[Kernel] Remove unused fake implementation (#55535)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
Jee Jee Li
2026-09-06 17:54:32 -07:00
committed by GitHub
parent 4df80187b5
commit 199cb9b964
39 changed files with 7 additions and 956 deletions
@@ -100,15 +100,6 @@ def pick_config(args: tuple[Any, ...], config_keys: list[CaseKey]) -> CaseKey |
return result
def fake_impl(
result: torch.Tensor, # [num_tokens, hidden_size]
input: torch.Tensor, # [num_tokens, hidden_size]
scale: torch.Tensor, # [num_tokens, 1]
scale_ub: torch.Tensor | None = None, # scalar tensor
) -> None:
return
def baseline(
result: torch.Tensor, # [num_tokens, hidden_size]
input: torch.Tensor, # [num_tokens, hidden_size]
@@ -135,7 +126,6 @@ def baseline(
mutates_args=["result", "scale"],
config_picker=pick_config,
input_generator=generate_inputs,
fake_impl=fake_impl,
helion_settings=helion.Settings(
autotune_baseline_fn=baseline,
ignore_warnings=[helion.exc.TensorOperationInWrapper],
@@ -159,23 +159,6 @@ def pick_config(args: tuple[Any, ...], config_keys: list[CaseKey]) -> CaseKey |
return result
def fake_impl(
qkv: torch.Tensor, # [num_tokens, (num_heads_q+num_heads_k+num_heads_v)*head_dim]
num_heads_q: int,
num_heads_k: int,
num_heads_v: int,
head_dim: int,
eps: float,
q_weight: torch.Tensor,
k_weight: torch.Tensor,
cos_sin_cache: torch.Tensor, # [max_position, rotary_dim]
is_neox: bool,
position_ids: torch.Tensor, # [num_tokens],
forced_token_heads_per_warp: int = -1, # dummy
) -> None:
return
def baseline(
qkv: torch.Tensor, # [num_tokens, (num_heads_q+num_heads_k+num_heads_v)*head_dim]
num_heads_q: int,
@@ -216,7 +199,6 @@ def baseline(
mutates_args=["qkv"],
config_picker=pick_config,
input_generator=generate_inputs,
fake_impl=fake_impl,
helion_settings=helion.Settings(
autotune_baseline_fn=baseline,
autotune_baseline_atol=5e-2,
@@ -135,23 +135,6 @@ def pick_config(args: tuple[Any, ...], config_keys: list[CaseKey]) -> CaseKey |
return result
def fake_impl(
input: torch.Tensor, # [num_tokens, hidden_size]
output_q: torch.Tensor, # [num_tokens, hidden_size]
output_s: torch.Tensor, # [num_tokens, groups_per_row]
group_size: int,
eps: float,
fp8_min: float,
fp8_max: float,
scale_ue8m0: bool,
# Unused dummy args
# Kept for consistency with existing kernel interface
dummy_is_scale_transposed: bool = False,
dummy_is_tma_aligned: bool = False,
) -> None:
return
def baseline(
input: torch.Tensor, # [num_tokens, hidden_size]
output_q: torch.Tensor, # [num_tokens, hidden_size]
@@ -181,7 +164,6 @@ def baseline(
mutates_args=["output_q", "output_s"],
config_picker=pick_config,
input_generator=generate_inputs,
fake_impl=fake_impl,
helion_settings=helion.Settings(
autotune_baseline_fn=baseline,
),
@@ -119,18 +119,6 @@ def pick_config(args: tuple[Any, ...], config_keys: list[CaseKey]) -> CaseKey |
return result
def fake_impl(
result: torch.Tensor, # [num_tokens, hidden_size]
input: torch.Tensor, # [num_tokens, hidden_size]
weight: torch.Tensor, # [hidden_size]
scale: torch.Tensor, # [num_tokens, 1]
epsilon: float,
scale_ub: torch.Tensor | None = None, # []
residual: torch.Tensor | None = None, # [num_tokens, hidden_size]
) -> None:
return
def baseline(
result: torch.Tensor, # [num_tokens, hidden_size]
input: torch.Tensor, # [num_tokens, hidden_size]
@@ -179,7 +167,6 @@ def baseline(
mutates_args=["result", "scale", "residual"],
config_picker=pick_config,
input_generator=generate_inputs,
fake_impl=fake_impl,
helion_settings=helion.Settings(
autotune_baseline_fn=baseline,
ignore_warnings=[helion.exc.TensorOperationInWrapper],
@@ -153,20 +153,6 @@ def pick_config(args: tuple[Any, ...], config_keys: list[CaseKey]) -> CaseKey |
return result
def fake_impl(
result: torch.Tensor, # [num_tokens, hidden_size]
input: torch.Tensor, # [num_tokens, hidden_size]
weight: torch.Tensor, # [hidden_size]
scale: torch.Tensor, # [num_tokens, groups_per_row]
epsilon: float,
scale_ub: torch.Tensor | None, # []
residual: torch.Tensor | None, # [num_tokens, hidden_size]
group_size: int,
is_scale_transposed: bool, # dummy
) -> None:
return
def baseline(
result: torch.Tensor, # [num_tokens, hidden_size]
input: torch.Tensor, # [num_tokens, hidden_size]
@@ -219,7 +205,6 @@ def baseline(
mutates_args=["result", "scale", "residual"],
config_picker=pick_config,
input_generator=generate_inputs,
fake_impl=fake_impl,
helion_settings=helion.Settings(
autotune_baseline_fn=baseline,
ignore_warnings=[helion.exc.TensorOperationInWrapper],
@@ -139,17 +139,6 @@ def pick_config(args: tuple[Any, ...], config_keys: list[CaseKey]) -> CaseKey |
return result
def fake_impl(
out: torch.Tensor, # [num_tokens, intermediate_size]
input: torch.Tensor, # [num_tokens, 2 * intermediate_size]
scales: torch.Tensor, # [num_tokens, groups_per_row]
group_size: int,
scale_ub: torch.Tensor | None = None, # scalar tensor
is_scale_transposed: bool = False,
) -> None:
return
def baseline(
out: torch.Tensor, # [num_tokens, intermediate_size]
input: torch.Tensor, # [num_tokens, 2 * intermediate_size]
@@ -193,7 +182,6 @@ def baseline(
mutates_args=["out", "scales"],
config_picker=pick_config,
input_generator=generate_inputs,
fake_impl=fake_impl,
helion_settings=helion.Settings(
autotune_baseline_fn=baseline,
ignore_warnings=[helion.exc.TensorOperationInWrapper],
+7 -2
View File
@@ -38,6 +38,7 @@ Key Classes
from __future__ import annotations
import inspect
from collections.abc import Callable
from typing import Any
@@ -258,7 +259,7 @@ class HelionKernelWrapper:
self,
raw_kernel_func: Callable,
op_name: str,
fake_impl: Callable,
fake_impl: Callable | None,
config_picker: ConfigPicker,
mutates_args: list[str] | None = None,
helion_settings: helion.Settings | None = None,
@@ -445,8 +446,12 @@ def register_kernel(
f"Use a different op_name or check for duplicate registrations."
)
# A void-returning kernel that mutates its args needs no fake impl:
# PyTorch auto-generates a trivial one for mutable, no-return schemas.
returns_none = inspect.signature(kernel_func).return_annotation is None
final_fake_impl = fake_impl
if final_fake_impl is None:
if final_fake_impl is None and not (mutates_args and returns_none):
final_fake_impl = infer_fake_impl(kernel_func, helion_settings)
logger.debug(
"Auto-generated fake_impl for Helion kernel '%s'",