[Kernel] DeepEP dispatch-combine kernel integration (#18434)

Signed-off-by: Varun <vsundarr@redhat.com>
Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
This commit is contained in:
Varun Sundar Rabindranath
2025-06-03 12:30:02 -07:00
committed by GitHub
co-authored by Varun Sundar Rabindranath
parent 01eee40536
commit fa98d77773
23 changed files with 1950 additions and 122 deletions
View File
+188
View File
@@ -0,0 +1,188 @@
# SPDX-License-Identifier: Apache-2.0
"""
DeepEP test utilities
"""
import dataclasses
import importlib
import traceback
from typing import Callable, Optional
import torch
from torch.distributed import ProcessGroup
from torch.multiprocessing import (
spawn) # pyright: ignore[reportPrivateImportUsage]
from typing_extensions import Concatenate, ParamSpec
has_deep_ep = importlib.util.find_spec("deep_ep") is not None
if has_deep_ep:
from vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize import ( # noqa: E501
DeepEPHTPrepareAndFinalize)
from vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize import ( # noqa: E501
DeepEPLLPrepareAndFinalize)
## Parallel Processes Utils
P = ParamSpec("P")
@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,
)
## DeepEP specific utils
@dataclasses.dataclass
class DeepEPHTArgs:
num_local_experts: int
@dataclasses.dataclass
class DeepEPLLArgs:
max_tokens_per_rank: int
hidden_size: int
num_experts: int
use_fp8_dispatch: bool
def make_deepep_ht_a2a(pg: ProcessGroup,
pgi: ProcessGroupInfo,
dp_size: int,
ht_args: DeepEPHTArgs,
q_dtype: Optional[torch.dtype] = None,
block_shape: Optional[list[int]] = None):
import deep_ep
# high throughput a2a
num_nvl_bytes = 1024 * 1024 * 1024 # 1GB
num_rdma_bytes, low_latency_mode, num_qps_per_rank = 0, False, 1
buffer = deep_ep.Buffer(group=pg,
num_nvl_bytes=num_nvl_bytes,
num_rdma_bytes=num_rdma_bytes,
low_latency_mode=low_latency_mode,
num_qps_per_rank=num_qps_per_rank)
return DeepEPHTPrepareAndFinalize(buffer=buffer,
world_size=pgi.world_size,
rank=pgi.rank,
dp_size=dp_size,
rank_expert_offset=pgi.rank *
ht_args.num_local_experts,
quant_dtype=q_dtype,
block_shape=block_shape)
def make_deepep_ll_a2a(pg: ProcessGroup,
pgi: ProcessGroupInfo,
dp_size: int,
deepep_ll_args: DeepEPLLArgs,
q_dtype: Optional[torch.dtype] = None,
block_shape: Optional[list[int]] = None):
import deep_ep
# low-latency a2a
num_rdma_bytes = deep_ep.Buffer.get_low_latency_rdma_size_hint(
deepep_ll_args.max_tokens_per_rank, deepep_ll_args.hidden_size,
pgi.world_size, deepep_ll_args.num_experts)
buffer = deep_ep.Buffer(group=pg,
num_rdma_bytes=num_rdma_bytes,
low_latency_mode=True,
num_qps_per_rank=deepep_ll_args.num_experts //
pgi.world_size)
return DeepEPLLPrepareAndFinalize(
buffer=buffer,
world_size=pgi.world_size,
dp_size=dp_size,
max_tokens_per_rank=deepep_ll_args.max_tokens_per_rank,
quant_dtype=q_dtype,
use_fp8_dispatch=deepep_ll_args.use_fp8_dispatch,
)
def make_deepep_a2a(pg: ProcessGroup,
pgi: ProcessGroupInfo,
dp_size: int,
deepep_ht_args: Optional[DeepEPHTArgs],
deepep_ll_args: Optional[DeepEPLLArgs],
q_dtype: Optional[torch.dtype] = None,
block_shape: Optional[list[int]] = None):
if deepep_ht_args is not None:
assert deepep_ll_args is None
return make_deepep_ht_a2a(pg, pgi, dp_size, deepep_ht_args, q_dtype,
