[Refactor] Remove multiple dead code (#47329)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
@@ -400,7 +400,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
|
||||
"csrc/libtorch_stable/topk.cu"
|
||||
"csrc/libtorch_stable/mamba/selective_scan_fwd.cu"
|
||||
"csrc/libtorch_stable/cache_kernels.cu"
|
||||
"csrc/libtorch_stable/cache_kernels.cu"
|
||||
"csrc/libtorch_stable/cache_kernels_fused.cu"
|
||||
"csrc/libtorch_stable/custom_all_reduce.cu"
|
||||
"csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu")
|
||||
|
||||
@@ -162,13 +162,6 @@ class ConfigSet:
|
||||
config_key,
|
||||
)
|
||||
|
||||
def has_config(self, platform: str, config_key: CaseKey) -> bool:
|
||||
platform = platform.lower()
|
||||
platform_dict = self._configs.get(platform)
|
||||
if platform_dict is None:
|
||||
return False
|
||||
return config_key in platform_dict
|
||||
|
||||
|
||||
class ConfigManager:
|
||||
"""File-level configuration management for Helion kernels (global singleton)."""
|
||||
@@ -327,15 +320,3 @@ class ConfigManager:
|
||||
|
||||
logger.info("Saved config to: %s", platform_path)
|
||||
return platform_path
|
||||
|
||||
def config_exists(
|
||||
self,
|
||||
kernel_name: str,
|
||||
platform: str,
|
||||
config_key: CaseKey,
|
||||
) -> bool:
|
||||
platform_data = self._load_platform_file(kernel_name, platform)
|
||||
if not platform_data:
|
||||
return False
|
||||
target = dict(config_key)
|
||||
return any(entry["key"] == target for entry in platform_data)
|
||||
|
||||
@@ -331,46 +331,6 @@ def rmsnorm_fn(
|
||||
)
|
||||
|
||||
|
||||
class LayerNormGated(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size,
|
||||
eps: float = 1e-5,
|
||||
group_size: int | None = None,
|
||||
norm_before_gate: bool = True,
|
||||
device: torch.device | None = None,
|
||||
dtype: torch.dtype | None = None,
|
||||
):
|
||||
"""If group_size is not None, we do GroupNorm with each group having group_size elements.
|
||||
group_size=None is equivalent to group_size=hidden_size (i.e. there's only 1 group).
|
||||
"""
|
||||
|
||||
factory_kwargs = {"device": device, "dtype": dtype}
|
||||
super().__init__()
|
||||
self.eps = eps
|
||||
self.weight = nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
|
||||
self.bias = nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
|
||||
self.group_size = group_size
|
||||
self.norm_before_gate = norm_before_gate
|
||||
self.reset_parameters()
|
||||
|
||||
def reset_parameters(self):
|
||||
torch.nn.init.ones_(self.weight)
|
||||
torch.nn.init.zeros_(self.bias)
|
||||
|
||||
def forward(self, x, z=None):
|
||||
"""If z is not None, we do norm(x) * silu(z) if norm_before_gate, else norm(x * silu(z))"""
|
||||
return layernorm_fn(
|
||||
x,
|
||||
self.weight,
|
||||
self.bias,
|
||||
z=z,
|
||||
group_size=self.group_size,
|
||||
eps=self.eps,
|
||||
norm_before_gate=self.norm_before_gate,
|
||||
)
|
||||
|
||||
|
||||
class RMSNormGated(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
@@ -1,6 +1,5 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
from enum import Enum
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
import torch
|
||||
@@ -15,12 +14,6 @@ if TYPE_CHECKING:
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class FlashinferMoeBackend(Enum):
|
||||
TENSORRT_LLM = "TensorRT-LLM"
|
||||
CUTLASS = "CUTLASS"
|
||||
CUTEDSL = "CUTEDSL"
|
||||
|
||||
|
||||
def activation_to_flashinfer_int(activation: MoEActivation) -> int:
|
||||
return activation_to_flashinfer_type(activation).value
|
||||
|
||||
@@ -97,16 +90,6 @@ def rotate_weights_for_fi_trtllm_fp8_per_tensor_moe(
|
||||
)
|
||||
|
||||
|
||||
def is_flashinfer_supporting_global_sf(backend: FlashinferMoeBackend | None) -> bool:
|
||||
# TODO(shuw@nvidia): Update when new backends are added.
|
||||
backends_supporting_global_sf = (
|
||||
FlashinferMoeBackend.CUTLASS,
|
||||
FlashinferMoeBackend.TENSORRT_LLM,
|
||||
FlashinferMoeBackend.CUTEDSL,
|
||||
)
|
||||
return backend in backends_supporting_global_sf
|
||||
|
||||
|
||||
def convert_moe_weights_to_flashinfer_trtllm_block_layout(
|
||||
cache_permute_indices: dict[torch.Size, torch.Tensor],
|
||||
w13_weight: torch.Tensor,
|
||||
|
||||
@@ -16,43 +16,6 @@ from vllm.triton_utils import tl, triton
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
def apply_w8a8_block_int8_linear(
|
||||
input: torch.Tensor,
|
||||
weight: torch.Tensor,
|
||||
block_size: list[int],
|
||||
weight_scale: torch.Tensor,
|
||||
input_scale: torch.Tensor | None = None,
|
||||
bias: torch.Tensor | None = None,
|
||||
) -> torch.Tensor:
|
||||
assert input_scale is None
|
||||
# View input as 2D matrix for fp8 methods
|
||||
input_2d = input.view(-1, input.shape[-1])
|
||||
output_shape = [*input.shape[:-1], weight.shape[0]]
|
||||
|
||||
q_input, x_scale = per_token_group_quant_int8(input_2d, block_size[1])
|
||||
output = w8a8_block_int8_matmul(
|
||||
q_input, weight, x_scale, weight_scale, block_size, output_dtype=input.dtype
|
||||
)
|
||||
|
||||
if bias is not None:
|
||||
output = output + bias
|
||||
return output.to(dtype=input.dtype).view(*output_shape)
|
||||
|
||||
|
||||
def input_to_int8(
|
||||
x: torch.Tensor, dtype: torch.dtype = torch.int8
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""This function quantizes input values to int8 values with
|
||||
tensor-wise quantization."""
