Add GLM-5.2 NVFP4 B300 SGLang single-node agentic benchmarks - #2268
Merged
Merged
Claude / Claude Code Review
completed
Jul 18, 2026 in 9m 21s
Code review found 1 important issue
Found 5 candidates, confirmed 2. See review comments for details.
Details
| Severity | Count |
|---|---|
| 🔴 Important | 1 |
| 🟡 Nit | 1 |
| 🟣 Pre-existing | 0 |
| Severity | File:Line | Issue |
|---|---|---|
| 🔴 Important | configs/nvidia-master.yaml:9902-9915 |
GLM-5.2 dp-attn arm omits ep, defaulting EP_SIZE to 1 (no expert parallelism) |
| 🟡 Nit | benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh:1-5 |
PR title/description missing required Chinese translation |
Annotations
Check failure on line 9915 in configs/nvidia-master.yaml
claude / Claude Code Review
GLM-5.2 dp-attn arm omits ep, defaulting EP_SIZE to 1 (no expert parallelism)
The high-throughput arm (`{ tp: 8, dp-attn: true, ... }`) omits the `ep` key, so `generate_sweep_configs.py` defaults `EP_SIZE` to 1, and `glm5.2_fp4_b300_sglang.sh` unconditionally passes `--ep-size $EP_SIZE` to the server. This launches attention-DP (dp=8) with ep-size=1, i.e. no expert parallelism across the 8 DP ranks for this MoE model — every other `dp-attn: true` entry in this file (including the DSv4 recipe this PR says it copies) pairs it with an explicit `ep` equal to `tp`. Add `ep: 8`
Check warning on line 5 in benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh
claude / Claude Code Review
PR title/description missing required Chinese translation
This PR's title and description are English-only, violating the AGENTS.md rule that all PR titles/descriptions must be bilingual (title format '<English> / <中文标题>' plus a mirrored Chinese section in the body).
Loading