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
235 lines
8.3 KiB
Python
235 lines
8.3 KiB
Python
import ixformer._C as ops
|
|
import torch
|
|
|
|
__all__ = [
|
|
"marlin_w4a16",
|
|
"marlin_w4_weight_repack",
|
|
"marlin_w8a16",
|
|
"marlin_w8_weight_repack",
|
|
]
|
|
|
|
|
|
def marlin_w4a16(
|
|
inputs: torch.Tensor,
|
|
weights: torch.Tensor,
|
|
scales: torch.Tensor,
|
|
zeros: torch.Tensor,
|
|
bias: torch.Tensor = None, # TODO
|
|
group_size: int = -1,
|
|
format: str = "k16n32",
|
|
batch_first: bool = True,
|
|
outputs: torch.Tensor = None,
|
|
):
|
|
"""
|
|
Args:
|
|
inputs: (batch, m, k) if batch_first else (m, batch, k) torch.float16, torch.bfloat16
|
|
weights: (batch, k/16, n/32, 64) torch.int32
|
|
scales:
|
|
(batch, k_groups, n) format:k16n32 torch.float16, torch.bfloat16
|
|
(batch, n_groups, k) format:k16n32_grouped_n torch.float16, torch.bfloat16
|
|
zeros:
|
|
(batch, k_groups, n/8) format:k16n32 torch.int32
|
|
(batch, n_groups, k/8) format:k16n32_grouped_n torch.int32
|
|
group_size: int
|
|
group size of quant
|
|
format: str
|
|
describe format of weight
|
|
batch_first: bool
|
|
describe format of input and output
|
|
Returns:
|
|
outputs: (batch, m, n) if batch_first else (m, batch, n) torch.float16, torch.bfloat16
|
|
"""
|
|
if outputs is None:
|
|
batch, m = (
|
|
(inputs.shape[0], inputs.shape[1])
|
|
if batch_first
|
|
else (inputs.shape[1], inputs.shape[0])
|
|
)
|
|
if format.startswith("k16n32"):
|
|
n = weights.shape[2] * 32
|
|
outputs = torch.empty(
|
|
(batch, m, n) if batch_first else (m, batch, n),
|
|
dtype=inputs.dtype,
|
|
device=inputs.device,
|
|
)
|
|
|
|
ops.infer.marlin_w4a16(
|
|
outputs, inputs, weights, scales, zeros, bias, group_size, format, batch_first
|
|
)
|
|
return outputs
|
|
|
|
|
|
def marlin_w4_weight_repack(
|
|
weights: torch.Tensor,
|
|
scales: torch.Tensor = None,
|
|
zeros: torch.Tensor = None,
|
|
weight_format: str = "gptq",
|
|
reformat: str = "k16n32",
|
|
pack_order: str = "default",
|
|
repack_weight: torch.Tensor = None,
|
|
):
|
|
"""
|
|
Args:
|
|
weights:
|
|
(batch, k, n/8) weight_format:awq torch.int32
|
|
(batch, k/8, n) weight_format:gptq torch.int32
|
|
scales:
|
|
(batch, k_groups, n) format:k16n32 torch.float16, torch.bfloat16
|
|
(batch, n_groups, k) format:k16n32_grouped_n torch.float16, torch.bfloat16
|
|
zeros:
|
|
(batch, k_groups, n/8) format:k16n32 torch.int32
|
|
(batch, n_groups, k/8) format:k16n32_grouped_n torch.int32
|
|
weight_format: str
|
|
describe format of weight
|
|
reformat: str
|
|
describe format of repacked weight
|
|
pack_order: str
|
|
describe pack order on a pack unit
|
|
Returns:
|
|
repack_weight: (batch, k/16, n/32, 64) torch.int32
|
|
"""
|
|
assert weight_format in ["gptq", "gptq_grouped_n", "awq"]
|
|
assert reformat in ["k16n32", "k16n32_grouped_n"]
|
|
assert pack_order in ["default", "02461357", "01234567"]
|
|
|
|
if pack_order == "default":
|
|
default_order = {
|
|
"gptq": "01234567",
|
|
"awq": "02461357",
|
|
"gptq_grouped_n": "02461357",
|
|
}
|
|
pack_order = default_order[weight_format]
|
|
|
|
if weight_format.startswith("gptq"):
|
|
batch, pack_k, n = weights.shape
|
|
k = pack_k * 8
|
|
elif weight_format == "awq":
|
|
batch, k, pack_n = weights.shape
|
|
