tests : switch fusion baseline from TSV to CSV

Use comma-separated values like the rest of the project, keeping the
padded, aligned columns. Split on ',' and trim on parse. Rename the
committed baseline to MTL.csv (data unchanged, verified identical modulo
padding/separator). Update the ci/run.sh check path accordingly.

Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
This commit is contained in:
Georgi Gerganov
2026-09-03 17:19:50 +03:00
parent 91f065ac9c
commit 259faf94ec
4 changed files with 317 additions and 317 deletions
+1 -1
View File
@@ -346,7 +346,7 @@ function gg_run_test_fusion {
# generate the dummy models and run the fusion regression test
./build-ci-release/bin/test-llama-archs -o build-ci-fusion-models 2>&1
./build-ci-release/bin/test-fusion --models build-ci-fusion-models --device MTL0 --check tests/fusion/MTL.tsv 2>&1
./build-ci-release/bin/test-fusion --models build-ci-fusion-models --device MTL0 --check tests/fusion/MTL.csv 2>&1
fi
set +e
+300
View File
@@ -0,0 +1,300 @@
# test-fusion baseline for device MTL
# arch ,moe ,mode ,label , count
arcee ,0 ,prefill ,RMS_NORM+MUL , 5
arcee ,0 ,decode ,RMS_NORM+MUL , 80
arctic ,0 ,prefill ,RMS_NORM+MUL , 7
arctic ,0 ,decode ,RMS_NORM+MUL , 112
baichuan ,0 ,prefill ,RMS_NORM+MUL , 5
baichuan ,0 ,decode ,RMS_NORM+MUL , 80
bailingmoe ,1 ,prefill ,RMS_NORM+MUL , 5
bailingmoe ,1 ,prefill ,ADD+ADD , 2
bailingmoe ,1 ,decode ,RMS_NORM+MUL , 80
bailingmoe ,1 ,decode ,ADD+ADD , 32
bailingmoe2 ,1 ,prefill ,RMS_NORM+MUL , 9
bailingmoe2 ,1 ,prefill ,ADD+ADD , 1
bailingmoe2 ,1 ,decode ,RMS_NORM+MUL , 144
bailingmoe2 ,1 ,decode ,ADD+ADD , 16
bailingmoe3 ,1 ,prefill ,RMS_NORM+MUL , 8
bailingmoe3 ,1 ,prefill ,ADD+ADD , 1
bailingmoe3 ,1 ,prefill ,GATED_DELTA_NET+CPY , 1
bailingmoe3 ,1 ,decode ,RMS_NORM+MUL , 128
bailingmoe3 ,1 ,decode ,ADD+ADD , 16
bailingmoe3 ,1 ,decode ,GATED_DELTA_NET+CPY , 16
bloom ,0 ,prefill ,NORM+MUL+ADD , 6
bloom ,0 ,decode ,NORM+MUL+ADD , 96
chatglm ,0 ,prefill ,RMS_NORM+MUL , 5
chatglm ,0 ,decode ,RMS_NORM+MUL , 80
codeshell ,0 ,prefill ,NORM+MUL+ADD , 5
codeshell ,0 ,decode ,NORM+MUL+ADD , 80
cogvlm ,0 ,prefill ,RMS_NORM+MUL , 5
cogvlm ,0 ,decode ,RMS_NORM+MUL , 80
command-r ,0 ,prefill ,NORM+MUL , 3
command-r ,0 ,decode ,NORM+MUL , 48
dbrx ,0 ,prefill ,NORM+MUL , 5
dbrx ,0 ,decode ,NORM+MUL , 80
deci ,0 ,prefill ,RMS_NORM+MUL , 5
deci ,0 ,decode ,RMS_NORM+MUL , 80
deepseek ,0 ,prefill ,RMS_NORM+MUL , 5
deepseek ,0 ,prefill ,ADD+ADD , 1
deepseek ,0 ,decode ,RMS_NORM+MUL , 80
deepseek ,0 ,decode ,ADD+ADD , 16
deepseek2 ,0 ,prefill ,RMS_NORM+MUL , 9
deepseek2 ,0 ,prefill ,ADD+ADD , 1
deepseek2 ,0 ,decode ,RMS_NORM+MUL , 144
deepseek2 ,0 ,decode ,ADD+ADD , 16
deepseek32 ,0 ,prefill ,NORM+MUL+ADD , 2
deepseek32 ,0 ,prefill ,RMS_NORM+MUL , 9
deepseek32 ,0 ,prefill ,ADD+ADD , 1
deepseek32 ,0 ,decode ,NORM+MUL+ADD , 32
deepseek32 ,0 ,decode ,RMS_NORM+MUL , 144
