feat: do softmax for lmhead - #256

Open

HarmonyHu wants to merge 938 commits into

sophgo:masterfrom

HarmonyHu:master

This pull request is big! We're only showing the most recent 250 commits

Commits on Aug 22, 2025

feat: add new tpulang ops

liang.wang01
authored and
charlesxzb
committed
6f3b282
feat: support IndexputOp

tpu-mlir-vis-tools
authored and
charlesxzb
committed
0933464
fix: fix bm1688 llama2-7b bmrt_test cmp failed prob
charlesxzb
committed
57ebff3
fix: fix nodechip_interp

chang.zhao
authored and
charlesxzb
committed
f65a88f
chore: tpuc-opt supports set time limit
charlesxzb
committed
d54a962
fix: LayerGroup multi-thread bug

yifan.chen
authored and
charlesxzb
committed
096beef

Commits on Aug 25, 2025

feat: support share prompt
HarmonyHu
committed
ae46d2e
feat: make yolo segment mask to 0-255

jianmin.xu
authored and
charlesxzb
committed
da3c97e
feat: add w4a8 mode

shee-gao
authored and
WhoIsYu
committed
0d5ce82

Commits on Aug 26, 2025

set logic output scale to 1.0
WhoIsYu
committed
1768860
Add a new pattern to recognize rope

wwwwwu64
authored and
WhoIsYu
committed
260c1b5
Add eva02 block pattern in MatchPattern

wwwwwu64
authored and
WhoIsYu
committed
3188615
fix: pool2d cv181x

zijing.wan
authored and
charlesxzb
committed
54f50e6
fix: A16MatMul inference error
HarmonyHu
committed
4cc3a72
fix: walkaround group_size=1 w8a16 bug
HarmonyHu
committed
3b4bcae
feat: append fc_fix8b multi-core
PostMalone1998
committed
f9accc7

Commits on Aug 27, 2025

fix: qtable bm1684
zijing.wan
committed
7d193fa
fix: qtable bm1684
zijing.wan
committed
55c58b7
revert: A16MatMul keep sign
HarmonyHu
committed
2296dcf

Commits on Aug 28, 2025

fix: bmodel_checker in multi-core
jianmin.xu
committed
9875e8c
feat: optimize CycleCalculator for multi-core
yifan.chen
committed
4e0e435
fix: groupnorm addr check prob
charlesxzb
committed
f6eaa20
fix: switch group_size=0 to group_size=128
HarmonyHu
committed
c2392fd
fix: model-zoo dino model problem
WhoIsYu
committed
06aae97
fix: add eva02 pattern for model zoo eva02 224 and 336 model
WhoIsYu
committed
6716696
fix: [MLIR-615] model simi bug fixed

xu.yang
authored and
PostMalone1998
committed
05bf8b3

Commits on Aug 29, 2025

fix: A16MatMul w4a16 no group will switch to group 128
HarmonyHu
committed
ff3e3b2
fix: pool2d int8 cv181x

zijing.wan
authored and
charlesxzb
committed
2d40093
feat: support w4int8 model for cv184x
hanqing.zhang
committed
e6ef8f5

Commits on Sep 1, 2025

docs: update release info in disclaimer
charlesxzb
committed
383a3c3

Commits on Sep 2, 2025

feat: add weight = f32 in RMSNorm

romain-cpu
authored and
HarmonyHu
committed
56ab347
feat: support mlir_cut and mlir2onnx for final.mlir

hanqing.zhang
authored and
charlesxzb
committed
f9cd0a1
feat: Permute/RequantIntAxis multicore optimize

PostMalone1998
authored and
HarmonyHu
committed
e03f7f1

Commits on Sep 3, 2025

add llm_code_reviewer,using qwen3_coder as base model

jiadong.jiang
authored and
浩哲杨
committed
c2fecf4
feat: refine the visualization of multicore op's cmp_res in bmodel_checker

jianmin.xu
authored and
浩哲杨
committed
3eb3f62

Commits on Sep 4, 2025

add softmax correction

wwwwwu64
authored and
WhoIsYu
committed
7c23053
fix: TopK when axis != -1

Heng2024
authored and
逸凡陈
committed
bc761c0
fix: fix RMSNorm

romain-cpu
authored and
HarmonyHu
committed
e1aa2b2
feat: support glm4.1v

sgwzy22
authored and
HarmonyHu
committed
a4ad193
feat: create StructOptimize pass
yubing.shen
committed
f3dffb9

Commits on Sep 5, 2025

refactor: hik env adaption for openmp error

xu.yang
authored and
charlesxzb
committed
eff81a1
refactor: hik env adaption for unused variable error

xu.yang
authored and
charlesxzb
committed
ba37889
fix: MLIR-612 compile error

xu.yang
authored and
charlesxzb
committed
cfda87c
fix: [MLIR-606] compile fail bug fixed

xu.yang
authored and
charlesxzb
committed
7800734
fix: [MLIR-627] weight_shape changed in https://gerrit-ai.sophgo.vip:8443/#/c/149323/

xu.yang
authored and
charlesxzb
committed
01722f0

Commits on Sep 7, 2025

feat: add deformable_attention and fix roiExtractor
xu.yang
committed
3325000

Commits on Sep 8, 2025

feat: tpulang support merger_matmul

yifan.chen
authored and
旭杨
committed
c3688b6
fix: tpulang:multi_scale_deformable_attention

xiangzhou.ye
authored and
xu.yang
committed
4636559
remove yolov5s softmax correction sample

wwwwwu64
authored and
WhoIsYu
committed
c50ef00
feat: add new tpulang op

zijing.wan
authored and
charlesxzb
committed
4ec7875

Commits on Sep 9, 2025

fix: top cannonicalize pass may encounter unranked type
charlesxzb
committed
db175a0
fix: ConvertMatMulWithRightTranspose pattern bug
charlesxzb
committed
eb9aac0

Commits on Sep 10, 2025

feat: enable to choose shape_secs_search_strategy
yifan.chen
committed
1f9e75b
fix: PermuteRopeWeightReorderPattern bugfix
yifan.chen
committed
f1c61f4
fix: f16 dynamic model binary opds need host2device
yifan.chen
committed
30c0a32
fix: moveUnaryPermute function
yifan.chen
committed
e519c13
fix: LayerGroup bug caused by Slice & Reshape
yifan.chen
committed
4b4f350
feat: add llama3.2 vision

hengyang123
authored and
HarmonyHu
committed
546e2d3
feat: LlmConverter support bm1690
HarmonyHu
committed
414f203
test: gen_shell.py don't duplicate onnx
Heng2024
committed
449d8d6

Commits on Sep 11, 2025

feat: add comment language check for Chinese characters in pre-commit hook

Artlesbol
and
charlesxzb
committed
dfde801
fix: refine llm analysis
HarmonyHu
committed
06ee352
feat: remove warning for new onnxruntime version, add cuda param for model_eval
WhoIsYu
committed
fc31502

Commits on Sep 12, 2025

fix: fix profile module not found problem
charlesxzb
committed
51e910c

Commits on Sep 14, 2025

feat: support conv3d multi core
yiqun.huang
committed
73a6329

Commits on Sep 15, 2025

feat: add new upsample local op

xiangzhou.ye
authored and
charlesxzb
committed
58b30ac
fix: LayerGroup bug for multi-core cases
yifan.chen
committed
cb9ed17

Commits on Sep 16, 2025

fix: total secs lower bound

Heng2024
authored and
yifan.chen
committed
c1f08df
fix: gatherelements indices incorrect dtype
charlesxzb
committed
c574bf6
fix: Memory Overflow in Upsample
xiangzhou.ye
committed
91b19dd
fix: fix index of interp

chang.zhao
authored and
charlesxzb
committed
bac6885
feat: support reduce sum>65536
zijing.wan
committed
98284b0

Commits on Sep 17, 2025

feat: support MiniCPMV4

zhou.yang
authored and
HarmonyHu
committed
9446da3
fix: [MLIR-642] multi output graph broken in visual tools
WhoIsYu
committed
7a37fae
perf: optimize arg with n h trans

Heng2024
authored and
逸凡陈
committed
737616e
fix: [MLIR-642] find topk second output as shape ops
WhoIsYu
committed
0d84d1f

Commits on Sep 18, 2025

feat: bm1690 support io alone and dynamic bmodel
HarmonyHu
committed
de0eccc
feat: tpulang supports specifying quantization mode for specific op
charlesxzb
committed
e87a36a
feat: add log_level choice
charlesxzb
committed
99457f2

Commits on Sep 19, 2025

fix: fix and walk around mlir_cut bugs

hanqing.zhang
authored and
逸凡陈
committed
3c6383b
fix: fix interp

chang.zhao
authored and
逸凡陈
committed
8c1cbbc
fix: Add case for split gridsampler

xiangzhou.ye
authored and
逸凡陈
committed
ee608cc

Commits on Sep 20, 2025

fix: fix bm1690 topk error
HarmonyHu
committed
e03b68c

Commits on Sep 22, 2025

feat: add StructOptimize pass pattern for correlation

yubing.shen
authored and
逸凡陈
committed
1d92aa6
fix: a cv184x opparam bug
liang.wang01
committed
6fed9a3

Commits on Sep 23, 2025

chore: weekly backend update
charlesxzb
committed
05958e5
fix: multicore should be set before LayerGroup
zhou.yang
committed
c307fbd
fix: modify the order of judgment conditions
zhou.yang
committed
8b68a0e

Commits on Sep 24, 2025

fix: deform attn for 1688

xiangzhou.ye
authored and
旭杨
committed
e23e8e5
fix: fix MatMul precision loss in ToptoTpupass.

hanqing.zhang
authored and
charlesxzb
committed
46b9c7c
feat: more op support multi-branch in LayerGroup
yifan.chen
committed
a28a0f2
fix: cv184x profile failed
charlesxzb
committed
ec2eab4
fix: bmodel_checker bug

jianmin.xu
authored and
charlesxzb
committed
e795804
fix: upsample lowering

PostMalone1998
authored and
charlesxzb
committed
47aec01
fix: LgCache bug of dynamic subnet

yifan.chen
authored and
charlesxzb
committed
f399bba
fix: bm1688 vit-l multi-core compile
PostMalone1998
committed
3ee1d8c
fix: bmodel_dis multi-subnet overwrite
PostMalone1998
committed
f7aafa3
feat: rms_norm multi core
yiqun.huang
committed
6f7033a
fix: fix nodechip_bcbinary
yubing.shen
committed
b522938

Commits on Sep 25, 2025

fix: matmul hdim_is_batch

xiangzhou.ye
authored and
charlesxzb
committed
824e807
feat: concat multi core
yiqun.huang
committed
0000b30
feat: cv18xx support special in/out to be iomem
HarmonyHu
committed
80b3ad2
feat: add LayerGroupProfile pass
zhou.yang
committed
3681307

Commits on Sep 26, 2025

fix: compare_visualizer import bug
yifan.chen
committed
02d4a69
fix: [SOPHONSILK-540] combine pattern qtable
WhoIsYu
committed
b49cee5
fix: CycleCalculator for 1690 family
yifan.chen
committed
c5554a7
fix: cv184x not supports profile with layer info
charlesxzb
committed
82af9ed
fix: bm1688/bm1690 mutmul multicore bias dtype issue
PostMalone1998
committed
dcec8f9

Commits on Sep 28, 2025

fix: fattention_bf16.pl nan

hengyang123
authored and
HarmonyHu
committed
787f834
fix: cv184x profile config varible compute ability
PostMalone1998
committed
0b61a43
feat: do core parallel or not determined by cycle

Heng2024
authored and
逸凡陈
committed
a4e4b97
fix: rope's weights sometime need to hold in lmem
yifan.chen
committed
ac4f724
feat: support vlm do dynamic compile for vit
HarmonyHu
committed
cdd113b

Commits on Sep 29, 2025

feat: refactor multi-branch concerning code
yifan.chen
committed
4c75174
fix: fix MiniCPM4 bug

romain-cpu
authored and
HarmonyHu
committed
60bcdb7
fix: group conv error for mars3
Heng2024
committed
5bc3874
feat: support multi subnet for layer group cache and debugger

Heng2024
authored and
逸凡陈
committed
a88e5ae
fix: rope reload bug & multi-branch bug
yifan.chen
committed
95a001d
fix: lg_debugger error
Heng2024
committed
8442112

Commits on Sep 30, 2025

fix: bm1690 fc multicore rq param
PostMalone1998
committed
f72f8a4
docs: update release version info
charlesxzb
committed
ad7632f
fix: use set to avoid repeated add of layers
WhoIsYu
committed
75f1c8d

Commits on Oct 10, 2025

feat: support f16 for Arg in 88 and 84x
Heng2024
committed
142fdaa
fix: sigmoid and hardsigmoid active support fp16 for bm1688

