Skip to content

[FLYDSL] Add mixed slicek-splitk policies for flydsl hgemm #3279

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
xytpai wants to merge 6 commits into
base: main
Choose a base branch
from
+201 −211
Open

[FLYDSL] Add mixed slicek-splitk policies for flydsl hgemm #3279

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
8bf075d
add mixed slicek
xytpai May 20, 2026
01883e0
update code
xytpai May 20, 2026
cf2b63c
update spk limit
xytpai May 20, 2026
8e4a983
refine tune
xytpai May 20, 2026
a52feed
add tuned config for gptoss
xytpai May 20, 2026
e5d867a
add more cfgs
xytpai May 20, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions aiter/aot/flydsl/gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -210,6 +210,7 @@ def _compile_hgemm_to_cache(
split_k: int,
block_m_warps: int,
block_n_warps: int,
block_k_warps: int,
n_tile_repeat: int = 1,
persistent_n_tiles: int = 1,
waves_per_eu: int = 0,
Expand Down Expand Up @@ -257,6 +258,7 @@ def _compile_hgemm_to_cache(
split_k=split_k,
block_m_warps=block_m_warps,
block_n_warps=block_n_warps,
block_k_warps=block_k_warps,
n_tile_repeat=n_tile_repeat,
persistent_n_tiles=persistent_n_tiles,
waves_per_eu=waves_per_eu,
Expand Down
178 changes: 89 additions & 89 deletions aiter/configs/model_configs/gptoss_bf16_tuned_gemm.csv
View file
Open in desktop

Large diffs are not rendered by default.

73 changes: 40 additions & 33 deletions aiter/ops/flydsl/gemm_kernels.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,6 +48,7 @@ def _get_dtypes():
r"split_k(?P<split_k>\d+)_"
r"block_m_warp(?P<block_m_warps>\d+)_"
r"block_n_warp(?P<block_n_warps>\d+)_"
r"block_k_warp(?P<block_k_warps>\d+)_"
r"async_copy(?P<async_copy>True|False)_"
r"b_to_lds(?P<b_to_lds>True|False)_"
r"b_preshuffle(?P<b_preshuffle>True|False)_"
Expand All @@ -73,35 +74,28 @@ def _get_dtypes():
HGEMM_TILE_N_OPTIONS = (64, 128, 256)
HGEMM_TILE_K_OPTIONS = (64, 128, 256)
HGEMM_TILE_M_OPTIONS = (16, 32, 48, 64, 96, 128, 256)
HGEMM_BASE_SPLIT_K_OPTIONS = tuple(range(1, 9))
HGEMM_MAX_SPLIT_K = 8
KERNEL_CONFIG_VARIANTS = (
{
"block_m_warps": 1,
"block_n_warps": 2,
"b_to_lds": False,
},
{
"block_m_warps": 1,
"block_n_warps": 4,
"b_to_lds": False,
},
HGEMM_BASE_SPLIT_K_OPTIONS = tuple(range(1, 14))
HGEMM_MAX_SPLIT_K = 13
HGEMM_WARP_SHAPE_OPTIONS = [
(wm, wn, wk) for wm, wn, wk in product([1, 2, 4], repeat=3) if wm * wn * wk <= 8
]
KERNEL_CONFIG_VARIANTS = [
{
"block_m_warps": 2,
"block_n_warps": 2,
"block_m_warps": wm,
"block_n_warps": wn,
"block_k_warps": wk,
"b_to_lds": False,
},
{
"block_m_warps": 1,
"block_n_warps": 4,
"b_to_lds": True,
},
}
for wm, wn, wk in HGEMM_WARP_SHAPE_OPTIONS
] + [
{
"block_m_warps": 2,
"block_n_warps": 2,
"block_m_warps": wm,
"block_n_warps": wn,
"block_k_warps": wk,
"b_to_lds": True,
},
)
}
for wm, wn, wk in HGEMM_WARP_SHAPE_OPTIONS
]

_SPLITK_HGEMM_KERNELS: Dict[str, Dict] = {}

Expand Down Expand Up @@ -141,6 +135,7 @@ def flydsl_kernel_name(
split_k: int,
block_m_warp: int,
block_n_warp: int,
block_k_warp: int,
async_copy: bool,
b_to_lds: bool,
b_preshuffle: bool = False,
Expand All @@ -167,7 +162,7 @@ def flydsl_kernel_name(
name = (
f"flydsl_gemm{stage}_a{dtype}_w{dtype}_{out_dtype}_t{tile_m}x{tile_n}x{tile_k}"
)
name += f"_split_k{split_k}_block_m_warp{block_m_warp}_block_n_warp{block_n_warp}"
name += f"_split_k{split_k}_block_m_warp{block_m_warp}_block_n_warp{block_n_warp}_block_k_warp{block_k_warp}"
name += (
f"_async_copy{async_copy}_b_to_lds{b_to_lds}_b_preshuffle{b_preshuffle}"
f"_c_to_lds{c_to_lds}"
Expand Down Expand Up @@ -332,6 +327,7 @@ def _validate_hgemm_tiling(
stages: int,
block_m_warps: int,
block_n_warps: int,
block_k_warps: int,
b_to_lds: bool,
) -> None:
del m
Expand All @@ -340,10 +336,10 @@ def _validate_hgemm_tiling(
raise ValueError(
f"Tile sizes must be positive, got tile_m={tile_m}, tile_n={tile_n}, tile_k={tile_k}"
)
if block_m_warps < 1 or block_n_warps < 1:
if block_m_warps < 1 or block_n_warps < 1 or block_k_warps < 1:
raise ValueError(
"Warp tiling must be positive, got "
f"block_m_warps={block_m_warps}, block_n_warps={block_n_warps}"
f"block_m_warps={block_m_warps}, block_n_warps={block_n_warps}, block_k_warps={block_k_warps}"
)
if tile_k < 32:
raise ValueError(
Expand Down Expand Up @@ -396,7 +392,7 @@ def _validate_hgemm_tiling(
f">= tile_k={tile_k} and % tile_k == 0"
)

