[Build] Avoid building too many extensions (#1624)
This commit is contained in:
@@ -4,7 +4,7 @@ from typing import Optional
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from vllm import activation_ops
|
||||
from vllm._C import ops
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
from vllm.model_executor.parallel_utils.parallel_state import (
|
||||
get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
|
||||
@@ -26,7 +26,7 @@ class SiluAndMul(nn.Module):
|
||||
d = x.shape[-1] // 2
|
||||
output_shape = (x.shape[:-1] + (d, ))
|
||||
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
|
||||
activation_ops.silu_and_mul(out, x)
|
||||
ops.silu_and_mul(out, x)
|
||||
return out
|
||||
|
||||
|
||||
@@ -34,7 +34,7 @@ class NewGELU(nn.Module):
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
out = torch.empty_like(x)
|
||||
activation_ops.gelu_new(out, x)
|
||||
ops.gelu_new(out, x)
|
||||
return out
|
||||
|
||||
|
||||
@@ -42,7 +42,7 @@ class FastGELU(nn.Module):
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
out = torch.empty_like(x)
|
||||
activation_ops.gelu_fast(out, x)
|
||||
ops.gelu_fast(out, x)
|
||||
return out
|
||||
|
||||
|
||||
|
||||
@@ -7,8 +7,8 @@ from xformers import ops as xops
|
||||
from xformers.ops.fmha.attn_bias import (BlockDiagonalCausalMask,
|
||||
LowerTriangularMaskWithTensorBias)
|
||||
|
||||
from vllm import attention_ops
|
||||
from vllm import cache_ops
|
||||
from vllm._C import ops
|
||||
from vllm._C import cache_ops
|
||||
from vllm.model_executor.input_metadata import InputMetadata
|
||||
from vllm.model_executor.layers.rotary_embedding import get_rope
|
||||
|
||||
@@ -163,7 +163,7 @@ class PagedAttention(nn.Module):
|
||||
max_num_partitions == 1 or num_seqs * num_heads > 512)
|
||||
if use_v1:
|
||||
# Run PagedAttention V1.
|
||||
attention_ops.paged_attention_v1(
|
||||
ops.paged_attention_v1(
|
||||
output,
|
||||
query,
|
||||
key_cache,
|
||||
@@ -190,7 +190,7 @@ class PagedAttention(nn.Module):
|
||||
device=output.device,
|
||||
)
|
||||
max_logits = torch.empty_like(exp_sums)
|
||||
attention_ops.paged_attention_v2(
|
||||
ops.paged_attention_v2(
|
||||
output,
|
||||
exp_sums,
|
||||
max_logits,
|
||||
|
||||
@@ -4,7 +4,7 @@ from typing import Optional, Tuple, Union
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from vllm import layernorm_ops
|
||||
from vllm._C import ops
|
||||
|
||||
|
||||
class RMSNorm(nn.Module):
|
||||
@@ -29,7 +29,7 @@ class RMSNorm(nn.Module):
|
||||
residual: Optional[torch.Tensor] = None,
|
||||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
||||
if residual is not None:
|
||||
layernorm_ops.fused_add_rms_norm(
|
||||
ops.fused_add_rms_norm(
|
||||
x,
|
||||
residual,
|
||||
self.weight.data,
|
||||
@@ -37,7 +37,7 @@ class RMSNorm(nn.Module):
|
||||
)
|
||||
return x, residual
|
||||
out = torch.empty_like(x)
|
||||
layernorm_ops.rms_norm(
|
||||
ops.rms_norm(
|
||||
out,
|
||||
x,
|
||||
self.weight.data,
|
||||
|
||||
@@ -3,7 +3,7 @@ from typing import Any, Dict, List, Optional
|
||||
import torch
|
||||
from torch.nn.parameter import Parameter
|
||||
|
||||
from vllm import quantization_ops
|
||||
from vllm._C import ops
|
||||
from vllm.model_executor.layers.linear import (LinearMethodBase,
|
||||
set_weight_attrs)
|
||||
from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
|
||||
@@ -151,8 +151,7 @@ class AWQLinearMethod(LinearMethodBase):
|
||||
pack_factor = self.quant_config.pack_factor
|
||||
out_shape = (x.shape[:-1] + (qweight.shape[-1] * pack_factor, ))
|
||||
reshaped_x = x.reshape(-1, x.shape[-1])
|
||||
out = quantization_ops.awq_gemm(reshaped_x, qweight, scales, qzeros,
|
||||
pack_factor)
|
||||
out = ops.awq_gemm(reshaped_x, qweight, scales, qzeros, pack_factor)
|
||||
if bias is not None:
|
||||
out = out + bias
|
||||
return out.reshape(out_shape)
|
||||
|
||||
@@ -3,7 +3,7 @@ from typing import Any, Dict, List, Optional
|
||||
import torch
|
||||
from torch.nn.parameter import Parameter
|
||||
|
||||
from vllm import quantization_ops
|
||||
from vllm._C import ops
|
||||
from vllm.model_executor.layers.linear import (LinearMethodBase,
|
||||
set_weight_attrs)
|
||||
from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
|
||||
@@ -116,8 +116,7 @@ class SqueezeLLMLinearMethod(LinearMethodBase):
|
||||
reshaped_x = x.reshape(-1, x.shape[-1])
|
||||
# NOTE: The output tensor should be zero-initialized.
|
||||
out = torch.zeros(out_shape, device="cuda", dtype=torch.float16)
|
||||
quantization_ops.squeezellm_gemm(reshaped_x, qweight, out,
|
||||
lookup_table)
|
||||
ops.squeezellm_gemm(reshaped_x, qweight, out, lookup_table)
|
||||
|
||||
if bias is not None:
|
||||
out = out + bias
|
||||
|
||||
@@ -27,7 +27,7 @@ from typing import Any, Dict, Optional, Tuple, Union
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from vllm import pos_encoding_ops
|
||||
from vllm._C import ops
|
||||
|
||||
|
||||
class RotaryEmbedding(nn.Module):
|
||||
@@ -87,11 +87,10 @@ class RotaryEmbedding(nn.Module):
|
||||
query: torch.Tensor,
|
||||
key: torch.Tensor,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
# pos_encoding_ops.rotary_embedding() is an in-place operation that
|
||||
# ops.rotary_embedding() is an in-place operation that
|
||||
# updates the query and key tensors.
|
||||
pos_encoding_ops.rotary_embedding(positions, query, key,
|
||||
self.head_size, self.cos_sin_cache,
|
||||
self.is_neox_style)
|
||||
ops.rotary_embedding(positions, query, key, self.head_size,
|
||||
self.cos_sin_cache, self.is_neox_style)
|
||||
return query, key
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user