2178f2829b
[Speculative Decoding] Support suffix decoding ( #6403 )
...
* support suffix decoding
2026-02-26 11:42:05 +08:00
6d3fede240
[OP][Feature] 统一 limit_thinking_content_length CUDA 算子，支持回复长度限制与注入序列 ( #6493 )
...
* Initial plan
* Migrate PRs #6311 , #6129 , #6305 to develop and merge unit tests
Co-authored-by: yuanlehome <23653004+yuanlehome@users.noreply.github.com >
* fix
* update
* fix
* fix ci
* fix ci
* Initial plan
* test: add test_chat_with_response_max_tokens to test_EB_VL_Lite_serving.py
Co-authored-by: yuanlehome <23653004+yuanlehome@users.noreply.github.com >
* test: add disable-thinking case to test_chat_with_response_max_tokens
Co-authored-by: yuanlehome <23653004+yuanlehome@users.noreply.github.com >
* test: add both reasoning_max_tokens and response_max_tokens case
Co-authored-by: yuanlehome <23653004+yuanlehome@users.noreply.github.com >
* fix ci
* fix ci
* fix ci
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com >
Co-authored-by: yuanlehome <23653004+yuanlehome@users.noreply.github.com >
2026-02-25 21:36:50 +08:00
51f812aaa4
fix empty get_padding_offset ( #6462 )
2026-02-12 12:34:23 +08:00
983be007f5
[Feature]support swa & sink Based on appendattn ( #6410 )
...
* support swa & sink Based on appendattn
2026-02-10 18:28:03 +08:00
8b1dd0f360
fix bug ( #6422 )
2026-02-10 14:58:50 +08:00
c776d483e4
[BugFix]fix handle 4 return values from noaux_tc_redundant op ( #6384 )
...
* fix: handle 4 return values from noaux_tc_redundant op
The noaux_tc_redundant CUDA op is defined with 4 outputs in PD_BUILD_STATIC_OP:
- output_tensor (scores)
- topk_values
- topk_indices
- tokens_per_expert_stats_list_out (inplace updated)
The Python code was only unpacking 3 values, causing:
ValueError: too many values to unpack (expected 3)
This fix correctly unpacks all 4 return values, ignoring the inplace
updated tensor which is the same as the input tokens_per_expert_stats_list.
Co-Authored-By: Claude (Claude Opus 4.5) <noreply@anthropic.com >
* fix: make noaux_tc_redundant return 4 values to match OP definition
The PD_BUILD_STATIC_OP defines 4 outputs but the function only returned 3,
causing inconsistent behavior across different Paddle framework versions.
This fix explicitly returns 4 values:
- scores (inplace modified)
- topk_values
- topk_indices
- tokens_per_expert_stats_list (inplace modified via atomicAdd)
Co-Authored-By: Claude (Claude Opus 4.5) <noreply@anthropic.com >
---------
Co-authored-by: Claude (Claude Opus 4.5) <noreply@anthropic.com >
2026-02-09 13:17:47 +08:00
2b4748de4f
[MTP] refactor MTP pre_process ( #6358 )
2026-02-09 10:47:15 +08:00
d6b3c722c1
[KVCache] Storage cache supports c8 model ( #6298 )
...
* Refine cache transfer manager
* Storage cache supports c8 model
2026-02-06 12:01:17 +08:00
e3fb8796b4
Remove MTP rebuil_padding useless code ( #6336 )
2026-02-05 16:28:44 +08:00
29a313a402
[Optimization] Support FA2/FA3/FA4 with attn_mask_q ( #6354 )
...
* support FA4 sm100
* flash attn backend support mask
* flash attn backend run flashmask correct
* add test for flash_attn_backend and flash_attn_func
* check
* add test for fa4
* requirements.txt add fa4 whl
* check test on sm100
* fix CI conflict
* add enable_torch_proxy for flash_mask
* lazy import fa4
* check
* fix tests import
* check test_load_mpt import
2026-02-05 14:39:00 +08:00
36547cfdb3
[Feature] FD_USE_PHI_FP8_QUANT ( #6320 )
...
* add ut
* add use_fd_quant env
* rm mask_per_token_quant
* add make ops list
* USE_FD_FP8_QUANT -> FD_USE_PHI_FP8_QUANT 默认是true
* modify comments
* use bool type
* Add function declaration
2026-02-03 22:33:03 -08:00
6225439778
add PADDLE_ENFORCE ( #6321 )
2026-02-04 10:47:19 +08:00
c745a22420
[Feature] Support Ernie FP8 on sm100 ( the fixed version) ( #6304 )
2026-02-03 17:47:38 +08:00
8277b95fa6
remove speculate_get_padding_offset op ( #6308 )
2026-02-03 15:18:12 +08:00
2ada119a38
[Optimize] optimize mask_quant & swiglu ( #6222 )
...
* optimize mask_quant op speed up 1.5
* fix calculate sequence
* add fused
* rm log
* push kernel code
* add ut
* accuracy ok
* add ue8m0
* add ut
* add merge develop
* rm ut of mask_per_token_quant
2026-02-02 13:52:38 +08:00
030647521a
[Metax] adapt to the latest develop ( #6282 )
2026-01-29 23:21:20 -08:00
6c685c9474
Revert "[Feature] Support Ernie FP8 on sm100 ( #5593 )" ( #6275 )
...
This reverts commit eb80724b71
2026-01-30 11:22:01 +08:00
eb80724b71
[Feature] Support Ernie FP8 on sm100 ( #5593 )
...
* Deepgemm暂时可用版本
* dense部分 e8m0 ok
* EB模型E8M0跑通的版本
* code check
* support 21b-tp2, dev_paddle
* 单机4.5T ep OK的版本
* 修复删除的代码,单机4.5T ep(非cudagraph)
* eb tp
* Support SM100 block-wise FP8 inference
* refine codes, support deepgemm on sm100
* add thirdparty PFCC/DeepGEMM
* fix ep decode
* 使用deepep ue8m0, 解决精度问题
* 修复FP8 TP精度
* Deepgemm升级适配Hopper逻辑
* add ue8m0 kernel
* add ue8m0 kernel
* fix custom_ops/gpu_ops/cpp_extensions.cc
* eb 输出正常
* eb5 text is right
* 目测精度一致
* 自测精度对齐
* 替换masked_per_token_quant, ep精度OK
* 性能提升约30%
* 暂时跑通ep但是有问题
* 自测一致
* rm test fun
* fix ep event
* 图优化算子更新Deepgemm
* fix build
* 暂时绕过deepgemm CI编译问题
* 根据SM区分deepgemm版本
* remove useless code
---------
Co-authored-by: ckl117 <ckl117@163.com >
Co-authored-by: K11OntheBoat <“ruianmaidanglao@163.com ”>
Co-authored-by: fxyfxy777 <fxyfxy777@163.com >
2026-01-29 13:49:54 +08:00
7da5f54fb3
[CI] Add unit test for swap_layout && remove unit test of splitwise_scheduler ( #6250 )
...
* Add unit test for swap_layout
* remove splitwise_scheduler test
2026-01-28 19:20:20 +08:00
7d6c87c29e
[Others] Support constrained decoding when enable_thinking is false ( #6248 )
...
* support constrained decoding when enable_thinking is false
* fix
* fix
* fix
2026-01-28 00:05:17 -08:00
27f8799f04
[Model Runner] Refactor execute_model for GPU async scheduling ( #6176 )
2026-01-28 14:19:33 +08:00
ce06c6dfb3
[BugFix] Fix token_penalty kernel ( #6069 )
...
* fix token_penalty kernel
* try to fix xpu
* fix xpu
* fix unit test
2026-01-28 12:03:05 +08:00
aa57864c5b
remove unneeded para from flash_mask_attention ( #6218 )
2026-01-27 14:04:27 +08:00
b3627b59f8
[Bug Fix] fix mask attention ( #6216 )
2026-01-26 07:46:26 -08:00
adc69c15d0
[Model Runner] Prepare token count and move FA3 initialization into the graph ( #6170 )
...
* prepare for token num and put FA3 init in graph
2026-01-26 12:16:57 +08:00
0966df78dc
[Others] remove stop_nums ( #6182 )
2026-01-26 12:12:47 +08:00
309c7d9764
router support divided roolout ( #6150 )
2026-01-22 10:39:39 +08:00
f4902fe42d
[BugFix] fix wint2 ( #6109 )
...
* fix
* fix
* fix
2026-01-20 21:46:21 +08:00
a4144e0b8e
[Optimization] Avoid unnecessary penalty computation ( #6078 )
2026-01-19 15:24:12 +08:00
bda38aa519
[Speculative Decoding] Support MTP for GLM-4.5-Air ( #6047 )
...
* glm mtp
* add spec neox partial rope
2026-01-16 14:35:24 +08:00
4c92035f2d
[Feature] Unify fp8 block_wise quant ops ( #5991 )
...
* quant stash
* blockwise_quant
* precommit
* rm tensor.cut
* tp ok
* add swiglu
* rm outdate code
* fix activate ut
* change baseline
* fix baseline error
2026-01-15 05:50:37 -08:00
49617d9832
[Feature]Support tag phase token enforce generation ( #6034 )
...
* support tag phase token enforce generation
* optimize note and some feature
* fix sampler unit test
---------
Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com >
2026-01-15 03:59:55 -08:00
6619298b50
【Optim】Optimize grid dimensions using max_tokens_per_expert for MoE models ( #6007 )
...
* update w4afp8
* build.sh ok
* support cuda_graph
* fix
* add test
* fix max_tokens_per_expert
* >=70
* fix
* compute_max_tokens_from_prefix_sum in w4afp8
* compute_max_tokens use cub
2026-01-15 19:18:42 +08:00
e10b51b8c6
[Feature] get_output_kv_signal blocking read mode & send_first_token ( #5836 )
...
* get_output_kv_signal blocking read mode
* send first token before recycle
* xpu get_output_kv_signal blocking read mode
---------
Co-authored-by: Jiang-Jia-Jun <163579578+Jiang-Jia-Jun@users.noreply.github.com >
2026-01-15 14:11:03 +08:00
74d0f1c01f
[Optim] Robust sync status when preempted happens ( #5796 )
...
* [Bug fix] Sync status for caching output cache
* fix
* fix
* fix bug
* fix
* fix
* support xpu
* fix
* fix
* fix
* fix
* fix
* fix ci
* fix ci
* fix xpu
---------
Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com >
2026-01-14 12:07:33 +08:00
ad8d05a8de
[Optimization] Do not compute ATTN padding part in In Cuda graph mode ( #5985 )
2026-01-13 11:32:27 +08:00
223b2f5d86
Support setting communication groups in custom_allreduce and the all-to-all\transpose fused operator during the decoding phase. ( #5917 )
2026-01-12 14:09:39 +08:00
17ef3920f3
remove decoder_num_blocks_device memset ( #5982 )
2026-01-10 21:22:06 +08:00
b8d9daa785
MLA clean code ( #5979 )
2026-01-10 21:05:00 +08:00
00a01ae024
[Feature] Support redundant expert for eplb ( #5918 )
...
* [BugFix] support redundant expert for eplb
* support redundant expert for eplb
* support redundant expert for eplb
* update
* fix ci eplb
2026-01-09 17:13:24 +08:00
a1fc4e249e
[Bugfix] Fix mtp logprob hang problem when include stop_seq ( #5927 )
...
* fix mtp logprob hang when include stop_seq
2026-01-08 14:21:24 +08:00
2be8656c29
[BugFix] fix mtp split kv attetion ( #5920 )
...
* [BugFix] fix mtp split kv attetion
* clean code
* clean code
2026-01-07 04:07:31 -08:00
a76e8ae40c
[Feature] support rdma pd dy-c8 ( #5788 )
...
* add rdma pd dy-c8
* update code
2026-01-07 14:55:25 +08:00
f15df1ec89
Revert cuda check ( #5915 )
...
* commit
* commit
2026-01-07 14:40:18 +08:00
59523b27de
opt w4afp8 ( #5853 )
2026-01-07 12:22:35 +08:00
83ae59431e
[BugFix] fix BatchMLAWithPagedKVCacheKernel O_tmp ( #5895 )
2026-01-06 15:39:06 +08:00
5e729bc2ba
[OPs] ep_moe_expert_dispatch.cu dispatch num_experts_per_rank 5 ( #5890 )
2026-01-06 10:39:35 +08:00
ab553b3b8b
revert cuda_check ( #5883 )
2026-01-05 20:51:31 +08:00
1d3ae7c024
[BugFix] fix w4afp8 tp=8 ( #5868 )
...
* fix w4afp8 tp=8
* fix
2026-01-05 18:59:02 +08:00
ac39c0f887
support fa3 qwen-vl rope ( #5869 )
2026-01-05 15:29:34 +08:00
GoldPancake