deepseek32 ,0 ,decode ,ADD+ADD , 16
deepseek4 ,0 ,prefill ,RMS_NORM+MUL , 20
deepseek4 ,0 ,decode ,RMS_NORM+MUL , 320
dots1 ,0 ,prefill ,RMS_NORM+MUL , 9
dots1 ,0 ,prefill ,ADD+ADD , 1
dots1 ,0 ,decode ,RMS_NORM+MUL , 144
dots1 ,0 ,decode ,ADD+ADD , 16
dream ,0 ,prefill ,RMS_NORM+MUL , 5
dream ,0 ,decode ,RMS_NORM+MUL , 80
ernie4_5-moe ,1 ,prefill ,RMS_NORM+MUL , 5
ernie4_5-moe ,1 ,prefill ,ADD+ADD , 1
ernie4_5-moe ,1 ,decode ,RMS_NORM+MUL , 80
ernie4_5-moe ,1 ,decode ,ADD+ADD , 16
ernie4_5 ,0 ,prefill ,RMS_NORM+MUL , 5
ernie4_5 ,0 ,decode ,RMS_NORM+MUL , 80
exaone ,0 ,prefill ,RMS_NORM+MUL , 5
exaone ,0 ,decode ,RMS_NORM+MUL , 80
exaone4 ,0 ,prefill ,RMS_NORM+MUL , 5
exaone4 ,0 ,prefill ,RMS_NORM+MUL+ADD , 4
exaone4 ,0 ,decode ,RMS_NORM+MUL , 80
exaone4 ,0 ,decode ,RMS_NORM+MUL+ADD , 64
falcon ,0 ,prefill ,NORM+MUL+ADD , 5
falcon ,0 ,prefill ,ADD+ADD , 2
falcon ,0 ,decode ,NORM+MUL+ADD , 80
falcon ,0 ,decode ,ADD+ADD , 32
falcon-h1 ,0 ,prefill ,RMS_NORM+MUL , 9
falcon-h1 ,0 ,prefill ,ADD+ADD , 2
falcon-h1 ,0 ,decode ,RMS_NORM+MUL , 144
falcon-h1 ,0 ,decode ,ADD+ADD , 32
gemma ,0 ,prefill ,RMS_NORM+MUL , 5
gemma ,0 ,decode ,RMS_NORM+MUL , 80
gemma2 ,0 ,prefill ,RMS_NORM+MUL , 5
gemma2 ,0 ,prefill ,RMS_NORM+MUL+ADD , 4
gemma2 ,0 ,decode ,RMS_NORM+MUL , 80
gemma2 ,0 ,decode ,RMS_NORM+MUL+ADD , 64
glm-dsa ,0 ,prefill ,NORM+MUL+ADD , 2
glm-dsa ,0 ,prefill ,RMS_NORM+MUL , 9
glm-dsa ,0 ,prefill ,ADD+ADD , 1
glm-dsa ,0 ,decode ,NORM+MUL+ADD , 32
glm-dsa ,0 ,decode ,RMS_NORM+MUL , 144
glm-dsa ,0 ,decode ,ADD+ADD , 16
glm4 ,0 ,prefill ,RMS_NORM+MUL , 5
glm4 ,0 ,prefill ,RMS_NORM+MUL+ADD , 4
glm4 ,0 ,decode ,RMS_NORM+MUL , 80
glm4 ,0 ,decode ,RMS_NORM+MUL+ADD , 64
glm4moe ,1 ,prefill ,RMS_NORM+MUL , 9
glm4moe ,1 ,prefill ,ADD+ADD , 1
glm4moe ,1 ,decode ,RMS_NORM+MUL , 144
glm4moe ,1 ,decode ,ADD+ADD , 16
gpt-oss ,0 ,prefill ,RMS_NORM+MUL , 5
gpt-oss ,0 ,decode ,RMS_NORM+MUL , 80
gpt2 ,0 ,prefill ,NORM+MUL+ADD , 5
gpt2 ,0 ,decode ,NORM+MUL+ADD , 80
gptneox ,0 ,prefill ,NORM+MUL+ADD , 5
gptneox ,0 ,decode ,NORM+MUL+ADD , 80
granite ,0 ,prefill ,RMS_NORM+MUL , 5
granite ,0 ,decode ,RMS_NORM+MUL , 80
granite ,0 ,prefill ,RMS_NORM+MUL , 5
granite ,0 ,decode ,RMS_NORM+MUL , 80
granitehybrid ,0 ,prefill ,RMS_NORM+MUL , 6
granitehybrid ,0 ,decode ,RMS_NORM+MUL , 96
granitemoe ,1 ,prefill ,RMS_NORM+MUL , 5
granitemoe ,1 ,decode ,RMS_NORM+MUL , 80
granitemoe ,1 ,prefill ,RMS_NORM+MUL , 5
granitemoe ,1 ,decode ,RMS_NORM+MUL , 80
grok ,0 ,prefill ,RMS_NORM+MUL , 5
grok ,0 ,prefill ,RMS_NORM+MUL+ADD , 4
grok ,0 ,decode ,RMS_NORM+MUL , 80
grok ,0 ,decode ,RMS_NORM+MUL+ADD , 64
grovemoe ,1 ,prefill ,RMS_NORM+MUL , 9
grovemoe ,1 ,prefill ,ADD+ADD , 2
grovemoe ,1 ,decode ,RMS_NORM+MUL , 144
grovemoe ,1 ,decode ,ADD+ADD , 32
hunyuan-dense ,0 ,prefill ,RMS_NORM+MUL , 9
hunyuan-dense ,0 ,decode ,RMS_NORM+MUL , 144
