EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[refactor] refactor weight trans nz and transpose (#4878)

Browse Source 
### What this PR does / why we need it?

Now `VLLM_ASCEND_ENABLE_NZ` will have three options:
0: disable nz;
1: only quant case enable nz;
2: enable nz as long as possible;

And `VLLM_ASCEND_ENABLE_NZ`=1 by default.

All cases are shown in the table below:

|  | W4A4 | W4A8 | W8A8 | fp16/bf16 | fp32 |
|---|---|---|---|---|---|
| trans nz | can't support nz | trans nz by default | trans nz by
default | trans nz when VLLM_ASCEND_ENABLE_NZ is 2 | can't support nz |
| transpose | only support not transpose case | only support transpose
case | only support transpose case | linear: only support not transpose
case<br>gmm: only support transpose case | same to fp16/bf16 |

Some exceptional cases:
1. MLAPO op need to do some additional processing on the weights,
including trans nz. If use MLAPO op, some weight will be transformed to
nz forcely;
2. MLA/SFA's weight `W_UV` will be used by op
`torch.ops._C_ascend.batch_matmul_transpose`, and this op can't support
nz currently;

### Does this PR introduce _any_ user-facing change?
Now fp16/bf16 weight will not trans nz by default.

### How was this patch tested?

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: zzzzwwjj <1183291235@qq.com>
This commit is contained in:
zzzzwwjj
2025-12-19 14:27:24 +08:00
committed by GitHub
parent ea8f544ce7
commit cc23067f1e
19 changed files with 156 additions and 255 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
vllm_ascend/quantization/w8a8_dynamic.py
View File

@@ -29,7 +29,7 @@ from vllm_ascend.ascend_forward_context import MoECommType
from vllm_ascend.distributed.parallel_state import get_mc2_group
from vllm_ascend.flash_common3_context import get_flash_common3_context
from vllm_ascend.ops.fused_moe.experts_selector import select_experts
from vllm_ascend.utils import ACL_FORMAT_FRACTAL_NZ, is_enable_nz
from vllm_ascend.utils import maybe_trans_nz
class AscendW8A8DynamicLinearMethod:
@@ -37,7 +37,7 @@ class AscendW8A8DynamicLinearMethod:
"""
def __init__(self):
self.transpose_weight = True
pass
@staticmethod
def get_weight(input_size: int, output_size: int,
@@ -91,12 +91,9 @@ class AscendW8A8DynamicLinearMethod:
return output
def process_weights_after_loading(self, layer):
if self.transpose_weight:
layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()
layer.weight.data = layer.weight.data.transpose(0, 1).contiguous()
# cast quantized weight tensors in NZ format for higher inference speed
if is_enable_nz():
layer.weight.data = torch_npu.npu_format_cast(
layer.weight.data, ACL_FORMAT_FRACTAL_NZ)
layer.weight.data = maybe_trans_nz(layer.weight.data)
layer.weight_scale.data = layer.weight_scale.data.flatten()
layer.weight_scale_fp32 = layer.weight_scale.data.to(torch.float32)
layer.weight_offset.data = layer.weight_offset.data.flatten()
@@ -107,8 +104,6 @@ class AscendW8A8DynamicFusedMoEMethod:
"""
def __init__(self):
self.transpose_weight = True
self.ep_group = get_ep_group()
vllm_config = get_current_vllm_config()
@@ -270,14 +265,12 @@ class AscendW8A8DynamicFusedMoEMethod:
mc2_mask=kwargs.get("mc2_mask", None))
def process_weights_after_loading(self, layer):
if self.transpose_weight:
layer.w13_weight.data = layer.w13_weight.data.transpose(
1, 2).contiguous()
layer.w2_weight.data = layer.w2_weight.data.transpose(
1, 2).contiguous()
if is_enable_nz():
torch_npu.npu_format_cast_(layer.w13_weight, ACL_FORMAT_FRACTAL_NZ)
torch_npu.npu_format_cast_(layer.w2_weight, ACL_FORMAT_FRACTAL_NZ)
layer.w13_weight.data = layer.w13_weight.data.transpose(
1, 2).contiguous()
layer.w2_weight.data = layer.w2_weight.data.transpose(1,
2).contiguous()
layer.w13_weight.data = maybe_trans_nz(layer.w13_weight.data)
layer.w2_weight.data = maybe_trans_nz(layer.w2_weight.data)
layer.w13_weight_scale.data = layer.w13_weight_scale.data.view(
layer.w13_weight_scale.data.shape[0], -1)
layer.w13_weight_scale_fp32 = layer.w13_weight_scale.data.to(