block_shape)
assert deepep_ll_args is not None
return make_deepep_ll_a2a(pg, pgi, dp_size, deepep_ll_args, q_dtype)
@@ -0,0 +1,371 @@
# SPDX-License-Identifier: Apache-2.0
"""
Test DeepEP + DeepGEMM integration
"""
import dataclasses
import importlib
from typing import Optional
import pytest
import torch.distributed
from torch.distributed import ProcessGroup
from typing_extensions import ParamSpec
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
from vllm.model_executor.layers.fused_moe.modular_kernel import (
FusedMoEModularKernel)
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
per_token_group_quant_fp8)
from vllm.platforms import current_platform
from .deepep_utils import ProcessGroupInfo, parallel_launch
has_deep_ep = importlib.util.find_spec("deep_ep") is not None
try:
import deep_gemm
has_deep_gemm = True
except ImportError:
has_deep_gemm = False
if has_deep_ep:
from vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize import ( # noqa: E501
DeepEPHTPrepareAndFinalize)
from .deepep_utils import DeepEPHTArgs, make_deepep_a2a
if has_deep_gemm:
from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
DeepGemmExperts)
requires_deep_ep = pytest.mark.skipif(
not has_deep_ep,
reason="Requires deep_ep kernels",
)
requires_deep_gemm = pytest.mark.skipif(
not has_deep_gemm,
reason="Requires deep_gemm kernels",
)
P = ParamSpec("P")
def per_block_cast_to_fp8(
x: torch.Tensor,
block_size_n: int = 128) -> tuple[torch.Tensor, torch.Tensor]:
assert x.dim() == 2
m, n = x.shape
x_padded = torch.zeros(
(deep_gemm.ceil_div(m, 128) * 128,
deep_gemm.ceil_div(n, block_size_n) * block_size_n),
dtype=x.dtype,
device=x.device)
x_padded[:m, :n] = x
x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, block_size_n)
x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
x_scaled_sub = x_scaled.view_as(x_padded)[:m, :n].contiguous()
scales = (x_amax / 448.0).view(x_view.size(0), x_view.size(2))
return x_scaled_sub, scales
def make_block_quant_fp8_weights(
e: int,
n: int,
k: int,
block_size: list[int],
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Return weights w1, w2, w1q, w2q, w1_scale, w2_scale
"""
dtype = torch.bfloat16
fp8_info = torch.finfo(torch.float8_e4m3fn)
fp8_max, fp8_min = fp8_info.max, fp8_info.min
w1_bf16 = torch.randn((e, 2 * n, k), dtype=dtype) / 10
w1_bf16 = w1_bf16.clamp(min=fp8_min, max=fp8_max).to(dtype=dtype)
w2_bf16 = torch.randn((e, k, n), dtype=dtype) / 10
w2_bf16 = w2_bf16.clamp(min=fp8_min, max=fp8_max).to(dtype=dtype)
block_n, block_k = block_size[0], block_size[1]
n_tiles_w1 = ((2 * n) + block_n - 1) // block_n
k_tiles_w1 = (k + block_k - 1) // block_k
n_tiles_w2 = (k + block_n - 1) // block_n
k_tiles_w2 = (n + block_k - 1) // block_k
w1 = torch.empty_like(w1_bf16, dtype=torch.float8_e4m3fn)
w2 = torch.empty_like(w2_bf16, dtype=torch.float8_e4m3fn)
w1_s = torch.empty((e, n_tiles_w1, k_tiles_w1),
device="cuda",
dtype=torch.float32)
w2_s = torch.empty((e, n_tiles_w2, k_tiles_w2),
device="cuda",
dtype=torch.float32)
assert w1_s.shape == (e, (2 * n + 127) // 128, (k + 127) // 128)
assert (w2.shape[-2] + block_n - 1) // block_n == w2_s.shape[-2]
for i in range(e):
w1[i], w1_s[i] = per_block_cast_to_fp8(w1_bf16[i])
