[Perf][Feat] Add generic cuteDSL LL FP32 router (GEMM) (#44343)

Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Roberto L. Castro
2026-10-13 12:38:46 -04:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 6849ca696a
commit 7fad7db330
7 changed files with 1447 additions and 101 deletions
@@ -0,0 +1,457 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import cutlass
import cutlass.cute as cute
from cuda.bindings.driver import CUstream
from cutlass import const_expr
class LLFp32WDotprod:
"""FP32-weight router GEMM kernel based on CTA-local dot products.
This kernel computes C[M, N] = A[M, K] @ B[N, K]^T for bf16/fp16/fp32
activations, fp32 weights, and fp32 output. It launches one CTA per
small output-column tile, distributes K across CTA threads with vectorized
loads, accumulates one fp32 dot product per token/expert, and reduces
through warp shuffles plus shared memory.
:param k: Compile-time K dimension specialized into the generated kernel.
:type k: int
:param bs: Threads per CTA and K-stripe width used by the reduction.
:type bs: int
:param main_vec_width: Elements loaded per thread in the main loop.
:type main_vec_width: int
:param tail_vec_width: Elements loaded per thread in the vector tail.
:type tail_vec_width: int
:param use_pdl: Whether to launch with Programmatic Dependent Launch.
:type use_pdl: bool
:note: Supported A/B data types:
- BFloat16/Float32
- Float16/Float32
- Float32/Float32
:note: Supported accumulator data types:
- Float32
:note: Supported C data types:
- Float32
:note: Constraints:
- Inputs must be contiguous row-major 2D tensors.
:note: K is handled as vectorized main/tail loops plus scalar remainder.
:compile-key: ``(M, K, bs, epb, act_dtype)`` selects the token count,
hidden size, CTA thread/K-stripe width, experts per CTA, and activation
input type.
"""
def __init__(
self,
m: int,
k: int,
bs: int = 128,
main_vec_width: int = 8,
tail_vec_width: int = 4,
token_groups: int = 1,
epb: int = 1,
use_pdl: bool = False,
):
"""Initialize the dot-product kernel configuration.
This configuration fixes the CTA thread count, reduction warp count,
vector widths, and K-loop decomposition used by the generated kernel.
:param m: Token count M selected by the compile key.
:type m: int
:param k: Hidden size K used to specialize the K-loop decomposition.
:type k: int
:param bs: Threads per CTA and K-stripe width for the reduction.
:type bs: int
:param main_vec_width: Elements loaded per thread in the main loop.
:type main_vec_width: int
:param tail_vec_width: Elements loaded per thread in the vector tail.
:type tail_vec_width: int
:param use_pdl: Whether to launch with Programmatic Dependent Launch.
:type use_pdl: bool
"""
self.m = m
self.m_per_group = m // token_groups
self.bs = bs
self.main_vec_width = main_vec_width
self.tail_vec_width = tail_vec_width
self.token_groups = token_groups
self.epb = epb
self.use_pdl = use_pdl
self.num_warps = bs // cute.arch.WARP_SIZE
self._init_k_tiles(k)
def _vectorized_elems(self, k_extent: int, vec_width: int) -> int:
vector_tile = vec_width * self.bs
return (k_extent // vector_tile) * vector_tile
def _init_k_tiles(self, k: int) -> None:
"""Split K into vector loops, scalar rounds, and ragged tail."""
