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
156 lines
4.7 KiB
Python
156 lines
4.7 KiB
Python
import ixformer._C as ops
|
|
import torch
|
|
|
|
__all__ = ["wui4a16_gemm", "wui4a16_gemv", "wui4a16", "ref_wui4a16"]
|
|
|
|
|
|
def dequant_weight(tensor, scales, zeros, block_size):
|
|
"""
|
|
tensor: (oc/2, ic)
|
|
scales: (oc, ic/group_size)
|
|
zeros: (oc, ic/group_size)
|
|
"""
|
|
dtype = scales.dtype
|
|
left = tensor >> 4
|
|
right = tensor << 4 >> 4
|
|
left, right = right, left
|
|
ret = torch.cat((left, right), dim=-1).reshape(-1, left.size(-1))
|
|
ret_shape = ret.size()
|
|
ret = ret.view(-1, block_size)
|
|
ret = scales.view(-1, 1) * (ret - zeros.view(-1, 1))
|
|
ret = ret.reshape(ret_shape).to(dtype=dtype)
|
|
return ret
|
|
|
|
|
|
def ref_wui4a16(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
zeros: "torch.Tensor",
|
|
bias: "torch.Tensor" = None,
|
|
group_size: int = -1,
|
|
format: str = "NN",
|
|
only_return_weight: bool = False,
|
|
):
|
|
"""
|
|
format = TN,NN
|
|
group_size = TN(128),NN(128, 32)
|
|
input : bfloat16|fp16 (bs, ic)
|
|
qweights : int32 NN: (ic, oc // 8) TN:(oc, ic // 8)
|
|
scales : bfloat16|fp16 (ic // group_size, oc)
|
|
zeros : int32 (ic // group_size, oc // 8)
|
|
bias : bfloat16|fp16 (oc, )
|
|
output : bfloat16|fp16 (bs, oc)
|
|
"""
|
|
|
|
def unpack_tensor(x, pack_num=8, order_map=None):
|
|
if order_map is None:
|
|
order_map = [0, 1, 2, 3, 4, 5, 6, 7]
|
|
unit = 32 // pack_num
|
|
rows, cols = x.shape
|
|
res = torch.zeros((rows, cols * pack_num), dtype=torch.int32, device=x.device)
|
|
for col in range(cols):
|
|
for k in range(pack_num):
|
|
res[:, col * pack_num + order_map[k]] = (x[:, col] >> (unit * k)) & 0xF
|
|
return res
|
|
|
|
scales = scales.t().contiguous()
|
|
if format == "NN":
|
|
zeros = unpack_tensor(zeros, order_map=[0, 2, 4, 6, 1, 3, 5, 7])
|
|
zeros = zeros.t().contiguous()
|
|
qweights = unpack_tensor(qweights, order_map=[0, 2, 4, 6, 1, 3, 5, 7])
|
|
qweights = qweights.t().contiguous()
|
|
else:
|
|
zeros = unpack_tensor(zeros)
|
|
zeros = zeros.t().contiguous()
|
|
qweights = unpack_tensor(qweights)
|
|
output_dim, input_dim = qweights.shape
|
|
qweights = qweights.view(output_dim, input_dim // group_size, group_size)
|
|
zeros = zeros.view(output_dim, input_dim // group_size, 1)
|
|
scales = scales.view(output_dim, input_dim // group_size, 1)
|
|
|
|
qweights = (qweights - zeros) * scales
|
|
qweights = qweights.view(output_dim, input_dim)
|
|
if only_return_weight:
|
|
return qweights
|
|
output = torch.nn.functional.linear(inputs, qweights.to(inputs.dtype))
|
|
return output, qweights
|
|
|
|
|
|
def wui4a16_gemm(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
zeros: "torch.Tensor",
|
|
bias: "torch.Tensor" = None,
|
|
group_size: int = 128,
|
|
format: str = "NN",
|
|
):
|
|
output_shape = inputs.shape[:-1] + (scales.shape[1],)
|
|
|
|
output = ops.infer.wui4a16_gemm(
|
|
inputs, qweights, scales, zeros, bias, group_size, format
|
|
)
|
|
return output.view(output_shape)
|
|
|
|
|
|
def wui4a16_gemv(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
zeros: "torch.Tensor",
|
|
bias: "torch.Tensor" = None,
|
|
group_size: int = 128,
|
|
format: str = "NN",
|
|
):
|
|
output_shape = inputs.shape[:-1] + (scales.shape[1],)
|
|
|
|
output = ops.infer.wui4a16_gemv(
|
|
inputs, qweights, scales, zeros, bias, group_size, format
|
|
)
|
|
return output.view(output_shape)
|
|
|
|
|
|
def wui4a16(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
zeros: "torch.Tensor",
|
|
bias: "torch.Tensor" = None,
|
|
group_size: int = 128,
|
|
format: str = "NN",
|
|
):
|
|
"""
|
|
format = TN,NN
|
|
group_size = TN(128),NN(128, 32)
|
|
input : bfloat16|fp16 (bs, ic)
|
|
qweights : int32 NN: (ic, oc // 8) TN:(oc, ic // 8)
|
|
scales : bfloat16|fp16 (ic // group_size, oc)
|
|
zeros : int32 (ic // group_size, oc // 8)
|
|
bias : bfloat16|fp16 (oc, )
|
|
output : bfloat16|fp16 (bs, oc)
|
|
支持条件 : NN: oc % 8 == 0 && ic % group_size == 0 && ic % 2 == 0
|
|
TN: oc % 2 == 0 && ic % group_size == 0
|
|
"""
|
|
batch = inputs.numel() // inputs.shape[-1]
|
|
if batch <= 1:
|
|
return wui4a16_gemv(
|
|
inputs=inputs,
|
|
qweights=qweights,
|
|
scales=scales,
|
|
zeros=zeros,
|
|
bias=bias,
|
|
group_size=group_size,
|
|
format=format,
|
|
)
|
|
else:
|
|
return wui4a16_gemm(
|
|
inputs=inputs,
|
|
qweights=qweights,
|
|
scales=scales,
|
|
zeros=zeros,
|
|
bias=bias,
|
|
group_size=group_size,
|
|
format=format,
|
|
)
|