mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2026-04-23 00:17:25 +08:00
xiaoxiaohehe001
00a01ae024
* [BugFix] support redundant expert for eplb * support redundant expert for eplb * support redundant expert for eplb * update * fix ci eplb
504 lines
16 KiB
Python
504 lines
16 KiB
Python
# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
from __future__ import annotations
|
|
|
|
from types import SimpleNamespace
|
|
|
|
import paddle
|
|
import pytest
|
|
|
|
from fastdeploy.config import MoEPhase
|
|
from fastdeploy.model_executor.layers.moe import ep
|
|
|
|
|
|
class FakeConfig:
|
|
def __init__(self, nvl_base: int, rdma_base: int):
|
|
self.nvl_base = nvl_base
|
|
self.rdma_base = rdma_base
|
|
|
|
def get_nvl_buffer_size_hint(self, hidden_bytes: int, world_size: int) -> int:
|
|
return hidden_bytes + self.nvl_base + world_size
|
|
|
|
def get_rdma_buffer_size_hint(self, hidden_bytes: int, world_size: int) -> int:
|
|
return hidden_bytes + self.rdma_base + world_size
|
|
|
|
|
|
class FakeBuffer:
|
|
def __init__(self, group, nvl_bytes, rdma_bytes, low_latency_mode=False, num_qps_per_rank=None):
|
|
self.group = group
|
|
self.nvl_bytes = nvl_bytes
|
|
self.rdma_bytes = rdma_bytes
|
|
self.low_latency_mode = low_latency_mode
|
|
self.num_qps_per_rank = num_qps_per_rank
|
|
self.num_sms = None
|
|
self.cleaned = None
|
|
self.dispatch_args = None
|
|
self.combine_args = None
|
|
self.barrier_called = False
|
|
self._dispatch_hook_called = False
|
|
self._combine_handle = None
|
|
|
|
@classmethod
|
|
def get_dispatch_config(cls, _world_size):
|
|
return FakeConfig(1, 2)
|
|
|
|
@classmethod
|
|
def get_combine_config(cls, _world_size):
|
|
return FakeConfig(3, 4)
|
|
|
|
@classmethod
|
|
def get_low_latency_rdma_size_hint(cls, *_args):
|
|
return 1000
|
|
|
|
@classmethod
|
|
def get_low_latency_rdma_size_hint_two_stage(cls, *_args):
|
|
return 2000
|
|
|
|
@classmethod
|
|
def get_low_latency_nvl_size_hint_two_stage(cls, *_args):
|
|
return 3000
|
|
|
|
def set_num_sms(self, num_sms: int):
|
|
self.num_sms = num_sms
|
|
|
|
def clean_low_latency_buffer(self, *args):
|
|
self.cleaned = ("single", args)
|
|
|
|
def clean_low_latency_two_stage_buffer(self, *args):
|
|
self.cleaned = ("two_stage", args)
|
|
|
|
def barrier_all(self):
|
|
self.barrier_called = True
|
|
|
|
def get_dispatch_layout(self, *args, **kwargs):
|
|
num_tokens_per_rank = paddle.to_tensor([1], dtype="int32")
|
|
num_tokens_per_rdma_rank = paddle.to_tensor([1], dtype="int32")
|
|
num_tokens_per_expert = paddle.to_tensor([1], dtype="int32")
|
|
is_token_in_rank = paddle.to_tensor([1], dtype="bool")
|
|
event = "dispatch_event"
|
|
return (
|
|
num_tokens_per_rank,
|
|
num_tokens_per_rdma_rank,
|
|
num_tokens_per_expert,
|
|
is_token_in_rank,
|
|
event,
|
|
)
|
|
|
|
def dispatch(self, **kwargs):
|
|
self.dispatch_args = kwargs
|
|
return "dispatch_result"
|
|
|
|
def combine(self, **kwargs):
|
|
self.combine_args = kwargs
|
|
return "combined", None, "combine_event"
|
|
|
|
def low_latency_dispatch(self, *_args, **_kwargs):
|
|
def _dispatch_hook():
|
|
self._dispatch_hook_called = True
|
|
|
|
