Modularize fused experts and integrate PPLX kernels (#15956)

This commit is contained in:
bnellnm
2025-05-14 13:11:54 -07:00
committed by GitHub
parent 418d2f8bfb
commit f9c069c85e
42 changed files with 3830 additions and 660 deletions
+114
View File
@@ -0,0 +1,114 @@
# SPDX-License-Identifier: Apache-2.0
from dataclasses import dataclass
import pytest
import torch
import triton.language as tl
from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
invoke_moe_batched_triton_kernel)
@dataclass
class BatchedMMConfig:
dtype: torch.dtype
num_experts: int
max_tokens_per_expert: int
K: int
N: int
@dataclass
class BatchedMMTensors:
A: torch.Tensor # [E, max_tokens, K]
B: torch.Tensor # [E, K, N] - column major
C: torch.Tensor # [E, max_tokens, N]
num_expert_tokens: torch.Tensor # [E]
@staticmethod
def make_tensors(config: BatchedMMConfig):
A = torch.randn(
(config.num_experts, config.max_tokens_per_expert, config.K),
device="cuda",
dtype=config.dtype) / 10
B = torch.randn((config.num_experts, config.N, config.K),
device="cuda",
dtype=config.dtype)
C = torch.zeros(
(config.num_experts, config.max_tokens_per_expert, config.N),
device="cuda",
dtype=config.dtype)
num_expert_tokens = torch.randint(low=0,
high=config.max_tokens_per_expert,
size=(config.num_experts, ),
device="cuda",
dtype=torch.int32)
return BatchedMMTensors(A, B, C, num_expert_tokens)
def ref_impl(A: torch.Tensor, B: torch.Tensor, C: torch.Tensor,
num_expert_tokens: torch.Tensor) -> torch.Tensor:
num_expert_tokens_cpu = num_expert_tokens.clone()
num_expert_tokens_cpu = num_expert_tokens_cpu.to(device="cpu")
num_experts = num_expert_tokens.size(0)
for e in range(num_experts):
num_tokens = num_expert_tokens_cpu[e]
C[e, :num_tokens, :] = A[e, :num_tokens, :] @ B[e].transpose(0, 1)
return C
@pytest.mark.parametrize("num_experts", [16, 32])
@pytest.mark.parametrize("max_tokens_per_expert",
[32, 64, 128, 192, 224, 256, 512])
@pytest.mark.parametrize("K", [128, 256, 1024])
@pytest.mark.parametrize("N", [128, 256, 512, 1024])
@pytest.mark.parametrize("dtype",
[torch.float32, torch.float16, torch.bfloat16])
def test_batched_mm(num_experts: int, max_tokens_per_expert: int, K: int,
N: int, dtype: torch.dtype):
config = BatchedMMConfig(dtype, num_experts, max_tokens_per_expert, K, N)
tensors = BatchedMMTensors.make_tensors(config)
test_output = tensors.C
ref_output = test_output.clone()
compute_tl_dtype = {
torch.float16: tl.float16,
torch.bfloat16: tl.bfloat16,
torch.float32: tl.float32
}[test_output.dtype]
invoke_moe_batched_triton_kernel(
tensors.A,
tensors.B,
test_output,
tensors.num_expert_tokens,
compute_tl_dtype,
# Quantization data
None,
None,
None,
# Quantization schemes
False,
False,
False,
config={
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 16
})
ref_output = ref_impl(tensors.A, tensors.B, ref_output,
tensors.num_expert_tokens)
rtol, atol = {
torch.float16: (6e-2, 6e-2),
torch.bfloat16: (6e-2, 6e-2),
torch.float32: (1e-2, 1e-2),
}[test_output.dtype]
torch.testing.assert_close(test_output, ref_output, atol=atol, rtol=rtol)
+21 -25
View File
@@ -30,6 +30,11 @@ MNK_FACTORS = [
(224, 3072, 1536),
]
vllm_config = VllmConfig(parallel_config=ParallelConfig(
pipeline_parallel_size=1))
vllm_config.scheduler_config.max_num_seqs = 128
vllm_config.scheduler_config.max_model_len = 8192
@dataclasses.dataclass
class MOETensors:
@@ -190,7 +195,7 @@ def run_8_bit(moe_tensors: MOETensors8Bit,
'w1_q': moe_tensors.w1_q.transpose(1, 2), # type: ignore[union-attr]
'w2_q': moe_tensors.w2_q.transpose(1, 2), # type: ignore[union-attr]
'topk_weights': topk_weights,
'topk_ids_': topk_ids,
'topk_ids': topk_ids,
'ab_strides1': moe_tensors.ab_strides1,
'c_strides1': moe_tensors.c_strides1,
'ab_strides2': moe_tensors.ab_strides2,
@@ -231,18 +236,15 @@ def test_cutlass_moe_8_bit_no_graph(
per_out_ch: bool,
):
current_platform.seed_everything(7)
with set_current_vllm_config(
VllmConfig(parallel_config=ParallelConfig(
pipeline_parallel_size=1))):
with set_current_vllm_config(vllm_config):
mt = MOETensors8Bit.make_moe_tensors_8bit(m, k, n, e, per_act_token,
per_out_ch)
score = torch.randn((m, e), device="cuda", dtype=torch.half)
topk_weights, topk_ids = fused_topk(mt.a,
score,
topk,
renormalize=False)
topk_weights, topk_ids, _ = fused_topk(mt.a,
score,
topk,
renormalize=False)
