Branches · educg-net-48294-3190806 / 2025pra_WaitingForYou-2962 · GitLab

zjy-test7

333986a4 · 加了老师72.4的分支 · Jul 15, 2026
xlang-blockm64-kv-int8-native

c1f4aaec · Add native ROCm INT8 KV cache path · Jul 14, 2026
zjy-test9

55f89d69 · disable prefix caching for single-request · Jul 14, 2026
zjy-test8

22e97a27 · rocm: fix gfx936 recognition + enable ROCM_ATTN · Jul 14, 2026
xlang-blockm64-aiter-moe-shared-fusion

350191a3 · Enable isolated AITER MoE shared expert fusion · Jul 14, 2026
zjy-test6

a0128114 · 没测过但几乎不会成功的32 · Jul 14, 2026
zjy-test5

95e0e0d4 · 不可能成真的量化 · Jul 14, 2026
xlang-blockm64-rocm-moe-prefill-tile

28348449 · Tune ROCm MoE prefill tiles · Jul 13, 2026
xlang-blockm64-online-moe-int8

3a9b4d9e · Allow internal experts int8 experiment · Jul 13, 2026
xlang-blockm64-online-w8a8

d5bfa91d · Avoid AMD SMI lookup during W8A8 graph capture · Jul 13, 2026
xlang-blockm64-rocm-native-paged

8827f3c5 · Enable ROCm split attention for native paged decode · Jul 13, 2026
xlang-blockm64-gdn-kkt-stage1

6f701c10 · Expand ROCm GDN KKT kernel staging autotune · Jul 13, 2026
xlang-blockm64-gdn-o-stage1

848e0b32 · Expand ROCm GDN output kernel staging autotune · Jul 13, 2026
xlang-blockm64-gdn-decode-warp2

e96c92d9 · Tune ROCm GDN packed decode parallelism · Jul 13, 2026
zjy-test1

298c544f · test1修为64 · Jul 13, 2026
xlang-blockm64-gdn-h-warp8

7a2bb1fd · Expand ROCm GDN state kernel autotuning · Jul 13, 2026
zjy-test3

eb3b4c64 · 解决_init_scale_buffers没被调用 · Jul 13, 2026
xlang-blockm64-gdnchunk32

616b99a9 · Tune ROCm GDN prefill chunk size · Jul 12, 2026
xlang-blockm64-batch8192

ac2ef161 · Tune ROCm prefill token budget for long prompts · Jul 12, 2026
xlang-blockm64-aiter

56a5057c · Enable ROCm AITER by default · Jul 12, 2026