self.k_main_elems = self._vectorized_elems(k, self.main_vec_width)
self.k_after_main = k - self.k_main_elems
self.k_tail_elems = self._vectorized_elems(
self.k_after_main, self.tail_vec_width
)
self.k_done_all = self.k_main_elems + self.k_tail_elems
self.scalar_rem = k - self.k_done_all
self.ks_full = self.scalar_rem // self.bs
self.ks_part = self.scalar_rem % self.bs
self.k_scalar_full = self.ks_full * self.bs
self.k_part_offset = self.k_done_all + self.k_scalar_full
self.main_tiles = self.k_main_elems // (self.main_vec_width * self.bs)
self.tail_tiles = self.k_tail_elems // (self.tail_vec_width * self.bs)
@cute.jit
def _vector_dotprod(
self,
acc: cute.Tensor,
tA: cute.Tensor,
tB: cute.Tensor,
M: cutlass.Constexpr,
epb: cutlass.Constexpr,
num_tiles: cutlass.Constexpr,
):
for tile in cutlass.range_constexpr(num_tiles):
bt0 = tB[0, None, tile]
br0 = cute.make_rmem_tensor_like(bt0)
cute.autovec_copy(bt0, br0)
br0_f32 = br0.load().to(cutlass.Float32)
if const_expr(epb == 2):
bt1 = tB[1, None, tile]
br1 = cute.make_rmem_tensor_like(bt1)
cute.autovec_copy(bt1, br1)
br1_f32 = br1.load().to(cutlass.Float32)
for m in cutlass.range_constexpr(M):
at = tA[m, None, tile]
ar = cute.make_rmem_tensor_like(at)
cute.autovec_copy(at, ar)
vec_width: cutlass.Constexpr = cute.size(ar)
for v in cutlass.range_constexpr(vec_width):
av = ar[v].to(cutlass.Float32)
acc[m, 0] = acc[m, 0] + av * br0_f32[v]
if const_expr(epb == 2):
acc[m, 1] = acc[m, 1] + av * br1_f32[v]
def _make_thread_vector_slice(
self,
gA_vec: cute.Tensor,
gB_vec: cute.Tensor,
tidx: cutlass.Int32,
bs: cutlass.Constexpr,
):
# A: (M, K_TILE, K_LANE, K_VEC); B: (epb, K_TILE, K_LANE, K_VEC).
tA = cute.logical_divide(gA_vec, (None, (None, bs)))
tB = cute.logical_divide(gB_vec, (None, (None, bs)))
return tA[None, (None, (tidx, None))], tB[None, (None, (tidx, None))]
def _make_m_slice(
self,
gX: cute.Tensor,
m_offset: cutlass.Int32,
m_extent: cutlass.Constexpr,
):
return cute.local_tile(
cute.domain_offset((m_offset, 0), gX),
(m_extent, cute.size(gX, mode=[1])),
(0, 0),
)
def _make_n_tile(
self,
gX: cute.Tensor,
n_tile_idx: cutlass.Int32,
n_extent: cutlass.Constexpr,
):
return cute.local_tile(
gX,
(n_extent, cute.size(gX, mode=[1])),
(n_tile_idx, 0),
)
def _make_sm_m_slice(
self,
sm: cute.Tensor,
m_offset: cutlass.Int32,
m_extent: cutlass.Constexpr,
epb: cutlass.Constexpr,
num_warps: cutlass.Constexpr,
):
return cute.local_tile(
cute.domain_offset((m_offset, 0, 0), sm),
(m_extent, epb, num_warps),
(0, 0, 0),
)
def _make_k_slice(
self,
gX: cute.Tensor,
k_offset: cutlass.Constexpr,
k_extent: cutlass.Constexpr,
):
k_layout_extent: cutlass.Constexpr = (
1 if const_expr(k_extent == 0) else k_extent
)
if const_expr(k_offset == 0):
return cute.local_tile(
gX, (cute.size(gX, mode=[0]), k_layout_extent), (0, 0)
)
return cute.local_tile(
cute.domain_offset((0, k_offset), gX),
(cute.size(gX, mode=[0]), k_layout_extent),
(0, 0),
)
@cute.jit
def _compute_token_group(
self,
gA: cute.Tensor,
gB: cute.Tensor,
sm: cute.Tensor,
n_tile_idx: cutlass.Int32,
tidx: cutlass.Int32,
local_wid: cutlass.Int32,
m_per_group: cutlass.Constexpr,
m_offset: cutlass.Int32,
main_vec_width: cutlass.Constexpr,
tail_vec_width: cutlass.Constexpr,
bs: cutlass.Constexpr,
epb: cutlass.Constexpr,
num_warps: cutlass.Constexpr,
k_main_elems: cutlass.Constexpr,
k_tail_elems: cutlass.Constexpr,
k_done_all: cutlass.Constexpr,
ks_full: cutlass.Constexpr,
ks_part: cutlass.Constexpr,
k_scalar_full: cutlass.Constexpr,
k_part_offset: cutlass.Constexpr,
main_tiles: cutlass.Constexpr,
tail_tiles: cutlass.Constexpr,
):