# Note that we are using the dequantized versions of the tensors.
# Using a, w1 and w2 directly results in minor output differences.
@@ -276,20 +278,17 @@ def test_cutlass_moe_8_bit_cuda_graph(
per_out_ch: bool,
):
current_platform.seed_everything(7)
with set_current_vllm_config(
VllmConfig(parallel_config=ParallelConfig(
pipeline_parallel_size=1))):
with set_current_vllm_config(vllm_config):
dtype = torch.half
mt = MOETensors8Bit.make_moe_tensors_8bit(m, k, n, e, per_act_token,
per_out_ch)
score = torch.randn((m, e), device="cuda", dtype=dtype)
topk_weights, topk_ids = fused_topk(mt.a,
score,
topk,
renormalize=False)
topk_weights, topk_ids, _ = fused_topk(mt.a,
score,
topk,
renormalize=False)
# Note that we are using the dequantized versions of the tensors.
# Using a, w1 and w2 directly results in minor output differences.
@@ -334,18 +333,15 @@ def test_cutlass_moe_8_bit_EP(
ep_size: int,
):
current_platform.seed_everything(7)
with set_current_vllm_config(
VllmConfig(parallel_config=ParallelConfig(
pipeline_parallel_size=1))):
with set_current_vllm_config(vllm_config):
mt = MOETensors8Bit.make_moe_tensors_8bit(m, k, n, e, per_act_token,
per_out_channel)
score = torch.randn((m, e), device="cuda", dtype=torch.half)
topk_weights, topk_ids = fused_topk(mt.a,
score,
topk,
renormalize=False)
topk_weights, topk_ids, _ = fused_topk(mt.a,
score,
topk,
renormalize=False)
# Note that we are using the dequantized versions of the tensors.
# Using a, w1 and w2 directly results in minor output differences.
+50 -41
View File
@@ -12,6 +12,7 @@ from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
import vllm.model_executor.layers.fused_moe # noqa
from tests.kernels.utils import opcheck, stack_and_dev, torch_moe
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.model_executor.layers.fused_moe import fused_moe
from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
from vllm.model_executor.layers.fused_moe.moe_torch_iterative import (
@@ -32,6 +33,10 @@ NUM_EXPERTS = [8, 64]
EP_SIZE = [1, 4]
TOP_KS = [2, 6]
vllm_config = VllmConfig()
vllm_config.scheduler_config.max_num_seqs = 128
vllm_config.scheduler_config.max_model_len = 8192
@pytest.mark.parametrize("m", [1, 33, 64, 222, 1024 * 128])
@pytest.mark.parametrize("n", [128, 1024, 2048])
@@ -70,31 +75,33 @@ def test_fused_moe(
else:
e_map = None
torch_output = torch_moe(a, w1, w2, score, topk, e_map)
iterative_output = iterative_moe(a,
w1,
w2,
score,
topk,
global_num_experts=e,
expert_map=e_map,
renormalize=False)
with set_current_vllm_config(vllm_config):
torch_output = torch_moe(a, w1, w2, score, topk, e_map)
iterative_output = iterative_moe(a,
w1,
w2,
score,
topk,
global_num_experts=e,
expert_map=e_map,
renormalize=False)
# Pad the weight if moe padding is enabled
if padding:
w1 = F.pad(w1, (0, 128), "constant", 0)[..., 0:-128]
torch.cuda.empty_cache()
w2 = F.pad(w2, (0, 128), "constant", 0)[..., 0:-128]
torch.cuda.empty_cache()
# Pad the weight if moe padding is enabled
if padding:
w1 = F.pad(w1, (0, 128), "constant", 0)[..., 0:-128]
torch.cuda.empty_cache()
w2 = F.pad(w2, (0, 128), "constant", 0)[..., 0:-128]
torch.cuda.empty_cache()
triton_output = fused_moe(a,
w1,
w2,
score,
topk,
global_num_experts=e,
expert_map=e_map,
renormalize=False)
triton_output = fused_moe(a,
w1,
w2,
score,
topk,
global_num_experts=e,
expert_map=e_map,
renormalize=False)
torch.testing.assert_close(triton_output, torch_output, atol=2e-2, rtol=0)
torch.testing.assert_close(iterative_output,
torch_output,
@@ -115,7 +122,6 @@ def test_fused_moe(
def test_fused_moe_wn16(m: int, n: int, k: int, e: int, topk: int,
ep_size: int, dtype: torch.dtype, group_size: int,
has_zp: bool, weight_bits: int):
print(m, n, k, e, topk, dtype, group_size, has_zp, weight_bits)
a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10
@@ -194,22 +200,24 @@ def test_fused_moe_wn16(m: int, n: int, k: int, e: int, topk: int,
else:
e_map = None
triton_output = fused_moe(a,
w1_qweight,
w2_qweight,
score,
topk,
renormalize=False,
use_int4_w4a16=weight_bits == 4,
use_int8_w8a16=weight_bits == 8,
global_num_experts=e,
expert_map=e_map,
w1_scale=w1_scales,
w2_scale=w2_scales,
w1_zp=w1_qzeros if has_zp else None,