w2[i], w2_s[i] = per_block_cast_to_fp8(w2_bf16[i])
return w1, w2, w1_s, w2_s
@dataclasses.dataclass
class TestConfig:
topk: int
m: int
k: int
n: int
num_experts: int
block_size: list[int]
@dataclasses.dataclass
class TestTensors:
rank_tokens: torch.Tensor # all ranks make this many tokens
rank_token_scales: Optional[torch.Tensor]
topk: torch.Tensor
topk_weights: torch.Tensor
config: TestConfig
@staticmethod
def make(config: TestConfig, rank) -> "TestTensors":
dtype = torch.bfloat16
topk, m, k, block_size = (config.topk, config.m, config.k,
config.block_size)
fp8_info = torch.finfo(torch.float8_e4m3fn)
fp8_max, fp8_min = fp8_info.max, fp8_info.min
rank_tokens = torch.randn(
(m, k), device=torch.cuda.current_device(), dtype=dtype) / 10.0
rank_tokens = rank_tokens.clamp(min=fp8_min, max=fp8_max)
block_k = block_size[1]
_, rank_token_scales = per_token_group_quant_fp8(rank_tokens, block_k)
topk_ids = torch.randint(
low=0,
high=config.num_experts,
size=(m, topk),
device=torch.cuda.current_device()).to(dtype=torch.int64)
topk_weights = torch.randn(topk_ids.shape,
dtype=torch.float32,
device=torch.cuda.current_device())
return TestTensors(rank_tokens=rank_tokens,
rank_token_scales=rank_token_scales,
topk=topk_ids,
topk_weights=topk_weights,
config=config)
def make_modular_kernel(pg: ProcessGroup, pgi: ProcessGroupInfo, dp_size: int,
num_local_experts: int, q_dtype: Optional[torch.dtype],
block_shape: list[int]) -> FusedMoEModularKernel:
a2a: DeepEPHTPrepareAndFinalize = make_deepep_a2a(
pg=pg,
pgi=pgi,
dp_size=dp_size,
deepep_ht_args=DeepEPHTArgs(num_local_experts=num_local_experts),
deepep_ll_args=None,
q_dtype=q_dtype,
block_shape=block_shape)
fused_experts = DeepGemmExperts()
mk = FusedMoEModularKernel(prepare_finalize=a2a,
fused_experts=fused_experts)
return mk
def deep_ep_moe_impl(pg: ProcessGroup, pgi: ProcessGroupInfo, dp_size: int,
test_tensors: TestTensors, w1: torch.Tensor,
w2: torch.Tensor, w1_scale: Optional[torch.Tensor],
w2_scale: Optional[torch.Tensor],
num_experts: int) -> torch.Tensor:
num_local_experts = w1.size(0)
def build_expert_map():
num_local_experts = w1.size(0)
expert_map = torch.full((num_experts, ),
fill_value=-1,
dtype=torch.int32)
s = pgi.rank * num_local_experts
e = s + num_local_experts
expert_map[s:e] = torch.tensor(list(range(num_local_experts)))
return expert_map.to(device=torch.cuda.current_device(),
dtype=torch.int32)
q_dtype = torch.float8_e4m3fn
# Make modular kernel
mk: FusedMoEModularKernel = make_modular_kernel(
pg, pgi, dp_size, num_local_experts, q_dtype,
test_tensors.config.block_size)
a1_scale = test_tensors.rank_token_scales
out = mk.forward(hidden_states=test_tensors.rank_tokens,
w1=w1,
w2=w2,
topk_weights=test_tensors.topk_weights,
topk_ids=test_tensors.topk,
inplace=False,
activation="silu",
global_num_experts=num_experts,
expert_map=build_expert_map(),
w1_scale=w1_scale,
w2_scale=w2_scale,
w1_zp=None,
w2_zp=None,
a1_scale=a1_scale,
a2_scale=None,
apply_router_weight_on_input=False)
return out
def triton_impl(a: torch.Tensor, topk_ids: torch.Tensor,
topk_weights: torch.Tensor, w1: torch.Tensor, w2: torch.Tensor,
w1_scale: torch.Tensor, w2_scale: torch.Tensor,
a1_scale: torch.Tensor, block_shape: list[int]):
