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

[main][Feature] Support deepseek w4a8 quantization (#2172)

Browse Source 
### What this PR does / why we need it?
Supports Deepseek-R1 w4a8 quantization.
Since R1 w4a8 uses mixed quantization, only the MOE layer uses
w4a8_dynamic quantization, so we added the w4a8_dynamic.py file, which
includes the AscendW4A8DynamicFusedMoEMethod class.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
Adding ut case in `tests/ut/quantization/test_w4a8_dynamic.py` and
`tests/ut/quantization/test_quantizer.py`
Adding e2e case in
`tests/e2e/multicard/test_offline_inference_distributed.py::test_models_distributed_DeepSeek_W4A8DYNAMIC`
to test deepseek w4a8_dynamic quantized model

#### 1.How to get weights using Modelslim
##### Installation steps
Use the branch master, the commit id is:
298e175d69b3b855111a1e09bbe2fcd12fdb4e24
git clone https://gitee.com/ascend/msit.git
cd msit/msmodelslim
bash install.sh

##### The required transformers environment
transformers>=4.48.2

##### Generate w4a8 weights
cd /example/DeepSeek
Command reference: msmodelslim/example/DeepSeek/README.md Execute the
[pre-check](https://gitee.com/ascend/msit/blob/master/msmodelslim/example/DeepSeek/README.md#%E8%BF%90%E8%A1%8C%E5%89%8D%E5%BF%85%E6%A3%80)
and [DeepSeek-R1 w4a8 mix
quantization](https://gitee.com/ascend/msit/blob/master/msmodelslim/example/DeepSeek/README.md#deepseek-r1-w4a8-%E6%B7%B7%E5%90%88%E9%87%8F%E5%8C%96%E5%89%8D%E4%B8%89%E5%B1%82-mlpw8a8-dynamic-%E9%87%8F%E5%8C%96mla%E5%85%B1%E4%BA%AB%E4%B8%93%E5%AE%B6w8a8%E9%87%8F%E5%8C%96%E8%B7%AF%E7%94%B1%E4%B8%93%E5%AE%B6w4a8-dynamic%E9%87%8F%E5%8C%96)
chapter
Reference command：python3 quant_deepseek_w4a8.py --model_path {Original
weight path} --save_path {Generate weight path} --mindie_format

##### Adapt to vllm-ascend
Since mindie_format generates mindie format, some adaptation
modifications are needed for vllm-ascend to use it:
`quant_model_description_w8a8_dynamic.json` rename to
`quant_model_description.json`, and add `"group_size": 256`
Modification in `config.json`：`"model_type":deepseekv2` is changed to
`"model_type":deepseek_v3`; `quantization_config` is removed;
tips:The group_size and weights match. If the w4a8 weights are not
generated using msmodelslim, you can check the group_size in
quantization_config in config.json.

#### 2.How to run w4a8
##### a.How to run eager mode
export VLLM_USE_V1=1 # v1

python -m vllm.entrypoints.openai.api_server --model=$1
--trust-remote-code -tp $2 -dp $3 --enable_expert_parallel
--quantization ascend --port $4 --max-model-len $5 --max-num-seqs $6
--enforce-eager
eg: python -m vllm.entrypoints.openai.api_server
--model=/weightpath/w4a8_4_layer --trust-remote-code -tp 4 -dp 4
--enable_expert_parallel --quantization ascend --port 8002
--max-model-len 5120 --max-num-seqs 128 --enforce-eager

##### b.How to run graph mode
export VLLM_USE_V1=1 # v1
export HCCL_BUFFSIZE=1024

python -m vllm.entrypoints.openai.api_server --model=$1
--trust-remote-code -tp $2 -dp $3 --enable_expert_parallel
--quantization ascend --port $4 --max-model-len $5
--additional_config='{"ascend_scheduler_config":{"enabled":true},"torchair_graph_config":{"enabled":true}}'
eg: python -m vllm.entrypoints.openai.api_server
--model=/weight/dsr1_w4a8_vllm --trust-remote-code -tp 4 -dp 4
--enable_expert_parallel --quantization ascend --port 8002
--max-model-len 5120
--additional_config='{"ascend_scheduler_config":{"enabled":true},"torchair_graph_config":{"enabled":true}}'


- vLLM version: v0.10.0
- vLLM main:
c494f96fbc

---------

Signed-off-by: Wang Kunpeng <1289706727@qq.com>
This commit is contained in:
Wang Kunpeng
2025-08-06 10:17:44 +08:00
committed by GitHub
parent e31b31f9c3
commit 8a59367d0c
9 changed files with 483 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