return "packed", "count", ("handle",), None, _dispatch_hook
|
|
|
|
def low_latency_dispatch_two_stage(self, *_args, **_kwargs):
|
|
def _dispatch_hook():
|
|
self._dispatch_hook_called = True
|
|
|
|
return "packed", "count", None, ("handle",), None, _dispatch_hook
|
|
|
|
def low_latency_combine(self, *_args, **_kwargs):
|
|
handle = _args[3]
|
|
self._combine_handle = handle
|
|
return "combined", None, None
|
|
|
|
def low_latency_combine_two_stage(self, *_args, **_kwargs):
|
|
handle = _args[3]
|
|
self._combine_handle = handle
|
|
return "combined", None, None
|
|
|
|
|
|
class FakeDeepEP:
|
|
Buffer = FakeBuffer
|
|
|
|
|
|
def _patch_deep_ep(monkeypatch):
|
|
monkeypatch.setattr(ep, "deep_ep", FakeDeepEP, raising=False)
|
|
|
|
|
|
def test_deepep_buffer_manager_calls_engine(monkeypatch):
|
|
class FakeEngine:
|
|
def __init__(self):
|
|
self.cleared = False
|
|
self.created = False
|
|
|
|
def clear_deep_ep_buffer(self):
|
|
self.cleared = True
|
|
|
|
def create_deep_ep_buffer(self):
|
|
self.created = True
|
|
|
|
engine = FakeEngine()
|
|
ep.DeepEPBufferManager.set_engine(engine)
|
|
ep.DeepEPBufferManager.clear_buffer()
|
|
ep.DeepEPBufferManager.recreate_buffer()
|
|
|
|
assert engine.cleared is True
|
|
assert engine.created is True
|
|
|
|
|
|
def test_deepep_buffer_mixed_two_stage_buffers(monkeypatch):
|
|
_patch_deep_ep(monkeypatch)
|
|
group = SimpleNamespace(world_size=2)
|
|
buffer = ep.DeepEPBuffer(
|
|
group=group,
|
|
hidden_size=4,
|
|
num_experts=8,
|
|
ep_size=2,
|
|
num_max_dispatch_tokens_per_rank=2,
|
|
splitwise_role="mixed",
|
|
moe_phase=MoEPhase("prefill"),
|
|
use_internode_ll_two_stage=True,
|
|
top_k=2,
|
|
)
|
|
|
|
assert buffer.num_nvl_bytes >= 3000
|
|
assert buffer.num_rdma_bytes >= 2000
|
|
|
|
buffer.create_buffer()
|
|
assert buffer.deepep_buffer.low_latency_mode is True
|
|
assert buffer.deepep_buffer.num_qps_per_rank == 24
|
|
assert buffer.deepep_buffer.num_sms == 14
|
|
|
|
buffer.clean_low_latency_buffer()
|
|
assert buffer.deepep_buffer.cleaned[0] == "two_stage"
|
|
|
|
buffer.clear_buffer()
|
|
assert buffer.deepep_buffer is None
|
|
|
|
|
|
def test_deepep_buffer_decode_low_latency_buffer(monkeypatch):
|
|
_patch_deep_ep(monkeypatch)
|
|
group = SimpleNamespace(world_size=4)
|
|
buffer = ep.DeepEPBuffer(
|
|
group=group,
|
|
hidden_size=8,
|
|
num_experts=16,
|
|
ep_size=16,
|
|
num_max_dispatch_tokens_per_rank=1,
|
|
splitwise_role="prefill",
|
|
moe_phase=MoEPhase("decode"),
|
|
use_internode_ll_two_stage=False,
|
|
top_k=4,
|
|
)
|
|
|
|
buffer.create_buffer()
|
|
assert buffer.deepep_buffer.low_latency_mode is True
|
|
assert buffer.deepep_buffer.num_qps_per_rank == 2
|
|
|
|
buffer.clean_low_latency_buffer()
|
|
assert buffer.deepep_buffer.cleaned[0] == "single"
|
|
|
|
|
|
def test_deepep_buffer_prefill_and_invalid_phase(monkeypatch):
|
|
_patch_deep_ep(monkeypatch)
|
|
group = SimpleNamespace(world_size=2)
|
|
buffer = ep.DeepEPBuffer(
|
|
group=group,
|
|
hidden_size=4,
|
|
num_experts=8,
|
|
ep_size=2,
|
|
num_max_dispatch_tokens_per_rank=2,
|
|
splitwise_role="prefill",
|
|
moe_phase=MoEPhase("prefill"),
|
|
use_internode_ll_two_stage=False,
