34 lines
1.5 KiB
Python
34 lines
1.5 KiB
Python
import importlib.util
|
|
import unittest
|
|
|
|
try:
|
|
import torch
|
|
except ImportError:
|
|
torch = None
|
|
|
|
GPU = torch is not None and torch.cuda.is_available() and importlib.util.find_spec("vllm") is not None
|
|
|
|
|
|
@unittest.skipUnless(GPU, "needs vLLM and a GPU")
|
|
class KernelMatch(unittest.TestCase):
|
|
def test_cpu_experts_compute_what_the_vllm_kernel_computes(self):
|
|
from vllm.config import VllmConfig, set_current_vllm_config
|
|
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
|
|
from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
|
|
|
|
from prain.tiered import cpu_experts
|
|
|
|
torch.manual_seed(0)
|
|
tokens, hidden, inter, experts = 8, 256, 128, 4
|
|
x = torch.randn(tokens, hidden, dtype=torch.bfloat16)
|
|
w13 = torch.randn(experts, 2 * inter, hidden, dtype=torch.bfloat16) * 0.1
|
|
w2 = torch.randn(experts, hidden, inter, dtype=torch.bfloat16) * 0.1
|
|
topk_ids = torch.stack([torch.randperm(experts)[:2] for _ in range(tokens)]).int()
|
|
topk_weights = torch.rand(tokens, 2)
|
|
with set_current_vllm_config(VllmConfig()):
|
|
gpu = fused_experts(x.cuda(), w13.cuda(), w2.cuda(), topk_weights.cuda(),
|
|
topk_ids.cuda(), activation=MoEActivation.GELU_TANH,
|
|
global_num_experts=experts).float().cpu()
|
|
cpu = cpu_experts(x, topk_weights, topk_ids, set(range(experts)), lambda e: (w13[e], w2[e]))
|
|
torch.testing.assert_close(cpu, gpu, rtol=3e-2, atol=3e-2)
|