block_threads = block_m_warps * block_n_warps * 64
block_threads = block_m_warps * block_n_warps * block_k_warps * 64
ldg_vec_size = 8
block_vecs = ldg_vec_size * block_threads
block_mk_size = tile_m * tile_k
Expand Down Expand Up @@ -457,6 +453,7 @@ def _normalize_registry_config(
split_k: int,
block_m_warps: int,
block_n_warps: int,
block_k_warps: int,
b_to_lds: bool,
) -> Optional[Dict]:
config = {
Expand All @@ -468,6 +465,7 @@ def _normalize_registry_config(
"split_k": int(split_k),
"block_m_warps": int(block_m_warps),
"block_n_warps": int(block_n_warps),
"block_k_warps": int(block_k_warps),
"async_copy": KERNEL_ASYNC_COPY,
"b_to_lds": bool(b_to_lds),
"b_preshuffle": False,
Expand All @@ -490,6 +488,7 @@ def _normalize_registry_config(
stages=FIXED_STAGE,
block_m_warps=config["block_m_warps"],
block_n_warps=config["block_n_warps"],
block_k_warps=config["block_k_warps"],
b_to_lds=config["b_to_lds"],
)
except ValueError:
Expand Down Expand Up @@ -519,6 +518,7 @@ def _parse_hgemm_kernel_params(name: str) -> Optional[Dict]:
"split_k": int(m.group("split_k")),
"block_m_warps": int(m.group("block_m_warps")),
"block_n_warps": int(m.group("block_n_warps")),
"block_k_warps": int(m.group("block_k_warps")),
"async_copy": m.group("async_copy") == "True",
"b_to_lds": m.group("b_to_lds") == "True",
"b_preshuffle": m.group("b_preshuffle") == "True",
Expand Down Expand Up @@ -582,6 +582,7 @@ def get_flydsl_splitk_hgemm_kernels(
split_k=split_k,
block_m_warps=variant["block_m_warps"],
block_n_warps=variant["block_n_warps"],
block_k_warps=variant["block_k_warps"],
b_to_lds=variant["b_to_lds"],
)
if config is None:
Expand All @@ -599,6 +600,7 @@ def get_flydsl_splitk_hgemm_kernels(
config["split_k"],
config["block_m_warps"],
config["block_n_warps"],
config["block_k_warps"],
config["async_copy"],
config["b_to_lds"],
config["b_preshuffle"],
Expand Down Expand Up @@ -699,8 +701,9 @@ def _compile_flydsl_hgemm(
k: int,
*,
tile_k: int = 64,
block_m_warps: int = 1,
block_n_warps: int = 4,
block_m_warps: int = 2,
block_n_warps: int = 2,
block_k_warps: int = 1,
tile_m: int = 128,
tile_n: int = 128,
pack_n: int = 1,
Expand Down Expand Up @@ -740,6 +743,7 @@ def _compile_flydsl_hgemm(
stages=stages,
block_m_warps=block_m_warps,
block_n_warps=block_n_warps,
block_k_warps=block_k_warps,
b_to_lds=b_to_lds,
)
elif kernel_family == KERNEL_FAMILY_SMALL_M:
Expand Down Expand Up @@ -776,6 +780,7 @@ def _compile_flydsl_hgemm(
split_k=split_k,
block_m_warps=block_m_warps,
block_n_warps=block_n_warps,
block_k_warps=block_k_warps,
n_tile_repeat=n_tile_repeat,
persistent_n_tiles=persistent_n_tiles,
waves_per_eu=waves_per_eu,
Expand Down Expand Up @@ -835,8 +840,9 @@ def flydsl_hgemm(
tile_k: int = 64,
pack_n: int = 1,
split_k: int = 1,
block_m_warps: int = 1,
block_n_warps: int = 4,
block_m_warps: int = 2,
block_n_warps: int = 2,
block_k_warps: int = 1,
n_tile_repeat: int = 1,
persistent_n_tiles: int = 1,
waves_per_eu: int = 0,
Expand Down Expand Up @@ -887,6 +893,7 @@ def flydsl_hgemm(
tile_k=tile_k,
block_m_warps=block_m_warps,
block_n_warps=block_n_warps,
block_k_warps=block_k_warps,
tile_m=tile_m,
tile_n=tile_n,
pack_n=pack_n,
Expand Down
6 changes: 4 additions & 2 deletions aiter/ops/flydsl/kernels/hgemm_dispatch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,8 +20,9 @@ def compile_flydsl_hgemm_kernel(
tile_k: int = 64,
pack_n: int = 1,
split_k: int = 1,
block_m_warps: int = 1,
block_n_warps: int = 4,
block_m_warps: int = 2,
block_n_warps: int = 2,
block_k_warps: int = 1,
n_tile_repeat: int = 1,
persistent_n_tiles: int = 1,
waves_per_eu: int = 0,
Expand Down Expand Up @@ -52,6 +53,7 @@ def compile_flydsl_hgemm_kernel(
SPLIT_K=split_k,
BLOCK_M_WARPS=block_m_warps,
BLOCK_N_WARPS=block_n_warps,
BLOCK_K_WARPS=block_k_warps,
B_TO_LDS=b_to_lds,
HAS_BIAS=has_bias,
)
Expand Down
Loading
Loading