EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Expert Parallelism (EP) Support for DeepSeek V3/R1 (#3602)

Browse Source 
Co-authored-by: laixin <xielx@shanghaitech.edu.cn>
Co-authored-by: HandH1998 <1335248067@qq.com>
Co-authored-by: laixin <q865809639@gmail.com>
This commit is contained in:
lukec
2025-02-26 18:29:37 +08:00
committed by GitHub
parent 3dc9ff3ce8
commit 21463e321a
3 changed files with 548 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
python/sglang/srt/layers/moe/ep_moe/kernels.py
View File

@@ -1,10 +1,17 @@
import logging
from typing import Optional
from typing import List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
_is_cuda = torch.cuda.is_available() and torch.version.cuda
if _is_cuda:
from sglang.srt.layers.quantization.fp8_kernel import (
sglang_per_token_group_quant_fp8,
)
logger = logging.getLogger(__name__)
@@ -218,12 +225,19 @@ def grouped_gemm_triton_kernel(
seg_indptr,
weight_indices,
m_num_tiles_indptr,
use_fp8_w8a8,
scale_a,
scale_b,
use_fp8_w8a8: tl.constexpr,
group_n: tl.constexpr,
group_k: tl.constexpr,
a_stride_0: tl.constexpr,
b_stride_0: tl.constexpr,
b_stride_1: tl.constexpr,
as_stride_0: tl.constexpr,
as_stride_1: tl.constexpr,
bs_stride_0: tl.constexpr,
bs_stride_2: tl.constexpr,
bs_stride_1: tl.constexpr,
BLOCK_SIZE_M: tl.constexpr,
BLOCK_SIZE_N: tl.constexpr,
BLOCK_SIZE_K: tl.constexpr,
@@ -260,6 +274,12 @@ def grouped_gemm_triton_kernel(
+ (n_range_start + offs_bn[:, None]) * b_stride_1
+ offs_k[None, :]
)
if group_k > 0 and group_n > 0:
a_scale_ptrs = scale_a + (m_range_start + offs_am[:, None]) * as_stride_0
offs_bsn = (n_range_start + offs_bn) // group_n
b_scale_ptrs = scale_b + (expert_id * bs_stride_0) + offs_bsn * bs_stride_1
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
a_tile = tl.load(
@@ -268,14 +288,23 @@ def grouped_gemm_triton_kernel(
b_tile = tl.load(
b_ptr, mask=offs_k[None, :] < (K - k * BLOCK_SIZE_K), other=0.0
)
accumulator = tl.dot(a_tile, b_tile.T, accumulator)
if group_k > 0 and group_n > 0:
k_start = k * BLOCK_SIZE_K
offs_ks = k_start // group_k
a_scale = tl.load(a_scale_ptrs + offs_ks * as_stride_1)
b_scale = tl.load(b_scale_ptrs + offs_ks * bs_stride_2)
accumulator += tl.dot(a_tile, b_tile.T) * a_scale * b_scale[None, :]
else:
accumulator = tl.dot(a_tile, b_tile.T, accumulator)
a_ptr += BLOCK_SIZE_K
b_ptr += BLOCK_SIZE_K
if use_fp8_w8a8:
if use_fp8_w8a8 and not (group_k > 0 and group_n > 0):
scale_a_value = tl.load(scale_a + expert_id)
scale_b_value = tl.load(scale_b + expert_id)
accumulator *= scale_a_value * scale_b_value
c_tile = accumulator.to(c_dtype)
offs_cm = m_range_start + tl.arange(0, BLOCK_SIZE_M)
@@ -307,14 +336,29 @@ def grouped_gemm_triton(
use_fp8_w8a8: bool = False,
scale_a: torch.Tensor = None,
scale_b: torch.Tensor = None,
block_shape: Optional[List[int]] = None,
):
assert weight_column_major == True # TODO: more
if use_fp8_w8a8:
if use_fp8_w8a8 and block_shape is None:
assert scale_a is not None and scale_b is not None
if block_shape is not None:
assert len(block_shape) == 2
block_n, block_k = block_shape[0], block_shape[1]
if _is_cuda:
a, scale_a = sglang_per_token_group_quant_fp8(a, block_k)
else:
a, scale_a = per_token_group_quant_fp8(a, block_k)
assert triton.cdiv(a.shape[-1], block_k) == scale_a.shape[-1]
assert triton.cdiv(b.shape[-2], block_n) == scale_b.shape[-2]
assert triton.cdiv(b.shape[-1], block_k) == scale_b.shape[-1]