Heng2024
authored and
逸凡陈
committed
ec1f9b3
feat: Slice/Pad multicore optimize

mlt
authored and
PostMalone1998
committed
aac2384

Commits on Oct 11, 2025

fix: c-dim local concat bug using tpu_bdc_cpy_cross_npu
yifan.chen
committed
2d5770c
fix: fix index buffer of interp

chang.zhao
authored and
逸凡陈
committed
0aa1977
fix: [AIF-1828] set to F32 for bm1684
WhoIsYu
committed
c379e4e
fix: [MLIR-658] check cali info before apply threshold spread
WhoIsYu
committed
894be7d

Commits on Oct 13, 2025

feat: support Qwen3VL
HarmonyHu
committed
442549e
fix: suitable for new ppl version
neocats
committed
0d05a61
fix: [CV184XSDK-689] rope int8 lowering

tpu-mlir-vis-tools
authored and
charlesxzb
committed
f17c0b0
fix: [AIF-1828] avoid huge bin numbers to reach system limit
WhoIsYu
committed
7b677a7

Commits on Oct 14, 2025

revert: f16 sigmoid

Heng2024
authored and
charlesxzb
committed
4df1744
fix: use io_alone for some model in llm
HarmonyHu
committed
135e319

Commits on Oct 15, 2025

fix: adjust Qwen3VL converter
HarmonyHu
committed
a20df03

Commits on Oct 16, 2025

fix: fix arg and yolodetection

liang.wang01
authored and
charlesxzb
committed
a5190d3
fix: conditional bug fixes for add Floor

yubing.shen
authored and
charlesxzb
committed
b306c1d
fix: bm1688 profile get freq
PostMalone1998
committed
ff98556
fix: [MLIR-669] add m2 image pattern and naive mix-precision setting
WhoIsYu
committed
3b08f83

Commits on Oct 17, 2025

feat: support yuv in calibration
WhoIsYu
committed
0ddd5f5
fix: mars3 GatherOp's index loss precissioin
yifan.chen
committed
f682ba4
fix: conbine shape_pattern searched fp_layers to simplify qtable search
boyu.liu
committed
8805f93

Commits on Oct 20, 2025

fix: correction of last change for shape_pattern
boyu.liu
committed
a82cb2d
Revert "fix: correction of last change for shape_pattern"
boyu.liu
committed
2157b14

Commits on Oct 21, 2025

feat: model_tool supports CPU peak estimate
charlesxzb
committed
75fb421
feat: add cuda calibration functions and code optimize
WhoIsYu
committed
17c2303
fix: Strengthen Rope logic
yubing.shen
committed
06bd28c

Commits on Oct 22, 2025

fix: add missing parameters
WhoIsYu
committed
66e9a0b
fix: layergroup error when strategy 2

Heng2024
authored and
逸凡陈
committed
0582dcd
fix: op could be inplace when it is the only user in group
zhou.yang
committed
11205ef

Commits on Oct 23, 2025

feat: add janus

hengyang123
authored and
HarmonyHu
committed
2825f98
fix: correct errors in former shape_pat code

boyu.liu
authored and
WhoIsYu
committed
07c2dc2
fix: support negative offset of sliceOp in CV18xx
dgyx
committed
5262818
fix: [MLIR-670]

tpu-mlir-vis-tools
authored and
charlesxzb
committed
92c4117

Commits on Oct 24, 2025

fix: disable lgcache when no_save

Heng2024
authored and
charlesxzb
committed
c319c39
fix: fix the fused_op split order in cali

boyu.liu
authored and
WhoIsYu
committed
5e709af
fix: Fix ppl integration issues

neocats
authored and
charlesxzb
committed
ab3ad6a
feat: Create a rearrangement logical operation

yubing.shen
authored and
逸凡陈
committed
a61ba93
fix: tpulang no_save compile error
charlesxzb
committed
77eeff7
feat: Support the codegen of bm1690 ops using ppl
neocats
committed
38ae3fb

Commits on Oct 25, 2025

refactor: Qwen3VL use static vit
HarmonyHu
committed
2675dbe

Commits on Oct 26, 2025

fix: GELU in 1688 should be f32
HarmonyHu
committed
504f27a

Commits on Oct 28, 2025

fix: cv184x Gelu does bf16 lowering
charlesxzb
committed
f5fab72

Commits on Oct 30, 2025

fix: bm1684 not support layergroup multi-branch case
yifan.chen
committed
30d8dad
fix: remove print, fix fp layers take effect in search qtable, and remove redundant sensitive layer search
WhoIsYu
committed
e955c45

