[9/N][warmup][DSv4] Migrate MHC TileLang kernels (#50178)

Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
Roberto L. Castro
2026-09-12 17:25:23 -07:00
committed by GitHub
co-authored by Codex
parent aed894c190
commit e52be1a62d
9 changed files with 1463 additions and 498 deletions
+296
View File
@@ -0,0 +1,296 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Validate JIT dispatch against pre-contract behavior."""
from typing import Any
import pytest
from vllm.platforms import current_platform
if not current_platform.is_cuda_alike():
pytest.skip("NVIDIA dispatch tests require CUDA", allow_module_level=True)
from vllm.model_executor.kernels.mhc.tilelang_kernels import (
HcHeadFusedTileLangKernel,
HcPrenormGemmTileLangKernel,
MhcFusedTileLangKernel,
MhcPostTileLangKernel,
MhcPreBigFuseTileLangKernel,
)
from vllm.model_executor.warmup import jit_warmup_tilelang_helper
@pytest.mark.parametrize(
("kwargs", "expected"),
[
(
dict(
num_tokens=64,
hc_hidden_size=4096,
hidden_size=2048,
hc_mult=2,
n_out=128,
),
(2048, 2, 128, 1024, 4, 1, False, 1),
),
(
dict(
num_tokens=1024,
hc_hidden_size=4096,
hidden_size=2048,
hc_mult=2,
n_out=128,
),
(2048, 2, 128, 512, 12, 1, True, 2),
),
(
dict(
num_tokens=64,
hc_hidden_size=4096,
hidden_size=2048,
hc_mult=2,
n_out=128,
n_thr=256,
tile_n=8,
n_splits=4,
),
(2048, 2, 128, 256, 8, 4, False, 1),
),
],
)
def test_hc_prenorm_gemm_dispatch_matches_legacy_runtime_config(
kwargs: dict[str, Any],
expected: tuple[int, int, int, int, int, int, bool, int],
) -> None:
kernel = HcPrenormGemmTileLangKernel()
assert kernel.dispatch(**kwargs) == kernel.CompileKey(*expected)
@pytest.mark.parametrize(
("is_broadcast", "use_norm_weight", "expected_use_norm", "expected_eps"),
[
(False, False, False, 0.0),
(False, True, True, 1.0e-5),
(True, False, True, 2.0e-5),
],
)
def test_mhc_pre_big_fuse_dispatch_matches_legacy_runtime_config(
is_broadcast: bool,
use_norm_weight: bool,
expected_use_norm: bool,
expected_eps: float,
) -> None:
kernel = MhcPreBigFuseTileLangKernel()
assert kernel.dispatch(
hidden_size=4096,
hc_mult=4,
n_splits=2,
is_broadcast=is_broadcast,
use_norm_weight=use_norm_weight,
rms_eps=1.0e-6,
hc_pre_eps=2.0e-6,
hc_sinkhorn_eps=3.0e-6,
hc_post_mult_value=0.5,
sinkhorn_repeat=3,
norm_eps=1.0e-5,
broadcast_norm_eps=2.0e-5,
) == kernel.CompileKey(
hidden_size=4096,
hc_mult=4,
n_splits=2,
use_norm_weight=expected_use_norm,
is_broadcast=is_broadcast,
rms_eps=1.0e-6,
hc_pre_eps=2.0e-6,
hc_sinkhorn_eps=3.0e-6,
hc_post_mult_value=0.5,
sinkhorn_repeat=3,
norm_eps=expected_eps,
)
@pytest.mark.parametrize(
("num_tokens", "hidden_size", "expected_n_splits", "expected_tile_n"),
[(4, 4096, 8, 2), (4, 8192, 4, 2), (8, 4096, 4, 3)],
)
def test_mhc_fused_dispatch_matches_legacy_runtime_config(
num_tokens: int,
hidden_size: int,
expected_n_splits: int,
expected_tile_n: int,
) -> None:
kernel = MhcFusedTileLangKernel()
assert kernel.dispatch(
num_tokens=num_tokens,
hidden_size=hidden_size,
hc_mult=4,
) == kernel.CompileKey(
hidden_size=hidden_size,
hc_mult=4,
n_splits=expected_n_splits,
tile_n=expected_tile_n,
)
@pytest.mark.parametrize(
("kernel", "compile_key"),
[
(
HcPrenormGemmTileLangKernel(),
HcPrenormGemmTileLangKernel.CompileKey(
hidden_size=2048,
hc_mult=2,
n_out=128,
n_thr=1024,
tile_n=4,
n_splits=1,
use_block_m=False,
block_m=1,
),
),
(
HcPrenormGemmTileLangKernel(),
HcPrenormGemmTileLangKernel.CompileKey(
hidden_size=2048,
hc_mult=2,