return fused_experts(
hidden_states=a,
w1=w1,
w2=w2,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=False,
use_fp8_w8a8=True,
w1_scale=w1_scale,
w2_scale=w2_scale,
a1_scale=a1_scale,
block_shape=block_shape,
# Make sure this is set to False so we
# dont end up comparing the same implementation.
allow_deep_gemm=False)
def _deep_ep_moe(
pgi: ProcessGroupInfo,
dp_size: int,
config: TestConfig,
w1: torch.Tensor,
w2: torch.Tensor,
w1_scale: torch.Tensor,
w2_scale: torch.Tensor,
):
current_platform.seed_everything(pgi.rank)
w1 = w1.to(device=torch.cuda.current_device())
w2 = w2.to(device=torch.cuda.current_device())
w1_scale = w1_scale.to(device=torch.cuda.current_device())
w2_scale = w2_scale.to(device=torch.cuda.current_device())
pg = torch.distributed.new_group(list(range(pgi.world_size)))
test_tensors = TestTensors.make(config, pgi.rank)
block_shape = [
w1.size(1) // w1_scale.size(1),
w1.size(2) // w1_scale.size(2)
]
with set_current_vllm_config(VllmConfig()):
# Reference
triton_moe = triton_impl(a=test_tensors.rank_tokens,
topk_ids=test_tensors.topk,
topk_weights=test_tensors.topk_weights,
w1=w1,
w2=w2,
w1_scale=w1_scale,
w2_scale=w2_scale,
a1_scale=test_tensors.rank_token_scales,
block_shape=block_shape)
# Slice experts for this rank.
num_local_experts = config.num_experts // pgi.world_size
e_start = num_local_experts * pgi.rank
e_end = e_start + num_local_experts
w1_ep = w1[e_start:e_end]
w2_ep = w2[e_start:e_end]
w1_scale_ep = w1_scale[e_start:e_end]
w2_scale_ep = w2_scale[e_start:e_end]
deepep_moe = deep_ep_moe_impl(
pg,
pgi,
dp_size,
test_tensors,
w1_ep,
w2_ep,
w1_scale_ep,
w2_scale_ep,
config.num_experts,
)
torch.testing.assert_close(
triton_moe,
deepep_moe,
atol=6e-2,
rtol=6e-2,
)
MNKs = [
(8, 128, 128),
(8, 128, 512),
(8, 512, 512),
(3, 1024, 2048),
(32, 128, 1024),
(45, 512, 2048),
(64, 1024, 1024),
(129, 128, 256),
(129, 1024, 2048),
(222, 1024, 2048),
]
@pytest.mark.parametrize("mnk", MNKs)
@pytest.mark.parametrize("num_experts", [32])
@pytest.mark.parametrize("topk", [2, 6])
@pytest.mark.parametrize("world_dp_size", [(2, 1)])
@requires_deep_ep
@requires_deep_gemm
def test_deep_ep_moe(mnk: tuple[int, int, int], num_experts: int, topk: int,
world_dp_size: tuple[int, int]):
m, n, k = mnk
current_platform.seed_everything(7)
if topk > num_experts:
pytest.skip(f"Skipping test: topk={topk} > E={num_experts}")
block_m = deep_gemm.get_m_alignment_for_contiguous_layout()
block_size = [block_m, block_m]
world_size, dp_size = world_dp_size
config = TestConfig(
topk=topk,
m=m,
k=k,
n=n,
num_experts=num_experts,
block_size=block_size,
)
w1, w2, w1_scale, w2_scale = make_block_quant_fp8_weights(
num_experts, n, k, block_size)
parallel_launch(world_size, _deep_ep_moe, dp_size, config, w1, w2,
w1_scale, w2_scale)
+459
View File
@@ -0,0 +1,459 @@
# SPDX-License-Identifier: Apache-2.0
"""