|
|
top_k=2,
|
|
)
|
|
|
|
buffer.create_buffer()
|
|
assert buffer.deepep_buffer.low_latency_mode is True
|
|
assert buffer.deepep_buffer.num_qps_per_rank == 24
|
|
|
|
buffer.barrier_all()
|
|
assert buffer.deepep_buffer.barrier_called is True
|
|
assert buffer.get_buffer() is buffer.deepep_buffer
|
|
|
|
invalid_phase_buffer = ep.DeepEPBuffer(
|
|
group=group,
|
|
hidden_size=4,
|
|
num_experts=8,
|
|
ep_size=2,
|
|
num_max_dispatch_tokens_per_rank=2,
|
|
splitwise_role="prefill",
|
|
moe_phase=MoEPhase("unknown"),
|
|
use_internode_ll_two_stage=False,
|
|
top_k=2,
|
|
)
|
|
with pytest.raises(ValueError, match="Unknown generation phase"):
|
|
invalid_phase_buffer.create_buffer()
|
|
|
|
|
|
def test_deepep_engine_combine_rewrites_handle_and_errors(monkeypatch):
|
|
_patch_deep_ep(monkeypatch)
|
|
group = SimpleNamespace(world_size=1)
|
|
engine = ep.DeepEPEngine(
|
|
num_max_dispatch_tokens_per_rank=1,
|
|
hidden_size=4,
|
|
num_experts=2,
|
|
ep_size=1,
|
|
ep_rank=0,
|
|
splitwise_role="prefill",
|
|
moe_phase=MoEPhase("decode"),
|
|
group=group,
|
|
)
|
|
|
|
hidden_states = paddle.randn([1, 4], dtype="float32")
|
|
topk_idx = paddle.zeros([1, 1], dtype="int64")
|
|
topk_weights = paddle.ones([1, 1], dtype="float32")
|
|
handle = ("src", "layout", 4, 2)
|
|
engine.low_latency_combine(hidden_states, topk_idx, topk_weights, handle)
|
|
|
|
assert engine.deepep_engine._combine_handle == handle
|
|
assert len(engine.deepep_engine._combine_handle) == 4
|
|
|
|
engine.buffer.deepep_buffer = None
|
|
with pytest.raises(RuntimeError, match="DeepEP buffer not initialized"):
|
|
engine.low_latency_dispatch(hidden_states, topk_idx, None)
|
|
|
|
|
|
def test_prefill_runner_dispatch_and_combine(monkeypatch):
|
|
_patch_deep_ep(monkeypatch)
|
|
|
|
class FakeEngine:
|
|
def __init__(self, *args, **kwargs):
|
|
self.async_finish = True
|
|
self.ep_config = "ep_config"
|
|
self.deepep_engine = FakeBuffer(SimpleNamespace(world_size=1), 1, 1)
|
|
|
|
def clean_low_latency_buffer(self):
|
|
self.deepep_engine.cleaned = ("single", ())
|
|
|
|
def clear_deep_ep_buffer(self):
|
|
self.deepep_engine = None
|
|
|
|
def create_deep_ep_buffer(self):
|
|
self.deepep_engine = FakeBuffer(SimpleNamespace(world_size=1), 1, 1)
|
|
|
|
monkeypatch.setattr(ep, "DeepEPEngine", FakeEngine)
|
|
|
|
ep.EPPrefillRunner.set_allocate_on_comm_stream(True)
|
|
ep.EPPrefillRunner.set_allocate_on_comm_stream(True)
|
|
|
|
runner = ep.EPPrefillRunner(
|
|
top_k=2,
|
|
hidden_size=4,
|
|
num_experts=2,
|
|
splitwise_role="prefill",
|
|
num_max_dispatch_tokens_per_rank=1,
|
|
)
|
|
x = paddle.randn([2, 4], dtype="float32")
|
|
topk_idx = paddle.zeros([2, 2], dtype="int64")
|
|
topk_weights = paddle.ones([2, 2], dtype="float32")
|
|
|
|
dispatch_result = runner.dispatch(x, topk_idx, topk_weights, expert_alignment=8)
|
|
assert dispatch_result == "dispatch_result"
|
|
assert runner.ep_engine.deepep_engine.dispatch_args["allocate_on_comm_stream"] is True
|
|
assert runner.ep_engine.deepep_engine.dispatch_args["expert_alignment"] == 8
|
|
|
|
combined, event = runner.combine(x, ("handle",), topk_weights)
|
|
assert combined == "combined"