Commits on Oct 31, 2025

fix: update manual
WhoIsYu
committed
ce54fe1
fix: fix bugs in final qtable
boyu.liu
committed
269c4a5
fix: fix bcbinary rearrangement bug

yubing.shen
authored and
charlesxzb
committed
b6656c2
fix: fix bug of tune_num=0 in search qtable
boyu.liu
committed
683d212
refactor: FAttention add keep_dims attr
HarmonyHu
committed
4335376

Commits on Nov 3, 2025

docs: upload quant.drawio

wwwwwu64
authored and
WhoIsYu
committed
217f193
fix: bug fix for the interp operator in the DHW dimension in 5 dimensions

yubing.shen
authored and
逸凡陈
committed
92ae111

Commits on Nov 4, 2025

fix: mars3 global depthwse deconv bug

yifan.chen
authored and
charlesxzb
committed
9f29863
chore: switch reg device to run llm code review
charlesxzb
committed
f972dcf

Commits on Nov 7, 2025

fix: io needs to be placed first in io_alone mode
charlesxzb
committed
d109f9c
fix: fix fattn head slice error
HarmonyHu
committed
c4ea2a4

Commits on Nov 9, 2025

fix: update bm1684x/bm1688 backend
HarmonyHu
committed
9af4d06
fix: remove 8ch interleave code
HarmonyHu
committed
80127c4

Commits on Nov 10, 2025

fix: avoid memory leak when repeat call to invoke
WhoIsYu
committed
d642b63
fix: fix the conv bug in bias_correction

boyu.liu
authored and
WhoIsYu
committed
5a6bbab
fix: [MLIR-693] opt qtable, add function to search qtable based on existing qtable
WhoIsYu
committed
a3da6f0
fix: fix bugs of sq and smc

wwwwwu64
authored and
WhoIsYu
committed
b96d471
fix: tpulang no_save mode block all log prob
charlesxzb
committed
4b4738a
fix: TpuReshapeReorderPattern bug
charlesxzb
committed
e3974d5
fix: mutil_defrom_attn for more case

xiangzhou.ye
authored and
逸凡陈
committed
9c8b771
fix: cv184x interp input coord use FP32/UINT32

liang.wang01
authored and
charlesxzb
committed
096e115

Commits on Nov 11, 2025

fix: [CV184XSDK-721]

tpu-mlir-vis-tools
authored and
charlesxzb
committed
96ad53a
feat: jira ID [CV184X-766]

jianmin.xu
authored and
charlesxzb
committed
e6b0203

Commits on Nov 13, 2025

fix: align dtype in set_data()
jianmin.xu
committed
64419e6
feat: add more info in AddPostprocess

jianmin.xu
authored and
charlesxzb
committed
4bdb29a
fix: inplaceOp liveRange

xiangzhou.ye
authored and
逸凡陈
committed
00e2317
feat: Add same_addr param in model_deploy

24octopus
authored and
HarmonyHu
committed
37177ac
feat: Adapt CV184x LSTM BF16
yubing.shen
committed
5fe4773
fix: ReduceOp not support dynamic local codgen
yifan.chen
committed
5a0ff08

Commits on Nov 14, 2025

fix: tpulang bug fix:
xu.yang
committed
ec679ff
fix: no_save error in afterlayergroup pass
charlesxzb
committed
12fb1f5
fix: add LSTM cv184x test_onnx
yubing.shen
committed
bdf5fb7

Commits on Nov 17, 2025

feat: add new rewriter pattern

xiangzhou.ye
authored and
逸凡陈
committed
4186fef
fix: dynamic codegen bug of UpSample
yifan.chen
committed
fd0207f
fix: more op support dynamic mlir inference
yifan.chen
committed
a530280

Commits on Nov 18, 2025

feat: add HdimIsBatch mm pattern cases

tpu-mlir-vis-tools
authored and
charlesxzb
committed
7956a08
fix: disable group_before.txt

Heng2024
authored and
charlesxzb
committed
e34afc0

Commits on Nov 19, 2025

feat: support bm1690e

qian.zheng
authored and
charlesxzb
committed
4aadfba
feat: apply qtable to more search functions
WhoIsYu
committed
aa6825f
fix: add missing shape parameter in set_tensor
WhoIsYu
committed
0901bfc
feat: update manual on calibration
WhoIsYu
committed
4e70627

