Skip to content

perf(EP): tune EPv1 MI300X EP8 for DeepSeek-V4-Pro / Kimi-K3, key con… - #539

Open
kudomcho wants to merge 1 commit into
ROCm:mainfrom
kudomcho:perf/mi300x-topk-tuning-v2
Open

perf(EP): tune EPv1 MI300X EP8 for DeepSeek-V4-Pro / Kimi-K3, key con…#539
kudomcho wants to merge 1 commit into
ROCm:mainfrom
kudomcho:perf/mi300x-topk-tuning-v2

Conversation

@kudomcho

Copy link
Copy Markdown
Contributor

…figs by topk

Measured block×warp sweep (~387 geometries per combo) on 8× MI300X for DeepSeek-V4-Pro (top-6) and Kimi-K3 (top-16) at hidden 7168, EP8.

Dispatch: bf16 + fp8_e4m3_fnuz, 7 token counts (64-4096), 2 topk values -> 28 rules (14 per topk).

Combine: bf16 x (fp8_direct_cast, none) x zero-copy, 7 token counts, 2 topk values -> 28 rules (14 per topk).

All 8 sweeps completed with 0 failures.

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

…figs by topk

Measured block×warp sweep (~387 geometries per combo) on 8× MI300X for
DeepSeek-V4-Pro (top-6) and Kimi-K3 (top-16) at hidden 7168, EP8.

Dispatch: bf16 + fp8_e4m3_fnuz, 7 token counts (64-4096), 2 topk values
-> 28 rules (14 per topk).

Combine: bf16 x (fp8_direct_cast, none) x zero-copy, 7 token counts,
2 topk values -> 28 rules (14 per topk).

All 8 sweeps completed with 0 failures.

Co-Authored-By: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant