Files
sglang/python/sglang/srt/layers/moe/cutlass_moe.py

374 lines
14 KiB
Python
Executable File

"""CUTLASS based Fused MoE kernels."""
import functools
import json
import logging
import os
from typing import Any, Callable, Dict, List, Optional, Tuple
import torch
from sglang.srt.layers.moe.cutlass_moe_params import CutlassMoEParams
from sglang.srt.layers.utils import is_sm90_supported, is_sm100_supported
from sglang.srt.utils import is_cuda
_is_cuda = is_cuda()
if _is_cuda:
import sgl_kernel
from sgl_kernel import (
apply_shuffle_mul_sum,
cutlass_fp4_group_mm,
fp8_blockwise_scaled_grouped_mm,
prepare_moe_input,
scaled_fp4_experts_quant,
shuffle_rows,
silu_and_mul,
)
def cutlass_fused_experts_fp8(
a: torch.Tensor,
w1_q: torch.Tensor,
w2_q: torch.Tensor,
w1_scale: torch.Tensor,
w2_scale: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
a1_strides: torch.Tensor,
c1_strides: torch.Tensor,
a2_strides: torch.Tensor,
c2_strides: torch.Tensor,
workspace: torch.Tensor,
a_ptrs: torch.Tensor,
b_ptrs: torch.Tensor,
out_ptrs: torch.Tensor,
a_scales_ptrs: torch.Tensor,
b_scales_ptrs: torch.Tensor,
expert_offsets: torch.Tensor,
problem_sizes1: torch.Tensor,
problem_sizes2: torch.Tensor,
use_fp8_blockscale: bool = True,
) -> torch.Tensor:
"""Performs Fused MoE computation using CUTLASS-like kernels with FP8 weights and activations.
This function implements a Mixture of Experts (MoE) layer with a SwiGLU/SiLU
activation, leveraging custom kernels likely derived from CUTLASS principles
for grouped matrix multiplication (`fp8_blockwise_scaled_grouped_mm`) and
data preparation (`prepare_moe_input`, `silu_and_mul`).
It handles per-token routing, quantizes input activations to FP8 with
per-token scales, performs the expert computations using FP8 GEMMs with
pre-quantized FP8 weights (per-block scales), applies the SiLU activation,
and combines the results weighted by the router scores.
Args:
a (torch.Tensor): Input activations. Shape: `(m, k)`, where `m` is the total
number of tokens and `k` is the hidden size. Expected dtype: `torch.half`
or `torch.bfloat16`.
w1_q (torch.Tensor): Pre-quantized FP8 weight tensor for the first GEMM
(up-projection part of SwiGLU). Expected shape: `(E, k, n*2)`, where
`E` is the number of experts, `k` is the hidden size, and `n*2` is the
intermediate size (`I`). Expected dtype: `torch.float8_e4m3fn`.
Note: This shape implies weights are stored as (num_experts, hidden_size, intermediate_size).
w2_q (torch.Tensor): Pre-quantized FP8 weight tensor for the second GEMM
(down-projection). Expected shape: `(E, n, k)`, where `n` is half the
intermediate size (`I // 2`). Expected dtype: `torch.float8_e4m3fn`.
Note: This shape implies weights are stored as (num_experts, intermediate_size // 2, hidden_size).
w1_scale (torch.Tensor): Scales corresponding to `w1_q` (per-block scales).
Shape: `(E, num_blocks_n, num_blocks_k)`. Dtype: `torch.float32`.
w2_scale (torch.Tensor): Scales corresponding to `w2_q` (per-block scales).
Shape: `(E, num_blocks_k, num_blocks_n)`. Dtype: `torch.float32`.
topk_weights (torch.Tensor): Router weights for the selected top-k experts
for each token. Shape: `(m, topk)`. Dtype should ideally match `a`.
topk_ids (torch.Tensor): Indices of the selected top-k experts for each token.
Shape: `(m, topk)`. Dtype: `torch.int32`.
a1_strides (torch.Tensor): Stride information for the first GEMM's 'a' input.
Passed directly to the underlying kernel. Expected shape `(E,)`, dtype `torch.int64`.
Note: Its exact usage within `fp8_blockwise_scaled_grouped_mm` needs clarification
as it's passed as both a_stride and b_stride in the first call.
c1_strides (torch.Tensor): Stride information for the first GEMM's 'c' output.
Passed directly to the underlying kernel. Expected shape `(E,)`, dtype `torch.int64`.
a2_strides (torch.Tensor): Stride information for the second GEMM's 'a' input.
Passed directly to the underlying kernel. Expected shape `(E,)`, dtype `torch.int64`.
Note: Its exact usage within `fp8_blockwise_scaled_grouped_mm` needs clarification
as it's passed as both a_stride and b_stride in the second call.
c2_strides (torch.Tensor): Stride information for the second GEMM's 'c' output.