# Each CTA reuses A for epb adjacent experts.
acc = cute.make_rmem_tensor((m_per_group, epb), cutlass.Float32)
acc.fill(0.0)
gA_group = self._make_m_slice(gA, m_offset, m_per_group)
gB_epb = self._make_n_tile(gB, n_tile_idx, epb)
sm_group = self._make_sm_m_slice(sm, m_offset, m_per_group, epb, num_warps)
if const_expr(k_main_elems > 0):
gA_main = self._make_k_slice(gA_group, 0, k_main_elems)
gB_main = self._make_k_slice(gB_epb, 0, k_main_elems)
gA_vec = cute.logical_divide(gA_main, (None, main_vec_width))
gB_vec = cute.logical_divide(gB_main, (None, main_vec_width))
tA, tB = self._make_thread_vector_slice(gA_vec, gB_vec, tidx, bs)
self._vector_dotprod(acc, tA, tB, m_per_group, epb, main_tiles)
if const_expr(k_tail_elems > 0):
gA_tail = self._make_k_slice(gA_group, k_main_elems, k_tail_elems)
gB_tail = self._make_k_slice(gB_epb, k_main_elems, k_tail_elems)
gA_tail_vec = cute.logical_divide(gA_tail, (None, tail_vec_width))
gB_tail_vec = cute.logical_divide(gB_tail, (None, tail_vec_width))
tA_t, tB_t = self._make_thread_vector_slice(
gA_tail_vec, gB_tail_vec, tidx, bs
)
self._vector_dotprod(acc, tA_t, tB_t, m_per_group, epb, tail_tiles)
if const_expr(ks_full > 0):
gA_scalar = self._make_k_slice(gA_group, k_done_all, k_scalar_full)
gB_scalar = self._make_k_slice(gB_epb, k_done_all, k_scalar_full)
gA_scalar_vec = cute.logical_divide(gA_scalar, (None, 1))
gB_scalar_vec = cute.logical_divide(gB_scalar, (None, 1))
tA_s, tB_s = self._make_thread_vector_slice(
gA_scalar_vec, gB_scalar_vec, tidx, bs
)
self._vector_dotprod(acc, tA_s, tB_s, m_per_group, epb, ks_full)
if const_expr(ks_part > 0):
gA_part = self._make_k_slice(gA_group, k_part_offset, ks_part)
gB_part = self._make_k_slice(gB_epb, k_part_offset, ks_part)
if tidx < ks_part:
for local_m in cutlass.range_constexpr(m_per_group):
av = gA_part[local_m, tidx].to(cutlass.Float32)
for e in cutlass.range_constexpr(epb):
bv = gB_part[e, tidx].to(cutlass.Float32)
acc[local_m, e] = acc[local_m, e] + av * bv
for local_m in cutlass.range_constexpr(m_per_group):
for e in cutlass.range_constexpr(epb):
acc[local_m, e] = cute.arch.warp_reduction_sum(acc[local_m, e])
with cute.arch.elect_one():
for local_m in cutlass.range_constexpr(m_per_group):
for e in cutlass.range_constexpr(epb):
sm_group[local_m, e, local_wid] = acc[local_m, e]
@cute.jit
def _store_outputs(
self,
gC: cute.Tensor,
sm: cute.Tensor,
n_tile_idx: cutlass.Int32,
N_dim: cutlass.Int32,
tidx: cutlass.Int32,
group_idx: cutlass.Int32,
M: cutlass.Constexpr,
epb: cutlass.Constexpr,
):
n_tile_layout = cute.make_layout(
(cute.ceil_div(N_dim, epb), epb), stride=(epb, 1)
)
n_coords = cute.make_tensor(0, n_tile_layout)
if group_idx == 0 and tidx < M:
for e in cutlass.range_constexpr(epb):
n = n_coords[n_tile_idx, e]
if n < N_dim:
partials = sm[tidx, e, None].load()
gC[tidx, n] = partials.reduce(
cute.ReductionOp.ADD,
init_val=cutlass.Float32(0.0),
reduction_profile=0,
)
@cute.jit
def __call__(
self,
gA: cute.Tensor,
gB: cute.Tensor,
gC: cute.Tensor,
N_dim: cutlass.Int32,
stream: CUstream,
):
"""Execute the dot-product GEMM operation in steps:
- Launch one CTA per small output-column tile with ``bs`` threads.