hunyuan-moe ,1 ,prefill ,RMS_NORM+MUL , 9
hunyuan-moe ,1 ,prefill ,ADD+ADD , 2
hunyuan-moe ,1 ,decode ,RMS_NORM+MUL , 144
hunyuan-moe ,1 ,decode ,ADD+ADD , 32
hunyuan_vl ,0 ,prefill ,RMS_NORM+MUL , 9
hunyuan_vl ,0 ,decode ,RMS_NORM+MUL , 144
hy_v3 ,0 ,prefill ,RMS_NORM+MUL , 9
hy_v3 ,0 ,prefill ,ADD+ADD , 2
hy_v3 ,0 ,decode ,RMS_NORM+MUL , 144
hy_v3 ,0 ,decode ,ADD+ADD , 32
internlm2 ,0 ,prefill ,RMS_NORM+MUL , 5
internlm2 ,0 ,decode ,RMS_NORM+MUL , 80
jais ,0 ,prefill ,NORM+MUL+ADD , 5
jais ,0 ,decode ,NORM+MUL+ADD , 80
jais2 ,0 ,prefill ,NORM+MUL+ADD , 5
jais2 ,0 ,decode ,NORM+MUL+ADD , 80
jamba ,0 ,prefill ,RMS_NORM+MUL , 8
jamba ,0 ,decode ,RMS_NORM+MUL , 128
kimi-k3 ,0 ,prefill ,RMS_NORM+MUL , 17
kimi-k3 ,0 ,prefill ,GATED_DELTA_NET+CPY , 1
kimi-k3 ,0 ,decode ,RMS_NORM+MUL , 272
kimi-k3 ,0 ,decode ,GATED_DELTA_NET+CPY , 16
kimi-linear ,0 ,prefill ,RMS_NORM+MUL , 7
kimi-linear ,0 ,prefill ,ADD+ADD , 1
kimi-linear ,0 ,prefill ,GATED_DELTA_NET+CPY , 1
kimi-linear ,0 ,decode ,RMS_NORM+MUL , 112
kimi-linear ,0 ,decode ,ADD+ADD , 16
kimi-linear ,0 ,decode ,GATED_DELTA_NET+CPY , 16
lfm2 ,0 ,prefill ,RMS_NORM+MUL , 7
lfm2 ,0 ,decode ,RMS_NORM+MUL , 112
lfm2moe ,1 ,prefill ,RMS_NORM+MUL , 7
lfm2moe ,1 ,decode ,RMS_NORM+MUL , 112
llada ,0 ,prefill ,RMS_NORM+MUL , 5
llada ,0 ,decode ,RMS_NORM+MUL , 80
llada-moe ,1 ,prefill ,RMS_NORM+MUL , 9
llada-moe ,1 ,decode ,RMS_NORM+MUL , 144
llama ,0 ,prefill ,RMS_NORM+MUL , 5
llama ,0 ,decode ,RMS_NORM+MUL , 80
llama ,0 ,prefill ,RMS_NORM+MUL , 5
llama ,0 ,decode ,RMS_NORM+MUL , 80
llama4 ,0 ,prefill ,RMS_NORM+MUL , 9
llama4 ,0 ,prefill ,ADD+ADD , 2
llama4 ,0 ,decode ,RMS_NORM+MUL , 144
llama4 ,0 ,decode ,ADD+ADD , 32
maincoder ,0 ,prefill ,RMS_NORM+MUL , 9
maincoder ,0 ,decode ,RMS_NORM+MUL , 144
mamba ,0 ,prefill ,RMS_NORM+MUL , 3
mamba ,0 ,decode ,RMS_NORM+MUL , 48
mamba2 ,0 ,prefill ,RMS_NORM+MUL , 5
mamba2 ,0 ,decode ,RMS_NORM+MUL , 80
minicpm ,0 ,prefill ,RMS_NORM+MUL , 5
minicpm ,0 ,decode ,RMS_NORM+MUL , 80
minicpm ,0 ,prefill ,RMS_NORM+MUL , 5
minicpm ,0 ,decode ,RMS_NORM+MUL , 80
minicpm3 ,0 ,prefill ,RMS_NORM+MUL , 9
minicpm3 ,0 ,decode ,RMS_NORM+MUL , 144
minimax-01 ,0 ,prefill ,RMS_NORM+MUL , 6
minimax-01 ,0 ,decode ,RMS_NORM+MUL , 96
minimax-m2 ,0 ,prefill ,RMS_NORM+MUL , 9
minimax-m2 ,0 ,decode ,RMS_NORM+MUL , 144
minimax-m3 ,0 ,prefill ,RMS_NORM+MUL , 11
minimax-m3 ,0 ,prefill ,ADD+ADD , 1
minimax-m3 ,0 ,decode ,RMS_NORM+MUL , 176
minimax-m3 ,0 ,decode ,ADD+ADD , 16
mistral3 ,0 ,prefill ,RMS_NORM+MUL , 5
mistral3 ,0 ,decode ,RMS_NORM+MUL , 80
mistral3 ,0 ,prefill ,RMS_NORM+MUL , 5
mistral3 ,0 ,decode ,RMS_NORM+MUL , 80
mistral4 ,0 ,prefill ,RMS_NORM+MUL , 9
mistral4 ,0 ,prefill ,ADD+ADD , 1
mistral4 ,0 ,decode ,RMS_NORM+MUL , 144
mistral4 ,0 ,decode ,ADD+ADD , 16
mpt ,0 ,prefill ,NORM+MUL+ADD , 5
mpt ,0 ,decode ,NORM+MUL+ADD , 80
nanbeige ,0 ,prefill ,RMS_NORM+MUL , 5