Passed directly to the underlying kernel. Expected shape `(E,)`, dtype `torch.int64`.
workspace (torch.Tensor): Reusable workspace for the underlying kernel.
a_ptrs (torch.Tensor): Pointers container for calculating offsets of the input activations for each expert.
b_ptrs (torch.Tensor): Pointers container for calculating offsets of the input weights for each expert.
out_ptrs (torch.Tensor): Pointers container for calculating offsets of the output activations for each expert.
a_scales_ptrs (torch.Tensor): Pointers container for calculating offsets of the input scales for each expert.
b_scales_ptrs (torch.Tensor): Pointers container for calculating offsets of the input scales for each expert.
use_fp8_blockscale (bool, optional): Flag indicating usage of FP8 with
block scaling. Currently, only `True` is supported. Defaults to `True`.
Returns:
torch.Tensor: The computed MoE layer output. Shape: `(m, k)`, dtype matches `a`.
Raises:
AssertionError: If input shapes, dtypes, or flags are inconsistent or unsupported.
NotImplementedError: If CUDA is not available or `sgl_kernel` is not properly installed.
"""
assert use_fp8_blockscale, "Only support fp8 blockscale for now"
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
assert w1_q.dtype == torch.float8_e4m3fn
assert w2_q.dtype == torch.float8_e4m3fn
assert a.shape[1] == w1_q.shape[1], "Hidden size mismatch w1"
assert w1_q.shape[2] == w2_q.shape[1] * 2, "Hidden size mismatch w2"
assert w1_q.shape[0] == w2_q.shape[0], "Expert number mismatch"
assert w1_q.shape[0] == w2_q.shape[0], "Weights expert number mismatch"
assert w1_q.shape[0] == w1_scale.shape[0], "w1 scales expert number mismatch"
assert w1_q.shape[0] == w2_scale.shape[0], "w2 scales expert number mismatch"
assert a.dtype in [torch.half, torch.bfloat16], "Invalid output dtype"
if is_cuda:
from sglang.srt.layers.quantization.fp8_kernel import (
per_group_transpose,
per_token_group_quant_fp8_hopper_moe_mn_major,
sglang_per_token_group_quant_fp8,
)
out_dtype = a.dtype
num_experts = w1_q.size(0)
m = a.size(0)
k = w1_q.size(1)
n = w2_q.size(1)
topk = topk_ids.size(1)
device = a.device
a_map = torch.empty((topk_ids.numel()), dtype=torch.int32, device=device)
c_map = torch.empty((topk_ids.numel()), dtype=torch.int32, device=device)
prepare_moe_input(
topk_ids,
expert_offsets,
problem_sizes1,
problem_sizes2,
a_map,
c_map,
num_experts,
n,
k,
)
a_q, a1_scale = sglang_per_token_group_quant_fp8(a, 128)
rep_a_q = shuffle_rows(a_q, a_map, (m * topk, k))
rep_a1_scales = shuffle_rows(a1_scale, a_map, (m * topk, int(k / 128)))
c1 = torch.empty((m * topk, n * 2), device=device, dtype=out_dtype)
c2 = torch.empty((m * topk, k), device=device, dtype=out_dtype)
a_sf_layout = torch.empty((num_experts, 5), device=device, dtype=torch.int)
w_sf_layout = torch.empty((num_experts, 5), device=device, dtype=torch.int)
fp8_blockwise_scaled_grouped_mm(
c1,
a_ptrs,
b_ptrs,
out_ptrs,
a_scales_ptrs,
b_scales_ptrs,
rep_a_q,
w1_q,
rep_a1_scales,
w1_scale,
a1_strides,
a1_strides,
c1_strides,
a_sf_layout,
w_sf_layout,
problem_sizes1,
expert_offsets[:-1],
workspace,
)
intermediate = torch.empty((m * topk, n), device=device, dtype=out_dtype)
silu_and_mul(c1, intermediate)
intemediate_q, a2_scale = sglang_per_token_group_quant_fp8(intermediate, 128)
fp8_blockwise_scaled_grouped_mm(
c2,
a_ptrs,
b_ptrs,
out_ptrs,
a_scales_ptrs,
b_scales_ptrs,
intemediate_q,
w2_q,
a2_scale,
w2_scale,
a2_strides,
a2_strides,
c2_strides,
a_sf_layout,
w_sf_layout,
problem_sizes2,
expert_offsets[:-1],
workspace,
)
result = torch.empty((m, k), device=device, dtype=out_dtype)
apply_shuffle_mul_sum(c2, result, c_map, topk_weights.to(out_dtype))
return result
FLOAT4_E2M1_MAX = 6.0
FLOAT8_E4M3_MAX = 448.0
def cutlass_moe_fp4(
a: torch.Tensor,
a1_gscale: torch.Tensor,
w1_fp4: torch.Tensor,
w1_blockscale: torch.Tensor,
w1_alphas: torch.Tensor,
a2_gscale: torch.Tensor,
w2_fp4: torch.Tensor,
w2_blockscale: torch.Tensor,
w2_alphas: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
params: CutlassMoEParams,
apply_router_weight_on_input: bool = False,
):
"""
MoE implementation for FP4 Inputs
# Gemm 1
a: Input tensor: [m, k] (half/bfloat16)
a1_gscale: Activation scale per expert: [e] (float32)
w1(gate up) (not an argument to cutlass_moe_fp4): [e, 2 * n, k]
w1_fp4: [e, 2 * n, k // 2], dtype: torch.uint8 (stacked fp4: E2M1)
(Note: `n` is the up projection output dim, `k` is the input dim in
full precision)
w1_blockscale: [e, 2 * n, k // block_size] (float8_e4m3)
(Block size = 16 for NVFP4)
# Gemm 2
a2_gscale: Activation scale per expert: [e]
w2(down projection) (not an argument to cutlass_moe_fp4): [e, k, n]
w2_fp4: [e, k, n // 2], dtype: torch.uint8 (stacked E2M1)
w2_blockscale: [e, k, n // block_size], dtype: float8_e4m3
Strides for activations, weights and output in logical number of elements.
