[Bugfix] Allow vllm to still work if triton is not installed. (#6786)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
This commit is contained in:
Thomas Parnell
2024-07-29 14:51:27 -07:00
committed by GitHub
parent 7f8d612d24
commit 9a7e2d0534
13 changed files with 65 additions and 37 deletions
@@ -1,14 +1,22 @@
from vllm.model_executor.layers.fused_moe.fused_moe import (
fused_experts, fused_moe, fused_topk, get_config_file_name, grouped_topk)
from vllm.model_executor.layers.fused_moe.layer import (FusedMoE,
FusedMoEMethodBase)
from vllm.triton_utils import HAS_TRITON
__all__ = [
"fused_moe",
"fused_topk",
"fused_experts",
"get_config_file_name",
"grouped_topk",
"FusedMoE",
"FusedMoEMethodBase",
]
if HAS_TRITON:
from vllm.model_executor.layers.fused_moe.fused_moe import (
fused_experts, fused_moe, fused_topk, get_config_file_name,
grouped_topk)
__all__ += [
"fused_moe",
"fused_topk",
"fused_experts",
"get_config_file_name",
"grouped_topk",
]
+1 -13
View File
@@ -1,4 +1,3 @@
import math
from typing import Optional, Tuple
import torch
@@ -6,21 +5,10 @@ import triton
import triton.language as tl
from vllm.model_executor.layers.ops.rand import seeded_uniform
from vllm.triton_utils.sample import get_num_triton_sampler_splits
_EPS = 1e-6
# This is a hardcoded limit in Triton (max block size).
MAX_TRITON_N_COLS = 131072
def get_num_triton_sampler_splits(n_cols: int) -> int:
"""Get the number of splits to use for Triton sampling.
Triton has a limit on the number of columns it can handle, so we need to
split the tensor and call the kernel multiple times if it's too large.
"""
return math.ceil(n_cols / MAX_TRITON_N_COLS)
def _multi_split_sample(
probs: torch.Tensor,
@@ -6,8 +6,7 @@ from torch.nn.parameter import Parameter
from vllm import _custom_ops as ops
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
fused_moe)
from vllm.model_executor.layers.fused_moe import FusedMoE, FusedMoEMethodBase
from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
UnquantizedLinearMethod)
from vllm.model_executor.layers.quantization.base_config import (
@@ -404,6 +403,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
num_expert_group: Optional[int] = None,
topk_group: Optional[int] = None) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe import fused_moe
return fused_moe(x,
layer.w13_weight,
layer.w2_weight,
+5 -1
View File
@@ -6,7 +6,11 @@ from typing import Dict, List, Optional, Tuple
import torch
import torch.nn as nn
from vllm.model_executor.layers.ops.sample import sample as sample_triton
from vllm.triton_utils import HAS_TRITON
if HAS_TRITON:
from vllm.model_executor.layers.ops.sample import sample as sample_triton
from vllm.model_executor.sampling_metadata import (SamplingMetadata,
SamplingTensors,
SequenceGroupToSample)