nanbeige ,0 ,decode ,RMS_NORM+MUL , 80
nemotron ,0 ,prefill ,NORM+MUL+ADD , 5
nemotron ,0 ,decode ,NORM+MUL+ADD , 80
nemotron_h ,0 ,prefill ,RMS_NORM+MUL , 5
nemotron_h ,0 ,decode ,RMS_NORM+MUL , 80
nemotron_h_moe ,1 ,prefill ,RMS_NORM+MUL , 5
nemotron_h_moe ,1 ,decode ,RMS_NORM+MUL , 80
olmoe ,1 ,prefill ,RMS_NORM+MUL , 9
olmoe ,1 ,decode ,RMS_NORM+MUL , 144
openelm ,0 ,prefill ,RMS_NORM+MUL , 9
openelm ,0 ,decode ,RMS_NORM+MUL , 144
orion ,0 ,prefill ,NORM+MUL+ADD , 5
orion ,0 ,decode ,NORM+MUL+ADD , 80
paddleocr ,0 ,prefill ,RMS_NORM+MUL , 5
paddleocr ,0 ,decode ,RMS_NORM+MUL , 80
pangu-embedded ,0 ,prefill ,RMS_NORM+MUL , 5
pangu-embedded ,0 ,decode ,RMS_NORM+MUL , 80
phi2 ,0 ,prefill ,NORM+MUL+ADD , 3
phi2 ,0 ,prefill ,ADD+ADD , 2
phi2 ,0 ,decode ,NORM+MUL+ADD , 48
phi2 ,0 ,decode ,ADD+ADD , 32
phi3 ,0 ,prefill ,RMS_NORM+MUL , 5
phi3 ,0 ,decode ,RMS_NORM+MUL , 80
phimoe ,1 ,prefill ,RMS_NORM+MUL+ADD , 5
phimoe ,1 ,decode ,RMS_NORM+MUL+ADD , 80
plamo ,0 ,prefill ,RMS_NORM+MUL , 3
plamo ,0 ,prefill ,ADD+ADD , 2
plamo ,0 ,decode ,RMS_NORM+MUL , 48
plamo ,0 ,decode ,ADD+ADD , 32
plamo2 ,0 ,prefill ,RMS_NORM+MUL , 11
plamo2 ,0 ,prefill ,RMS_NORM+MUL+ADD , 3
plamo2 ,0 ,decode ,RMS_NORM+MUL , 176
plamo2 ,0 ,decode ,RMS_NORM+MUL+ADD , 48
pockettts ,0 ,prefill ,NORM+MUL+ADD , 5
pockettts ,0 ,decode ,NORM+MUL+ADD , 80
qwen ,0 ,prefill ,RMS_NORM+MUL , 5
qwen ,0 ,decode ,RMS_NORM+MUL , 80
qwen2 ,0 ,prefill ,RMS_NORM+MUL , 5
qwen2 ,0 ,decode ,RMS_NORM+MUL , 80
qwen2moe ,1 ,prefill ,RMS_NORM+MUL , 5
qwen2moe ,1 ,prefill ,ADD+ADD , 2
qwen2moe ,1 ,decode ,RMS_NORM+MUL , 80
qwen2moe ,1 ,decode ,ADD+ADD , 32
qwen2vl ,0 ,prefill ,RMS_NORM+MUL , 5
qwen2vl ,0 ,decode ,RMS_NORM+MUL , 80
qwen3 ,0 ,prefill ,RMS_NORM+MUL , 9
qwen3 ,0 ,decode ,RMS_NORM+MUL , 144
qwen35 ,0 ,prefill ,RMS_NORM+MUL , 8
qwen35 ,0 ,prefill ,GATED_DELTA_NET+CPY , 1
qwen35 ,0 ,decode ,RMS_NORM+MUL , 128
qwen35 ,0 ,decode ,GATED_DELTA_NET+CPY , 16
qwen35moe ,1 ,prefill ,RMS_NORM+MUL , 8
qwen35moe ,1 ,prefill ,ADD+ADD , 2
qwen35moe ,1 ,prefill ,GATED_DELTA_NET+CPY , 1
qwen35moe ,1 ,decode ,RMS_NORM+MUL , 128
qwen35moe ,1 ,decode ,ADD+ADD , 32
qwen35moe ,1 ,decode ,GATED_DELTA_NET+CPY , 16
qwen3moe ,1 ,prefill ,RMS_NORM+MUL , 9
qwen3moe ,1 ,decode ,RMS_NORM+MUL , 144
qwen3next ,0 ,prefill ,RMS_NORM+MUL , 8
qwen3next ,0 ,prefill ,ADD+ADD , 2
qwen3next ,0 ,prefill ,GATED_DELTA_NET+CPY , 1
qwen3next ,0 ,decode ,RMS_NORM+MUL , 128
qwen3next ,0 ,decode ,ADD+ADD , 32
qwen3next ,0 ,decode ,GATED_DELTA_NET+CPY , 16
qwen3tts ,0 ,prefill ,RMS_NORM+MUL , 9
qwen3tts ,0 ,decode ,RMS_NORM+MUL , 144
qwen3vl ,0 ,prefill ,RMS_NORM+MUL , 9
qwen3vl ,0 ,decode ,RMS_NORM+MUL , 144
qwen3vlmoe ,1 ,prefill ,RMS_NORM+MUL , 9
qwen3vlmoe ,1 ,decode ,RMS_NORM+MUL , 144
qwen4exp ,0 ,prefill ,RMS_NORM+MUL , 5
qwen4exp ,0 ,prefill ,ADD+ADD+ADD , 5
qwen4exp ,0 ,prefill ,GATED_DELTA_NET+CPY , 1
qwen4exp ,0 ,decode ,RMS_NORM+MUL , 80
qwen4exp ,0 ,decode ,ADD+ADD+ADD , 80
qwen4exp ,0 ,decode ,GATED_DELTA_NET+CPY , 16
refact ,0 ,prefill ,RMS_NORM+MUL , 5
refact ,0 ,decode ,RMS_NORM+MUL , 80
refact ,0 ,prefill ,RMS_NORM+MUL , 5
refact ,0 ,decode ,RMS_NORM+MUL , 80
rnd1 ,0 ,prefill ,RMS_NORM+MUL , 9
rnd1 ,0 ,decode ,RMS_NORM+MUL , 144
seed_oss ,0 ,prefill ,RMS_NORM+MUL , 5
seed_oss ,0 ,decode ,RMS_NORM+MUL , 80
smallthinker ,0 ,prefill ,RMS_NORM+MUL , 5
smallthinker ,0 ,decode ,RMS_NORM+MUL , 80
smollm3 ,0 ,prefill ,RMS_NORM+MUL , 5
smollm3 ,0 ,decode ,RMS_NORM+MUL , 80
stablelm ,0 ,prefill ,NORM+MUL , 4
stablelm ,0 ,prefill ,NORM+MUL+ADD , 5
stablelm ,0 ,decode ,NORM+MUL , 64
stablelm ,0 ,decode ,NORM+MUL+ADD , 80
starcoder ,0 ,prefill ,NORM+MUL+ADD , 5
starcoder ,0 ,decode ,NORM+MUL+ADD , 80
starcoder2 ,0 ,prefill ,NORM+MUL+ADD , 5
starcoder2 ,0 ,decode ,NORM+MUL+ADD , 80
talkie ,0 ,prefill ,ADD+ADD , 2
talkie ,0 ,decode ,ADD+ADD , 32
xverse ,0 ,prefill ,RMS_NORM+MUL , 5
xverse ,0 ,decode ,RMS_NORM+MUL , 80
1 # test-fusion baseline for device MTL
2 # arch ,moe ,mode ,label , count
3 arcee ,0 ,prefill ,RMS_NORM+MUL , 5
4 arcee ,0 ,decode ,RMS_NORM+MUL , 80
5 arctic ,0 ,prefill ,RMS_NORM+MUL , 7
6 arctic ,0 ,decode ,RMS_NORM+MUL , 112
7 baichuan ,0 ,prefill ,RMS_NORM+MUL , 5
8 baichuan ,0 ,decode ,RMS_NORM+MUL , 80
9 bailingmoe ,1 ,prefill ,RMS_NORM+MUL , 5
10 bailingmoe ,1 ,prefill ,ADD+ADD , 2
11 bailingmoe ,1 ,decode ,RMS_NORM+MUL , 80
12 bailingmoe ,1 ,decode ,ADD+ADD , 32
13 bailingmoe2 ,1 ,prefill ,RMS_NORM+MUL , 9
14 bailingmoe2 ,1 ,prefill ,ADD+ADD , 1
15 bailingmoe2 ,1 ,decode ,RMS_NORM+MUL , 144
16 bailingmoe2 ,1 ,decode ,ADD+ADD , 16
17 bailingmoe3 ,1 ,prefill ,RMS_NORM+MUL , 8
18 bailingmoe3 ,1 ,prefill ,ADD+ADD , 1
19 bailingmoe3 ,1 ,prefill ,GATED_DELTA_NET+CPY , 1
20 bailingmoe3 ,1 ,decode ,RMS_NORM+MUL , 128
21 bailingmoe3 ,1 ,decode ,ADD+ADD , 16
22 bailingmoe3 ,1 ,decode ,GATED_DELTA_NET+CPY , 16
23 bloom ,0 ,prefill ,NORM+MUL+ADD , 6
24 bloom ,0 ,decode ,NORM+MUL+ADD , 96
25 chatglm ,0 ,prefill ,RMS_NORM+MUL , 5
26 chatglm ,0 ,decode ,RMS_NORM+MUL , 80
27 codeshell ,0 ,prefill ,NORM+MUL+ADD , 5
28 codeshell ,0 ,decode ,NORM+MUL+ADD , 80
29 cogvlm ,0 ,prefill ,RMS_NORM+MUL , 5
30 cogvlm ,0 ,decode ,RMS_NORM+MUL , 80
31 command-r ,0 ,prefill ,NORM+MUL , 3
32 command-r ,0 ,decode ,NORM+MUL , 48
33 dbrx ,0 ,prefill ,NORM+MUL , 5
34 dbrx ,0 ,decode ,NORM+MUL , 80
35 deci ,0 ,prefill ,RMS_NORM+MUL , 5
36 deci ,0 ,decode ,RMS_NORM+MUL , 80
37 deepseek ,0 ,prefill ,RMS_NORM+MUL , 5
38 deepseek ,0 ,prefill ,ADD+ADD , 1
39 deepseek ,0 ,decode ,RMS_NORM+MUL , 80
40 deepseek ,0 ,decode ,ADD+ADD , 16
41 deepseek2 ,0 ,prefill ,RMS_NORM+MUL , 9
42 deepseek2 ,0 ,prefill ,ADD+ADD , 1
43 deepseek2 ,0 ,decode ,RMS_NORM+MUL , 144
44 deepseek2 ,0 ,decode ,ADD+ADD , 16
45 deepseek32 ,0 ,prefill ,NORM+MUL+ADD , 2