w2_zp=w2_qzeros if has_zp else None,
block_shape=[0, group_size])
torch_output = torch_moe(a, w1_ref, w2_ref, score, topk, e_map)
with set_current_vllm_config(vllm_config):
triton_output = fused_moe(a,
w1_qweight,
w2_qweight,
score,
topk,
renormalize=False,
use_int4_w4a16=weight_bits == 4,
use_int8_w8a16=weight_bits == 8,
global_num_experts=e,
expert_map=e_map,
w1_scale=w1_scales,
w2_scale=w2_scales,
w1_zp=w1_qzeros if has_zp else None,
w2_zp=w2_qzeros if has_zp else None,
block_shape=[0, group_size])
torch_output = torch_moe(a, w1_ref, w2_ref, score, topk, e_map)
torch.testing.assert_close(triton_output, torch_output, atol=2e-2, rtol=0)
@@ -515,7 +523,8 @@ def test_fused_marlin_moe(
topk_weights, topk_ids, _ = fused_topk(a, score, topk, False)
torch_output = torch_moe(a, w_ref1, w_ref2, score, topk, e_map)
with set_current_vllm_config(vllm_config):
torch_output = torch_moe(a, w_ref1, w_ref2, score, topk, e_map)
marlin_output = torch.ops.vllm.fused_marlin_moe(
a,
+691
View File
@@ -0,0 +1,691 @@
# SPDX-License-Identifier: Apache-2.0
"""Tests for the MOE layers.
Run `pytest tests/kernels/test_pplx_moe.py`.
"""
import dataclasses
import os
import traceback
from typing import Callable, Optional
import pytest
import torch
try:
from pplx_kernels import AllToAll
from pplx_kernels.nvshmem import (nvshmem_alloc_empty_unique_id,
nvshmem_finalize, nvshmem_get_unique_id,
nvshmem_init)
has_pplx = True
except ImportError:
has_pplx = False
from torch.multiprocessing import (
spawn) # pyright: ignore[reportPrivateImportUsage]
from typing_extensions import Concatenate, ParamSpec
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import override_config
from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
BatchedExperts, BatchedPrepareAndFinalize, BatchedTritonExperts)
from vllm.model_executor.layers.fused_moe.fused_moe import (fused_topk,
get_default_config)
from vllm.model_executor.layers.fused_moe.modular_kernel import (
FusedMoEModularKernel)
from vllm.platforms import current_platform
PPLX_PREPARE_COMBOS = [(4, 128, 128), (32, 1024, 512), (64, 1024, 512),
(222, 2048, 1024)]
PPLX_MOE_COMBOS = [
(1, 128, 128),
(2, 128, 512),
(3, 1024, 2048),
(32, 128, 1024),
(45, 512, 2048),
(64, 1024, 1024),
(222, 1024, 2048),
]
NUM_EXPERTS = [8, 64]
EP_SIZE = [1, 4]
TOP_KS = [1, 2, 6]
vllm_config = VllmConfig()
vllm_config.scheduler_config.max_num_seqs = 128
vllm_config.scheduler_config.max_model_len = 8192
P = ParamSpec("P")
requires_pplx = pytest.mark.skipif(
not has_pplx,
reason="Requires PPLX kernels",
)
@dataclasses.dataclass
class ProcessGroupInfo:
world_size: int
world_local_size: int
rank: int
node_rank: int
local_rank: int
device: torch.device
def _worker_parallel_launch(
local_rank: int,
world_size: int,
world_local_size: int,
node_rank: int,
init_method: str,
worker: Callable[Concatenate[ProcessGroupInfo, P], None],
*args: P.args,
**kwargs: P.kwargs,
) -> None:
rank = node_rank * world_local_size + local_rank
torch.cuda.set_device(local_rank)
device = torch.device("cuda", local_rank)
torch.distributed.init_process_group(
backend="cpu:gloo,cuda:nccl",
init_method=init_method,
rank=rank,
world_size=world_size,
device_id=device,
)
barrier = torch.tensor([rank], device=device)
torch.distributed.all_reduce(barrier)
try:
worker(
ProcessGroupInfo(
world_size=world_size,
world_local_size=world_local_size,
rank=rank,
node_rank=node_rank,
local_rank=local_rank,
device=device,
),
*args,
**kwargs,
)
except Exception as ex:
print(ex)
traceback.print_exc()
raise
finally:
torch.distributed.destroy_process_group()
def parallel_launch(
world_size: int,
worker: Callable[Concatenate[ProcessGroupInfo, P], None],
*args: P.args,
**kwargs: P.kwargs,
) -> None:
assert not kwargs
spawn(
_worker_parallel_launch,
args=(
world_size,
world_size,
0,
"tcp://localhost:29500",
worker,
) + args,
nprocs=world_size,
join=True,
)
def parallel_launch_from_env(
worker: Callable[Concatenate[ProcessGroupInfo, P], None],
*args: P.args,
**kwargs: P.kwargs,
) -> None:
"""
Launches a worker function in parallel across all processes in the current
environment. The environment must have the following variables set:
- WORLD_SIZE: The total number of processes.
