Commit Graph
100 Commits
Author SHA1 Message Date
fzyzcjy a291439a59 Support logprobs in two-batch overlap (#7709) 2025-07-05 19:05:32 -07:00
fzyzcjy 0c9c6c75a8 Move files related to EPLB (#7580) 2025-06-29 15:39:38 -07:00
fzyzcjy 00c7b1ad07 Let EP prefill support new DeepGEMM (#7310) 2025-06-28 01:45:30 -07:00
fzyzcjy 82eccae44e Let ep_scatter support arbitrary strides / ue8m0 format (#7309) 2025-06-28 01:38:33 -07:00
fzyzcjy 9efb2993da Tiny add logs for expert location updater (#7308) 2025-06-27 19:12:33 -07:00
fzyzcjy 3d7cdb2ebd Fix MTP error when enabling two-batch overlap (#7569) 2025-06-26 15:40:54 -07:00
fzyzcjy edc21cc8ae Tiny add logging for GC (#7406) 2025-06-22 12:40:02 +08:00
fzyzcjy 650127a173 Reintroduce tiny fix sampler error when prob is not contiguous (#7354) 2025-06-19 09:39:34 -07:00
fzyzcjy 9c6a0656a3 Fix profiler error when there are idle passes (#7003) 2025-06-18 10:55:01 -07:00
fzyzcjy 91a066ec6a Tiny remove comments about DeepEP on H20 (#7234) 2025-06-16 09:13:57 -07:00
fzyzcjy 349bb2c92a Fix error when disabling new DeepGEMM (#7198) 2025-06-14 19:24:54 -07:00
fzyzcjy 0b8939bc2c Fix NCCL 2.27.3 not in docker image (#7195) 2025-06-14 18:36:11 -07:00
fzyzcjy b57d87c297 Fix shared experts fusion + weight requant (#7177) 2025-06-14 02:35:18 -07:00
fzyzcjy 93cec4335f Support new DeepGEMM (#7172) 2025-06-13 23:00:17 -07:00
fzyzcjy b4c41f7276 Refactor DeepGEMM integration (#7150) 2025-06-13 20:41:03 -07:00
fzyzcjy 8b8f2e7463 Support new DeepGEMM input format in silu_and_mul_masked_post_quant_fwd (#7153) 2025-06-13 20:40:24 -07:00
fzyzcjy 5b1afa7814 Re-quantize DeepSeek model weights to support DeepGEMM new input format (#7156) 2025-06-13 15:57:45 -07:00
fzyzcjy c49c1d9226 Remove 200us slow concat kernel (part 2: srt) (#7020) 2025-06-13 15:19:31 -07:00
fzyzcjy 0f1dfa1efe Tiny add sanity checks for DeepGEMM inputs (#7157) 2025-06-13 14:36:27 -07:00
fzyzcjy e3ec6bf4b6 Minor speed up block_quant_dequant (#6814) 2025-06-13 14:32:46 -07:00
fzyzcjy b04df75acd Fix DeepEP error in some environments (#7154) 2025-06-13 14:27:25 -07:00
fzyzcjy bec3e48402 Support new DeepGEMM format in per token group quant (part 2: srt) (#7155) 2025-06-13 14:25:40 -07:00
fzyzcjy 5c66c4424f Support new DeepGEMM format in per token group quant (#7146) 2025-06-13 02:00:22 -07:00
fzyzcjy aa46ed34d2 Remove 200us slow concat kernel (part 1: kernel) (#7145) 2025-06-13 01:58:29 -07:00
fzyzcjy da47621ccc Minor speedup topk postprocessing (#7058) 2025-06-13 00:50:18 -07:00
fzyzcjy b02df20a8d Minor add utility to read expert distribution recorder output (#7134) 2025-06-12 16:50:05 -07:00
fzyzcjy fe55947acd Add sanity checks when a test file is not added to CI (#6947) 2025-06-10 12:34:57 -07:00
fzyzcjy 19995dd78e Tiny fix cutlass_mla_get_workspace_size stub incorrect signature (#7057) 2025-06-10 12:27:57 -07:00
fzyzcjy f6ebba537a Support both approximate and exact expert distribution collection (#6964) 2025-06-09 20:56:17 -07:00
fzyzcjy e58423b2b9 Fix cutlass MLA gets almost zero accuracy (#6998) 2025-06-09 10:16:29 -07:00
fzyzcjy de1350ea20 Minor remove one kernel for DeepSeek (#6977) 2025-06-08 17:41:35 -07:00
fzyzcjy 86fe943bc3 Fix expert distribution dumping causes OOM (#6967) 2025-06-08 17:41:14 -07:00
fzyzcjy 2fc1299562 Remove unnecessary kernels of num_token_non_padded (#6965) 2025-06-08 05:09:17 -07:00