285
vllm_ascend/quantization/w4a8_dynamic.py
View File

@@ -15,11 +15,22 @@
# limitations under the License.
#
from typing import Any, Dict, Optional
from typing import Any, Callable, Dict, Optional
import numpy as np
import torch
import torch_npu
from vllm.config import get_current_vllm_config
from vllm.distributed import get_ep_group
from vllm.forward_context import get_forward_context
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.ascend_forward_context import FusedMoEState
from vllm_ascend.distributed.parallel_state import get_mc2_group
from vllm_ascend.ops.fused_moe import select_experts
from vllm_ascend.quantization.w8a8_dynamic import (fused_experts_with_all2all,
fused_experts_with_mc2)
from vllm_ascend.torchair.utils import npu_stream_switch, npu_wait_tensor
class AscendW4A8DynamicLinearMethod:
@@ -111,3 +122,275 @@ class AscendW4A8DynamicLinearMethod:
layer.register_parameter("weight_scale_bias", param)
layer.weight.data = torch_npu.npu_convert_weight_to_int4pack(
layer.weight.data.to(torch.int32))
class AscendW4A8DynamicFusedMoEMethod:
"""FusedMoe method for Ascend W4A8_DYNAMIC.
"""
def __init__(self):
self.transpose_weight = True
self.ep_group = get_ep_group()
ascend_config = get_ascend_config()
self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
try:
device_group = get_mc2_group().device_group
# TODO: Try local_rank = ep_group.rank_in_group
local_rank = torch.distributed.get_rank(group=device_group)
backend = device_group._get_backend(torch.device("npu"))
self.moe_all_to_all_group_name = backend.get_hccl_comm_name(
local_rank)
except AttributeError:
self.moe_all_to_all_group_name = ""
@staticmethod
def get_weight(num_experts: int, intermediate_size_per_partition: int,
hidden_sizes: int,
params_dtype: torch.dtype) -> Dict[str, Any]:
param_dict = {}
param_dict["w13_weight"] = torch.empty(num_experts,
2 *
intermediate_size_per_partition,
hidden_sizes,
dtype=torch.int8)
param_dict["w2_weight"] = torch.empty(num_experts,
hidden_sizes,
intermediate_size_per_partition,
dtype=torch.int8)
return param_dict
@staticmethod
def get_dynamic_quant_param(num_experts: int,
intermediate_size_per_partition: int,
hidden_sizes: int,
params_dtype: torch.dtype) -> Dict[str, Any]:
param_dict = {}
config = get_current_vllm_config()
group_size = config.quant_config.quant_description.get(
"group_size", 256)
param_dict["w13_weight_scale"] = torch.empty(
num_experts,
2 * intermediate_size_per_partition,
1,
dtype=params_dtype)
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 // group_size,
dtype=params_dtype)
param_dict["w13_weight_offset_second"] = torch.empty(
num_experts,
2 * intermediate_size_per_partition,
hidden_sizes // group_size,
dtype=params_dtype)
param_dict["w2_weight_scale"] = torch.empty(num_experts,
hidden_sizes,
1,
dtype=params_dtype)
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 // group_size,
dtype=params_dtype)
param_dict["w2_weight_offset_second"] = torch.empty(
num_experts,
hidden_sizes,
intermediate_size_per_partition // group_size,
dtype=params_dtype)
return param_dict
def apply(
self,
layer: torch.nn.Module,
x: torch.Tensor,
router_logits: torch.Tensor,
top_k: int,
renormalize: bool,
use_grouped_topk: bool = False,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
topk_group: Optional[int] = None,
num_expert_group: Optional[int] = None,
custom_routing_function: Optional[Callable] = None,
scoring_func: str = "softmax",
e_score_correction_bias: Optional[torch.Tensor] = None,
is_prefill: bool = True,
enable_force_load_balance: bool = True,
log2phy: torch.Tensor = None,
global_redundant_expert_num: int = 0,
shared_experts: Optional[Any] = None,
quantized_x_for_share: Optional[Any] = None,
dynamic_scale_for_share: Optional[Any] = None,
**kwargs,
) -> torch.Tensor:
assert router_logits.shape[
1] == global_num_experts, "Number of global experts mismatch"
# NOTE: now npu_moe_gating_top_k can only support `group_count=256` pattern
if global_num_experts == 256:
topk_weights, topk_ids, _ = torch_npu.npu_moe_gating_top_k(
router_logits,
k=top_k, # topk currently is 8
bias=e_score_correction_bias,
k_group=topk_group, # fix: 4
group_count=num_expert_group, # fix 8
group_select_mode=
1, # 0: the maximum in the group; 1: topk2.sum(fix)
renorm=0, # 0: softmax->topk(fix); 1: topk->softmax
norm_type=1, # 0: softmax; 1: sigmoid(fix)
# out_flag=False, # todo new api; should the third output be output
# y2_flag=False, # old api; should the third output be output
routed_scaling_factor=1,