- WORLD_LOCAL_SIZE: The number of processes on the current node.
- NODE_RANK: The rank of the current
- MASTER_ADDR: The address of the master process.
- MASTER_PORT: The port of the master process.
"""
assert not kwargs
world_size = int(os.environ["WORLD_SIZE"])
world_local_size = int(os.environ["WORLD_LOCAL_SIZE"])
node_rank = int(os.environ["NODE_RANK"])
assert "MASTER_ADDR" in os.environ
assert "MASTER_PORT" in os.environ
spawn(
_worker_parallel_launch,
args=(
world_size,
world_local_size,
node_rank,
"env://",
worker,
) + args,
nprocs=world_local_size,
join=True,
)
def torch_prepare(
a: torch.Tensor,
topk_ids: torch.Tensor,
num_experts: int,
max_num_tokens: Optional[int] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
assert topk_ids.dim() == 2
assert topk_ids.shape[0] == a.shape[0]
num_tokens, hidden_dim = a.shape
topk = topk_ids.shape[1]
tokens_per_expert = torch.bincount(topk_ids.view(-1),
minlength=num_experts)
assert tokens_per_expert.numel() == num_experts
if max_num_tokens is None:
max_num_tokens = int(tokens_per_expert.max().item())
b_a = torch.zeros((num_experts, max_num_tokens, hidden_dim),
dtype=a.dtype,
device=a.device)
token_counts = torch.zeros(num_experts, dtype=torch.int, device=a.device)
for token in range(num_tokens):
for j in range(topk):
expert_id = topk_ids[token, j]
idx = token_counts[expert_id]
b_a[expert_id, idx:idx + 1, :] = a[token, :]
token_counts[expert_id] = token_counts[expert_id] + 1
return b_a, tokens_per_expert
def torch_finalize(b_out: torch.Tensor, topk_weight: torch.Tensor,
topk_ids: torch.Tensor) -> torch.Tensor:
num_tokens = topk_ids.shape[0]
num_experts = b_out.shape[0]
K = b_out.shape[-1]
out = torch.zeros((num_tokens, K), dtype=b_out.dtype, device=b_out.device)
expert_counts = torch.zeros(num_experts,
dtype=torch.int,
device=b_out.device)
for token in range(num_tokens):
expert_ids = topk_ids[token]
for i in range(expert_ids.numel()):
expert_id = expert_ids[i]
idx = expert_counts[expert_id]
out[token, :] = out[token, :] + b_out[expert_id, idx:idx +
1, :] * topk_weight[token, i]
expert_counts[expert_id] = expert_counts[expert_id] + 1
return out
def torch_batched_moe(
a: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
topk_weight: torch.Tensor,
topk_ids: torch.Tensor,
) -> torch.Tensor:
num_experts = w1.shape[0]
b_a, tokens_per_expert = torch_prepare(a, topk_ids, num_experts)
assert b_a.dim() == 3
num_tokens, topk = topk_ids.shape
_, max_num_tokens, K = b_a.shape
assert num_experts == b_a.shape[0] and w2.shape[1] == K
out = torch.zeros((num_experts, max_num_tokens, K),
dtype=b_a.dtype,
device=b_a.device)
tmp = torch.empty((max_num_tokens, w1.shape[1] // 2),
dtype=b_a.dtype,
device=b_a.device)
for expert in range(num_experts):
num = tokens_per_expert[expert]
if num > 0:
torch.ops._C.silu_and_mul(
tmp[:num], b_a[expert, :num, :] @ w1[expert].transpose(0, 1))
out[expert, :num, :] = tmp[:num] @ w2[expert].transpose(0, 1)
return torch_finalize(out, topk_weight, topk_ids)
def batched_moe(
a: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
topk_weight: torch.Tensor,
topk_ids: torch.Tensor,
) -> torch.Tensor:
num_experts = w1.shape[0]
fused_experts = FusedMoEModularKernel(
BatchedPrepareAndFinalize(a.shape[0], world_size=1, dp_size=1, rank=0),
BatchedExperts(max_num_tokens=a.shape[0], dp_size=1, world_size=1))
return fused_experts(a, w1, w2, topk_weight, topk_ids, num_experts)