fzyzcjy f5599ef124 Refactor global_server_args_dict (#6866) 2025-06-07 03:10:35 -07:00
fzyzcjy c499591ac8 Add canary for EPLB rebalancing (#6895) 2025-06-07 03:09:33 -07:00
fzyzcjy d5c097a2f9 Tiny re-introduce profile id logging (#6912) 2025-06-07 02:32:50 -07:00
fzyzcjy 2f715f51cc Minor compile fused topk (#6944) 2025-06-07 01:40:38 -07:00
fzyzcjy 61ce91ed28 Tiny support customize DeepEP max dispatch tokens per rank (#6934) 2025-06-06 17:18:35 -07:00
fzyzcjy 35b65cf0ca Use deepgemm instead of triton for fused_qkv_a_proj_with_mqa (#6890) 2025-06-05 11:37:05 -07:00
fzyzcjy 0166403c20 Support Blackwell DeepEP docker images (#6868) 2025-06-05 00:07:53 -07:00
fzyzcjy bcf66ef3e1 Tiny allow profiler API to auto create directory (#6865) 2025-06-05 00:07:03 -07:00
fzyzcjy 0de5e7d40f Support layerwise rebalancing experts (#6851) 2025-06-05 00:05:52 -07:00
fzyzcjy 72a110f664 Tiny update error hints (#6846) 2025-06-05 00:05:28 -07:00
fzyzcjy 5aff1e9392 Fix Qwen3MoE missing token padding optimization (#6820) 2025-06-05 00:04:59 -07:00
fzyzcjy b6d0ce9f78 Minor add metrics to expert location updater (#6816) 2025-06-02 23:59:11 -07:00
fzyzcjy 0ea330ca34 Fix wrong weight reference in dynamic EPLB (#6818) 2025-06-02 23:26:04 -07:00
fzyzcjy df7f61ee7d Speed up rebalancing when using non-static dispatch algorithms (#6812) 2025-06-02 11:18:17 -07:00
fzyzcjy ef21729c1d Fix profiles do not have consistent names (#6811) 2025-06-02 11:17:22 -07:00
fzyzcjy f5159315b2 Add simple utility to dump tensors for debugging (#6815) 2025-06-02 11:15:31 -07:00
fzyzcjy 6d7b6696d4 Tiny fix EPLB assertion about rebalancing period and recorder window size (#6813) 2025-06-02 11:13:33 -07:00
fzyzcjy 6376b632eb Tiny log prefill time (#6780) 2025-06-02 10:28:27 -07:00
fzyzcjy e05e29d178 Refactor CustomOp to avoid confusing bugs (#5382) 2025-06-02 10:27:36 -07:00
fzyzcjy 2c3b71d678 Improve EPLB logical to physical dispatch map (#6727) 2025-05-29 19:23:54 -07:00
fzyzcjy 6df81e8a39 Support tuning DeepEP configs (#6742) 2025-05-29 08:12:22 -07:00
fzyzcjy 3ab7d9b55e Support picking variants of EPLB algorithms (#6728) 2025-05-29 08:12:01 -07:00
fzyzcjy 7e5071c92a Super tiny enable sole usage of expert distribution metrics and update doc (#6680) 2025-05-29 08:11:38 -07:00
fzyzcjyandCheng Wan 31589e177e Speed up when having padding tokens two-batch overlap (#6668)
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
2025-05-28 16:00:58 -07:00
fzyzcjy ae6a5b2950 Minor refactor two-batch overlap (#6682) 2025-05-28 15:54:17 -07:00
fzyzcjy 4839999b76 Overlap two kernels in DeepSeek with communication (#6711) 2025-05-28 15:53:51 -07:00
fzyzcjy 541a985f85 Fuse routed_scaling_factor in DeepSeek (#6710) 2025-05-28 15:53:37 -07:00
fzyzcjy 87068b5cc7 Support gathering expert distribution details (#6665) 2025-05-27 15:32:59 -07:00
fzyzcjy a564e001b5 Fix DeepEP error in Qwen 3 MoE models (#6673) 2025-05-27 15:12:54 -07:00
fzyzcjy 673ff668f7 Speed up expert location update (#6661) 2025-05-27 10:00:09 -07:00
fzyzcjy 447be24228 Fix OOM when updating expert locations (#6660) 2025-05-27 09:59:53 -07:00
fzyzcjy 1a8f5f6836 Super tiny rename environment variable (#6648) 2025-05-26 21:01:16 -07:00
fzyzcjy 32cd707002 Support TP in attention for two batch overlap (#6634) 2025-05-26 20:28:12 -07:00
fzyzcjy ebd1ed49d4 Tiny refactor communicator (#6646) 2025-05-26 20:24:17 -07:00