# TODO: adjust config or tune kernel
# Reduce block size to prevent L40 shared memory overflow.
config = {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
}
@@ -338,12 +382,19 @@ def grouped_gemm_triton(
seg_indptr,
weight_indices,
m_num_tiles_indptr,
use_fp8_w8a8,
scale_a,
scale_b,
use_fp8_w8a8,
0 if block_shape is None else block_shape[0],
0 if block_shape is None else block_shape[1],
a.stride(0),
b.stride(0),
b.stride(1),
scale_a.stride(0) if scale_a is not None and scale_a.ndim == 2 else 0,
scale_a.stride(1) if scale_a is not None and scale_a.ndim == 2 else 0,
scale_b.stride(0) if scale_b is not None and scale_b.ndim >= 2 else 0,
scale_b.stride(2) if scale_b is not None and scale_b.ndim == 3 else 0,
scale_b.stride(1) if scale_b is not None and scale_b.ndim >= 2 else 0,
**config,
)
return c

155
python/sglang/srt/layers/moe/ep_moe/layer.py
View File

@@ -17,6 +17,7 @@ from sglang.srt.layers.moe.ep_moe.kernels import (
run_moe_ep_preproess,
silu_and_mul_triton_kernel,
)
from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoEMethodBase
from sglang.srt.layers.moe.topk import select_experts
from sglang.srt.layers.quantization.base_config import (
@@ -61,6 +62,7 @@ class GroupedGemmRunner(torch.nn.Module):
use_fp8_w8a8: bool = False,
scale_a: torch.Tensor = None,
scale_b: torch.Tensor = None,
block_shape: Optional[List[int]] = None,
):
if self.use_flashinfer:
# TODO: flashinfer
@@ -87,6 +89,7 @@ class GroupedGemmRunner(torch.nn.Module):
use_fp8_w8a8,
scale_a,
scale_b,
block_shape=block_shape,
)
return c
@@ -147,12 +150,20 @@ class EPMoE(torch.nn.Module):
if quant_config is None:
self.quant_method: Optional[QuantizeMethodBase] = UnquantizedEPMoEMethod()
self.use_fp8_w8a8 = False
self.use_block_quant = False
self.block_shape = None
self.activation_scheme = None
else:
self.quant_method: Optional[QuantizeMethodBase] = Fp8EPMoEMethod(
quant_config
)
self.use_fp8_w8a8 = True
self.use_block_quant = getattr(self.quant_method, "block_quant", False)
self.block_shape = (
self.quant_method.quant_config.weight_block_size
if self.use_block_quant
else None
)
self.fp8_dtype = torch.float8_e4m3fn
self.activation_scheme = quant_config.activation_scheme
@@ -173,7 +184,8 @@ class EPMoE(torch.nn.Module):
if self.grouped_gemm_runner is None:
self.grouped_gemm_runner = GroupedGemmRunner(
hidden_states.device, use_flashinfer=False # TODO: use flashinfer
hidden_states.device,
use_flashinfer=False, # TODO: use flashinfer
)
topk_weights, topk_ids = select_experts(
@@ -195,9 +207,13 @@ class EPMoE(torch.nn.Module):
gateup_input = torch.empty(
(int(hidden_states.shape[0] * self.top_k), hidden_states.shape[1]),
device=hidden_states.device,