- Keep one FP32 accumulator per token/expert in each thread.
- Traverse K with vectorized main/tail, scalar, and ragged-tail loops
from the precomputed K decomposition.
- Reduce each token accumulator first within the warp, then across
warps through shared memory, and store ``C[:, n]``.
:param gA: Input tensor A with shape ``[M, K]``.
:type gA: cute.Tensor
:param gB: Input tensor B with shape ``[N, K]``.
:type gB: cute.Tensor
:param gC: Output tensor C with shape ``[M, N]``.
:type gC: cute.Tensor
:param N_dim: Output column count used for the launch grid.
:type N_dim: cutlass.Int32
:param stream: CUDA stream for asynchronous execution.
:type stream: CUstream
"""
self.kernel(
gA,
gB,
gC,
N_dim,
self.m,
self.m_per_group,
self.main_vec_width,
self.tail_vec_width,
self.bs,
self.num_warps,
self.token_groups,
self.epb,
self.k_main_elems,
self.k_tail_elems,
self.k_done_all,
self.ks_full,
self.ks_part,
self.k_scalar_full,
self.k_part_offset,
self.main_tiles,
self.tail_tiles,
).launch(
grid=[cute.ceil_div(N_dim, self.epb), 1, 1],
block=[self.bs, self.token_groups, 1],
smem=self.m * self.epb * 4 * self.num_warps,
stream=stream,
use_pdl=self.use_pdl,
min_blocks_per_mp=1,
)
@cute.kernel
def kernel(
self,
gA: cute.Tensor,
gB: cute.Tensor,
gC: cute.Tensor,
N_dim: cutlass.Int32,
M: cutlass.Constexpr,
m_per_group: cutlass.Constexpr,
main_vec_width: cutlass.Constexpr,
tail_vec_width: cutlass.Constexpr,
bs: cutlass.Constexpr,
num_warps: cutlass.Constexpr,
token_groups: cutlass.Constexpr,
epb: cutlass.Constexpr,
k_main_elems: cutlass.Constexpr,
k_tail_elems: cutlass.Constexpr,
k_done_all: cutlass.Constexpr,
ks_full: cutlass.Constexpr,
ks_part: cutlass.Constexpr,
k_scalar_full: cutlass.Constexpr,
k_part_offset: cutlass.Constexpr,
main_tiles: cutlass.Constexpr,
tail_tiles: cutlass.Constexpr,
):
tidx, group_idx, _ = cute.arch.thread_idx()
n_tile_idx, _, _ = cute.arch.block_idx()
local_wid = tidx // cute.arch.WARP_SIZE
if const_expr(self.use_pdl):
cute.arch.griddepcontrol_wait()
smem_red_layout = cute.make_layout(
(M, epb, num_warps), stride=(epb * num_warps, num_warps, 1)
)
smem = cutlass.utils.SmemAllocator()
sm = smem.allocate_tensor(cutlass.Float32, smem_red_layout, byte_alignment=16)
if const_expr(token_groups == 1):
m_offset: cutlass.Constexpr = 0
else:
m_offset = group_idx * m_per_group
self._compute_token_group(
gA,
gB,
sm,
n_tile_idx,
tidx,
local_wid,
m_per_group,
m_offset,
main_vec_width,
tail_vec_width,
bs,
epb,
num_warps,
k_main_elems,
k_tail_elems,
k_done_all,
ks_full,
ks_part,
k_scalar_full,
k_part_offset,
main_tiles,
tail_tiles,
)
cute.arch.sync_threads()
self._store_outputs(gC, sm, n_tile_idx, N_dim, tidx, group_idx, M, epb)
if const_expr(self.use_pdl):
cute.arch.griddepcontrol_launch_dependents()
def make_host_fp32w(
k_val: int, bs: int = 128, token_groups: int = 1, epb: int = 1, m: int = 1
):
return LLFp32WDotprod(m=m, k=k_val, bs=bs, token_groups=token_groups, epb=epb)
@@ -0,0 +1,120 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import logging
from typing import Any
import torch
logger = logging.getLogger(__name__)
_cutedsl_available: bool | None = None
_cute_ctx: tuple[Any, Any] | None = None
def is_cutedsl_available(kernel_name: str) -> bool:
global _cutedsl_available
if _cutedsl_available is not None:
return _cutedsl_available
try:
import cutlass # noqa: F401
import cutlass.cute # noqa: F401