Test deepep dispatch-combine logic
"""
import dataclasses
import importlib
from typing import Optional, Union
import pytest
import torch.distributed
from torch.distributed import ProcessGroup
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 TritonExperts
from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
BatchedTritonExperts)
from vllm.model_executor.layers.fused_moe.modular_kernel import (
FusedMoEModularKernel)
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
per_token_group_quant_fp8)
from vllm.platforms import current_platform
from .deepep_utils import ProcessGroupInfo, parallel_launch
has_deep_ep = importlib.util.find_spec("deep_ep") is not None
if has_deep_ep:
from vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize import ( # noqa: E501
DeepEPHTPrepareAndFinalize)
from vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize import ( # noqa: E501
DeepEPLLPrepareAndFinalize)
from .deepep_utils import DeepEPHTArgs, DeepEPLLArgs, make_deepep_a2a
requires_deep_ep = pytest.mark.skipif(
not has_deep_ep,
reason="Requires deep_ep kernels",
)
MAX_TOKENS_PER_RANK = 64
def make_weights(
e, n, k, dtype
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Return weights w1, w2, w1_scale, w2_scale
"""
if dtype in [torch.float16, torch.bfloat16]:
w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10
return w1, w2, None, None
# per-out-channel weight quantization
assert dtype == torch.float8_e4m3fn
w1 = torch.empty((e, 2 * n, k), device="cuda", dtype=torch.float16)
w2 = torch.empty((e, k, n), device="cuda", dtype=torch.float16)
n_b_scales = 2 * n
k_b_scales = k
w1_q = torch.empty_like(w1, dtype=dtype)
w2_q = torch.empty_like(w2, dtype=dtype)
w1_scale = torch.empty((e, n_b_scales, 1),
device="cuda",
dtype=torch.float32)
w2_scale = torch.empty((e, k_b_scales, 1),
device="cuda",
dtype=torch.float32)
for expert in range(e):
w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant(
w1[expert], use_per_token_if_dynamic=True)
w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant(
w2[expert], use_per_token_if_dynamic=True)
return w1_q, w2_q, w1_scale, w2_scale
@dataclasses.dataclass
class TestConfig:
dtype: torch.dtype
topk: int
m: int
k: int
n: int
num_experts: int
@dataclasses.dataclass
class TestTensors:
rank_tokens: torch.Tensor # all ranks make this many tokens
rank_token_scales: Optional[torch.Tensor]
topk: torch.Tensor
topk_weights: torch.Tensor
config: TestConfig
@staticmethod
def make(config: TestConfig, low_latency_mode: bool) -> "TestTensors":
# TODO (varun) - check that float16 works ?
assert config.dtype in [torch.bfloat16, torch.float8_e4m3fn]
token_dtype = (torch.bfloat16 if config.dtype == torch.float8_e4m3fn
else config.dtype)
rank_tokens = torch.randn(
(config.m, config.k), device="cuda", dtype=token_dtype) / 10
rank_token_scales = None
if config.dtype == torch.float8_e4m3fn:
# low_latency_mode kernels dont support per-token quant.
_, rank_token_scales = ops.scaled_fp8_quant(
rank_tokens, use_per_token_if_dynamic=not low_latency_mode)
topk = torch.randint(low=0,
high=config.num_experts,
size=(config.m, config.topk),
device="cuda").to(dtype=torch.int64)
topk_weights = torch.randn(topk.shape,
dtype=torch.float32,
device="cuda")
return TestTensors(rank_tokens=rank_tokens,
rank_token_scales=rank_token_scales,
topk=topk,
topk_weights=topk_weights,
config=config)
def make_modular_kernel(pg: ProcessGroup, pgi: ProcessGroupInfo,
low_latency_mode: bool, hidden_size: int, dp_size: int,
num_experts: int, num_local_experts: int,
q_dtype: Optional[torch.dtype],
use_fp8_dispatch: bool) -> FusedMoEModularKernel:
is_quantized = q_dtype is not None
ht_args: Optional[DeepEPHTArgs] = None
ll_args: Optional[DeepEPLLArgs] = None
if low_latency_mode:
ll_args = DeepEPLLArgs(max_tokens_per_rank=MAX_TOKENS_PER_RANK,
hidden_size=hidden_size,
num_experts=num_experts,
use_fp8_dispatch=use_fp8_dispatch)
else:
assert not use_fp8_dispatch, (
"FP8 Dispatch is valid only for low-latency kernels")
ht_args = DeepEPHTArgs(num_local_experts=num_local_experts)
a2a : Union[DeepEPHTPrepareAndFinalize, DeepEPLLPrepareAndFinalize] = \
make_deepep_a2a(pg = pg,
pgi = pgi,
dp_size = dp_size,
q_dtype = q_dtype,
block_shape = None,
deepep_ht_args = ht_args,
deepep_ll_args = ll_args)
if low_latency_mode:
fused_experts = BatchedTritonExperts(
max_num_tokens=MAX_TOKENS_PER_RANK,
world_size=pgi.world_size,
dp_size=dp_size,
use_fp8_w8a8=is_quantized,
use_int8_w8a8=False,
use_int8_w8a16=False,
use_int4_w4a16=False)
else:
fused_experts = TritonExperts(use_fp8_w8a8=is_quantized,
use_int8_w8a8=False,
use_int8_w8a16=False,
use_int4_w4a16=False,
per_channel_quant=False)
mk = FusedMoEModularKernel(prepare_finalize=a2a,
fused_experts=fused_experts)
return mk
def deep_ep_moe_impl(pg: ProcessGroup, pgi: ProcessGroupInfo,
low_latency_mode: bool, dp_size: int,
test_tensors: TestTensors, w1: torch.Tensor,
w2: torch.Tensor, w1_scale: Optional[torch.Tensor],
w2_scale: Optional[torch.Tensor], num_experts: int,
use_fp8_dispatch: bool) -> torch.Tensor:
num_local_experts = w1.size(0)
def build_expert_map():
num_local_experts = w1.size(0)
expert_map = torch.full((num_experts, ),
fill_value=-1,
dtype=torch.int32)
s = pgi.rank * num_local_experts
e = s + num_local_experts
expert_map[s:e] = torch.tensor(list(range(num_local_experts)))
return expert_map.to(device=torch.cuda.current_device(),
dtype=torch.int32)
hidden_size = test_tensors.rank_tokens.size(1)
is_quantized = w1.dtype == torch.float8_e4m3fn
q_dtype = None
if is_quantized:
q_dtype = torch.float8_e4m3fn
# Make modular kernel
mk: FusedMoEModularKernel = make_modular_kernel(pg, pgi, low_latency_mode,
hidden_size, dp_size,
num_experts,
num_local_experts, q_dtype,
use_fp8_dispatch)
out_hidden_states = torch.empty_like(test_tensors.rank_tokens)
total_num_tokens = test_tensors.rank_tokens.size(0)
def process_chunk(chunk_start, chunk_end, skip_result_store=False):
rank_tokens_chunk = test_tensors.rank_tokens[chunk_start:chunk_end]
topk_weights_chunk = test_tensors.topk_weights[chunk_start:chunk_end]
topk_chunk = test_tensors.topk[chunk_start:chunk_end]
rank_token_scales_chunk = test_tensors.rank_token_scales
