Skip to content

Commit cb24491

Browse files
committed
update pd readme & expert_map bug fix
Signed-off-by: liziyu <liziyu16@huawei.com>
1 parent f07c108 commit cb24491

File tree

3 files changed

+9
-11
lines changed

3 files changed

+9
-11
lines changed

examples/disaggregate_prefill_v1/README.md

Lines changed: 5 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -30,15 +30,14 @@ Execution Sequence
3030

3131
* Run prefill server P1 on first node
3232
```shell
33-
export HCCL_IF_IP=`hostname -I|awk -F " " '{print$1}'`
34-
export GLOO_SOCKET_IFNAME="eth0"
33+
export HCCL_IF_IP=172.19.32.175 # node ip
34+
export GLOO_SOCKET_IFNAME="eth0" # network card name
3535
export TP_SOCKET_IFNAME="eth0"
3636
export HCCL_SOCKET_IFNAME="eth0"
3737
export DISAGGREGATED_PREFILL_RANK_TABLE_PATH=/vllm-workspace/vllm-ascend/examples/disaggregate_prefill_v1/ranktable.json
3838
export OMP_PROC_BIND=false
3939
export OMP_NUM_THREADS=100
4040
export VLLM_USE_V1=1
41-
export VLLM_VERSION=0.9.1
4241
vllm serve /data01/deepseek_r1_w8a8_zhw \
4342
--host 0.0.0.0 \
4443
--port 20002 \
@@ -71,15 +70,14 @@ vllm serve /data01/deepseek_r1_w8a8_zhw \
7170

7271
* Run prefill server P2 on second node
7372
```shell
74-
export HCCL_IF_IP=`hostname -I|awk -F " " '{print$1}'`
73+
export HCCL_IF_IP=172.19.241.49
7574
export GLOO_SOCKET_IFNAME="eth0"
7675
export TP_SOCKET_IFNAME="eth0"
7776
export HCCL_SOCKET_IFNAME="eth0"
7877
export DISAGGREGATED_PREFILL_RANK_TABLE_PATH=/vllm-workspace/vllm-ascend/examples/disaggregate_prefill_v1/ranktable.json
7978
export OMP_PROC_BIND=false
8079
export OMP_NUM_THREADS=100
8180
export VLLM_USE_V1=1
82-
export VLLM_VERSION=0.9.1
8381
vllm serve /data01/deepseek_r1_w8a8_zhw \
8482
--host 0.0.0.0 \
8583
--port 20002 \
@@ -113,15 +111,14 @@ vllm serve /data01/deepseek_r1_w8a8_zhw \
113111

114112
* Run decode server d1 on third node
115113
```shell
116-
export HCCL_IF_IP=`hostname -I|awk -F " " '{print$1}'`
114+
export HCCL_IF_IP=172.19.123.51
117115
export GLOO_SOCKET_IFNAME="eth0"
118116
export TP_SOCKET_IFNAME="eth0"
119117
export HCCL_SOCKET_IFNAME="eth0"
120118
export DISAGGREGATED_PREFILL_RANK_TABLE_PATH=/vllm-workspace/vllm-ascend/examples/disaggregate_prefill_v1/ranktable.json
121119
export OMP_PROC_BIND=false
122120
export OMP_NUM_THREADS=100
123121
export VLLM_USE_V1=1
124-
export VLLM_VERSION=0.9.1
125122
vllm serve /data01/deepseek_r1_w8a8_zhw \
126123
--host 0.0.0.0 \
127124
--port 20002 \
@@ -154,15 +151,14 @@ vllm serve /data01/deepseek_r1_w8a8_zhw \
154151

155152
* Run decode server d2 on last node
156153
```shell
157-
export HCCL_IF_IP=`hostname -I|awk -F " " '{print$1}'`
154+
export HCCL_IF_IP=172.19.190.36
158155
export GLOO_SOCKET_IFNAME="eth0"
159156
export TP_SOCKET_IFNAME="eth0"
160157
export HCCL_SOCKET_IFNAME="eth0"
161158
export DISAGGREGATED_PREFILL_RANK_TABLE_PATH=/vllm-workspace/vllm-ascend/examples/disaggregate_prefill_v1/ranktable.json
162159
export OMP_PROC_BIND=false
163160
export OMP_NUM_THREADS=100
164161
export VLLM_USE_V1=1
165-
export VLLM_VERSION=0.9.1
166162
vllm serve /data01/deepseek_r1_w8a8_zhw \
167163
--host 0.0.0.0 \
168164
--port 20002 \

examples/disaggregate_prefill_v1/gen_ranktable.sh

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -8,7 +8,6 @@ while [[ $# -gt 0 ]]; do
88
case "$1" in
99
--ips)
1010
shift
11-
# 收集所有后续参数直到遇到下一个选项或结束
1211
while [[ $# -gt 0 && ! "$1" == --* ]]; do
1312
IPs+=("$1")
1413
shift

vllm_ascend/quantization/w8a8_dynamic.py

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -121,7 +121,10 @@ def fused_experts_with_mc2(
121121
if log2phy:
122122
topk_ids = log2phy[topk_ids]
123123
global_bs = 0
124-
moe_expert_num = len(expert_map) + global_redundant_expert_num
124+
if (expert_map is not None):
125+
moe_expert_num = len(expert_map) + global_redundant_expert_num
126+
else:
127+
moe_expert_num = global_redundant_expert_num
125128
# hidden_states = hidden_states.bfloat16()
126129
kwargs_mc2 = {
127130
"x": hidden_states,

0 commit comments

Comments
 (0)