_cutedsl_available = True
except ImportError:
_cutedsl_available = False
logger.info(
"cuteDSL (CUTLASS Python) not available, %s disabled",
kernel_name,
)
return _cutedsl_available
def cute_context():
global _cute_ctx
if _cute_ctx is not None:
return _cute_ctx
import cutlass.cute as cute
from cuda.bindings.driver import CUstream
_cute_ctx = (cute, CUstream)
return _cute_ctx
def current_cuda_stream():
_, CUstream = cute_context()
from vllm.utils.torch_utils import current_stream
return CUstream(current_stream().cuda_stream)
def use_pdl() -> bool:
from vllm.platforms import current_platform
return current_platform.is_arch_support_pdl()
def cutlass_dtype(dtype: torch.dtype):
from cutlass import BFloat16, Float16, Float32
if dtype == torch.bfloat16:
return BFloat16
if dtype == torch.float16:
return Float16
if dtype == torch.float32:
return Float32
raise ValueError(f"unsupported router GEMM dtype: {dtype}")
def make_fake_gemm_tensors(
*,
M,
K,
N,
a_dtype: torch.dtype,
b_dtype: torch.dtype,
divisibility: int,
):
from cutlass import Float32
from quack.compile_utils import make_fake_tensor
hidden_states = make_fake_tensor(
cutlass_dtype(a_dtype), (M, K), divisibility=divisibility
)
router_weight = make_fake_tensor(
cutlass_dtype(b_dtype), (N, K), divisibility=divisibility
)
output = make_fake_tensor(Float32, (M, N), divisibility=1)
return hidden_states, router_weight, output
def validate_common_gemm_inputs(
hidden_states: torch.Tensor,
router_weight: torch.Tensor,
output_dtype: torch.dtype,
*,
op_name: str,
k_multiple: int | None = None,
) -> None:
if hidden_states.dim() != 2 or router_weight.dim() != 2:
raise ValueError("hidden_states and router_weight must be 2D tensors")
if hidden_states.device.type != "cuda" or router_weight.device.type != "cuda":
raise ValueError("hidden_states and router_weight must have device_type=cuda")
if hidden_states.device != router_weight.device:
raise ValueError(
"hidden_states and router_weight must be on the same CUDA device"
)
if output_dtype != torch.float32:
raise ValueError(f"{op_name} only supports output_dtype=torch.float32")
if hidden_states.shape[1] != router_weight.shape[1]:
raise ValueError(
"hidden_states and router_weight must have matching K dimensions"
)
if k_multiple is not None and hidden_states.shape[1] % k_multiple != 0:
raise ValueError(f"{op_name} requires K to be divisible by {k_multiple}")
if not hidden_states.is_contiguous() or not router_weight.is_contiguous():
raise ValueError(
"hidden_states and router_weight must be contiguous row-major inputs"
)
@@ -11,24 +11,20 @@ from typing import Any, Literal
import torch
logger = logging.getLogger(__name__)
from ._ll_router_common import (
current_cuda_stream,
cute_context,
is_cutedsl_available,
make_fake_gemm_tensors,
use_pdl,
validate_common_gemm_inputs,
)
_cutedsl_available: bool | None = None
logger = logging.getLogger(__name__)
def is_available() -> bool:
global _cutedsl_available
if _cutedsl_available is not None:
return _cutedsl_available
try:
import cutlass # noqa: F401
import cutlass.cute # noqa: F401
_cutedsl_available = True
except ImportError:
_cutedsl_available = False
logger.info("cuteDSL (CUTLASS Python) not available, ll_bf16_gemm disabled")
return _cutedsl_available
return is_cutedsl_available("ll_bf16_gemm")
# Default configs
@@ -96,33 +92,6 @@ def _arch_tuned_configs() -> tuple[
