
cu_seqlens=cu_seqlens,
chunk_indices=chunk_indices,
T=T,
H=H,
BT=BT,
)
if BT == 16:
return Ad
Ai = torch.empty(B, T, H, BT, device=A.device, dtype=output_dtype)
merge_fn = (
merge_16x16_to_32x32_inverse_kernel
if BT == 32
else merge_16x16_to_64x64_inverse_kernel
)
chunk_indices = ( chunk_indices = (
prepare_chunk_indices(cu_seqlens, BT) if cu_seqlens is not None else None prepare_chunk_indices(cu_seqlens, BT) if cu_seqlens is not None else None
) )
NT = len(chunk_indices) if cu_seqlens is not None else triton.cdiv(T, 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
elif BT == 32:
merge_fn = merge_16x16_to_32x32_inverse_kernel
elif BT == 64:
merge_fn = merge_16x16_to_64x64_inverse_kernel
merge_fn[NT, B * H]( merge_fn[NT, B * H](
A=A, A=A,
Ad=Ad,
Ai=Ai, Ai=Ai,
cu_seqlens=cu_seqlens, cu_seqlens=cu_seqlens,
chunk_indices=chunk_indices, chunk_indices=chunk_indices,
T=T, T=T,
H=H, H=H,
BT=BT, BT=BT,
USE_TMA=is_tma_supported,
DOT_PRECISION=FLA_TRIL_PRECISION,
) )
return Ai return Ai

5
vllm/model_executor/layers/fla/ops/utils.py
View File

@ -150,6 +150,11 @@ is_nvidia_hopper = is_nvidia and (
or torch.cuda.get_device_capability()[0] >= 9 or torch.cuda.get_device_capability()[0] >= 9
) )
use_cuda_graph = is_nvidia and os.environ.get("FLA_USE_CUDA_GRAPH", "0") == "1" use_cuda_graph = is_nvidia and os.environ.get("FLA_USE_CUDA_GRAPH", "0") == "1"
is_gather_supported = hasattr(triton.language, "gather")
is_tma_supported = (is_nvidia and torch.cuda.get_device_capability(0)[0] >= 9) and (
hasattr(triton.language, "_experimental_make_tensor_descriptor")
or hasattr(triton.language, "make_tensor_descriptor")
)
def get_all_max_shared_mem(): def get_all_max_shared_mem():