Commit Graph

93 Commits

Author SHA1 Message Date
GoldPancake
076c30cb0f fix top_p_candidates and support separate setting of sampling params for mtp (#4189)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
CE Compile Job / Show Code Archive Output (push) Has been cancelled
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
CE Compile Job / CE_UPLOAD (push) Has been cancelled
* fix top_p_candidates

* For separate setting params for mtp

* delete print

* fix
2025-09-28 11:41:20 +08:00
Sunny-bot1
8a506500f3 [BugFix] Fix EP MoE prefill function (#4101) 2025-09-24 00:31:41 -07:00
Yuan Xiaolan
1e86418c4a optimize dy_cfp8's performance (#4145)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
CE Compile Job / Show Code Archive Output (push) Has been cancelled
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Co-authored-by: carryyu <569782149@qq.com>
2025-09-19 09:35:28 +08:00
Yuan Xiaolan
25aa2d94aa cp dynamic Cfp8 (#4120)
* supports dynamic Cfp8

* add unittest

* fix dynamic Cfp8 computing error

* fix Cfp8 for RL load

---------

Co-authored-by: carryyu <569782149@qq.com>
2025-09-17 11:55:47 +08:00
gaoziyuan
b6caf6e622 suppoort expert num 3 pre rank (#4133)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
CE Compile Job / Show Code Archive Output (push) Has been cancelled
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
CE Compile Job / CE_UPLOAD (push) Has been cancelled
2025-09-16 17:34:44 +08:00
chen
2485333f71 ep support logprob (#4089)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
CE Compile Job / Show Code Archive Output (push) Has been cancelled
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
CE Compile Job / CE_UPLOAD (push) Has been cancelled
2025-09-12 21:11:16 +08:00
freeliuzc
2f473ba966 [Feature][MTP]Support MTP for rl-model (#4009)
* qk norm for speculate decode C16

* support mtp in v1_scheduler mode

* support mtp rope_3d

* support mtp features

* add unit test && del some log

---------

Co-authored-by: yuanxiaolan <yuanxiaolan01@baidu.com>
Co-authored-by: xiaoxiaohehe001 <hiteezsf@163.com>
2025-09-10 13:34:37 +08:00
yangjianfengo1
9213a58a06 【Fix bug] w4afp8 的nblock固定为256,并且fa3的append attn 增加mask参数 (#3771) (#3835)
* fix w4afp8

* 增加集中式配置

* codestyle

* fix fa3 append attn
2025-09-03 19:36:45 +08:00
lizhenyun01
bed09ae8f8 fix mask_offset in append_attn (#3745)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
CE Compile Job / Show Code Archive Output (push) Has been cancelled
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Deploy GitHub Pages / deploy (push) Has been cancelled
* fix mask_offset in append_attn

* fix test
2025-08-31 15:03:16 +08:00
Sunny-bot1
fe5d09f9ee [FIX]Fix Machete compile via ENABLE_MACHETE (#3727)
* add ENABLE_MACHETE

* fix

* revert

* update

* pre_commit

* fix

* fix

---------

Co-authored-by: Ayakouji <yuhongh@qq.com>
Co-authored-by: Jiang-Jia-Jun <163579578+Jiang-Jia-Jun@users.noreply.github.com>
Co-authored-by: aquagull <hongyuh@qq.com>
2025-08-30 17:50:17 +08:00
yangjianfengo1
3754a9906d [Feature] block sparse attention (#3668)
* 支持稀疏attn

* fix bug

* code style

* fix moba attn get kv shape

* 修复a100编译

* codestyle

* code style

* code style

* code style

* fix conflict

* 增加单侧

* code style

* 增加eblite 加载时间

* fix bug

* for ci

* for ci

* for ci

* for ci

* 支持mlp block size 128

* 增加小算子单测

* fix 单测 mlp

* 将环境变量加入到config里面

* fix rollout config

* 修复显存

* add test server

* add test server

* fix mlp  最后一层使用full attn
2025-08-29 19:46:30 +08:00
Yuan Xiaolan
c71ee0831c add w4afp8 offline script (#3636) 2025-08-29 17:56:05 +08:00
co63oc
b6edd15d55 fix scaled_gemm_f8_i4_f16_weight_quantize input (#3685) 2025-08-29 11:04:04 +08:00
lifulll
72094d4d82 enable dcu ci (#3402) 2025-08-29 10:23:08 +08:00
Liumengyuan
e93d4cfcdd Add with_output version AppendAttention (#3302)
* get use_output from fd_config

