[Kernel] [Helion] [4/N] Add silu_mul_fp8 Helion kernel (#33373)

Signed-off-by: Yanan Cao <gmagogsfm@gmail.com>
This commit is contained in:
Yanan Cao
2026-02-12 18:13:12 -08:00
committed by GitHub
parent 4453ba8d9e
commit 96161fe978
7 changed files with 1002 additions and 4 deletions
+1
View File
@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Helion integration for vLLM."""
import vllm.kernels.helion.ops # noqa: F401 Auto-register all Helion ops
from vllm.kernels.helion.config_manager import (
ConfigManager,
ConfigSet,
-3
View File
@@ -104,9 +104,6 @@ class ConfigSet:
result[platform] = {}
for config_key, config in config_keys_dict.items():
# Convert helion.Config to dict using to_json() + json.loads()
import json
result[platform][config_key] = json.loads(config.to_json())
return result
@@ -0,0 +1,550 @@
{
"nvidia_h200": {
"intermediate_2048_batchsize_256": {
"block_sizes": [
64,
128
],
"loop_orders": [
[
0,
1
]
],
"flatten_loops": [
true
],
"l2_groupings": [
2
],
"range_unroll_factors": [
0
],
"range_num_stages": [
0
],
"range_multi_buffers": [
null
],
"range_flattens": [
null
],
"load_eviction_policies": [
"",
"",
""
],
"num_warps": 32,
"num_stages": 1,
"indexing": [
"pointer",
"tensor_descriptor",
"pointer",
"pointer"
],
"pid_type": "flat",
"range_warp_specializes": []
},
"intermediate_4096_batchsize_256": {
"block_sizes": [
16,
64
],
"loop_orders": [
[
0,
1
]
],
"flatten_loops": [
false
],
"l2_groupings": [
1
],
"range_unroll_factors": [
0
],
"range_num_stages": [
0
],
"range_multi_buffers": [
null
],
"range_flattens": [
null
],
"load_eviction_policies": [
"",
"",
""
],
"num_warps": 2,
"num_stages": 1,
"indexing": [
"pointer",
"pointer",
"pointer",
"pointer"
],
"pid_type": "flat",
"range_warp_specializes": []
},
"default": {
"block_sizes": [
1,
512
],
"loop_orders": [
[
1,
0
]
],
"flatten_loops": [
false
],
"l2_groupings": [
4
],
"range_unroll_factors": [
0
],
"range_num_stages": [
0
],
"range_multi_buffers": [
null
],
"range_flattens": [
null
],
"load_eviction_policies": [
"",
"first",
""
],
"num_warps": 8,
"num_stages": 2,
"indexing": [
"tensor_descriptor",
"tensor_descriptor",
"tensor_descriptor",
"pointer"
],
"pid_type": "flat",
"range_warp_specializes": []
}
},
"nvidia_h100_pcie": {
"intermediate_2048_batchsize_256": {
"block_sizes": [
1,
512
],
"loop_orders": [
[
1,
0
]
],
"flatten_loops": [
false
],
"l2_groupings": [
4
],
"range_unroll_factors": [
0
],
"range_num_stages": [
0
],
"range_multi_buffers": [
null
],
"range_flattens": [
null
],
"load_eviction_policies": [
"",
"first",
""
],
"num_warps": 8,
"num_stages": 2,
"indexing": [
"tensor_descriptor",
"tensor_descriptor",
"tensor_descriptor",
"pointer"
],
"pid_type": "flat",
"range_warp_specializes": []
},
"intermediate_4096_batchsize_256": {
"block_sizes": [
256,
128
],
"loop_orders": [
[
0,
1
]
],
"flatten_loops": [
true
],
"l2_groupings": [
1
],
"range_unroll_factors": [
2
],
"range_num_stages": [
3
],
"range_multi_buffers": [
false
],
"range_flattens": [
true
],
"load_eviction_policies": [
"last",
"last",
""
],
"num_warps": 32,
"num_stages": 3,
"indexing": [
"tensor_descriptor",
"tensor_descriptor",
"tensor_descriptor",
"pointer"
],
"pid_type": "persistent_blocked",
"range_warp_specializes": []
},
"default": {
"block_sizes": [
1,
512
],
"loop_orders": [
[
1,
0
]
],
"flatten_loops": [
false
],
"l2_groupings": [
4
],
"range_unroll_factors": [
0
],
"range_num_stages": [