n = pack_n * 8
|
|
|
|
repack_scales, repack_zeros = None, None
|
|
if reformat.startswith("k16n32"):
|
|
if repack_weight is None:
|
|
repack_weight = torch.empty(
|
|
(batch, k // 16, n // 32, 64),
|
|
dtype=torch.int32,
|
|
device=weights.device,
|
|
)
|
|
if scales is not None:
|
|
repack_scales = torch.empty_like(scales)
|
|
if zeros is not None:
|
|
repack_zeros = torch.empty_like(zeros)
|
|
|
|
ops.infer.marlin_w4_weight_repack(
|
|
weights,
|
|
repack_weight,
|
|
scales,
|
|
repack_scales,
|
|
zeros,
|
|
repack_zeros,
|
|
weight_format,
|
|
reformat,
|
|
pack_order,
|
|
)
|
|
|
|
if repack_scales is not None and repack_zeros is not None:
|
|
return repack_weight, repack_scales, repack_zeros
|
|
else:
|
|
return repack_weight
|
|
|
|
|
|
def marlin_w8a16(
|
|
inputs: torch.Tensor,
|
|
weights: torch.Tensor,
|
|
scales: torch.Tensor,
|
|
bias: torch.Tensor = None, # TODO
|
|
group_size: int = -1,
|
|
format: str = "k16n16",
|
|
batch_first: bool = True,
|
|
outputs: torch.Tensor = None,
|
|
):
|
|
"""
|
|
Args:
|
|
inputs: (batch, m, k) if batch_first else (m, batch, k) torch.float16, torch.bfloat16
|
|
weights: (batch, k/16, n/16, 64) torch.int32
|
|
scales:
|
|
(batch, k_groups, n) format:k16n16 torch.float32
|
|
(batch, n_groups, k) format:k16n16_grouped_n torch.float32
|
|
group_size: int
|
|
group size of quant
|
|
format: str
|
|
describe format of weight
|
|
batch_first: bool
|
|
describe format of input and output
|
|
Returns:
|
|
outputs: (batch, m, n) if batch_first else (m, batch, n) torch.float16, torch.bfloat16
|
|
"""
|
|
if outputs is None:
|
|
batch, m = (
|
|
(inputs.shape[0], inputs.shape[1])
|
|
if batch_first
|
|
else (inputs.shape[1], inputs.shape[0])
|
|
)
|
|
if format.startswith("k16n16"):
|
|
n = weights.shape[2] * 16
|
|
outputs = torch.empty(
|
|
(batch, m, n) if batch_first else (m, batch, n),
|
|
dtype=inputs.dtype,
|
|
device=inputs.device,
|
|
)
|
|
|
|
ops.infer.marlin_w8a16(
|
|
outputs, inputs, weights, scales, bias, group_size, format, batch_first
|
|
)
|
|
return outputs
|
|
|
|
|
|
def marlin_w8_weight_repack(
|
|
weights: torch.Tensor,
|
|
scales: torch.Tensor = None,
|
|
weight_format: str = "int8",
|
|
reformat: str = "k16n16",
|
|
):
|
|
"""
|
|
Args:
|
|
weights:
|
|
(batch, k, n) weight_format:int8 torch.int8
|
|
scales:
|
|
(batch, k_groups, n) format:k16n16 torch.float32
|
|
(batch, n_groups, k) format:k16n16_grouped_n torch.float32
|
|
weight_format: str
|
|
describe format of weight
|
|
reformat: str
|
|
describe format of repacked weight
|
|
Returns:
|
|
repack_weight:
|
|
(batch, k/16, n/16, 64) torch.int32
|
|
"""
|
|
assert weight_format in ["int8"]
|
|
assert reformat in ["k16n16", "k16n16_grouped_n"]
|
|
|
|
repack_scales = None
|
|
if weight_format == "int8":
|
|
batch, k, n = weights.shape
|
|
repack_weight = torch.empty(
|
|
(batch, k // 16, n // 16, 64),
|
|
dtype=torch.int32,
|
|
device=weights.device,
|
|
)
|
|
if scales is not None:
|
|
repack_scales = torch.empty_like(scales)
|
|
|
|
ops.infer.marlin_w8_weight_repack(
|
|
weights,
|
|
repack_weight,
|
|
scales,
|
|
repack_scales,
|
|
weight_format,
|
|
reformat,
|
|
)
|
|
|
|
if repack_scales is not None:
|
|
return repack_weight, repack_scales
|
|
else:
|
|
return repack_weight
|