46 deepseek32 ,0 ,prefill ,RMS_NORM+MUL , 9
47 deepseek32 ,0 ,prefill ,ADD+ADD , 1
48 deepseek32 ,0 ,decode ,NORM+MUL+ADD , 32
49 deepseek32 ,0 ,decode ,RMS_NORM+MUL , 144
50 deepseek32 ,0 ,decode ,ADD+ADD , 16
51 deepseek4 ,0 ,prefill ,RMS_NORM+MUL , 20
52 deepseek4 ,0 ,decode ,RMS_NORM+MUL , 320
53 dots1 ,0 ,prefill ,RMS_NORM+MUL , 9
54 dots1 ,0 ,prefill ,ADD+ADD , 1
55 dots1 ,0 ,decode ,RMS_NORM+MUL , 144
56 dots1 ,0 ,decode ,ADD+ADD , 16
57 dream ,0 ,prefill ,RMS_NORM+MUL , 5
58 dream ,0 ,decode ,RMS_NORM+MUL , 80
59 ernie4_5-moe ,1 ,prefill ,RMS_NORM+MUL , 5
60 ernie4_5-moe ,1 ,prefill ,ADD+ADD , 1
61 ernie4_5-moe ,1 ,decode ,RMS_NORM+MUL , 80
62 ernie4_5-moe ,1 ,decode ,ADD+ADD , 16
63 ernie4_5 ,0 ,prefill ,RMS_NORM+MUL , 5
64 ernie4_5 ,0 ,decode ,RMS_NORM+MUL , 80
65 exaone ,0 ,prefill ,RMS_NORM+MUL , 5
66 exaone ,0 ,decode ,RMS_NORM+MUL , 80
67 exaone4 ,0 ,prefill ,RMS_NORM+MUL , 5
68 exaone4 ,0 ,prefill ,RMS_NORM+MUL+ADD , 4
69 exaone4 ,0 ,decode ,RMS_NORM+MUL , 80
70 exaone4 ,0 ,decode ,RMS_NORM+MUL+ADD , 64
71 falcon ,0 ,prefill ,NORM+MUL+ADD , 5
72 falcon ,0 ,prefill ,ADD+ADD , 2
73 falcon ,0 ,decode ,NORM+MUL+ADD , 80
74 falcon ,0 ,decode ,ADD+ADD , 32
75 falcon-h1 ,0 ,prefill ,RMS_NORM+MUL , 9
76 falcon-h1 ,0 ,prefill ,ADD+ADD , 2
77 falcon-h1 ,0 ,decode ,RMS_NORM+MUL , 144
78 falcon-h1 ,0 ,decode ,ADD+ADD , 32
79 gemma ,0 ,prefill ,RMS_NORM+MUL , 5
80 gemma ,0 ,decode ,RMS_NORM+MUL , 80
81 gemma2 ,0 ,prefill ,RMS_NORM+MUL , 5
82 gemma2 ,0 ,prefill ,RMS_NORM+MUL+ADD , 4
83 gemma2 ,0 ,decode ,RMS_NORM+MUL , 80
84 gemma2 ,0 ,decode ,RMS_NORM+MUL+ADD , 64
85 glm-dsa ,0 ,prefill ,NORM+MUL+ADD , 2
86 glm-dsa ,0 ,prefill ,RMS_NORM+MUL , 9
87 glm-dsa ,0 ,prefill ,ADD+ADD , 1
88 glm-dsa ,0 ,decode ,NORM+MUL+ADD , 32
89 glm-dsa ,0 ,decode ,RMS_NORM+MUL , 144
90 glm-dsa ,0 ,decode ,ADD+ADD , 16
91 glm4 ,0 ,prefill ,RMS_NORM+MUL , 5
92 glm4 ,0 ,prefill ,RMS_NORM+MUL+ADD , 4
93 glm4 ,0 ,decode ,RMS_NORM+MUL , 80
94 glm4 ,0 ,decode ,RMS_NORM+MUL+ADD , 64
95 glm4moe ,1 ,prefill ,RMS_NORM+MUL , 9
96 glm4moe ,1 ,prefill ,ADD+ADD , 1
97 glm4moe ,1 ,decode ,RMS_NORM+MUL , 144
98 glm4moe ,1 ,decode ,ADD+ADD , 16
99 gpt-oss ,0 ,prefill ,RMS_NORM+MUL , 5
100 gpt-oss ,0 ,decode ,RMS_NORM+MUL , 80
101 gpt2 ,0 ,prefill ,NORM+MUL+ADD , 5
102 gpt2 ,0 ,decode ,NORM+MUL+ADD , 80
103 gptneox ,0 ,prefill ,NORM+MUL+ADD , 5
104 gptneox ,0 ,decode ,NORM+MUL+ADD , 80
105 granite ,0 ,prefill ,RMS_NORM+MUL , 5
106 granite ,0 ,decode ,RMS_NORM+MUL , 80
107 granite ,0 ,prefill ,RMS_NORM+MUL , 5
108 granite ,0 ,decode ,RMS_NORM+MUL , 80
109 granitehybrid ,0 ,prefill ,RMS_NORM+MUL , 6
110 granitehybrid ,0 ,decode ,RMS_NORM+MUL , 96
111 granitemoe ,1 ,prefill ,RMS_NORM+MUL , 5
112 granitemoe ,1 ,decode ,RMS_NORM+MUL , 80
113 granitemoe ,1 ,prefill ,RMS_NORM+MUL , 5
114 granitemoe ,1 ,decode ,RMS_NORM+MUL , 80