dtype=self.fp8_dtype if self.use_fp8_w8a8 else hidden_states.dtype,
dtype=(
self.fp8_dtype
if (self.use_fp8_w8a8 and not self.use_block_quant)
else hidden_states.dtype
),
)
if self.activation_scheme == "dynamic":
if self.activation_scheme == "dynamic" and not self.use_block_quant:
max_value = (
torch.max(hidden_states)
.repeat(self.num_experts_per_partition)
@@ -243,7 +259,12 @@ class EPMoE(torch.nn.Module):
weight_indices=weight_indices_cur_rank,
use_fp8_w8a8=self.use_fp8_w8a8,
scale_a=self.w13_input_scale,
scale_b=self.w13_weight_scale,
scale_b=(
self.w13_weight_scale_inv
if self.use_block_quant
else self.w13_weight_scale
),
block_shape=self.block_shape,
)
# Act
@@ -251,9 +272,13 @@ class EPMoE(torch.nn.Module):
gateup_output.shape[0],
gateup_output.shape[1] // 2,
device=gateup_output.device,
dtype=self.fp8_dtype if self.use_fp8_w8a8 else hidden_states.dtype,
dtype=(
self.fp8_dtype
if (self.use_fp8_w8a8 and not self.use_block_quant)
else hidden_states.dtype
),
)
if self.w2_input_scale is None:
if self.w2_input_scale is None and not self.use_block_quant:
self.w2_input_scale = torch.ones(
self.num_experts_per_partition,
dtype=torch.float32,
@@ -291,7 +316,12 @@ class EPMoE(torch.nn.Module):
weight_indices=weight_indices_cur_rank,
use_fp8_w8a8=self.use_fp8_w8a8,
scale_a=self.w2_input_scale,
scale_b=self.w2_weight_scale,
scale_b=(
self.w2_weight_scale_inv
if self.use_block_quant
else self.w2_weight_scale
),
block_shape=self.block_shape,
)
# PostReorder
@@ -358,7 +388,11 @@ class EPMoE(torch.nn.Module):
# Special case for fp8 scales.
if "scale" in weight_name:
self._load_fp8_scale(
param.data, loaded_weight, weight_name, shard_id, expert_id
param.data,
loaded_weight,
weight_name,
shard_id,
expert_id,
)
return
@@ -395,18 +429,33 @@ class EPMoE(torch.nn.Module):
param_data[expert_id] = loaded_weight
# Weight scales
elif "weight_scale" in weight_name:
if self.use_block_quant:
block_n, block_k = self.block_shape[0], self.block_shape[1]
if shard_id == "w1":
param_data[expert_id][
: (self.intermediate_size + block_n - 1) // block_n, :
] = loaded_weight
elif shard_id == "w3":
param_data[expert_id][
(self.intermediate_size + block_n - 1) // block_n :, :
] = loaded_weight
else: # w2
param_data[expert_id] = loaded_weight
# If we are in merged column case (gate_up_proj)
if shard_id in ("w1", "w3"):
# We have to keep the weight scales of w1 and w3 because
# we need to re-quantize w1/w3 weights after weight loading.
idx = 0 if shard_id == "w1" else 1
param_data[expert_id][idx] = loaded_weight
# If we are in the row parallel case (down_proj)
else:
param_data[expert_id] = loaded_weight
if shard_id in ("w1", "w3"):
