87a19d2d00
来源:
1. Chranos/ixformer (GitHub) → ixformer_sdk/ (230 files, 70K lines)
- inference/functions/vllm.py: vllm_moe_topk_softmax 完整实现 (2033 lines)
- inference/functions/moe.py: MoE ops 完整实现 (1380 lines)
- contrib/vllm_flash_attn/: FA2 Python 接口 (1018 lines)
- contrib/tgi/fused_moe.py: TGI fused MoE (429 lines)
- csrc/include/ixformer/: C++ kernel headers + cmake
2. Deep-Spark/xllm (GitHub) → upstream_ref/xllm_latest/ (+15 files)
- npu_torch/qwen3_5_decoder_layer_impl.cpp/.h
- npu_torch/qwen3_5_gated_delta_net.cpp/.h
- npu_torch/qwen3_next_*.cpp/.h (6 files)
- npu_torch/attention.cpp/.h + fused_moe.cpp/.h + CMakeLists.txt
- models/llm/qwen3_5.h + qwen3_5_mtp.h + qwen3_next.h
- models/vlm/qwen3_5.h
调用链完整性:
ixformer_sdk/inference/functions/vllm.py
→ ops.infer.moe_topk_softmax() (C++ 层)
→ 这就是 base 镜像 libixformer.so 里的实现
upstream_ref/xllm_latest/core/layers/ilu/fused_moe.cpp
→ ixformer::infer::topk_softmax() (直接 C++ 调用)
→ ixformer::infer::group_gemm() → 完整 7-step MoE pipeline
226 lines
6.1 KiB
Python
226 lines
6.1 KiB
Python
import math
|
|
from typing import List, Union, Optional
|
|
|
|
import ixformer._C as ops
|
|
import torch
|
|
from torch.autograd.function import Function, FunctionCtx
|
|
|
|
import ixformer
|
|
from ixformer.core import config
|
|
|
|
__all__ = [
|
|
"w8a16_gemm",
|
|
"w8a16_gemv",
|
|
"w8a16",
|
|
"ref_w8a16",
|
|
"wu8a16",
|
|
"ref_wu8a16",
|
|
]
|
|
|
|
|
|
def w8a16_gemv(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
group_size: int = -1,
|
|
format: str = "unknown",
|
|
output: Optional[torch.Tensor] = None
|
|
):
|
|
"""
|
|
w8a16 gemv 接口
|
|
input : bf16|fp16 (bs, ic)
|
|
qweights : int8 TN:(oc, ic) NN:(ic, oc)
|
|
scales : bf16|fp16 TN: 当groupsize为-1时, shape: (1, oc), 否则,shape: (ic/group_size, oc) NN:(1, oc)
|
|
TN 支持条件： ic % groupSize = 0, oc % 2 = 0, bs<=4
|
|
NN 支持条件： groupsize = -1 or groupsize = ic, oc % 4 = 0, bs<=4
|
|
"""
|
|
assert format in ["TN", "NN"]
|
|
assert len(qweights.shape) == 2
|
|
assert len(scales.shape) == 2
|
|
|
|
input_shape = list(inputs.shape)
|
|
inputs = inputs.view(-1, input_shape[-1])
|
|
|
|
if format == "TN":
|
|
output_shape = input_shape[:-1] + [qweights.shape[0]]
|
|
else:
|
|
output_shape = input_shape[:-1] + [qweights.shape[1]]
|
|
|
|
if output is None:
|
|
output = inputs.new_empty(output_shape).view(-1, output_shape[-1])
|
|
|
|
ops.infer.w8a16_gemv(output, inputs, qweights, scales, group_size, format)
|
|
return output.view(output_shape)
|
|
|
|
|
|
def w8a16_gemm(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
group_size: int = -1,
|
|
format: str = "TN",
|
|
persistent: int = 0,
|
|
output: Optional[torch.Tensor] = None
|
|
):
|
|
"""
|
|
w8a16 gemm 接口
|
|
1. group_size=-1 or group_size=ic
|
|
input : bf16|fp16 (bs, ic)
|
|
qweights : int8 TN:(oc, ic) NN:(ic, oc)
|
|
scales : bf16|fp16 (1, oc)
|
|
NN 支持条件： ic%64==0, oc%64==0
|
|
|
|
2. group_size=64
|
|
input : bf16|fp16 (bs, ic)
|
|
qweights : int8 TN:(oc, ic)
|
|
scales : bf16|fp16 (ic/64, oc)
|
|
TN 支持条件： oc%2==0, ic%64==0
|
|
NN 不支持
|
|
"""
|
|
|
|
assert format in ["TN", "NN"]
|
|
assert len(qweights.shape) == 2
|
|
assert len(scales.shape) == 2
|
|
|
|
input_shape = list(inputs.shape)
|
|
inputs = inputs.view(-1, input_shape[-1])
|
|
|
|
if format == "TN":
|
|
output_shape = input_shape[:-1] + [qweights.shape[0]]
|
|
else:
|
|
output_shape = input_shape[:-1] + [qweights.shape[1]]
|
|