# Note: same as torch_moe but with fused_topk factored out.
def torch_moe2(
a: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
topk_weight: torch.Tensor,
topk_ids: torch.Tensor,
) -> torch.Tensor:
M, K = a.shape
topk = topk_ids.shape[1]
a = a.view(M, -1, K).repeat(1, topk, 1).reshape(-1, K)
out = torch.zeros(M * topk, w2.shape[1], dtype=a.dtype, device=a.device)
num_experts = w1.shape[0]
for i in range(num_experts):
mask = (topk_ids == i).view(-1)
if mask.sum():
out[mask] = SiluAndMul()(
a[mask] @ w1[i].transpose(0, 1)) @ w2[i].transpose(0, 1)
return (out.view(M, -1, w2.shape[1]) *
topk_weight.view(M, -1, 1).to(out.dtype)).sum(dim=1)
@pytest.mark.parametrize("m", [1, 33, 64, 222])
@pytest.mark.parametrize("n", [128, 1024, 2048])
@pytest.mark.parametrize("k", [128, 512, 1024])
@pytest.mark.parametrize("e", NUM_EXPERTS)
@pytest.mark.parametrize("topk", TOP_KS)
@pytest.mark.parametrize("dtype", [torch.bfloat16])
def test_fused_moe_batched_experts(
m: int,
n: int,
k: int,
e: int,
topk: int,
dtype: torch.dtype,
):
current_platform.seed_everything(7)
a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10
score = torch.randn((m, e), device="cuda", dtype=dtype)
with set_current_vllm_config(vllm_config):
topk_weight, topk_ids, _ = fused_topk(a, score, topk, False)
baseline_output = torch_moe2(a, w1, w2, topk_weight, topk_ids)
torch_output = torch_batched_moe(a, w1, w2, topk_weight, topk_ids)
batched_output = batched_moe(a, w1, w2, topk_weight, topk_ids)
torch.testing.assert_close(baseline_output,
torch_output,
atol=2e-2,
rtol=0)
torch.testing.assert_close(baseline_output,
batched_output,
atol=2e-2,
rtol=0)
def rank_chunk(num: int, r: int, w: int) -> int:
rem = num % w
return (num // w) + (1 if r < rem else 0)
def chunk_by_rank(t: torch.Tensor, r: int, w: int) -> torch.Tensor:
chunk = rank_chunk(t.shape[0], r, w)
return t[(r * chunk):(r + 1) * chunk]
def pplx_prepare_finalize(pgi: ProcessGroupInfo, dp_size: int, a: torch.Tensor,
topk_weight: torch.Tensor, topk_ids: torch.Tensor,
num_experts: int) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe.pplx_prepare_finalize import (
PplxPrepareAndFinalize)
assert torch.cuda.current_device() == pgi.local_rank
topk = topk_ids.shape[1]
num_tokens, hidden_dim = a.shape
block_size = 128
device = pgi.device
rank = pgi.rank
world_size = pgi.world_size
max_num_tokens = rank_chunk(num_tokens, 0, world_size)
ata = AllToAll.internode(
max_num_tokens=max_num_tokens,
num_experts=num_experts,
experts_per_token=topk,
rank=rank,
world_size=world_size,
dp_size=dp_size,
hidden_dim=hidden_dim,
hidden_dim_bytes=hidden_dim * a.dtype.itemsize,
hidden_dim_scale_bytes=(0 if a.dtype.itemsize != 1 else
((hidden_dim + block_size - 1) // block_size *
torch.float32.itemsize)),
)
topk_ids = topk_ids.to(dtype=torch.uint32)
prepare_finalize = PplxPrepareAndFinalize(
ata,
max_num_tokens,
world_size,
rank,
dp_size,
a.dtype,
)
a_chunk = chunk_by_rank(a, rank, world_size).to(device)
chunk_topk_weight = chunk_by_rank(topk_weight, rank, world_size).to(device)
chunk_topk_ids = chunk_by_rank(topk_ids, rank, world_size).to(device)
b_a, b_a_scale, expert_num_tokens = prepare_finalize.prepare(
a_chunk,
None,
None,
chunk_topk_weight,
chunk_topk_ids,
