[Build] Avoid building too many extensions (#1624)
This commit is contained in:
@@ -3,7 +3,7 @@ import torch
|
||||
import torch.nn.functional as F
|
||||
from transformers.activations import get_activation
|
||||
|
||||
from vllm import activation_ops
|
||||
from vllm._C import ops
|
||||
|
||||
DTYPES = [torch.half, torch.bfloat16, torch.float]
|
||||
NUM_TOKENS = [7, 83, 2048] # Arbitrary values for testing
|
||||
@@ -31,7 +31,7 @@ def test_silu_and_mul(
|
||||
torch.cuda.manual_seed(seed)
|
||||
x = torch.randn(num_tokens, 2 * d, dtype=dtype, device="cuda")
|
||||
out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
|
||||
activation_ops.silu_and_mul(out, x)
|
||||
ops.silu_and_mul(out, x)
|
||||
ref_out = ref_silu_and_mul(x)
|
||||
assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
|
||||
|
||||
@@ -51,7 +51,7 @@ def test_gelu_new(
|
||||
torch.cuda.manual_seed(seed)
|
||||
x = torch.randn(num_tokens, d, dtype=dtype, device="cuda")
|
||||
out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
|
||||
activation_ops.gelu_new(out, x)
|
||||
ops.gelu_new(out, x)
|
||||
ref_out = get_activation("gelu_new")(x)
|
||||
assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
|
||||
|
||||
@@ -70,6 +70,6 @@ def test_gelu_fast(
|
||||
torch.cuda.manual_seed(seed)
|
||||
x = torch.randn(num_tokens, d, dtype=dtype, device="cuda")
|
||||
out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
|
||||
activation_ops.gelu_fast(out, x)
|
||||
ops.gelu_fast(out, x)
|
||||
ref_out = get_activation("gelu_fast")(x)
|
||||
assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
|
||||
|
||||
@@ -6,7 +6,7 @@ import torch
|
||||
from xformers import ops as xops
|
||||
from xformers.ops.fmha.attn_bias import BlockDiagonalCausalMask
|
||||
|
||||
from vllm import attention_ops
|
||||
from vllm._C import ops
|
||||
from vllm.utils import get_max_shared_memory_bytes
|
||||
|
||||
FLOAT32_BYTES = torch.finfo(torch.float).bits // 8
|
||||
@@ -165,7 +165,7 @@ def test_paged_attention(
|
||||
# Call the paged attention kernel.
|
||||
output = torch.empty_like(query)
|
||||
if version == "v1":
|
||||
attention_ops.paged_attention_v1(
|
||||
ops.paged_attention_v1(
|
||||
output,
|
||||
query,
|
||||
key_cache,
|
||||
@@ -194,7 +194,7 @@ def test_paged_attention(
|
||||
device=output.device,
|
||||
)
|
||||
max_logits = torch.empty_like(exp_sums)
|
||||
attention_ops.paged_attention_v2(
|
||||
ops.paged_attention_v2(
|
||||
output,
|
||||
exp_sums,
|
||||
max_logits,
|
||||
|
||||
@@ -3,7 +3,7 @@ import random
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from vllm import cache_ops
|
||||
from vllm._C import cache_ops
|
||||
|
||||
DTYPES = [torch.half, torch.bfloat16, torch.float]
|
||||
NUM_TOKENS = [83] # Arbitrary values for testing
|
||||
|
||||
@@ -2,7 +2,7 @@ import pytest
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from vllm import layernorm_ops
|
||||
from vllm._C import ops
|
||||
|
||||
DTYPES = [torch.half, torch.bfloat16, torch.float]
|
||||
HIDDEN_SIZES = [67, 768, 2048, 5120, 8192] # Arbitrary values for testing
|
||||
@@ -48,7 +48,7 @@ def test_rms_norm(
|
||||
ref = RefRMSNorm(hidden_size).to(dtype).cuda()
|
||||
|
||||
out = torch.empty_like(x)
|
||||
layernorm_ops.rms_norm(
|
||||
ops.rms_norm(
|
||||
out,
|
||||
x,
|
||||
ref.weight.data,
|
||||
|
||||
@@ -5,7 +5,7 @@ import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
from vllm import pos_encoding_ops
|
||||
from vllm._C import ops
|
||||
|
||||
IS_NEOX_STYLE = [True, False]
|
||||
DTYPES = [torch.half, torch.bfloat16, torch.float]
|
||||
@@ -145,7 +145,7 @@ def test_rotary_embedding(
|
||||
# Run the kernel. The kernel is in-place, so we need to clone the inputs.
|
||||
out_query = query.clone()
|
||||
out_key = key.clone()
|
||||
pos_encoding_ops.rotary_embedding(
|
||||
ops.rotary_embedding(
|
||||
positions,
|
||||
out_query,
|
||||
out_key,
|
||||
|
||||
Reference in New Issue
Block a user