* add clear TODO description

* add mask_offset para to align with develop

* fix bug

* fix use_output logic

* fix sot bug
2025-08-28 17:10:18 +08:00
yangjianfengo1
e81046fdad 【New Feature】集中式支持w4afp8 (#3644)
* 支持tp w4afp8

* code style
2025-08-28 10:53:24 +08:00
周周周
76513f6416 Support 45t fp8 8 GPU (#3659) 2025-08-28 10:52:53 +08:00
Sunny-bot1
479c8b85d3 [Optimize]support machete weight only gemm (#3561)
* support machete weight only gemm

* add generate

* update

* fix

* change file location

* add sm_version limit

* fix

* fix

* fix ci

* fix coverage

* fix xpu
2025-08-28 09:49:58 +08:00
Jiang-Jia-Jun
c694fa2879 Revert "[Feature] block sparse attention (#3209)" (#3647)
This reverts commit 646a0c2fd8.
2025-08-27 17:35:04 +08:00
xiaoxiaohehe001
ad319a87cc support fa3 rope3d (#3622) 2025-08-27 11:31:29 +08:00
yangjianfengo1
646a0c2fd8 [Feature] block sparse attention (#3209)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
CE Compile Job / Show Code Archive Output (push) Has been cancelled
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Deploy GitHub Pages / deploy (push) Has been cancelled
* 支持稀疏attn

* fix bug

* code style

* fix moba attn get kv shape

* 修复a100编译

* codestyle

* code style

* code style

* code style

* fix conflict

* 增加单侧

* code style

* 增加eblite 加载时间

* fix bug

* for ci

* for ci

* for ci

* for ci

* 支持mlp block size 128

* 增加小算子单测

* fix 单测 mlp

* 将环境变量加入到config里面

* fix rollout config
2025-08-26 07:16:04 -07:00
gaoziyuan
82e64b13e1 [NewFeature]Support dp multi api server && Fix some bug in mixed ep && merge develop (#3598)
* [Feature] update ep

* fix ci

* fix ci

* fix ci

* fix ci

* fix ci

* fix ci

* fix ci

* fix queue ports idx

* fix ci

* fix ci

* fix ci

* fix ci

* fix ci

* fix ci

* fix ci

* fix ci

* Update engine.py

* fix ci

* fix some bug in mixed ep

* add server fix and op fix

* rm some log

* fix code style

* ltd fix

* fix

* fix

* fix some bug

* fix bug

* fix bug

* fix style

* Update config.py

* Update splitwise_connector.py

* Update cache_messager.py

* Update __init__.py

* merge and fix

* Update engine.py

* Update common_engine.py

* Update run_ci_xpu.sh

* Update ernie_processor.py

* Update ernie_processor.py

---------

Co-authored-by: ltd0924 <ltd0924@sina.com>
Co-authored-by: ltd0924 <32387785+ltd0924@users.noreply.github.com>
2025-08-26 19:59:02 +08:00
lzy
d339df2e90 Supports DP+TP+EP hybrid parallel deployment strategy (#3489)
* Support DP+TP+EP hybrid parallel deployment strategy

* Support DP+TP+EP hybrid parallel deployment strategy

* fix conflict

* add moe_tp_ep function split_allgather_out

* del tp_group in moe_cutlass_backend

* for ci

* fix parallel_config for ci

* del log
2025-08-26 00:04:01 -07:00
freeliuzc
52eda7fdb3 [Feature][MTP]support new speculative decoding method named hybrid mtp with ngram (#3610) 2025-08-26 14:29:22 +08:00
xiaoxiaohehe001
70c75798a7 [NewFeatures] support noex rope3d (#3542)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
CE Compile Job / Show Code Archive Output (push) Has been cancelled
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Deploy GitHub Pages / deploy (push) Has been cancelled
* [NewFeatures] support noex rope3d

* [NewFeatures] support noex rope3d encoder
2025-08-26 11:44:57 +08:00
Yuan Xiaolan
9205c88da1 support w4afp8 EP inference (#3044)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
CE Compile Job / Show Code Archive Output (push) Has been cancelled
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Deploy GitHub Pages / deploy (push) Has been cancelled
2025-08-25 11:27:45 +08:00
freeliuzc
76759108c9 [Feature][SpeculativeDecoding]Support tree-attention (#3514)
* support tree-attention

* fix merge bug

* fix unit-test api

* fix merge bug
2025-08-22 13:36:41 +08:00
yangjianfengo1
e5aa7087db 【bug fix】修复w4a8编译慢 (#3510)
* 修复w4a8编译

* code style

* 修复tma copy
2025-08-21 18:50:14 +08:00
Ryan
bcdfc1d6b9 Add custom op declaration for all_reduce (#3473)
* add custom op declaration