eps=float(1e-20))
else:
topk_weights, topk_ids = select_experts(
hidden_states=x,
router_logits=router_logits,
top_k=top_k,
use_grouped_topk=use_grouped_topk,
renormalize=renormalize,
topk_group=topk_group,
num_expert_group=num_expert_group,
custom_routing_function=custom_routing_function,
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias,
)
fused_moe_state = get_forward_context().fused_moe_state
shared_gate_up, shared_dequant_scale = None, None
if shared_experts is not None and fused_moe_state == FusedMoEState.MC2:
with npu_stream_switch("moe_secondary", 0):
npu_wait_tensor(quantized_x_for_share, router_logits)
share_up_out, _ = shared_experts.gate_up_proj(
(quantized_x_for_share, dynamic_scale_for_share))
shared_gate_up, shared_dequant_scale = share_up_out[
0], share_up_out[1]
# this is a naive implementation for experts load balance so as
# to avoid accumulating too much tokens on a single rank.
# currently it is only activated when doing profile runs.
if enable_force_load_balance:
topk_ids = torch.randint_like(topk_ids, 0, global_num_experts)
topk_weights = topk_weights.to(x.dtype)
if fused_moe_state == FusedMoEState.MC2:
return fused_experts_with_mc2(
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_bias=layer.w13_scale_bias,
w2_scale_bias=layer.w2_scale_bias,
topk_weights=topk_weights,
topk_ids=topk_ids,
top_k=top_k,
expert_map=expert_map,
moe_all_to_all_group_name=self.moe_all_to_all_group_name,
log2phy=log2phy,
global_redundant_expert_num=global_redundant_expert_num,
shared_experts=shared_experts,
is_torchair=self.torchair_graph_enabled,
quantized_x_for_share=shared_gate_up,
dynamic_scale_for_share=shared_dequant_scale,
mc2_mask=kwargs.get("mc2_mask", None))
else:
# The current implementation of deepseek moe splits hidden_states
# according to tp_size before they are feed into fused_moe module.
# Therefore, all2all is needed no matter how dp/tp is set so as to
# dispatch/combine tokens.
return fused_experts_with_all2all(
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_bias=layer.w13_scale_bias,
w2_scale_bias=layer.w2_scale_bias,
topk_weights=topk_weights,
topk_ids=topk_ids,
top_k=top_k,
expert_map=expert_map,
ep_group=self.ep_group,
log2phy=log2phy,
global_redundant_expert_num=global_redundant_expert_num,
)
def process_scale(self, weight: torch.Tensor, scale, per_group_scale):
group_num, k, n = weight.shape
per_group_scale = per_group_scale.reshape(group_num, -1, n)
group_num, quantgroup_num, n = per_group_scale.shape
weight_high = weight.to(torch.float32).reshape([group_num, quantgroup_num, -1, n]) * \
per_group_scale.reshape([group_num, quantgroup_num, 1, n])
weight_high = weight_high.reshape([group_num, k, n])
bias = 8 * (weight_high.to(torch.float32) * scale).sum(axis=1)
scale_fp32 = (scale * per_group_scale).to(torch.float16).to(
torch.float32)
scale_fp32_np = scale_fp32.cpu().numpy()
scale_fp32_np.dtype = np.uint32
sscale_uint64 = np.zeros((group_num, quantgroup_num, n * 2),
dtype=np.uint32)
sscale_uint64[..., ::2] = scale_fp32_np
sscale_uint64_buffer = np.frombuffer(sscale_uint64.tobytes(),
dtype=np.int64).copy()
sscale_uint64_tensor = torch.from_numpy(sscale_uint64_buffer).reshape(
group_num, quantgroup_num, n)
sscale_uint64_tensor = sscale_uint64_tensor.npu()
return sscale_uint64_tensor, bias
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()
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_offset.data = layer.w13_weight_offset.data.view(
layer.w13_weight_offset.data.shape[0], -1)
layer.w2_weight_offset.data = layer.w2_weight_offset.data.view(
layer.w2_weight_offset.data.shape[0], -1)
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, bias = self.process_scale(
layer.w13_weight, layer.w13_weight_scale.data,
layer.w13_weight_scale_second.data)
param = torch.nn.Parameter(bias, requires_grad=False)
layer.register_parameter("w13_scale_bias", param)
layer.w2_weight_scale_second.data, bias1 = self.process_scale(
layer.w2_weight, layer.w2_weight_scale.data,
layer.w2_weight_scale_second.data)
param = torch.nn.Parameter(bias1, requires_grad=False)
layer.register_parameter("w2_scale_bias", param)
layer.w13_weight.data = torch_npu.npu_quantize(
layer.w13_weight.data.to(torch.float32),
torch.tensor([1.]).npu(), None, torch.quint4x2, -1, False)
layer.w2_weight.data = torch_npu.npu_quantize(
layer.w2_weight.data.to(torch.float32),
torch.tensor([1.]).npu(), None, torch.quint4x2, -1, False)