num_experts,
None,
False,
)
b_a = b_a * 1.5
out = torch.full(
(max_num_tokens, hidden_dim),
torch.nan,
dtype=a.dtype,
device=device,
)
prepare_finalize.finalize(
out,
b_a,
chunk_topk_weight,
chunk_topk_ids,
False,
)
torch.cuda.synchronize()
ata.destroy()
num_tokens = a_chunk.shape[0]
return out[:num_tokens]
def _pplx_prepare_finalize(
pgi: ProcessGroupInfo,
dp_size: int,
a: torch.Tensor,
score: torch.Tensor,
topk: torch.Tensor,
num_experts: int,
):
uid = nvshmem_get_unique_id(
) if pgi.rank == 0 else nvshmem_alloc_empty_unique_id()
torch.distributed.broadcast(uid, src=0)
nvshmem_init(uid, pgi.rank, pgi.world_size)
device = pgi.device
topk_weight, topk_ids, _ = fused_topk(a, score, topk, False)
k = a.shape[1]
a_rep = torch.repeat_interleave(a, topk, dim=0).to(device)
torch_output = (a_rep.view(-1, topk, k) * 1.5 *
topk_weight.view(-1, topk, 1).to(device)).sum(dim=1).to(
a.dtype)
pplx_output = pplx_prepare_finalize(pgi, dp_size, a, topk_weight, topk_ids,
num_experts)
torch_output = chunk_by_rank(torch_output, pgi.rank,
pgi.world_size).to(pplx_output.device)
torch.testing.assert_close(pplx_output, torch_output, atol=2e-2, rtol=0)
nvshmem_finalize()
# TODO (bnell): this test point does not work for odd M due to how the test is
# written, not due to limitations of the pplx kernels. The pplx_moe
# test below is able to deal with odd M.
@pytest.mark.parametrize("mnk", PPLX_PREPARE_COMBOS)
@pytest.mark.parametrize("e", NUM_EXPERTS)
@pytest.mark.parametrize("topk", TOP_KS)
@pytest.mark.parametrize("dtype", [torch.bfloat16])
@pytest.mark.parametrize("world_dp_size", [[2, 1]])
@requires_pplx
def test_pplx_prepare_finalize(
mnk: tuple[int, int, int],
e: int,
topk: int,
dtype: torch.dtype,
world_dp_size: tuple[int, int],
):
current_platform.seed_everything(7)
m, n, k = mnk
world_size, dp_size = world_dp_size
device = "cuda"
a = torch.randn((m, k), device=device, dtype=dtype) / 10
score = torch.randn((m, e), device=device, dtype=dtype)
parallel_launch(world_size, _pplx_prepare_finalize, dp_size, a, score,
topk, e)
def pplx_moe(
rank: int,
world_size: int,
dp_size: int,
a: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
topk_weight: torch.Tensor,
topk_ids: torch.Tensor,
use_compile: bool = True,
use_cudagraphs: bool = True,
) -> torch.Tensor:
from vllm.model_executor.layers.fused_moe.pplx_prepare_finalize import (
PplxPrepareAndFinalize)
device = torch.device("cuda", rank)
hidden_dim = a.shape[1]
num_experts = w1.shape[0]
block_size = 128
topk = topk_ids.shape[1]
max_num_tokens = rank_chunk(a.shape[0], 0, world_size)
ata = AllToAll.internode(
max_num_tokens=max_num_tokens,
num_experts=num_experts,
experts_per_token=topk,
rank=rank,
world_size=world_size,
dp_size=dp_size,
hidden_dim=hidden_dim,
hidden_dim_bytes=hidden_dim * a.dtype.itemsize,
hidden_dim_scale_bytes=(0 if a.dtype.itemsize != 1 else
((hidden_dim + block_size - 1) // block_size *
torch.float32.itemsize)),
)
topk_ids = topk_ids.to(dtype=torch.uint32)
prepare_finalize = PplxPrepareAndFinalize(
ata,
max_num_tokens,
world_size,
rank,
dp_size,
)
experts = BatchedTritonExperts(max_num_tokens=a.shape[0],
world_size=world_size,
dp_size=dp_size)
fused_experts = FusedMoEModularKernel(
prepare_finalize,
experts,
)