* roll back try except
2025-08-20 20:29:58 +08:00
yangjianfengo1
b047681c5d 【New Feature】支持Fp8 group Gemm 24稀疏 (#3463)
Some checks failed
Deploy GitHub Pages / deploy (push) Has been cancelled
Publish Job / publish_pre_check (push) Has been cancelled
Publish Job / print_publish_pre_check_outputs (push) Has been cancelled
Publish Job / FD-Clone-Linux (push) Has been cancelled
Publish Job / Show Code Archive Output (push) Has been cancelled
Publish Job / BUILD_SM8090 (push) Has been cancelled
Publish Job / BUILD_SM8689 (push) Has been cancelled
Publish Job / PADDLE_PYPI_UPLOAD_8090 (push) Has been cancelled
Publish Job / PADDLE_PYPI_UPLOAD_8689 (push) Has been cancelled
Publish Job / Run FastDeploy Unit Tests and Coverage (push) Has been cancelled
Publish Job / Run FastDeploy LogProb Tests (push) Has been cancelled
Publish Job / Extracted partial CE model tasks to run in CI. (push) Has been cancelled
Publish Job / Run Base Tests (push) Has been cancelled
Publish Job / Run Accuracy Tests (push) Has been cancelled
* 支持24稀疏

* code style

* 增加stmatrix 宏定义判断

* code style
2025-08-19 02:54:47 -07:00
lizexu123
32b39620bc [Code Simplification] remove cum_offsets (#3410)
Some checks failed
Deploy GitHub Pages / deploy (push) Has been cancelled
Publish Job / publish_pre_check (push) Has been cancelled
Publish Job / print_publish_pre_check_outputs (push) Has been cancelled
Publish Job / FD-Clone-Linux (push) Has been cancelled
Publish Job / Show Code Archive Output (push) Has been cancelled
Publish Job / BUILD_SM8090 (push) Has been cancelled
Publish Job / BUILD_SM8689 (push) Has been cancelled
Publish Job / PADDLE_PYPI_UPLOAD_8090 (push) Has been cancelled
Publish Job / PADDLE_PYPI_UPLOAD_8689 (push) Has been cancelled
Publish Job / Run FastDeploy Unit Tests and Coverage (push) Has been cancelled
Publish Job / Run FastDeploy LogProb Tests (push) Has been cancelled
Publish Job / Extracted partial CE model tasks to run in CI. (push) Has been cancelled
Publish Job / Run Base Tests (push) Has been cancelled
Publish Job / Run Accuracy Tests (push) Has been cancelled
2025-08-18 20:21:25 +08:00
Jundong Liu
ea4a3b479c [Excutor] Increase buffer size to prevent address corruption; add forward metadata debug tool (#3404)
* 修复buffer申请不够大,增加打印forwardmetadata的工具

* fix mistake

* Make CPU tensor in CPUPlace

* Add test about forward_meta_str and Add unitest_requirement

---------

Co-authored-by: RAM <gstian5555@outlook.com>
2025-08-18 16:14:09 +08:00
freeliuzc
a12d0bc549 [Feature][MTP]update multi-draft-token strategy (#3369)
* update multi-draft-token strategy

* fix format

---------

Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com>
2025-08-18 13:59:56 +08:00
Sunny-bot1
6c1f3ff897 topk_gating_softmax support bias (#3405) 2025-08-15 11:57:45 +08:00
chen
f0f00a6025 [OPs] Universal optimization and Fix early_stop cuda 700 (#3375)
Some checks failed
Deploy GitHub Pages / deploy (push) Has been cancelled
* delete nonzero

* delete setup_ops_base.py

* check if

* check gcp infer_seed.cpu()

* fix repetition_early_stopper_kernel cuda 700
2025-08-14 22:40:44 +08:00
lzy
1e06b9fa6d make append_attn supports mask_offset (#3138)
* make append_attn supports mask_offset

* add unittest
2025-08-14 03:40:55 -07:00
Sunny-bot1
2e7831185f [Optimize]Add norm_weights feature for topk_gating_softmax (#3372)
Some checks failed
Deploy GitHub Pages / deploy (push) Has been cancelled
2025-08-14 15:05:23 +08:00
Kane2011
b4fef2cf29 [MetaxGPU] Support FastDeploy on metax gpu (#3241)
* [MetaxGPU] Support FastDeploy on metax gpu

* Update metax_worker.py

1. change worker log;
2. remove custom allreduce, adapt it later;
3. remove cuda graph;