0
],
"range_multi_buffers": [
null
],
"range_flattens": [
null
],
"load_eviction_policies": [
"",
"first",
""
],
"num_warps": 8,
"num_stages": 2,
"indexing": [
"tensor_descriptor",
"tensor_descriptor",
"tensor_descriptor",
"pointer"
],
"pid_type": "flat",
"range_warp_specializes": []
}
},
"nvidia_h100_sxm5": {
"intermediate_2048_batchsize_256": {
"block_sizes": [
1,
512
],
"loop_orders": [
[
1,
0
]
],
"flatten_loops": [
false
],
"l2_groupings": [
4
],
"range_unroll_factors": [
0
],
"range_num_stages": [
0
],
"range_multi_buffers": [
null
],
"range_flattens": [
null
],
"load_eviction_policies": [
"",
"first",
""
],
"num_warps": 8,
"num_stages": 2,
"indexing": [
"tensor_descriptor",
"tensor_descriptor",
"tensor_descriptor",
"pointer"
],
"pid_type": "flat",
"range_warp_specializes": []
},
"intermediate_4096_batchsize_256": {
"block_sizes": [
256,
128
],
"loop_orders": [
[
0,
1
]
],
"flatten_loops": [
true
],
"l2_groupings": [
1
],
"range_unroll_factors": [
2
],
"range_num_stages": [
3
],
"range_multi_buffers": [
false
],
"range_flattens": [
true
],
"load_eviction_policies": [
"last",
"last",
""
],
"num_warps": 32,
"num_stages": 3,
"indexing": [
"tensor_descriptor",
"tensor_descriptor",
"tensor_descriptor",
"pointer"
],
"pid_type": "persistent_blocked",
"range_warp_specializes": []
},
"default": {
"block_sizes": [
1,
512
],
"loop_orders": [
[
1,
0
]
],
"flatten_loops": [
false
],
"l2_groupings": [
4
],
"range_unroll_factors": [
0
],
"range_num_stages": [
0
],
"range_multi_buffers": [
null
],
"range_flattens": [
null
],
"load_eviction_policies": [
"",
"first",
""
],
"num_warps": 8,
"num_stages": 2,
"indexing": [
"tensor_descriptor",
"tensor_descriptor",
"tensor_descriptor",
"pointer"
],
"pid_type": "flat",
"range_warp_specializes": []
}
},
"nvidia_h100": {
"intermediate_2048_batchsize_256": {
"block_sizes": [
1,
512
],
"loop_orders": [
[
1,
0
]
],
"flatten_loops": [
false
],
"l2_groupings": [
4
],
"range_unroll_factors": [
0
],
"range_num_stages": [
0
],
"range_multi_buffers": [
null
],
"range_flattens": [
null
],
"load_eviction_policies": [
"",
"first",
""
],
"num_warps": 8,
"num_stages": 2,
"indexing": [
"tensor_descriptor",
"tensor_descriptor",
"tensor_descriptor",
"pointer"
],
"pid_type": "flat",
"range_warp_specializes": []
},
"intermediate_4096_batchsize_256": {
"block_sizes": [
256,
128
],
"loop_orders": [
[
0,
1
]
],
"flatten_loops": [
true
],
"l2_groupings": [
1
],
"range_unroll_factors": [
2
],
"range_num_stages": [
3
],
"range_multi_buffers": [
false
],
"range_flattens": [
true
],
"load_eviction_policies": [
"last",
"last",
""
],
"num_warps": 32,
"num_stages": 3,
"indexing": [
"tensor_descriptor",
"tensor_descriptor",
"tensor_descriptor",
"pointer"
],
"pid_type": "persistent_blocked",
"range_warp_specializes": []
},
"default": {
"block_sizes": [
1,
512
],
"loop_orders": [
[
1,
0
]
],
"flatten_loops": [
false
],
"l2_groupings": [
4
],
"range_unroll_factors": [
0
],
"range_num_stages": [
0
],
"range_multi_buffers": [
null
],
"range_flattens": [
null
],
"load_eviction_policies": [
"",
"first",
""
],
"num_warps": 8,
"num_stages": 2,
"indexing": [
"tensor_descriptor",
"tensor_descriptor",
"tensor_descriptor",
"pointer"
],
"pid_type": "flat",
"range_warp_specializes": []
}
}
}
+11
View File
@@ -0,0 +1,11 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Auto-import all Helion op modules to trigger kernel registration."""
