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

[main][quantization] Support deepseek w4a8 per-channel quantization (#3011)

Browse Source 
### What this PR does / why we need it?
1.Support deepseek w4a8 per-channel quantization
2.The eager mode supports converting weights to the NZ format
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
#### How to get weights using Modelslim

##### Installation steps

git clone https://gitcode.com/Ascend/msit.git
cd msit/msmodelslim
bash install.sh

##### Generate w4a8 per-channel weights

cd /example/DeepSeek
Command reference: msmodelslim/example/DeepSeek/README.md

- vLLM version: v0.10.2
- vLLM main:
f225ea7dd9

---------

Signed-off-by: Wang Kunpeng <1289706727@qq.com>
This commit is contained in:
Wang Kunpeng
2025-09-27 21:01:16 +08:00
committed by GitHub
parent e9359bd8fa
commit 859e861d92
6 changed files with 299 additions and 196 deletions
Download Patch File Download Diff File Expand all files Collapse all files

107
vllm_ascend/torchair/quantization/torchair_w4a8_dynamic.py
View File

@@ -139,6 +139,8 @@ class TorchairAscendW4A8DynamicFusedMoEMethod:
vllm_config = get_current_vllm_config()
self.group_size = vllm_config.quant_config.quant_description.get(
"group_size", 256)
# NOTE: the weights are quantized from bf16 to int4 through a per-channel quantization process
self.is_per_channel_weight = self.group_size == 0
quant_version = vllm_config.quant_config.quant_description.get(
"version", "0")
# NOTE: new quantize weights: 2 int4 pack into int8
@@ -188,44 +190,45 @@ class TorchairAscendW4A8DynamicFusedMoEMethod:
num_experts,
2 * intermediate_size_per_partition,
1,
dtype=params_dtype)
dtype=torch.float32)
param_dict["w13_weight_offset"] = torch.empty(
num_experts,
2 * intermediate_size_per_partition,
1,
dtype=params_dtype)
param_dict["w13_weight_scale_second"] = torch.empty(
num_experts,
2 * intermediate_size_per_partition,
hidden_sizes // self.group_size,
dtype=params_dtype)
param_dict["w13_weight_offset_second"] = torch.empty(
num_experts,
2 * intermediate_size_per_partition,
hidden_sizes // self.group_size,
dtype=params_dtype)
dtype=torch.float32)
param_dict["w2_weight_scale"] = torch.empty(num_experts,
hidden_sizes,
1,
dtype=params_dtype)
dtype=torch.float32)
param_dict["w2_weight_offset"] = torch.empty(num_experts,
hidden_sizes,
1,
dtype=params_dtype)
param_dict["w2_weight_scale_second"] = torch.empty(
num_experts,
hidden_sizes,
intermediate_size_per_partition // self.group_size,
dtype=params_dtype)
param_dict["w2_weight_offset_second"] = torch.empty(
num_experts,
hidden_sizes,
intermediate_size_per_partition // self.group_size,
dtype=params_dtype)
dtype=torch.float32)
if not self.is_per_channel_weight:
param_dict["w13_weight_scale_second"] = torch.empty(
num_experts,
2 * intermediate_size_per_partition,
hidden_sizes // self.group_size,
dtype=torch.float32)
param_dict["w13_weight_offset_second"] = torch.empty(
num_experts,
2 * intermediate_size_per_partition,
hidden_sizes // self.group_size,
dtype=torch.float32)
param_dict["w2_weight_scale_second"] = torch.empty(
num_experts,
hidden_sizes,
intermediate_size_per_partition // self.group_size,
dtype=torch.float32)
param_dict["w2_weight_offset_second"] = torch.empty(
num_experts,
hidden_sizes,
intermediate_size_per_partition // self.group_size,
dtype=torch.float32)
if self.new_quant_version:
param_dict["w13_scale_bias"] = torch.empty(
@@ -318,8 +321,8 @@ class TorchairAscendW4A8DynamicFusedMoEMethod:
