Skip to content

[AMD][CI] Move the MI35x DeepSeek-V4-Flash nightly to the official 0731 checkpoint (blocked: triton FlashMLA accuracy regression) #36388

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

Draft
michaelzhang-ai wants to merge 7 commits into
base: main
Choose a base branch
from
+5 −1
Draft

[AMD][CI] Move the MI35x DeepSeek-V4-Flash nightly to the official 0731 checkpoint (blocked: triton FlashMLA accuracy regression) #36388

Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
9883451
[AMD][CI] Run the DeepSeek-V4-Flash nightly on the official 0731 chec…
cursoragent Aug 25, 2026
cfeacd6
[AMD][CI] Stop scheduling the DeepSeek-V3-0324 / V3.1 MI30x nightlies
cursoragent Aug 25, 2026
50cd8b5
[AMD][CI] Merge the four MI35x DeepSeek-V3.2 nightly jobs into two
cursoragent Aug 25, 2026
1ae99ef
[AMD][CI] Scope to the ROCm 7.2 nightly only
cursoragent Aug 25, 2026
d544336
[AMD][CI] Keep the DSV4-Flash FP8 test out of the diff
cursoragent Aug 25, 2026
0a2f424
[AMD][CI] Restore the V3-0324/V3.1 MI30x jobs; move that change to it…
cursoragent Aug 25, 2026
6d11d0f
[AMD][CI] Move the MI35x V3.2 job merge to its own PR
cursoragent Aug 27, 2026
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
6 changes: 5 additions & 1 deletion test/registered/amd/test_deepseek_v4_flash_fp4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,10 @@

Both tests share a single launched server.

The 0731 checkpoint bundles a DSpark draft head, but this test serves the target
only: the DSV4 loader skips `mtp.*` outside a draft worker, so the draft weights
never reach VRAM and `--mem-fraction-static 0.90` below stays valid.

Registry: nightly-amd-8-gpu-mi35x-deepseek-v4-flash suite
"""

Expand All @@ -31,7 +35,7 @@
)

DEEPSEEK_V4_FP4_MODEL_PATH = os.environ.get(
"DEEPSEEK_V4_FP4_MODEL_PATH", "deepseek-ai/DeepSeek-V4-Flash"
"DEEPSEEK_V4_FP4_MODEL_PATH", "deepseek-ai/DeepSeek-V4-Flash-0731"
)
SERVER_LAUNCH_TIMEOUT = 3600
FLASHMLA_BACKEND = os.environ.get("SGLANG_HACK_FLASHMLA_BACKEND", "unified_kv_triton")
Expand Down
Loading