Commits on Nov 20, 2025

fix: bm1684 conv failed prob
charlesxzb
committed
007c0be
feat: support auto-round for llm
HarmonyHu
committed
3f562b8
test: add bm1690e op and model test
charlesxzb
committed
bfcd3a5

Commits on Nov 21, 2025

feat: add addr_mode 'in_reuse'
yifan.chen
committed
2d5f0e5
fix: sample_head use dynamic
HarmonyHu
committed
e5384c4
fix: handle to all zero tensor
WhoIsYu
committed
3086fd0
feat: manual update
WhoIsYu
committed
455dddc

Commits on Nov 25, 2025

fix: change for igev pass

xiangzhou.ye
authored and
逸凡陈
committed
b265e4f
fix: fix ppl cv184x issue

neocats
authored and
charlesxzb
committed
f469e55
docs: update release record
charlesxzb
committed
b605a21
fix: fix cv184x codgen issue
neocats
committed
7fef01f

Commits on Nov 26, 2025

fix: avoid inf match in op name
WhoIsYu
committed
7e07def
feat: add profile test regression

yubing.shen
authored and
charlesxzb
committed
6bf1f98
feat: remove not used sophgo-mq version
WhoIsYu
committed
a8ee861
feat: support lightstereo

xiangzhou.ye
authored and
charlesxzb
committed
115c206

Commits on Nov 27, 2025

feat: qwen3vl support multi vit
HarmonyHu
committed
79b0c17

Commits on Nov 28, 2025

fix: io address should assign firstly for io_alone
HarmonyHu
committed
a0e5644

Commits on Nov 30, 2025

feat: support multi prefill input length
HarmonyHu
committed
593528c

Commits on Dec 1, 2025

refactor: uniform gen_mlirs and compiles in llm
HarmonyHu
committed
bd54ca3

Commits on Dec 2, 2025

refactor: make weights 4K aligned in bmodel file
HarmonyHu
committed
812fd67
fix: [BM1684X-8119]

tpu-mlir-vis-tools
authored and
charlesxzb
committed
37f5c80
fix: [MLIR-660]

jianmin.xu
authored and
charlesxzb
committed
b3ff967
feat: llm support compressed-tensors mode
HarmonyHu
committed
5a78925

Commits on Dec 3, 2025

fix: [CV186AHDEV-519] reference inference overflow if use F16
WhoIsYu
committed
ef85193
feat: inference mlir with cuda
WhoIsYu
committed
c32afcf
fix: update bm1690e backend

qian.zheng
authored and
charlesxzb
committed
0ae1b10

Commits on Dec 4, 2025

feat: add out_fixed mode in bmodel_checker

jianmin.xu
authored and
逸凡陈
committed
c09e0b6
feat: make all binary 4K aligned in bmodel file
HarmonyHu
committed
bc50baa

Commits on Dec 5, 2025

feat: model_tool can refresh bmodel to 4k aligned
HarmonyHu
committed
903c080
fix: fix LSTM cv184x int8 bug

yubing.shen
authored and
charlesxzb
committed
c5d7cc0
feat: MLIR-729
xu.yang
committed
c4b88d6

Commits on Dec 9, 2025

fix: time_fixed_subnet error

xiangzhou.ye
authored and
逸凡陈
committed
c71bf2d
fix: add chip support data types
WhoIsYu
committed
cb2bade
fix: postprocess shape

jianmin.xu
authored and
逸凡陈
committed
085f559

Commits on Dec 10, 2025

fix: tensor mul vector of matmul not work on this pass
WhoIsYu
committed
e9b738c
chore: update libbmrt
Korbin-chen
committed
447af21
feat: add more op to GROUP_SMALL_C
yifan.chen
committed
e94f184
fix: vit models deploy failed in w4a16/w8a16 mode

wwwwwu64
authored and
WhoIsYu
committed
3f1d70d
refactor: refine lora support
HarmonyHu
committed
98cfbff
feat: layer group method optimize
yifan.chen
committed
2cb943a

Commits on Dec 11, 2025

feat: support contiguous_halves rope mode

liang.wang01
authored and
HarmonyHu
committed
fba100a
refactor: refine lora path
HarmonyHu
committed
95062df
feat: do softmax for lmhead
HarmonyHu
committed
9847954