|
||||
iinfo = torch.iinfo(dtype)
|
||||
min_val, max_val = x.aminmax()
|
||||
amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
|
||||
int8_min, int8_max = iinfo.min, iinfo.max
|
||||
scale = int8_max / amax
|
||||
x_scl_sat = (x * scale).clamp(min=int8_min, max=int8_max)
|
||||
return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
|
||||
|
||||
|
||||
def block_dequant(
|
||||
x_q_block: torch.Tensor,
|
||||
x_s: torch.Tensor,
|
||||
|
||||
@@ -16,10 +16,6 @@ def query_machete_supported_quant_types(zero_points: bool) -> list[ScalarType]:
|
||||
return [scalar_types.uint4b8, scalar_types.uint8b128]
|
||||
|
||||
|
||||
def query_machete_supported_act_types(zero_points: bool) -> list[ScalarType]:
|
||||
return [torch.float16, torch.bfloat16]
|
||||
|
||||
|
||||
def query_machete_supported_group_sizes(act_type: torch.dtype) -> list[int]:
|
||||
"""
|
||||
Queries the supported group sizes for Machete based on the activation type.
|
||||
|
||||
@@ -738,71 +738,3 @@ def apply_gptq_marlin_linear(
|
||||
|
||||
output = marlin_unpad_output(output, output_size_per_partition, padded_n)
|
||||
return output.reshape(out_shape)
|
||||
|
||||
|
||||
def apply_awq_marlin_linear(
|
||||
input: torch.Tensor,
|
||||
weight: torch.Tensor,
|
||||
weight_scale: torch.Tensor,
|
||||
weight_zp: torch.Tensor,
|
||||
g_idx: torch.Tensor,
|
||||
g_idx_sort_indices: torch.Tensor,
|
||||
workspace: torch.Tensor,
|
||||
quant_type: ScalarType,
|
||||
output_size_per_partition: int,
|
||||
input_size_per_partition: int,
|
||||
input_global_scale: torch.Tensor | None = None,
|
||||
bias: torch.Tensor | None = None,
|
||||
use_fp32_reduce: bool = USE_FP32_REDUCE_DEFAULT,
|
||||
input_dtype: torch.dtype | None = None,
|
||||
) -> torch.Tensor:
|
||||
reshaped_x = input.reshape(-1, input.shape[-1])
|
||||
out_shape = input.shape[:-1] + (output_size_per_partition,)
|
||||
|
||||
padded_n, padded_k = marlin_repacked_nk(weight, quant_type.size_bits)
|
||||
reshaped_x = marlin_pad_dim(reshaped_x, input_size_per_partition, padded_k)
|
||||
|
||||
use_atomic_add = should_use_atomic_add_reduce(
|
||||
m=reshaped_x.size(0),
|
||||
n=padded_n,
|
||||
k=padded_k,
|
||||
device=input.device,
|
||||
dtype=input.dtype,
|
||||
)
|
||||
|
||||
a_scales = None
|
||||
if input_dtype == torch.int8:
|
||||
assert quant_type == scalar_types.uint4, (
|
||||
"W8A8-INT8 is not supported by marlin kernel."
|
||||
)
|
||||
reshaped_x, a_scales = marlin_quant_input(reshaped_x, input_dtype)
|
||||
a_scales = a_scales * input_global_scale
|
||||
elif input_dtype == torch.float8_e4m3fn:
|
||||
assert quant_type == scalar_types.uint4, (
|
||||
"INT8 weight + FP8 activation is not supported."
|
||||
)
|
||||
reshaped_x, a_scales = marlin_quant_input(reshaped_x, input_dtype)
|
||||
|
||||
output = ops.marlin_gemm(
|
||||
reshaped_x,
|
||||
None,
|
||||
weight,
|
||||
bias,
|
||||
weight_scale,
|
||||
a_scales,
|
||||
None,
|
||||
weight_zp,
|
||||
g_idx,
|
||||
g_idx_sort_indices,
|
||||
workspace,
|
||||
quant_type,
|
||||
size_m=reshaped_x.shape[0],
|
||||
size_n=padded_n,
|
||||
size_k=padded_k,
|
||||
use_atomic_add=use_atomic_add,
|
||||
use_fp32_reduce=use_fp32_reduce,
|
||||
is_zp_float=False,
|
||||
)
|
||||
|
||||
output = marlin_unpad_output(output, output_size_per_partition, padded_n)
|
||||
return output.reshape(out_shape)
|
||||
|
||||
Reference in New Issue
Block a user