# We have to keep the weight scales of w1 and w3 because
# we need to re-quantize w1/w3 weights after weight loading.
idx = 0 if shard_id == "w1" else 1
param_data[expert_id][idx] = loaded_weight
# If we are in the row parallel case (down_proj)
else:
param_data[expert_id] = loaded_weight
class UnquantizedEPMoEMethod(FusedMoEMethodBase, CustomOp):
def create_weights(
self,
layer: torch.nn.Module,
@@ -498,6 +547,7 @@ class Fp8EPMoEMethod(Fp8MoEMethod):
def __init__(self, quant_config: Fp8Config):
self.quant_config = quant_config
self.block_quant = self.quant_config.weight_block_size is not None
def create_weights(
self,
@@ -512,6 +562,29 @@ class Fp8EPMoEMethod(Fp8MoEMethod):
if self.quant_config.is_checkpoint_fp8_serialized:
params_dtype = torch.float8_e4m3fn
tp_size = get_tensor_model_parallel_world_size()
if self.block_quant:
block_n, block_k = (
self.quant_config.weight_block_size[0],
self.quant_config.weight_block_size[1],
)
# NOTE(HandH1998): To ensure proper alignment of the block-wise quantization scales, the output_size of the weights for both the gate and up layers must be divisible by block_n.
# Required by collum parallel or enabling merged weights
if intermediate_size % block_n != 0:
raise ValueError(
f"The output_size of gate's and up's weight = "
f"{intermediate_size} is not divisible by "
f"weight quantization block_n = {block_n}."
)
if tp_size > 1:
# Required by row parallel
if intermediate_size % block_k != 0:
raise ValueError(
f"The input_size of down's weight = "
f"{intermediate_size} is not divisible by "
f"weight quantization block_k = {block_k}."
)
# WEIGHTS
w13_weight = torch.nn.Parameter(
torch.empty(
@@ -538,21 +611,49 @@ class Fp8EPMoEMethod(Fp8MoEMethod):
set_weight_attrs(w2_weight, extra_weight_attrs)
# WEIGHT_SCALES
# Allocate 2 scales for w1 and w3 respectively.
w13_weight_scale = torch.nn.Parameter(
torch.ones(num_experts_per_partition, 2, dtype=torch.float32),
requires_grad=False,
)
layer.register_parameter("w13_weight_scale", w13_weight_scale)
if self.block_quant:
w13_weight_scale = torch.nn.Parameter(
torch.ones(
num_experts_per_partition,
2 * ((intermediate_size + block_n - 1) // block_n),
(hidden_size + block_k - 1) // block_k,
dtype=torch.float32,
),
requires_grad=False,
)
w2_weight_scale = torch.nn.Parameter(
torch.ones(
num_experts_per_partition,
(hidden_size + block_n - 1) // block_n,
(intermediate_size + block_k - 1) // block_k,
dtype=torch.float32,
),
requires_grad=False,
)
layer.register_parameter("w13_weight_scale_inv", w13_weight_scale)
layer.register_parameter("w2_weight_scale_inv", w2_weight_scale)
assert self.quant_config.activation_scheme == "dynamic"
else:
# WEIGHT_SCALES
# Allocate 2 scales for w1 and w3 respectively.
w13_weight_scale = torch.nn.Parameter(
torch.ones(num_experts_per_partition, 2, dtype=torch.float32),
requires_grad=False,
)
layer.register_parameter("w13_weight_scale", w13_weight_scale)
w2_weight_scale = torch.nn.Parameter(
torch.ones(num_experts_per_partition, dtype=torch.float32),
requires_grad=False,
)
layer.register_parameter("w2_weight_scale", w2_weight_scale)
w2_weight_scale = torch.nn.Parameter(
torch.ones(num_experts_per_partition, dtype=torch.float32),
requires_grad=False,
)
layer.register_parameter("w2_weight_scale", w2_weight_scale)
# Add the quantization method used (per tensor/grouped/channel)
# to ensure the weight scales are loaded in properly
extra_weight_attrs.update({"quant_method": "tensor"})
extra_weight_attrs.update(
{"quant_method": FusedMoeWeightScaleSupported.BLOCK.value}
if self.block_quant
else {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
)
# If loading fp8 checkpoint, pass the weight loaders.
# If loading an fp16 checkpoint, do not (we will quantize in
# process_weights_after_loading()