|
|
assert event == "combine_event"
|
|
|
|
|
|
def test_decoder_runner_dispatch_and_combine_hooks(monkeypatch):
|
|
_patch_deep_ep(monkeypatch)
|
|
|
|
class FakeEngine:
|
|
def __init__(self, *args, **kwargs):
|
|
self.dispatch_called = False
|
|
self.combine_called = False
|
|
self.two_stage_dispatch_called = False
|
|
self.two_stage_combine_called = False
|
|
|
|
def low_latency_dispatch(self, *args, **kwargs):
|
|
self.dispatch_called = True
|
|
return "recv", "count", ("handle",), lambda: self._mark_hook("dispatch")
|
|
|
|
def low_latency_dispatch_two_stage(self, *args, **kwargs):
|
|
self.two_stage_dispatch_called = True
|
|
return "recv", "count", ("handle",), lambda: self._mark_hook("dispatch")
|
|
|
|
def low_latency_combine(self, *args, **kwargs):
|
|
self.combine_called = True
|
|
return "combined", lambda: self._mark_hook("combine")
|
|
|
|
def low_latency_combine_two_stage(self, *args, **kwargs):
|
|
self.two_stage_combine_called = True
|
|
return "combined", lambda: self._mark_hook("combine")
|
|
|
|
def _mark_hook(self, name):
|
|
setattr(self, f"{name}_hook_called", True)
|
|
|
|
monkeypatch.setattr(ep, "DeepEPEngine", FakeEngine)
|
|
|
|
runner = ep.EPDecoderRunner(
|
|
top_k=2,
|
|
hidden_size=4,
|
|
num_experts=2,
|
|
splitwise_role="prefill",
|
|
num_max_dispatch_tokens_per_rank=1,
|
|
)
|
|
x = paddle.randn([1, 4], dtype="float32")
|
|
topk_idx = paddle.zeros([1, 1], dtype="int64")
|
|
topk_weights = paddle.ones([1, 1], dtype="float32")
|
|
|
|
recv_hidden, recv_count, handle = runner.dispatch(x, topk_idx, topk_weights)
|
|
assert recv_hidden == "recv"
|
|
assert recv_count == "count"
|
|
assert handle == ("handle",)
|
|
assert runner.ep_engine.dispatch_called is True
|
|
assert runner.ep_engine.dispatch_hook_called is True
|
|
|
|
combined = runner.combine(x, topk_idx, topk_weights, handle)
|
|
assert combined == "combined"
|
|
assert runner.ep_engine.combine_called is True
|
|
assert runner.ep_engine.combine_hook_called is True
|
|
|
|
runner_two_stage = ep.EPDecoderRunner(
|
|
top_k=2,
|
|
hidden_size=4,
|
|
num_experts=2,
|
|
splitwise_role="prefill",
|
|
num_max_dispatch_tokens_per_rank=1,
|
|
use_internode_ll_two_stage=True,
|
|
)
|
|
recv_hidden, recv_count, handle = runner_two_stage.dispatch(
|
|
x, topk_idx, topk_weights, expertwise_scale=None, use_fp8=True
|
|
)
|
|
assert recv_hidden == "recv"
|
|
assert runner_two_stage.ep_engine.two_stage_dispatch_called is True
|
|
|
|
combined = runner_two_stage.combine(x, topk_idx, topk_weights, handle, quant_group_size=64)
|
|
assert combined == "combined"
|
|
assert runner_two_stage.ep_engine.two_stage_combine_called is True
|
|
|
|
|
|
def test_eprunner_moe_select_noaux_tc_without_redundant(monkeypatch):
|
|
_patch_deep_ep(monkeypatch)
|
|
|
|
def fake_get_moe_scores(*_args, **_kwargs):
|
|
return "score", paddle.to_tensor([[0.5]]), paddle.to_tensor([[1]], dtype="int64")
|
|
|
|
from fastdeploy.model_executor.layers.moe import moe as moe_module
|
|
|
|
monkeypatch.setattr(moe_module, "get_moe_scores", fake_get_moe_scores, raising=True)
|
|
|
|
runner = ep.EPPrefillRunner(
|
|
top_k=2,
|
|
hidden_size=4,