return {}, {}
_cute_ctx = None
def _cute():
global _cute_ctx
if _cute_ctx is not None:
return _cute_ctx
import cutlass.cute as cute
from cuda.bindings.driver import CUstream
_cute_ctx = (cute, CUstream)
return _cute_ctx
def _stream():
_, CUstream = _cute()
from vllm.utils.torch_utils import current_stream
return CUstream(current_stream().cuda_stream)
def _use_pdl() -> bool:
from vllm.platforms import current_platform
return current_platform.is_arch_support_pdl()
class LLBf16Gemm:
@dataclass(frozen=True, slots=True)
class CompileKey:
@@ -165,16 +134,17 @@ class LLBf16Gemm:
@staticmethod
def _fake_gemm_tensors(*, M, K, N, divisibility: int):
from cutlass import BFloat16, Float32
from quack.compile_utils import make_fake_tensor
hidden_states = make_fake_tensor(BFloat16, (M, K), divisibility=divisibility)
router_weight = make_fake_tensor(BFloat16, (N, K), divisibility=divisibility)
output = make_fake_tensor(Float32, (M, N), divisibility=1)
return hidden_states, router_weight, output
return make_fake_gemm_tensors(
M=M,
K=K,
N=N,
a_dtype=torch.bfloat16,
b_dtype=torch.bfloat16,
divisibility=divisibility,
)
def _compile_splitk(self, compile_key: CompileKey) -> None:
cute, _ = _cute()
cute, _ = cute_context()
from ._ll_bf16_splitk import LLBf16SplitK
hidden_states, router_weight, output = self._fake_gemm_tensors(
@@ -189,14 +159,14 @@ class LLBf16Gemm:
num_stages=compile_key.num_stages,
num_dma_warps=4,
split_k=compile_key.split_k,
use_pdl=_use_pdl(),
use_pdl=use_pdl(),
)
compiled = cute.compile(
gemm,
hidden_states,
router_weight,
output,
_stream(),
current_cuda_stream(),
options="--enable-tvm-ffi",
)
self._splitk_cache[(compile_key.split_k, compile_key.num_stages)] = compiled
@@ -207,7 +177,7 @@ class LLBf16Gemm:
)
def _compile_dotprod(self, compile_key: CompileKey) -> None:
cute, _ = _cute()
cute, _ = cute_context()
from ._ll_bf16_dotprod import LLBf16Dotprod
N = cute.sym_int()
@@ -221,7 +191,7 @@ class LLBf16Gemm:
gemm = LLBf16Dotprod(
k=compile_key.K,
bs=compile_key.bs,
use_pdl=_use_pdl(),
use_pdl=use_pdl(),
prefetch_pdl_weights=self._prefetch_pdl_weights,
)
compiled = cute.compile(
@@ -232,7 +202,7 @@ class LLBf16Gemm:
compile_key.M,
compile_key.K,
1, # runtime N placeholder for fake-tensor compile
_stream(),
current_cuda_stream(),
options="--enable-tvm-ffi --ptxas-options -maxrregcount=64",
)
self._compiled_cache[(compile_key.M, compile_key.K, compile_key.bs)] = compiled
@@ -269,34 +239,19 @@ class LLBf16Gemm:
router_weight: torch.Tensor,
output_dtype: torch.dtype,
) -> None:
if hidden_states.dim() != 2 or router_weight.dim() != 2:
raise ValueError("hidden_states and router_weight must be 2D tensors")
if (
hidden_states.dtype != torch.bfloat16
or router_weight.dtype != torch.bfloat16
):
raise ValueError("hidden_states and router_weight must have dtype=bfloat16")
if hidden_states.device.type != "cuda" or router_weight.device.type != "cuda":
raise ValueError(
"hidden_states and router_weight must have device_type=cuda"
)
if hidden_states.device != router_weight.device:
raise ValueError(
"hidden_states and router_weight must be on the same CUDA device"
)
if output_dtype != torch.float32:
raise ValueError("ll_bf16_gemm only supports output_dtype=torch.float32")
if hidden_states.shape[1] != router_weight.shape[1]:
raise ValueError(
"hidden_states and router_weight must have matching K dimensions"
)