if rank_token_scales_chunk is not None and rank_token_scales_chunk.size(
0) == total_num_tokens:
# per act token
rank_token_scales_chunk = rank_token_scales_chunk[
chunk_start:chunk_end]
out = mk.forward(hidden_states=rank_tokens_chunk,
w1=w1,
w2=w2,
topk_weights=topk_weights_chunk,
topk_ids=topk_chunk,
inplace=False,
activation="silu",
global_num_experts=num_experts,
expert_map=build_expert_map(),
w1_scale=w1_scale,
w2_scale=w2_scale,
w1_zp=None,
w2_zp=None,
a1_scale=rank_token_scales_chunk,
a2_scale=None,
apply_router_weight_on_input=False)
if not skip_result_store:
out_hidden_states[chunk_start:chunk_end, :].copy_(
out, non_blocking=True)
max_num_tokens_per_dp = (MAX_TOKENS_PER_RANK
if low_latency_mode else total_num_tokens)
for chunk_start_ in range(0, total_num_tokens, max_num_tokens_per_dp):
chunk_start = chunk_start_
chunk_end = min(chunk_start + max_num_tokens_per_dp, total_num_tokens)
# clamp start and end
chunk_start = min(chunk_start, total_num_tokens - 1)
chunk_end = min(chunk_end, total_num_tokens)
process_chunk(chunk_start,
chunk_end,
skip_result_store=chunk_start_ >= total_num_tokens)
return out_hidden_states
def torch_moe_impl(test_tensors: TestTensors, w1: torch.Tensor,
w2: torch.Tensor, w1_scale: Optional[torch.Tensor],
w2_scale: Optional[torch.Tensor], using_fp8_dispatch: bool):
a, topk_ids, topk_weights = (test_tensors.rank_tokens, test_tensors.topk,
test_tensors.topk_weights)
if using_fp8_dispatch:
# The DeepEP implementation is requested to dispatch using FP8.
# For numerical stability for testing, emulate the fp8 dispatch by
# blockwise quant and de-quant.
a = test_tensors.rank_tokens
aq, aq_scale = per_token_group_quant_fp8(a, 128)
a = (aq.view(-1, 128).to(torch.float32) * aq_scale.view(-1, 1)).view(
a.shape).to(a.dtype)
is_quantized = w1.dtype == torch.float8_e4m3fn
a_dtype = a.dtype
if is_quantized:
w1 = w1.to(dtype=torch.float32) * w1_scale
w2 = w2.to(dtype=torch.float32) * w2_scale
a = a.to(dtype=torch.float32)
m, _ = a.shape
topk = topk_ids.size(1)
out = torch.zeros_like(a)
for i in range(m):
a_i = a[i]
o_i = out[i]
for j in range(topk):
e = topk_ids[i][j]
e_w = topk_weights[i][j]
w1_e = w1[e]
w2_e = w2[e]
o_i += (SiluAndMul()
(a_i @ w1_e.transpose(0, 1)) @ w2_e.transpose(0, 1)) * e_w
if is_quantized:
out = out.to(dtype=a_dtype)
return out
def _deep_ep_moe(
pgi: ProcessGroupInfo,
low_latency_mode: bool,
dp_size: int,
config: TestConfig,
w1: torch.Tensor,
w2: torch.Tensor,
w1_scale: Optional[torch.Tensor],
w2_scale: Optional[torch.Tensor],
use_fp8_dispatch: bool,
):
if not low_latency_mode:
assert not use_fp8_dispatch, (
"FP8 dispatch interface is available only in low-latency mode")
is_quantized = w1.dtype == torch.float8_e4m3fn
w1 = w1.to(device=torch.cuda.current_device())
w2 = w2.to(device=torch.cuda.current_device())
if is_quantized:
w1_scale = w1_scale.to( # type: ignore
device=torch.cuda.current_device())
w2_scale = w2_scale.to( # type: ignore
device=torch.cuda.current_device())
pg = torch.distributed.new_group(list(range(pgi.world_size)))
test_tensors = TestTensors.make(config, low_latency_mode)
with set_current_vllm_config(VllmConfig()):
# Reference
torch_combined = torch_moe_impl(test_tensors, w1, w2, w1_scale,
w2_scale, use_fp8_dispatch)