|
|
num_experts=2,
|
|
splitwise_role="prefill",
|
|
num_max_dispatch_tokens_per_rank=1,
|
|
)
|
|
|
|
layer = SimpleNamespace(
|
|
redundant_table_manger=None,
|
|
topk_method="noaux_tc",
|
|
n_group=1,
|
|
topk_group=1,
|
|
top_k=2,
|
|
routed_scaling_factor=1.0,
|
|
gate_correction_bias=None,
|
|
renormalize=False,
|
|
)
|
|
gate_out = paddle.randn([1, 4], dtype="float32")
|
|
|
|
topk_idx, topk_weights = runner.moe_select(layer, gate_out)
|
|
assert list(topk_idx.shape) == [1, 1]
|
|
assert list(topk_weights.shape) == [1, 1]
|
|
assert paddle.allclose(topk_idx, paddle.to_tensor([[1]], dtype="int64"))
|
|
assert paddle.allclose(topk_weights, paddle.to_tensor([[0.5]]))
|
|
|
|
|
|
def test_eprunner_moe_select_redundant_and_topk(monkeypatch):
|
|
_patch_deep_ep(monkeypatch)
|
|
|
|
def fake_redundant_topk_select(**_kwargs):
|
|
return paddle.to_tensor([[2]], dtype="int64"), paddle.to_tensor([[0.25]])
|
|
|
|
from fastdeploy.model_executor.ops import gpu as gpu_ops
|
|
|
|
monkeypatch.setattr(gpu_ops, "moe_redundant_topk_select", fake_redundant_topk_select, raising=True)
|
|
|
|
runner = ep.EPPrefillRunner(
|
|
top_k=2,
|
|
hidden_size=4,
|
|
num_experts=2,
|
|
splitwise_role="prefill",
|
|
num_max_dispatch_tokens_per_rank=1,
|
|
)
|
|
|
|
class FakeRedundantTableManager:
|
|
def get_ep_rank_to_expert_id_list_by_layer(self, _layer_idx):
|
|
return [0], paddle.to_tensor([0], dtype="int64"), [1], [1]
|
|
|
|
layer = SimpleNamespace(
|
|
redundant_table_manger=FakeRedundantTableManager(),
|
|
layer_idx=0,
|
|
topk_method="aux",
|
|
n_group=1,
|
|
topk_group=1,
|
|
top_k=2,
|
|
routed_scaling_factor=1.0,
|
|
gate_correction_bias=None,
|
|
fd_config=SimpleNamespace(eplb_config=SimpleNamespace(redundant_experts_num=0)),
|
|
)
|
|
gate_out = paddle.randn([1, 4], dtype="float32")
|
|
|
|
topk_idx, topk_weights = runner.moe_select(layer, gate_out)
|
|
assert list(topk_idx.shape) == [1, 1]
|
|
assert list(topk_weights.shape) == [1, 1]
|
|
assert paddle.allclose(topk_idx, paddle.to_tensor([[2]], dtype="int64"))
|
|
assert paddle.allclose(topk_weights, paddle.to_tensor([[0.25]]))
|
|
|
|
|
|
def test_eprunner_moe_select_topk_without_redundant(monkeypatch):
|
|
_patch_deep_ep(monkeypatch)
|
|
|
|
def fake_topk_select(*_args, **_kwargs):
|
|
return paddle.to_tensor([[3]], dtype="int64"), paddle.to_tensor([[0.75]])
|
|
|
|
from fastdeploy.model_executor.ops import gpu as gpu_ops
|
|
|
|
monkeypatch.setattr(gpu_ops, "moe_topk_select", fake_topk_select, raising=True)
|
|
|
|
runner = ep.EPPrefillRunner(
|
|
top_k=2,
|
|
hidden_size=4,
|
|
num_experts=2,
|
|
splitwise_role="prefill",
|
|
num_max_dispatch_tokens_per_rank=1,
|
|
)
|
|
|
|
layer = SimpleNamespace(
|
|
redundant_table_manger=None,
|
|
topk_method="aux",
|
|
gate_correction_bias=None,
|
|
)
|
|
gate_out = paddle.randn([1, 4], dtype="float32")
|
|
|
|
topk_idx, topk_weights = runner.moe_select(layer, gate_out)
|
|
assert list(topk_idx.shape) == [1, 1]
|
|
assert list(topk_weights.shape) == [1, 1]
|
|
assert paddle.allclose(topk_idx, paddle.to_tensor([[3]], dtype="int64"))
|
|
assert paddle.allclose(topk_weights, paddle.to_tensor([[0.75]]))
|