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
158 lines
4.0 KiB
Python
158 lines
4.0 KiB
Python
import ixformer._C as ops
|
|
import torch
|
|
|
|
__all__ = ["wi4a16_gemm", "wi4a16_gemv", "wi4a16", "ref_wi4a16"]
|
|
|
|
|
|
def dequant_weight(tensor, scales, zeros, block_size):
|
|
# from CPM
|
|
"""
|
|
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_wi4a16(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
zeros: "torch.Tensor",
|
|
group_size: int = -1,
|
|
format: str = "TN",
|
|
):
|
|
assert format in ["TN"]
|
|
weights = dequant_weight(
|
|
qweights,
|
|
scales.transpose(0, 1).contiguous(),
|
|
zeros.transpose(0, 1).contiguous(),
|
|
group_size,
|
|
)
|
|
output = torch.nn.functional.linear(inputs, weights.to(inputs.dtype))
|
|
return output
|
|
|
|
|
|
def wi4a16_gemm(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
zeros: "torch.Tensor",
|
|
group_size: int = -1,
|
|
format: str = "TN",
|
|
output=None,
|
|
):
|
|
"""
|
|
wi4a16 gemm 接口
|
|
支持条件:
|
|
format = TN
|
|
group_size = 128
|
|
input : fp16 (bs, ic)
|
|
qweights : int8 (oc/2, ic)
|
|
scales : fp16 (ic/group_size, oc)
|
|
zeros : fp16 (ic/group_size, oc)
|
|
TN 支持条件： oc % 2 == 0 && ic % 128 == 0
|
|
NN 支持条件： 不支持
|
|
"""
|
|
|
|
assert format in ["TN"]
|
|
assert len(qweights.shape) == 2
|
|
assert len(scales.shape) == 2
|
|
assert len(zeros.shape) == 2
|
|
|
|
input_shape = list(inputs.shape)
|
|
inputs = inputs.view(-1, input_shape[-1])
|
|
|
|
if output is None:
|
|
output_shape = input_shape[:-1] + [scales.shape[1]]
|
|
output = inputs.new_empty(output_shape).view(-1, output_shape[-1])
|
|
else:
|
|
output_shape = output.shape
|
|
|
|
ops.infer.wi4a16_gemm(output, inputs, qweights, scales, zeros, group_size, format)
|
|
return output.view(output_shape)
|
|
|
|
|
|
def wi4a16_gemv(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
zeros: "torch.Tensor",
|
|
group_size: int = -1,
|
|
format: str = "TN",
|
|
output=None,
|
|
):
|
|
"""
|
|
wi4a16 gemv 接口
|
|
支持条件:
|
|
format = TN
|
|
group_size = 128
|
|
input : bf16|fp16 (bs, ic)
|
|
qweights : int8 (oc/2, ic)
|
|
scales : bf16|fp16 (ic/group_size, oc)
|
|
zeros : bf16|fp16 (ic/group_size, oc)
|
|
TN 支持条件： oc % 2 == 0 && ic % 128 == 0
|
|
NN 支持条件： 不支持
|
|
"""
|
|
|
|
assert format in ["TN"]
|
|
assert len(qweights.shape) == 2
|
|
assert len(scales.shape) == 2
|
|
assert len(zeros.shape) == 2
|
|
|
|
input_shape = list(inputs.shape)
|
|
inputs = inputs.view(-1, input_shape[-1])
|
|
|
|
if output is None:
|
|
output_shape = input_shape[:-1] + [scales.shape[1]]
|
|
output = inputs.new_empty(output_shape).view(-1, output_shape[-1])
|
|
else:
|
|
output_shape = output.shape
|
|
|
|
ops.infer.wi4a16_gemv(output, inputs, qweights, scales, zeros, group_size, format)
|
|
return output.view(output_shape)
|
|
|
|
|
|
def wi4a16(
|
|
inputs: "torch.Tensor",
|
|
qweights: "torch.Tensor",
|
|
scales: "torch.Tensor",
|
|
zeros: "torch.Tensor",
|
|
group_size: int = -1,
|
|
format: str = "TN",
|
|
output=None,
|
|
):
|
|
input_shape = inputs.shape
|
|
inputs = inputs.view(-1, input_shape[-1])
|
|
bs = inputs.size(0)
|
|
inputs = inputs.view(input_shape)
|
|
if bs <= 1:
|
|
return wi4a16_gemv(
|
|
inputs=inputs,
|
|
qweights=qweights,
|
|
scales=scales,
|
|
zeros=zeros,
|
|
group_size=group_size,
|
|
format=format,
|
|
output=output,
|
|
)
|
|
else:
|
|
return wi4a16_gemm(
|
|
inputs=inputs,
|
|
qweights=qweights,
|
|
scales=scales,
|
|
zeros=zeros,
|
|
group_size=group_size,
|
|
format=format,
|
|
output=output,
|
|
)
|