分支列表

默认分支

0ba6b6aab7 · perf(mla): add opt-in packed Q decode for FP8 and FP16 (#1512) · 最后更新于 2026-09-12 09:04:01 +00:00

分支列表

e55bd9083e · Merge branch 'main' into feat/instanttensor-weight-loader-relanded · 最后更新于 2026-08-19 19:38:34 +00:00    frostbyte_neo

284
4

2b193c9086 · refactor(mla): declare latent_write_sanitizes above the methods, export the fused entry · 最后更新于 2026-08-19 15:27:50 +00:00    frostbyte_neo

284
10

099a54b977 · fix(kda): probe both ops the replay path launches · 最后更新于 2026-08-19 15:24:24 +00:00    frostbyte_neo

284
33

25b14d0162 · ci: add GB300 two-node AIME26 gates for K3-NVFP4 and K3-NVFP4+DSpark · 最后更新于 2026-08-19 11:35:47 +00:00    frostbyte_neo

286
1

406905cc92 · feat(zmq): materialize precomputed multimodal inputs on msgpack ingest · 最后更新于 2026-08-13 13:10:20 +00:00    frostbyte_neo

341
1

534e3aa9e1 · feat(zmq): piggyback a scheduler-load snapshot on slim output batches · 最后更新于 2026-08-13 12:28:09 +00:00    frostbyte_neo

341
1

4fd4ad0d27 · fix(layers): pass a 0-token batch through SiluAndMul without a launch · 最后更新于 2026-08-13 11:07:59 +00:00    frostbyte_neo

341
1

31bb248504 · perf(comm): route generic MNNVL AR patterns to upstream flashinfer · 最后更新于 2026-08-13 05:23:44 +00:00    frostbyte_neo

345
1

c4b6db207f · Merge branch 'main' into perf/k3-latent-moe-multicast-tail · 最后更新于 2026-08-11 15:59:40 +00:00    frostbyte_neo

364
42

107ffb23e1 · Merge branch 'main' into zmq-msgpack-dp · 最后更新于 2026-08-11 14:19:19 +00:00    frostbyte_neo

364
2

e7e276e175 · test(deepseek-v4): stub deep_gemm so the BF16 indexer test runs on AMD · 最后更新于 2026-08-11 13:38:59 +00:00    frostbyte_neo

368
16

874cb314e8 · perf(amd): fold weighted DSA queries on gfx950 · 最后更新于 2026-08-11 10:29:39 +00:00    frostbyte_neo

364
2

3aad031c27 · perf(amd): optimize gfx1250 MLA serving · 最后更新于 2026-08-11 09:55:38 +00:00    frostbyte_neo

364
1

039f3cdf2e · refactor(amd): split gfx950 mxfp4 decode kernels · 最后更新于 2026-08-10 22:49:04 +00:00    frostbyte_neo

375
1

b4f4ae5e08 · fix(runtime): avoid ephemeral rendezvous ports · 最后更新于 2026-08-10 20:22:09 +00:00    frostbyte_neo

376
1

43ebc5b3f4 · perf(amd): optimize gfx1250 attention serving kernels · 最后更新于 2026-08-10 19:45:33 +00:00    frostbyte_neo

378
1

639cfb47b2 · add file trigger · 最后更新于 2026-08-10 17:15:09 +00:00    frostbyte_neo

383
1

6522ea2a81 · fix(kernel): restore gpt-oss MXFP4 MoE on H100 via Marlin expert bias · 最后更新于 2026-08-10 14:05:44 +00:00    frostbyte_neo

382
1

7f3c2d4487 · ci: recover stale torch extension builds · 最后更新于 2026-08-10 11:10:41 +00:00    frostbyte_neo

378
1

a95731c8c8 · ci: bound kernel unit-test commands · 最后更新于 2026-08-10 10:05:47 +00:00    frostbyte_neo

378
1