The activations & output stride is the number of elements to the next row.
The weights stride is the number of elements to the next row per expert.
For example, if the weight is [e, n, k], then the b_stride is a tensor of
shape [e] with each element being k. Similarly for activations, if the
shape is [m, k], then the a_stride has shape [e] with each value k.
Similarly for output, if the output is [m, n], then the c_stride is a
tensor of shape [e] with each element being k.
Note: cutlass_fp4_group_mm is designed to accept the strides of
activations and weights to be the same, so it is passed in as a single
tensor.
ab_strides_13: [e] dtype: int64 [Gemm 1: Activation / Weight strides]
ab_strides_2: [e] dtype: int64 [Gemm 2: Activation / Weight strides]
c_strides_13: [e] dtype: int64 [Gemm 1: Output Strides]
c_strides_2: [e] dtype: int64 [Gemm 1: Output Strides]
topk_weights: [m, topk] dtype: float8
topk_ids: [m, topk] dtype: float8
m, n, k: Unquantized weight shapes, dtype: int
e: number of experts for the current rank, dtype: int
assumes that topk < k < n to satisfy - up/down projection expectations.
"""
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
assert w1_fp4.dtype == torch.uint8, "weight 1 must be uint8"
assert w2_fp4.dtype == torch.uint8, "weight 2 must be uint8"
assert (
w1_fp4.ndim == 3
and w2_fp4.ndim == 3
and w1_blockscale.ndim == 3
and w2_blockscale.ndim == 3
), "All Weights must be of rank 3 for cutlass_moe_fp4"
m_a, k_a = a.shape
e_w1, nx2_w1, half_k_w1 = w1_fp4.shape
e_w2, k_w2, half_n_w2 = w2_fp4.shape
assert e_w1 == e_w2 and e_w1 == params.num_experts, (
"Number of experts must match",
" between weights.",
)
assert (
k_a // 2 == half_k_w1 and params.hidden_size == k_w2
), "Hidden size mismatch between a, w1 and w2"
assert (
nx2_w1 == params.intermediate_size_per_partition * 2
and half_n_w2 == params.intermediate_size_per_partition // 2
), ("mismatch in " "expected `n`")
assert 2 * half_k_w1 == k_w2, "Hidden size mismatch w2 and w1"
assert a.dtype in [torch.half, torch.bfloat16], "Invalid input dtype"
out_dtype = a.dtype
num_topk = topk_ids.shape[1]
device = a.device
a_map = torch.empty((topk_ids.numel()), dtype=torch.int32, device=device)
c_map = torch.empty((topk_ids.numel()), dtype=torch.int32, device=device)
prepare_moe_input(
topk_ids,
params.expert_offsets,
params.problem_sizes1,
params.problem_sizes2,
a_map,
c_map,
params.num_experts,
params.intermediate_size_per_partition,
params.hidden_size,
params.blockscale_offsets,
)
rep_a_fp4, rep_a_blockscale = scaled_fp4_experts_quant(
a,
a1_gscale,
params.expert_offsets,
params.blockscale_offsets,
num_topk,
expert_map=a_map,
)
c1 = cutlass_fp4_group_mm(
rep_a_fp4,
w1_fp4,
rep_a_blockscale,
w1_blockscale,
w1_alphas,
out_dtype,
device,
params.to_gemm1_args(),
)
del rep_a_fp4, rep_a_blockscale
# hidden size dimension is split to one halfpytho sized tensor.
intermediate = torch.empty(
(m_a * num_topk, w1_fp4.shape[1] // 2), device=device, dtype=out_dtype
)
silu_and_mul(c1, intermediate)
int_fp4, int_blockscale = scaled_fp4_experts_quant(
intermediate,
a2_gscale,
params.expert_offsets,
params.blockscale_offsets,
num_topk,
)
c2 = cutlass_fp4_group_mm(
int_fp4,
w2_fp4,
int_blockscale,
w2_blockscale,
w2_alphas,
out_dtype,
device,
params.to_gemm2_args(),
)
del int_fp4, int_blockscale
c2 = shuffle_rows(c2, c_map, (m_a * num_topk, params.hidden_size))
c2 = c2.view(m_a, num_topk, params.hidden_size)
if not apply_router_weight_on_input:
c2 = c2 * topk_weights.view(m_a, num_topk, 1).to(out_dtype)
return c2.sum(dim=1).to(out_dtype)