# Splice experts for this rank.
num_local_experts = config.num_experts // pgi.world_size
e_start = num_local_experts * pgi.rank
e_end = e_start + num_local_experts
w1_ep = w1[e_start:e_end]
w2_ep = w2[e_start:e_end]
w1_scale_ep, w2_scale_ep = None, None
if is_quantized:
w1_scale_ep = w1_scale[e_start:e_end] # type: ignore
w2_scale_ep = w2_scale[e_start:e_end] # type: ignore
deepep_combined = deep_ep_moe_impl(
pg,
pgi,
low_latency_mode,
dp_size,
test_tensors,
w1_ep,
w2_ep,
w1_scale_ep,
w2_scale_ep,
config.num_experts,
use_fp8_dispatch,
)
torch.testing.assert_close(
torch_combined,
deepep_combined,
atol=6e-2,
rtol=6e-2,
)
MNKs = [
(1, 128, 128),
(2, 128, 512),
(3, 1024, 2048),
(32, 128, 1024),
(45, 512, 2048),
(64, 1024, 1024),
(222, 1024, 2048),
]
DTYPES = [torch.bfloat16, torch.float8_e4m3fn]
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("mnk", MNKs)
@pytest.mark.parametrize("num_experts", [32])
@pytest.mark.parametrize("topk", [6])
@pytest.mark.parametrize("world_dp_size", [(2, 1)])
@requires_deep_ep
def test_deep_ep_moe(dtype: torch.dtype, mnk: tuple[int, int, int],
num_experts: int, topk: int, world_dp_size: tuple[int,
int]):
low_latency_mode = False
use_fp8_dispatch = False
m, n, k = mnk
current_platform.seed_everything(7)
world_size, dp_size = world_dp_size
config = TestConfig(dtype=dtype,
topk=topk,
m=m,
k=k,
n=n,
num_experts=num_experts)
w1, w2, w1_scale, w2_scale = make_weights(num_experts, n, k, dtype)
parallel_launch(world_size, _deep_ep_moe, low_latency_mode, dp_size,
config, w1, w2, w1_scale, w2_scale, use_fp8_dispatch)
MNKs = [
(1, 128, 2560),
(2, 128, 2560),
(3, 1024, 2560),
(32, 128, 2560),
(45, 512, 2560),
(64, 1024, 2560),
(222, 1024, 2560),
]
DTYPES = [torch.float8_e4m3fn, torch.bfloat16]
USE_FP8_DISPATCH = [True, False]
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("mnk", MNKs)
@pytest.mark.parametrize("num_experts", [32])
@pytest.mark.parametrize("topk", [6])
@pytest.mark.parametrize("world_dp_size", [(2, 1)])
@pytest.mark.parametrize("use_fp8_dispatch", USE_FP8_DISPATCH)
@requires_deep_ep
def test_low_latency_deep_ep_moe(dtype: torch.dtype, mnk: tuple[int, int, int],
num_experts: int, topk: int,
world_dp_size: tuple[int, int],
use_fp8_dispatch: bool):
low_latency_mode = True
m, n, k = mnk
if (low_latency_mode
and k not in DeepEPLLPrepareAndFinalize.SUPPORTED_HIDDEN_SIZES):
pytest.skip(
f"Skipping test as hidden size {k} is not in list of supported "
f"hidden sizes {DeepEPLLPrepareAndFinalize.SUPPORTED_HIDDEN_SIZES}"
)
current_platform.seed_everything(7)
world_size, dp_size = world_dp_size
config = TestConfig(dtype=dtype,
topk=topk,
m=m,
k=k,
n=n,
num_experts=num_experts)
w1, w2, w1_scale, w2_scale = make_weights(num_experts, n, k, dtype)
parallel_launch(world_size, _deep_ep_moe, low_latency_mode, dp_size,
config, w1, w2, w1_scale, w2_scale, use_fp8_dispatch)