|
|
if output is None:
|
|
output = inputs.new_empty(output_shape).view(-1, output_shape[-1])
|
|
|
|
ops.infer.w8a16_gemm(
|
|
output, inputs, qweights, scales, group_size, format, persistent
|
|
)
|
|
return output.view(output_shape)
|
|
|
|
|
|
def dequant(qweight, scales, group_size):
|
|
IC, OC = qweight.shape
|
|
weight = qweight.t().reshape(OC, -1, group_size).to(
|
|
torch.float32
|
|
) * scales.t().unsqueeze(-1)
|
|
return weight.reshape(OC, IC)
|
|
|
|
|
|
def ref_w8a16(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
group_size: int = -1,
|
|
format: str = "TN",
|
|
):
|
|
if group_size == -1:
|
|
group_size = inputs.shape[1]
|
|
if format == "TN":
|
|
weights = dequant(qweights.transpose(0, 1), scales, group_size)
|
|
elif format == "NN":
|
|
weights = dequant(qweights, scales, group_size)
|
|
return torch.nn.functional.linear(inputs, weights.to(inputs.dtype))
|
|
|
|
|
|
def w8a16(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
group_size: int = -1,
|
|
format: str = "TN",
|
|
output: Optional[torch.Tensor] = None,
|
|
persistent: int = 0,
|
|
):
|
|
input_shape = inputs.shape
|
|
inputs = inputs.view(-1, input_shape[-1])
|
|
bs = inputs.size(0)
|
|
inputs = inputs.view(input_shape)
|
|
if bs <= config.IXFORMER_GEMV_THRESHOLD:
|
|
return w8a16_gemv(
|
|
inputs=inputs,
|
|
qweights=qweights,
|
|
scales=scales,
|
|
group_size=group_size,
|
|
format=format,
|
|
output=output
|
|
)
|
|
else:
|
|
return w8a16_gemm(
|
|
inputs=inputs,
|
|
qweights=qweights,
|
|
scales=scales,
|
|
group_size=group_size,
|
|
format=format,
|
|
output=output,
|
|
persistent=persistent
|
|
)
|
|
|
|
|
|
def ref_wu8a16(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
zeros: "torch.Tensor",
|
|
group_size: int = -1,
|
|
format: str = "TN",
|
|
):
|
|
assert format in ["TN"]
|
|
assert len(qweights.shape) == 2
|
|
assert len(scales.shape) == 2
|
|
org_w_shape = qweights.shape
|
|
scales = scales.transpose(0, 1).flatten().view(-1, 1)
|
|
zeros = zeros.transpose(0, 1).flatten().view(-1, 1)
|
|
|
|
if group_size != -1:
|
|
qweights = qweights.reshape(-1, group_size)
|
|
w = (qweights - zeros) * scales
|
|
w = w.reshape(org_w_shape)
|
|
output = torch.matmul(inputs, w.t())
|
|
return output
|
|
|
|
|
|
def wu8a16(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
zeros: "torch.Tensor",
|
|
group_size: int = -1,
|
|
format: str = "TN",
|
|
persistent: int = 0,
|
|
):
|
|
"""
|
|
http://confluence.iluvatar.ai:8090/display/SW/cuinferCustomGemm+Interface+Doc
|
|
wu8a16 非对称量化 gemm 接口
|
|
1. group_size=-1
|
|
input : bf16|fp16 (bs, ic)
|
|
qweights : uint8 TN:(oc, ic)
|
|
scales : bf16|fp16 (1,oc)
|
|
zeros : bf16|fp16 (1,oc)
|
|
TN 支持条件： ic % 64 == 0
|
|
NN 不支持
|
|
|
|
2. group_size=64
|
|
input : bf16|fp16 (bs, ic)
|
|
qweights : int8 TN:(oc, ic)
|
|
scales : bf16|fp16 (ic/64, oc)
|
|
zeros : bf16|fp16 (ic/64, oc)
|
|
TN 支持条件： oc % 2 == 0 && ic % 64 == 0
|
|
NN 不支持
|
|
"""
|
|
assert format in ["TN"]
|
|
assert len(qweights.shape) == 2
|
|
assert len(scales.shape) == 2
|
|
|
|
input_shape = list(inputs.shape)
|
|
inputs = inputs.view(-1, input_shape[-1])
|
|
|
|
output_shape = input_shape[:-1] + [qweights.shape[0]]
|
|
|
|
output = inputs.new_empty(output_shape).view(-1, output_shape[-1])
|
|
|
|
ops.infer.wu8a16_gemm(
|
|
output, inputs, qweights, scales, zeros, group_size, format, persistent
|
|
)
|
|
return output.view(output_shape)
|