115 grok ,0 ,prefill ,RMS_NORM+MUL , 5
116 grok ,0 ,prefill ,RMS_NORM+MUL+ADD , 4
117 grok ,0 ,decode ,RMS_NORM+MUL , 80
118 grok ,0 ,decode ,RMS_NORM+MUL+ADD , 64
119 grovemoe ,1 ,prefill ,RMS_NORM+MUL , 9
120 grovemoe ,1 ,prefill ,ADD+ADD , 2
121 grovemoe ,1 ,decode ,RMS_NORM+MUL , 144
122 grovemoe ,1 ,decode ,ADD+ADD , 32
123 hunyuan-dense ,0 ,prefill ,RMS_NORM+MUL , 9
124 hunyuan-dense ,0 ,decode ,RMS_NORM+MUL , 144
125 hunyuan-moe ,1 ,prefill ,RMS_NORM+MUL , 9
126 hunyuan-moe ,1 ,prefill ,ADD+ADD , 2
127 hunyuan-moe ,1 ,decode ,RMS_NORM+MUL , 144
128 hunyuan-moe ,1 ,decode ,ADD+ADD , 32
129 hunyuan_vl ,0 ,prefill ,RMS_NORM+MUL , 9
130 hunyuan_vl ,0 ,decode ,RMS_NORM+MUL , 144
131 hy_v3 ,0 ,prefill ,RMS_NORM+MUL , 9
132 hy_v3 ,0 ,prefill ,ADD+ADD , 2
133 hy_v3 ,0 ,decode ,RMS_NORM+MUL , 144
134 hy_v3 ,0 ,decode ,ADD+ADD , 32
135 internlm2 ,0 ,prefill ,RMS_NORM+MUL , 5
136 internlm2 ,0 ,decode ,RMS_NORM+MUL , 80
137 jais ,0 ,prefill ,NORM+MUL+ADD , 5
138 jais ,0 ,decode ,NORM+MUL+ADD , 80
139 jais2 ,0 ,prefill ,NORM+MUL+ADD , 5
140 jais2 ,0 ,decode ,NORM+MUL+ADD , 80
141 jamba ,0 ,prefill ,RMS_NORM+MUL , 8
142 jamba ,0 ,decode ,RMS_NORM+MUL , 128
143 kimi-k3 ,0 ,prefill ,RMS_NORM+MUL , 17
144 kimi-k3 ,0 ,prefill ,GATED_DELTA_NET+CPY , 1
145 kimi-k3 ,0 ,decode ,RMS_NORM+MUL , 272
146 kimi-k3 ,0 ,decode ,GATED_DELTA_NET+CPY , 16
147 kimi-linear ,0 ,prefill ,RMS_NORM+MUL , 7
148 kimi-linear ,0 ,prefill ,ADD+ADD , 1
149 kimi-linear ,0 ,prefill ,GATED_DELTA_NET+CPY , 1
150 kimi-linear ,0 ,decode ,RMS_NORM+MUL , 112
151 kimi-linear ,0 ,decode ,ADD+ADD , 16
152 kimi-linear ,0 ,decode ,GATED_DELTA_NET+CPY , 16
153 lfm2 ,0 ,prefill ,RMS_NORM+MUL , 7
154 lfm2 ,0 ,decode ,RMS_NORM+MUL , 112
155 lfm2moe ,1 ,prefill ,RMS_NORM+MUL , 7
156 lfm2moe ,1 ,decode ,RMS_NORM+MUL , 112
157 llada ,0 ,prefill ,RMS_NORM+MUL , 5
158 llada ,0 ,decode ,RMS_NORM+MUL , 80
159 llada-moe ,1 ,prefill ,RMS_NORM+MUL , 9
160 llada-moe ,1 ,decode ,RMS_NORM+MUL , 144
161 llama ,0 ,prefill ,RMS_NORM+MUL , 5
162 llama ,0 ,decode ,RMS_NORM+MUL , 80
163 llama ,0 ,prefill ,RMS_NORM+MUL , 5
164 llama ,0 ,decode ,RMS_NORM+MUL , 80
165 llama4 ,0 ,prefill ,RMS_NORM+MUL , 9
166 llama4 ,0 ,prefill ,ADD+ADD , 2
167 llama4 ,0 ,decode ,RMS_NORM+MUL , 144
168 llama4 ,0 ,decode ,ADD+ADD , 32
169 maincoder ,0 ,prefill ,RMS_NORM+MUL , 9
170 maincoder ,0 ,decode ,RMS_NORM+MUL , 144
171 mamba ,0 ,prefill ,RMS_NORM+MUL , 3
172 mamba ,0 ,decode ,RMS_NORM+MUL , 48
173 mamba2 ,0 ,prefill ,RMS_NORM+MUL , 5
174 mamba2 ,0 ,decode ,RMS_NORM+MUL , 80
175 minicpm ,0 ,prefill ,RMS_NORM+MUL , 5
176 minicpm ,0 ,decode ,RMS_NORM+MUL , 80
177 minicpm ,0 ,prefill ,RMS_NORM+MUL , 5
178 minicpm ,0 ,decode ,RMS_NORM+MUL , 80
179 minicpm3 ,0 ,prefill ,RMS_NORM+MUL , 9
180 minicpm3 ,0 ,decode ,RMS_NORM+MUL , 144
181 minimax-01 ,0 ,prefill ,RMS_NORM+MUL , 6
182 minimax-01 ,0 ,decode ,RMS_NORM+MUL , 96