hidden_states=x,
w1=layer.w13_weight,
w2=layer.w2_weight,
w1_scale=layer.w13_weight_scale_second,
w2_scale=layer.w2_weight_scale_second,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
w1_scale_bias=layer.w13_scale_bias,
w2_scale_bias=layer.w2_scale_bias,
topk_weights=topk_weights,
@@ -343,8 +346,8 @@ class TorchairAscendW4A8DynamicFusedMoEMethod:
hidden_states=x,
w1=layer.w13_weight,
w2=layer.w2_weight,
w1_scale=layer.w13_weight_scale_second,
w2_scale=layer.w2_weight_scale_second,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
w1_scale_bias=layer.w13_scale_bias,
w2_scale_bias=layer.w2_scale_bias,
topk_weights=topk_weights,
@@ -357,6 +360,14 @@ class TorchairAscendW4A8DynamicFusedMoEMethod:
)
def process_scale(self, weight: torch.Tensor, scale, per_group_scale):
scale = scale.transpose(1, 2).contiguous()
if self.is_per_channel_weight:
scale_np = scale.cpu().numpy()
scale_np.dtype = np.uint32
scale_uint64_tensor = torch.from_numpy(scale_np.astype(
np.int64)).npu()
return scale_uint64_tensor, None
per_group_scale = per_group_scale.transpose(1, 2).contiguous()
group_num, k, n = weight.shape
# the weight of the new version is reduced by half by pack n, so it needs to be restored
if self.new_quant_version:
@@ -399,13 +410,10 @@ class TorchairAscendW4A8DynamicFusedMoEMethod:
def pack_to_int32(self, weight: torch.Tensor):
if self.new_quant_version:
group_num, k, n = weight.shape
assert n % 4 == 0, "the last dim of weight needs to be divided by 4"
packed_n = n // 4
# pack 4 int8(int4*2) to int32, because in pytorch, we need to use int32 to represent int4
packed_weight = torch.from_numpy(
np.frombuffer(weight.cpu().numpy().tobytes(), dtype=np.int32))
return packed_weight.reshape(group_num, k, packed_n).npu()
assert weight.shape[
-1] % 4 == 0, "the last dim of weight needs to be divided by 4"
return weight.view(torch.int32).contiguous()
else:
return torch_npu.npu_quantize(weight.to(torch.float32),
torch.tensor([1.]).npu(), None,
@@ -417,21 +425,22 @@ class TorchairAscendW4A8DynamicFusedMoEMethod:
1, 2).contiguous()
layer.w2_weight.data = layer.w2_weight.data.transpose(
1, 2).contiguous()
layer.w13_weight_scale.data = layer.w13_weight_scale.data.transpose(
1, 2).contiguous()
layer.w2_weight_scale.data = layer.w2_weight_scale.data.transpose(
1, 2).contiguous()
layer.w13_weight_scale_second.data = layer.w13_weight_scale_second.data.transpose(
1, 2).contiguous()
layer.w2_weight_scale_second.data = layer.w2_weight_scale_second.data.transpose(
1, 2).contiguous()
layer.w13_weight_scale_second.data, w13_bias = self.process_scale(
w13_weight_scale_second = layer.w13_weight_scale_second.data if hasattr(
layer, "w13_weight_scale_second") else None
w2_weight_scale_second = layer.w2_weight_scale_second.data if hasattr(
layer, "w2_weight_scale_second") else None
layer.w13_weight_scale.data, w13_bias = self.process_scale(
layer.w13_weight, layer.w13_weight_scale.data,
layer.w13_weight_scale_second.data)
layer.w2_weight_scale_second.data, w2_bias = self.process_scale(
w13_weight_scale_second)
layer.w2_weight_scale.data, w2_bias = self.process_scale(
layer.w2_weight, layer.w2_weight_scale.data,
layer.w2_weight_scale_second.data)
w2_weight_scale_second)
if hasattr(layer, "w13_weight_scale_second"):
# scale_second is no longer used, release this part of the memory
del layer.w13_weight_scale_second
del layer.w2_weight_scale_second
del layer.w13_weight_offset_second
del layer.w2_weight_offset_second
self.update_bias(layer, w13_bias, w2_bias)