# Kernels use vectorized bf16 loads and require 16-byte row alignment.
if hidden_states.shape[1] % 8 != 0:
raise ValueError("ll_bf16_gemm requires K to be divisible by 8")
if not hidden_states.is_contiguous() or not router_weight.is_contiguous():
raise ValueError(
"hidden_states and router_weight must be contiguous row-major inputs"
)
validate_common_gemm_inputs(
hidden_states,
router_weight,
output_dtype,
op_name="ll_bf16_gemm",
k_multiple=8,
)
def __call__(
self,
@@ -320,7 +275,7 @@ class LLBf16Gemm:
self.compile(compile_key)
kernel = self._compiled_cache[dotprod_cache_key]
stream = _stream()
stream = current_cuda_stream()
output = torch.empty(M, N, dtype=output_dtype, device=hidden_states.device)
if compile_key.backend == "splitk":
kernel(hidden_states, router_weight, output, stream, 1.0)
@@ -0,0 +1,197 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
import logging
import math
from collections.abc import Iterable
from dataclasses import dataclass
from typing import Any
import torch
from vllm.model_executor.warmup.jit_warmup import VllmJitKernel, zip_inputs
from ._ll_router_common import (
current_cuda_stream,
cute_context,
is_cutedsl_available,
make_fake_gemm_tensors,
use_pdl,
validate_common_gemm_inputs,
)
logger = logging.getLogger(__name__)
def is_available() -> bool:
return is_cutedsl_available("ll_fp32w_gemm")
_DEFAULT_DOTPROD_CONFIG = (192, 1, 1)
_DEFAULT_DOTPROD_GROUPED_CONFIG = (192, 2, 1)
_TUNED_DOTPROD_CONFIGS: dict[tuple[int, int], dict[int, tuple[int, int, int]]] = {
(6144, 128): {
**{
M: (384, 1, 1)
for M in (
*range(1, 6),
*range(9, 14, 2),
17,
*range(23, 32, 2),
)
},
**{M: (256, 1, 1) for M in (7, 15, *range(19, 22, 2))},
**{M: (384, 2, 1) for M in range(6, 12, 2)},
**{M: (192, 2, 1) for M in (*range(12, 26, 2), 32)},
**{M: (128, 2, 1) for M in range(26, 32, 2)},
},
(6144, 256): {
**{M: (384, 1, 1) for M in range(1, 4)},
**{M: (128, 1, 1) for M in (*range(4, 7, 2), 11)},
**{M: (128, 1, 2) for M in range(5, 11, 2)},
**{M: (128, 2, 2) for M in range(8, 33, 2)},
**{M: (256, 1, 2) for M in range(13, 33, 2)},
},
}
_SUPPORTED_ACTIVATION_DTYPES = (torch.bfloat16, torch.float16, torch.float32)
class LLFp32WGemm(VllmJitKernel["LLFp32WGemm.CompileKey"]):