# Note: workers with the same dp_rank must use the exact same inputs.
a_chunk = chunk_by_rank(a, rank, world_size).to(device)
chunk_topk_weight = chunk_by_rank(topk_weight, rank, world_size).to(device)
chunk_topk_ids = chunk_by_rank(topk_ids, rank, world_size).to(device)
# Chunking weights like this only works for batched format
w1_chunk = chunk_by_rank(w1, rank, world_size).to(device)
w2_chunk = chunk_by_rank(w2, rank, world_size).to(device)
if use_compile:
_fused_experts = torch.compile(fused_experts,
backend='inductor',
fullgraph=True)
else:
_fused_experts = fused_experts
out = _fused_experts(a_chunk,
w1_chunk,
w2_chunk,
chunk_topk_weight,
chunk_topk_ids,
global_num_experts=num_experts)
if use_cudagraphs:
out.fill_(0)
stream = torch.cuda.Stream()
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph, stream=stream):
out = _fused_experts(a_chunk,
w1_chunk,
w2_chunk,
chunk_topk_weight,
chunk_topk_ids,
global_num_experts=num_experts)
torch.cuda.synchronize()
graph.replay()
torch.cuda.synchronize()
ata.destroy()
return out
def _batched_moe(pgi, dp_size, a, w1, w2, topk_weight, topk_ids):
assert torch.cuda.current_device() == pgi.local_rank
num_experts = w1.shape[0]
device = pgi.device
rank = pgi.rank
world_size = pgi.world_size
max_num_tokens = rank_chunk(a.shape[0], 0, world_size)
prepare_finalize = BatchedPrepareAndFinalize(
max_num_tokens=max_num_tokens,
world_size=world_size,
dp_size=dp_size,
rank=rank,
)
experts = BatchedExperts(max_num_tokens=a.shape[0],
world_size=1,
dp_size=1)
fused_experts = FusedMoEModularKernel(
prepare_finalize,
experts,
)
# Note: workers with the same dp_rank must use the exact same inputs.
a_chunk = chunk_by_rank(a, rank, world_size).to(device)
chunk_topk_weight = chunk_by_rank(topk_weight, rank, world_size).to(device)
chunk_topk_ids = chunk_by_rank(topk_ids, rank, world_size).to(device)
out = fused_experts(
a_chunk,
# Chunking weights like this only works for batched format
chunk_by_rank(w1, rank, world_size).to(device),
chunk_by_rank(w2, rank, world_size).to(device),
chunk_topk_weight,
chunk_topk_ids,
global_num_experts=num_experts)
return out
def _pplx_moe(
pgi: ProcessGroupInfo,
dp_size: int,
a: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
score: torch.Tensor,
topk: int,
):
uid = nvshmem_get_unique_id(
) if pgi.rank == 0 else nvshmem_alloc_empty_unique_id()
torch.distributed.broadcast(uid, src=0)
nvshmem_init(uid, pgi.rank, pgi.world_size)
m, k = a.shape
e, _, n = w2.shape
moe_config = get_default_config(m, e, n, k, topk, a.dtype, False)
with set_current_vllm_config(vllm_config), override_config(moe_config):
topk_weight, topk_ids, _ = fused_topk(a, score, topk, False)
torch_output = torch_moe2(a, w1, w2, topk_weight, topk_ids)
pplx_output = pplx_moe(pgi.rank, pgi.world_size, dp_size, a, w1, w2,
topk_weight, topk_ids)
# TODO (bnell): fix + re-enable
#batched_output = _batched_moe(pgi, dp_size, a, w1, w2, topk_weight,
# topk_ids)
torch_output = chunk_by_rank(torch_output, pgi.rank,
pgi.world_size).to(pplx_output.device)
torch.testing.assert_close(pplx_output, torch_output, atol=2e-2, rtol=0)
#torch.testing.assert_close(batched_output, torch_output, atol=2e-2, rtol=0)
nvshmem_finalize()
@pytest.mark.parametrize("mnk", PPLX_MOE_COMBOS)
@pytest.mark.parametrize("e", NUM_EXPERTS)
@pytest.mark.parametrize("topk", TOP_KS)
@pytest.mark.parametrize("dtype", [torch.bfloat16])
@pytest.mark.parametrize("world_dp_size", [[2, 1]])
@requires_pplx
def test_pplx_moe(
mnk: tuple[int, int, int],
e: int,
topk: int,
dtype: torch.dtype,
world_dp_size: tuple[int, int],
):
current_platform.seed_everything(7)
m, n, k = mnk
world_size, dp_size = world_dp_size
a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10
score = torch.randn((m, e), device="cuda", dtype=dtype)
parallel_launch(world_size, _pplx_moe, dp_size, a, w1, w2, score, topk)
+20 -14
View File
@@ -7,6 +7,7 @@ import pytest
import torch
from vllm import _custom_ops as ops
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import fused_moe
