bench: extend MHC benchmarks with DeepSeek V4 production shapes #3

NiseAzaka · 2026-08-04T17:53:43+08:00

NiseAzaka commented

2026-08-04 17:53:43 +08:00

bench: extend MHC benchmarks with DeepSeek V4
production shapes
Align benchmark cases with production deployment targets :

DeepSeek V4 (arXiv:2606.19348 + officialconfig.json): V4-Pro hidden_size=7168, V4-Flash hidden_size=4096, hc_mult=4;
mHC paper Table 5 (arXiv:2512.24880): legacy tiers 1280/1920/2560;
production batch = num_tokens: Decode 32/128/512, Prefill 1024/4096/8192.

bench: extend MHC benchmarks with DeepSeek V4 production shapes Align benchmark cases with production deployment targets : DeepSeek V4 (arXiv:2606.19348 + officialconfig.json): V4-Pro hidden_size=7168, V4-Flash hidden_size=4096, hc_mult=4; mHC paper Table 5 (arXiv:2512.24880): legacy tiers 1280/1920/2560; production batch = num_tokens: Decode 32/128/512, Prefill 1024/4096/8192.

NiseAzaka added 1 commit 2026-08-04 17:53:44 +08:00

auto-label / apply-label (pull_request) Has been cancelled Details

f910eabd01 bench: extend MHC benchmarks with DeepSeek V4

production shapes
 Align benchmark cases with production deployment targets :

 - DeepSeek V4 (arXiv:2606.19348 + officialconfig.json):
    V4-Pro hidden_size=7168, V4-Flash hidden_size=4096, hc_mult=4;
 - mHC paper Table 5 (arXiv:2512.24880): legacy tiers 1280/1920/2560;
 - production batch = num_tokens: Decode 32/128/512, Prefill 1024/4096/8192.

eryi merged commit 15b9c38fdb into group4/stable

2026-08-04 19:17:05 +08:00

eryi referenced this issue from a commit

2026-08-04 19:17:07 +08:00

Merge pull request 'bench: extend MHC benchmarks with DeepSeek V4 production shapes' (#3) from zjs/dev1 into group4/stable

Sign in to join this conversation.

No reviewers

No Label

No Milestone

No project

No Assignees

1 Participants

Notifications

Due Date

The due date is invalid or out of range. Please use the format 'yyyy-mm-dd'.

No due date set.

Dependencies

No dependencies set.

Reference: eryi/TileOPs-Metax#3