* Update __init__.py

1. remove metax's key work comment

* Update __init__.py

1. remove metax's key word comment;
2. add fused_moe_kernel_paddle import

---------

Co-authored-by: yongqiangma <xing.wo@163.com>
2025-08-13 11:11:54 +08:00
Ryan
ed6bff215a fix custom op order rms_norm_eps (#3348) 2025-08-13 10:12:49 +08:00
Sunny-bot1
8224b21525 Refactor moe_topk_select op to use apply_norm_weight as a template parameter (#3345)
* Refactor moe_topk_select op to use apply_norm_weight as a template parameter

* update test
2025-08-13 08:44:16 +08:00
Jundong Liu
1e4968e810 [Excutor] Fixed the issue of CUDA graph execution failure caused by different branches during decoding (#3223)
Some checks failed
Deploy GitHub Pages / deploy (push) Has been cancelled
* 彻底解决解码切块问题

* update C8 and C4 kernel

* fix problem

* fix with pre-commit

* retain branch for mtp
2025-08-09 07:37:19 +08:00
gaoziyuan
a799d14df1 [Bugfix] Fix model accuracy in some ops (#3231)
* fix noaux_tc op

* fix

* update

* fix qk norm

* fix linear for prequant loader

* test

* fix

* fix

* rm some print

* fix noaux_tc op

* test

* Fix the confused enable_early_stop when only set early_stop_config (#3214)

* fix the confused early_stop_config when only set early_stop_config

* pre-commit

* write a general method

* Add ci case for min token and max token (#3229)

Co-authored-by: xujing43 <xujing43@baidu.com>

* add some evil cases (#3240)

* add repitation early stop cases

* add repitation early stop cases

* add bad cases

* add bad cases

* add evil cases

* qwen3_moe (#3084)

* [Feature] support seed parameter (#3161)

* support seed

* fix

* add SamplingMetadata seed test

* The next_tokens values are inconsistent!

* add air and rejection seed test

* fix

* add SamplingParams seed test

* fix seed=0

* Default to defualt

* fix

* fix args_utils

* fix review

* fix review

* fix

* fix

* add xpu,gcu,iluvatar support seed

* fix

* 【Fix Bug】 修复 fa3 支持集中式bug (#3235)

* fix fa3 集中式bug

* 增加qknorm参数

* fix qk norm

* fix

* update

* fix linear for prequant loader

* fix

* fix

* rm some print

* fix

* fix moe init weight&scale

* fix moe init weight&scale

---------

Co-authored-by: bukejiyu <395822456@qq.com>
Co-authored-by: yuanxiaolan <yuanxiaolan01@baidu.com>
Co-authored-by: Zero Rains <linjunlu@zerorains.top>
Co-authored-by: xjkmfa <108254620+xjkmfa@users.noreply.github.com>
Co-authored-by: xujing43 <xujing43@baidu.com>
Co-authored-by: Divano <dddivano@outlook.com>
Co-authored-by: bukejiyu <52310069+bukejiyu@users.noreply.github.com>
Co-authored-by: lizexu123 <39205361+lizexu123@users.noreply.github.com>
Co-authored-by: yangjianfengo1 <125249383+yangjianfengo1@users.noreply.github.com>
Co-authored-by: qingqing01 <dangqingqing@baidu.com>
2025-08-08 17:30:37 +08:00
yzwu
fbdd6b0663 [Iluvatar GPU] Optimze attention and moe performance (#3234) 2025-08-08 10:51:24 +08:00
yangjianfengo1
89397516a8 [New Feature] Support W4Afp8 MoE GroupGemm (#3171)
* init

* 增加多线程编译

* fix bug

* fix bug

* code style

* 增加fp16

* 将print替换成assert

* 修复stmatrix

* 减小单测shape

* 减小单测shape
2025-08-06 10:34:05 +08:00
Yuan Xiaolan
7ce00e597c support qk norm (#3145) 2025-08-05 16:46:14 +08:00
Yuan Xiaolan
af543b7f0f revise get_moe_scores (#3164) 2025-08-05 16:43:07 +08:00
yangjianfengo1
40f7f3e0d8 [New Feature] fa3 支持flash mask (#3184)
* 支持flash mask

* 修改test_flash_mask

* 修改test.sh
2025-08-05 12:20:48 +08:00
RichardWooSJTU
e39159f3bd Add switch to apply fine-grained per token quant fp8 (#3192)
Co-authored-by: yuanxiaolan <yuanxiaolan01@baidu.com>
2025-08-04 19:54:03 -07:00
AIbin
22fe695f1c 【Inference Optimize】Support automatic generation of marlin kernel (#3149)
* Support automatic generation of marlin kernel
2025-08-01 22:43:18 +08:00
yangjianfengo1
64d7a3194d 集中式支持fa3 (#3112) 2025-08-01 18:03:36 +08:00