Skip to content

test: reduce Kimi K3 B300 MegaMoE memory pressure #41228

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
alphabetc1 wants to merge 1 commit into from
+10 −1
Closed

test: reduce Kimi K3 B300 MegaMoE memory pressure #41228

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 10 additions & 1 deletion test/registered/e2e/models/test_kimi_k3_b300.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,8 @@
DSPARK_DRAFT_MODEL = "RadixArk/Kimi-K3-DSpark"
MEGAMOE_URL = "http://0.0.0.0:30000"
MODEL_LOADER_EXTRA_CONFIG = '{"enable_multithread_load": true, "num_threads": 12}'
MEGAMOE_MEM_FRACTION_STATIC = "0.82"
MEGAMOE_MAX_RUNNING_REQUESTS = "24"
MEGAMOE_ENV = {
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK": "8320",
}
Expand Down Expand Up @@ -101,7 +103,14 @@ def setUpClass(cls):
"--dcp-size",
"8",
"--mem-fraction-static",
"0.85",
MEGAMOE_MEM_FRACTION_STATIC,
# Keep the CI-sized GSM8K load below the DSPARK default of 48
# so MegaMoE has enough headroom for FlashInfer autotune and
# decode CUDA graph capture on B300.
"--max-running-requests",
MEGAMOE_MAX_RUNNING_REQUESTS,
"--cuda-graph-max-bs-decode",
MEGAMOE_MAX_RUNNING_REQUESTS,
"--model-loader-extra-config",
MODEL_LOADER_EXTRA_CONFIG,
"--reasoning-parser",
Expand Down
Loading