n_out=128,
n_thr=512,
tile_n=12,
n_splits=1,
use_block_m=False,
block_m=1,
),
),
(
HcPrenormGemmTileLangKernel(),
HcPrenormGemmTileLangKernel.CompileKey(
hidden_size=2048,
hc_mult=2,
n_out=128,
n_thr=512,
tile_n=12,
n_splits=1,
use_block_m=True,
block_m=2,
),
),
(
MhcPreBigFuseTileLangKernel(),
MhcPreBigFuseTileLangKernel.CompileKey(
hidden_size=4096,
hc_mult=4,
n_splits=2,
use_norm_weight=True,
is_broadcast=False,
rms_eps=1.0e-6,
hc_pre_eps=2.0e-6,
hc_sinkhorn_eps=3.0e-6,
hc_post_mult_value=0.5,
sinkhorn_repeat=3,
norm_eps=1.0e-5,
),
),
(
MhcPreBigFuseTileLangKernel(),
MhcPreBigFuseTileLangKernel.CompileKey(
hidden_size=4096,
hc_mult=4,
n_splits=1,
use_norm_weight=False,
is_broadcast=False,
rms_eps=1.0e-6,
hc_pre_eps=2.0e-6,
hc_sinkhorn_eps=3.0e-6,
hc_post_mult_value=0.5,
sinkhorn_repeat=3,
norm_eps=0.0,
),
),
(
MhcPreBigFuseTileLangKernel(),
MhcPreBigFuseTileLangKernel.CompileKey(
hidden_size=4096,
hc_mult=4,
n_splits=2,
use_norm_weight=True,
is_broadcast=True,
rms_eps=1.0e-6,
hc_pre_eps=2.0e-6,
hc_sinkhorn_eps=3.0e-6,
hc_post_mult_value=0.5,
sinkhorn_repeat=3,
norm_eps=1.0e-5,
),
),
(
MhcPostTileLangKernel(),
MhcPostTileLangKernel.CompileKey(hidden_size=4096, hc_mult=4),
),
(
MhcFusedTileLangKernel(),
MhcFusedTileLangKernel.CompileKey(
hidden_size=4096,
hc_mult=4,
n_splits=4,
tile_n=3,
),
),
(
MhcFusedTileLangKernel(),
MhcFusedTileLangKernel.CompileKey(
hidden_size=4096,
hc_mult=4,
n_splits=8,
tile_n=2,
),
),
(
MhcFusedTileLangKernel(),
MhcFusedTileLangKernel.CompileKey(
hidden_size=8192,
hc_mult=4,
n_splits=4,
tile_n=2,
),
),
(
HcHeadFusedTileLangKernel(),
HcHeadFusedTileLangKernel.CompileKey(
hidden_size=4096,
hc_mult=4,
rms_eps=1.0e-6,
hc_eps=2.0e-6,
),
),
],
)
def test_tilelang_warmup_inputs_reproduce_compile_key(
monkeypatch: pytest.MonkeyPatch,
kernel: Any,
compile_key: Any,
) -> None:
compiled: list[Any] = []
single_kernel = isinstance(
kernel,
(MhcPostTileLangKernel, MhcFusedTileLangKernel, HcHeadFusedTileLangKernel),
)
expected_kernel = kernel.kernel() if single_kernel else kernel.kernel(compile_key)
if single_kernel:
monkeypatch.setattr(
kernel,
"dispatch",
lambda **kwargs: pytest.fail("single-kernel launch must not dispatch"),
)
monkeypatch.setattr(
jit_warmup_tilelang_helper,
"compile_tilelang",
lambda jit_impl, *args, **kwargs: compiled.append(jit_impl),
)
kernel.compile(compile_key)
assert compiled == [expected_kernel]
+11 -2
View File
@@ -10,10 +10,12 @@ import torch.nn.functional as F
import vllm.model_executor.kernels.mhc # noqa: F401
import vllm.model_executor.layers.mhc as mhc_layers
from vllm.model_executor.kernels.mhc.tilelang import (
_tilelang_hc_prenorm_gemm,
_torch_hc_prenorm_gemm,
mhc_pre_delayed_tilelang,
)
from vllm.model_executor.kernels.mhc.tilelang_kernels import (
_HC_PRENORM_GEMM_TILELANG_KERNEL,
)
from vllm.model_executor.kernels.mhc.torch import mhc_pre_delayed_torch
from vllm.model_executor.kernels.mhc.triton import hc_collapse_triton
from vllm.model_executor.layers.mhc import (
@@ -485,7 +487,14 @@ def test_hc_prenorm_gemm_tilelang(num_tokens, hidden_size):
sqrsum = torch.empty_like(sqrsum_ref)
_torch_hc_prenorm_gemm(x, fn, out_ref, sqrsum_ref)
_tilelang_hc_prenorm_gemm(x, fn, out, sqrsum, hidden_size, hc_mult)
_HC_PRENORM_GEMM_TILELANG_KERNEL(
x,
fn,
out,
sqrsum,
hidden_size,
hc_mult,
)
torch.testing.assert_close(out, out_ref, atol=1e-5, rtol=1e-4)
torch.testing.assert_close(sqrsum, sqrsum_ref, atol=1e-2, rtol=1e-6)