# Dot-prod is specialized on M/K, tile shape, and activation dtype.
@dataclass(frozen=True, slots=True)
class CompileKey:
m: int
k: int
bs: int
a_dtype: torch.dtype
token_groups: int = 1
epb: int = 1
@staticmethod
def kernel(compile_key: CompileKey) -> Any:
from ._ll_fp32w_dotprod import LLFp32WDotprod
return LLFp32WDotprod(
m=compile_key.m,
k=compile_key.k,
bs=compile_key.bs,
token_groups=compile_key.token_groups,
epb=compile_key.epb,
use_pdl=use_pdl(),
)
def dispatch( # type: ignore[override]
self,
*,
M: int,
K: int,
N: int,
a_dtype: torch.dtype,
) -> CompileKey:
default_config = (
_DEFAULT_DOTPROD_GROUPED_CONFIG
if N <= 128 and M >= 12 and M % 2 == 0
else _DEFAULT_DOTPROD_CONFIG
)
tuned_configs = _TUNED_DOTPROD_CONFIGS.get((K, N))
config = (
tuned_configs.get(M, default_config) if tuned_configs else default_config
)
return self.CompileKey(
m=M,
k=K,
bs=config[0],
a_dtype=a_dtype,
token_groups=config[1],
epb=config[2],
)
def get_warmup_keys(
self,
*,
shapes: Iterable[tuple[int, int]],
m_values: Iterable[int],
a_dtypes: Iterable[torch.dtype] = (torch.bfloat16,),
) -> list[CompileKey]:
return self._trace_dispatch(self.dispatch)(
zip_inputs(*(dict(K=K, N=N) for K, N in shapes)),
M=tuple(m_values),
a_dtype=tuple(a_dtypes),
)
def compile(self, compile_key: CompileKey) -> None:
if compile_key in self._compiled_cache:
return
if compile_key.token_groups not in (1, 2):
raise ValueError("ll_fp32w_gemm supports token_groups 1 or 2")
if compile_key.epb not in (1, 2):
raise ValueError("ll_fp32w_gemm supports epb 1 or 2")
if compile_key.m % compile_key.token_groups != 0:
raise ValueError("M must be divisible by token_groups")
cute, _ = cute_context()
N = cute.sym_int()
stride_divisibility = math.gcd(8, compile_key.k)
hidden_states, router_weight, output = make_fake_gemm_tensors(
M=compile_key.m,
K=compile_key.k,
N=N,
a_dtype=compile_key.a_dtype,
b_dtype=torch.float32,
divisibility=stride_divisibility,
)
compiled = cute.compile(
self.kernel(compile_key),
hidden_states,
router_weight,
output,
1, # runtime N placeholder for fake-tensor compile
current_cuda_stream(),
options="--enable-tvm-ffi",
)
self._compiled_cache[compile_key] = compiled
@staticmethod
def _validate_inputs(
hidden_states: torch.Tensor,
router_weight: torch.Tensor,
output_dtype: torch.dtype,
) -> None:
if hidden_states.dtype not in _SUPPORTED_ACTIVATION_DTYPES:
raise ValueError("hidden_states must have dtype bf16, fp16, or fp32")
if router_weight.dtype != torch.float32:
raise ValueError("router_weight must have dtype=float32")
validate_common_gemm_inputs(
hidden_states,
router_weight,
output_dtype,
op_name="ll_fp32w_gemm",
)
def __call__(
self,
hidden_states: torch.Tensor, # [M, K] bf16/fp16/fp32
router_weight: torch.Tensor, # [N, K] fp32
output_dtype: torch.dtype = torch.float32,
) -> torch.Tensor: # [M, N] fp32
self._validate_inputs(hidden_states, router_weight, output_dtype)
M, K = hidden_states.shape
N = router_weight.shape[0]
compile_key = self.dispatch(M=M, K=K, N=N, a_dtype=hidden_states.dtype)
compiled = self._get_or_compile(
compile_key,
runtime_context={"M": M, "K": K, "N": N},
)
stream = current_cuda_stream()
output = torch.empty(M, N, dtype=output_dtype, device=hidden_states.device)
compiled(hidden_states, router_weight, output, N, stream)
return output
LL_FP32W_GEMM_KERNEL = LLFp32WGemm()