Skip to content

[Recipe][Docs] Add MXFP4 intermediate variable and scheduler delay advice for Kimi-K2 #1566

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

Merged
valarLip merged 1 commit into from
Jul 12, 2026
Merged

[Recipe][Docs] Add MXFP4 intermediate variable and scheduler delay advice for Kimi-K2 #1566

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
2 changes: 2 additions & 0 deletions recipes/Kimi-K2.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,7 @@ This command follows the native ATOM benchmark configuration for `amd/Kimi-K2.7-
export AITER_QUICK_REDUCE_QUANTIZATION=INT4
export AITER_USE_FLYDSL_MOE_SORTING=1
export AITER_AR_1STAGE_MAX_KB=512
export AITER_MXFP4_INTERMEDIATE=1

HSA_NO_SCRATCH_RECLAIM=1 python -m atom.entrypoints.openai_server \
--model amd/Kimi-K2.7-Code-MXFP4 \
Expand Down Expand Up @@ -77,6 +78,7 @@ python -m atom.entrypoints.openai_server \
- Kimi-K2.5 and Kimi-K2.7-Code use a DeepseekV3-style architecture with MLA attention, so they leverage the same optimized kernels (MLA, FP8 KV cache, etc.) as DeepSeek models.
- For the MXFP4 variant, `HSA_NO_SCRATCH_RECLAIM=1` is recommended for stability.
- Non-MoE layers (attention, shared experts, dense MLPs) remain in BF16 for all quantized variants.
- For large concs, `--scheduler-delay-factor 1` is recommended for boost throughputs.

## Performance baseline

Expand Down
Loading