import importlib
import pkgutil
# Automatically import all submodules so that @register_kernel
# decorators execute and register ops with torch.ops.vllm_helion.
for _module_info in pkgutil.iter_modules(__path__):
importlib.import_module(f"{__name__}.{_module_info.name}")
+100
View File
@@ -0,0 +1,100 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
import torch
from vllm.logger import init_logger
from vllm.utils.import_utils import has_helion
if not has_helion():
raise ImportError(
"silu_mul_fp8 Helion kernel requires helion to be installed. "
"Install it with: pip install helion"
)
import helion.language as hl
from vllm.kernels.helion.register import register_kernel
logger = init_logger(__name__)
@register_kernel # type: ignore[misc]
def silu_mul_fp8(input: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
original_shape = input.shape
two_d = hl.specialize(original_shape[-1])
d = two_d // 2
output_shape = original_shape[:-1] + (d,)
input_2d = input.view(-1, original_shape[-1])
m = input_2d.shape[0]
# TODO(gmagogsfm): Support for more float8 subtypes (e4m3fnuz, e5m2) coming
out = torch.empty((m, d), device=input.device, dtype=torch.float8_e4m3fn)
input_part_a = input_2d[:, :d]
input_part_b = input_2d[:, d:]
assert scale.numel() == 1, "Scale must be a scalar Tensor"
for tile_m, tile_n in hl.tile([m, d]):
a_vals = input_part_a[tile_m, tile_n]
silu_result = torch.nn.functional.silu(a_vals)
b_vals = input_part_b[tile_m, tile_n]
result = silu_result * b_vals
result_f32 = result.to(torch.float32)
scale_val = hl.load(scale, [0])
inv_scale = 1.0 / scale_val
result_scaled = result_f32 * inv_scale
out[tile_m, tile_n] = result_scaled.to(out.dtype)
return out.view(output_shape)
@silu_mul_fp8.register_config_picker # type: ignore[misc]
def pick_silu_mul_fp8_config(
args: tuple[Any, ...], config_keys: list[str]
) -> str | None:
if not config_keys:
return None
input_tensor, scale = args
intermediate_size = input_tensor.shape[-1] // 2
# TODO(gmagosfm): Rerun autotuning to capture config for
# other batch sizes.
target_key = f"intermediate_{intermediate_size}_batchsize_256"
if target_key in config_keys:
return target_key
intermediate_sizes = []
for key in config_keys:
if key.startswith("intermediate_") and "_batchsize_256" in key:
try:
size_str = key.split("_")[1]
size = int(size_str)
intermediate_sizes.append((abs(size - intermediate_size), key))
except (ValueError, IndexError):
continue
if intermediate_sizes:
_, best_key = min(intermediate_sizes)
logger.debug(
"No exact config for intermediate_size=%d, using closest match: %s",
intermediate_size,
best_key,
)
return best_key
if "default" in config_keys:
return "default"
return None
def silu_mul_fp8_baseline(input: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
output_shape = input.shape[:-1] + (input.shape[-1] // 2,)
out = torch.empty(output_shape, dtype=torch.float8_e4m3fn, device=input.device)
torch.ops._C.silu_and_mul_quant(out, input, scale)
return out