from vllm.platforms import current_platform
@@ -15,6 +16,10 @@ if current_platform.get_device_capability() < (9, 0):
pytest.skip("FP8 Triton requires CUDA 9.0 or higher",
allow_module_level=True)
vllm_config = VllmConfig()
vllm_config.scheduler_config.max_num_seqs = 128
vllm_config.scheduler_config.max_model_len = 8192
def native_w8a8_per_token_matmul(A, B, As, Bs, output_dtype=torch.float16):
"""Matrix multiplication function that supports per-token input
@@ -137,20 +142,21 @@ def test_w8a8_fp8_fused_moe(M, N, K, E, topk, dtype, seed):
w2_s = torch.rand(E, K, device=w2_fp32.device) * factor_for_scale
score = torch.randn((M, E), dtype=dtype)
ref_out = torch_w8a8_per_column_moe(a, w1, w2, w1_s, w2_s, score, topk)
out = fused_moe(
a,
w1,
w2,
score,
topk,
renormalize=False,
use_fp8_w8a8=True, # using fp8
per_channel_quant=True,
w1_scale=w1_s,
w2_scale=w2_s,
block_shape=None, # Not using block quantization
)
with set_current_vllm_config(vllm_config):
ref_out = torch_w8a8_per_column_moe(a, w1, w2, w1_s, w2_s, score, topk)
out = fused_moe(
a,
w1,
w2,
score,
topk,
renormalize=False,
use_fp8_w8a8=True, # using fp8
per_channel_quant=True,
w1_scale=w1_s,
w2_scale=w2_s,
block_shape=None, # Not using block quantization
)
# Check results
rel_diff = (torch.mean(
+10 -10
View File
@@ -11,7 +11,7 @@ from vllm.config import VllmConfig, set_current_vllm_config
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import fused_moe
from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
deep_gemm_moe_fp8)
_valid_deep_gemm_shape, deep_gemm_moe_fp8)
from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
from vllm.model_executor.layers.fused_moe.moe_align_block_size import (
moe_align_block_size)
@@ -30,6 +30,10 @@ if current_platform.get_device_capability() < (9, 0):
pytest.skip("FP8 Triton requires CUDA 9.0 or higher",
allow_module_level=True)
vllm_config = VllmConfig()
vllm_config.scheduler_config.max_num_seqs = 128
vllm_config.scheduler_config.max_model_len = 8192
# Test configurations
DTYPES = [torch.bfloat16] # [torch.half, torch.bfloat16, torch.float32]
NUM_TOKENS = [7, 83, 2048]
@@ -210,7 +214,6 @@ def test_w8a8_block_fp8_fused_moe(M, N, K, E, topk, block_size, dtype, seed):
score = torch.randn((M, E), dtype=dtype)
# Set the context to avoid lots of warning spam.
vllm_config = VllmConfig()
with set_current_vllm_config(vllm_config):
out = fused_moe(
a,
@@ -258,6 +261,7 @@ def per_block_cast_to_fp8(
@pytest.mark.parametrize(
"M,N,K,block_size,out_dtype,seed",
itertools.product(M, N, K, BLOCK_SIZE, OUT_DTYPES, SEEDS))
@pytest.mark.skipif(not dg_available, reason="DeepGemm kernels not available.")
@torch.inference_mode()
def test_w8a8_block_fp8_deep_gemm_matmul(M, N, K, block_size, out_dtype, seed):
# only aligned sizes
@@ -381,15 +385,11 @@ def test_w8a8_block_fp8_deep_gemm_fused_moe(M, N, K, E, topk, seed):
block_size = [block_m, block_m]
dtype = torch.bfloat16
# only aligned sizes
if (N % block_m != 0 or K % block_m != 0 or topk > E):
pytest.skip(
f"Skipping test; bad size m={M}, n={N}, k={K}, topk={topk}, E={E}")
if topk > E:
pytest.skip(f"Skipping test: topk={topk} > E={E}")
if N <= 512:
pytest.skip("Skipping N <= 512 until performance issues solved.")
vllm_config = VllmConfig()
if not _valid_deep_gemm_shape(M, N, K):
pytest.skip(f"Skipping test: invalid size m={M}, n={N}, k={K}")
torch.manual_seed(seed)
fp8_info = torch.finfo(torch.float8_e4m3fn)
@@ -18,6 +18,10 @@ if current_platform.get_device_capability() < (7, 0):
pytest.skip("INT8 Triton requires CUDA 7.0 or higher",
allow_module_level=True)
vllm_config = VllmConfig()
vllm_config.scheduler_config.max_num_seqs = 128
vllm_config.scheduler_config.max_model_len = 8192
# For test
def native_per_token_group_quant_int8(x,
@@ -174,7 +178,6 @@ def test_w8a8_block_int8_fused_moe(M, N, K, E, topk, block_size, dtype, seed):
score = torch.randn((M, E), dtype=dtype)
# Set the context to avoid lots of warning spam.
vllm_config = VllmConfig()
with set_current_vllm_config(vllm_config):
out = fused_moe(
a,