fzyzcjy ca95556c76 Tiny fix sampler error when prob is not contiguous (#6639) 2025-05-26 19:19:08 -07:00
fzyzcjy 0ca3e56802 Tiny fix missing expert location dispatch info (#6620) 2025-05-26 08:58:31 -07:00
fzyzcjy 5c7aa00976 Fix EPLB algorithm fail to run when using 3 nodes for prefill (#6629) 2025-05-26 08:43:24 -07:00
fzyzcjy fe386acae6 Automatically configure for EPLB-related args (#6628) 2025-05-26 08:42:49 -07:00
fzyzcjy 501efc3d36 Tiny fix CI (#6611) 2025-05-25 23:36:34 -07:00
fzyzcjy 84147254c9 Tiny fix lint CI does not trigger on master (#6609) 2025-05-25 22:47:03 -07:00
fzyzcjy 6bebef60a7 Support accurate length control for bench serving (#6594) 2025-05-25 22:46:23 -07:00
fzyzcjyandyizhang2077 25be63d0b2 Auto handle PD disaggregation in bench_serving (#6587)
Co-authored-by: yizhang2077 <1109276519@qq.com>
2025-05-25 22:41:27 -07:00
fzyzcjy d502dae0f0 Tiny change killall_sglang.sh (#6596) 2025-05-25 22:36:51 -07:00
fzyzcjy 93e53f6e0b Logging and minor fixes to two batch overlap and EPLB (#6595) 2025-05-25 22:36:40 -07:00
fzyzcjy a191a0e47c Improve performance of two batch overlap in some imbalanced cases (#6593) 2025-05-25 22:36:18 -07:00
fzyzcjy 8c7279c24e Fix profiling will crash the server when using num_steps (#6586) 2025-05-25 22:36:02 -07:00
fzyzcjy 0ca1811715 Support fake perfectly balanced EP dispatch algorithm (#6571) 2025-05-25 22:35:51 -07:00
fzyzcjy 2c3a6fe1de Fix bench_serving does not support changing warmup requests (#6439) 2025-05-25 22:35:36 -07:00
fzyzcjy 5ccf8fe1a0 Hint users when weight update timeouts (#6570) 2025-05-25 09:13:17 -07:00
fzyzcjy 0d47788025 Support overlapping two batches (#4068) 2025-05-24 17:39:07 -07:00
fzyzcjy f456037396 Utilize static dispatching for communicator (#6577) 2025-05-24 17:34:35 -07:00
fzyzcjy b2388433be Add back DeepSeek non-TBO branches (#6578) 2025-05-24 17:34:00 -07:00
fzyzcjy a38376fa99 Refactor attention into multiple stages (#6477) 2025-05-24 17:33:25 -07:00
fzyzcjy c4831e2fcf Fix accuracy is zero when enabling moe-dense-tp-size as in large scale EP (#6567) 2025-05-24 00:27:10 -07:00
fzyzcjy 7a80f56513 Support dynamically rebalancing experts using EPLB (#6469) 2025-05-21 23:13:21 -07:00
fzyzcjy 9484eba4ad Support logging expert balancedness metrics (#6482) 2025-05-21 23:05:33 -07:00
fzyzcjy fc992a09f9 Support updating expert locations dynamically (#6388) 2025-05-21 21:59:33 -07:00
fzyzcjy 969660c762 Recover from corrupted cache file in bench serving (#6510) 2025-05-21 17:13:54 -07:00
fzyzcjy 55f6005f53 Fix bench_one_batch_server (#6503) 2025-05-21 11:08:17 -07:00
fzyzcjy 7222e1dacc Let bench_one_batch_server use sharegpt data to make expert distribution more natural (#5573) 2025-05-21 02:08:43 -07:00
fzyzcjy 505eec4dc9 Tiny make Lint CI show diff (#6445) 2025-05-21 02:06:25 -07:00
fzyzcjy ccfe5c009d Support redundant experts in expert parallel (#6461) 2025-05-21 02:05:53 -07:00
fzyzcjy a071dc4084 Tiny add stage assertions to DeepEPDispatcher to avoid misuse (#6467) 2025-05-21 02:05:05 -07:00
fzyzcjy a40aecc5a3 Fix num_qps_per_rank computation when providing custom DeepEP configuration (#6468) 2025-05-21 02:04:33 -07:00
fzyzcjy d6e1d28c8a Refactor DeepSeek attention dispatching (#6476) 2025-05-21 02:03:39 -07:00
fzyzcjy 13feffd082 Fix master CI for DeepSeek (#6447) 2025-05-20 00:31:42 -07:00
fzyzcjy e98afbe042 Support dispatching logical to physical experts (#6385) 2025-05-19 22:13:55 -07:00