Repository navigation
Commit 1903073
committed
fix: disable GLM-5.2 GB300 draft MoE FP8 conversion
Set SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE=0 so the NextN/MTP draft keeps its shipped
precision, and use the 1800 s scheduler watchdog of the GLM-5.2 B200/GB200
recipes for disaggregated prefill and decode.
设置 SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE=0,使 NextN/MTP draft 保持原始发布精度;
分离式 prefill 和 decode 采用 GLM-5.2 B200/GB200 配方的 1800 秒调度 watchdog。
Co-authored-by: Wenyao Gao <edwingao28@users.noreply.github.com>1 parent 5097c7d commit 1903073
3 files changed
Lines changed: 16 additions & 2 deletions
File tree
- inferencex-e2e
- benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb300-fp4/agentx
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
48 | 48 | | |
49 | 49 | | |
50 | 50 | | |
51 | | - | |
| 51 | + | |
52 | 52 | | |
53 | 53 | | |
54 | 54 | | |
| |||
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
87 | 87 | | |
88 | 88 | | |
89 | 89 | | |
| 90 | + | |
90 | 91 | | |
91 | 92 | | |
92 | 93 | | |
| |||
106 | 107 | | |
107 | 108 | | |
108 | 109 | | |
109 | | - | |
| 110 | + | |
110 | 111 | | |
111 | 112 | | |
112 | 113 | | |
| |||
147 | 148 | | |
148 | 149 | | |
149 | 150 | | |
| 151 | + | |
150 | 152 | | |
151 | 153 | | |
152 | 154 | | |
| |||
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
8995 | 8995 | | |
8996 | 8996 | | |
8997 | 8997 | | |
| 8998 | + | |
| 8999 | + | |
| 9000 | + | |
| 9001 | + | |
| 9002 | + | |
| 9003 | + | |
| 9004 | + | |
| 9005 | + | |
| 9006 | + | |
| 9007 | + | |
| 9008 | + | |
| 9009 | + | |
0 commit comments