183 minimax-m2 ,0 ,prefill ,RMS_NORM+MUL , 9
184 minimax-m2 ,0 ,decode ,RMS_NORM+MUL , 144
185 minimax-m3 ,0 ,prefill ,RMS_NORM+MUL , 11
186 minimax-m3 ,0 ,prefill ,ADD+ADD , 1
187 minimax-m3 ,0 ,decode ,RMS_NORM+MUL , 176
188 minimax-m3 ,0 ,decode ,ADD+ADD , 16
189 mistral3 ,0 ,prefill ,RMS_NORM+MUL , 5
190 mistral3 ,0 ,decode ,RMS_NORM+MUL , 80
191 mistral3 ,0 ,prefill ,RMS_NORM+MUL , 5
192 mistral3 ,0 ,decode ,RMS_NORM+MUL , 80
193 mistral4 ,0 ,prefill ,RMS_NORM+MUL , 9
194 mistral4 ,0 ,prefill ,ADD+ADD , 1
195 mistral4 ,0 ,decode ,RMS_NORM+MUL , 144
196 mistral4 ,0 ,decode ,ADD+ADD , 16
197 mpt ,0 ,prefill ,NORM+MUL+ADD , 5
198 mpt ,0 ,decode ,NORM+MUL+ADD , 80
199 nanbeige ,0 ,prefill ,RMS_NORM+MUL , 5
200 nanbeige ,0 ,decode ,RMS_NORM+MUL , 80
201 nemotron ,0 ,prefill ,NORM+MUL+ADD , 5
202 nemotron ,0 ,decode ,NORM+MUL+ADD , 80
203 nemotron_h ,0 ,prefill ,RMS_NORM+MUL , 5
204 nemotron_h ,0 ,decode ,RMS_NORM+MUL , 80
205 nemotron_h_moe ,1 ,prefill ,RMS_NORM+MUL , 5
206 nemotron_h_moe ,1 ,decode ,RMS_NORM+MUL , 80
207 olmoe ,1 ,prefill ,RMS_NORM+MUL , 9
208 olmoe ,1 ,decode ,RMS_NORM+MUL , 144
209 openelm ,0 ,prefill ,RMS_NORM+MUL , 9
210 openelm ,0 ,decode ,RMS_NORM+MUL , 144
211 orion ,0 ,prefill ,NORM+MUL+ADD , 5
212 orion ,0 ,decode ,NORM+MUL+ADD , 80
213 paddleocr ,0 ,prefill ,RMS_NORM+MUL , 5
214 paddleocr ,0 ,decode ,RMS_NORM+MUL , 80
215 pangu-embedded ,0 ,prefill ,RMS_NORM+MUL , 5
216 pangu-embedded ,0 ,decode ,RMS_NORM+MUL , 80
217 phi2 ,0 ,prefill ,NORM+MUL+ADD , 3
218 phi2 ,0 ,prefill ,ADD+ADD , 2
219 phi2 ,0 ,decode ,NORM+MUL+ADD , 48
220 phi2 ,0 ,decode ,ADD+ADD , 32
221 phi3 ,0 ,prefill ,RMS_NORM+MUL , 5
222 phi3 ,0 ,decode ,RMS_NORM+MUL , 80
223 phimoe ,1 ,prefill ,RMS_NORM+MUL+ADD , 5
224 phimoe ,1 ,decode ,RMS_NORM+MUL+ADD , 80
225 plamo ,0 ,prefill ,RMS_NORM+MUL , 3
226 plamo ,0 ,prefill ,ADD+ADD , 2
227 plamo ,0 ,decode ,RMS_NORM+MUL , 48
228 plamo ,0 ,decode ,ADD+ADD , 32
229 plamo2 ,0 ,prefill ,RMS_NORM+MUL , 11
230 plamo2 ,0 ,prefill ,RMS_NORM+MUL+ADD , 3
231 plamo2 ,0 ,decode ,RMS_NORM+MUL , 176
232 plamo2 ,0 ,decode ,RMS_NORM+MUL+ADD , 48
233 pockettts ,0 ,prefill ,NORM+MUL+ADD , 5
234 pockettts ,0 ,decode ,NORM+MUL+ADD , 80
235 qwen ,0 ,prefill ,RMS_NORM+MUL , 5
236 qwen ,0 ,decode ,RMS_NORM+MUL , 80
237 qwen2 ,0 ,prefill ,RMS_NORM+MUL , 5
238 qwen2 ,0 ,decode ,RMS_NORM+MUL , 80
239 qwen2moe ,1 ,prefill ,RMS_NORM+MUL , 5
240 qwen2moe ,1 ,prefill ,ADD+ADD , 2
241 qwen2moe ,1 ,decode ,RMS_NORM+MUL , 80
242 qwen2moe ,1 ,decode ,ADD+ADD , 32
243 qwen2vl ,0 ,prefill ,RMS_NORM+MUL , 5
244 qwen2vl ,0 ,decode ,RMS_NORM+MUL , 80
245 qwen3 ,0 ,prefill ,RMS_NORM+MUL , 9
246 qwen3 ,0 ,decode ,RMS_NORM+MUL , 144
247 qwen35 ,0 ,prefill ,RMS_NORM+MUL , 8
248 qwen35 ,0 ,prefill ,GATED_DELTA_NET+CPY , 1
249 qwen35 ,0 ,decode ,RMS_NORM+MUL , 128
250 qwen35 ,0 ,decode ,GATED_DELTA_NET+CPY , 16
251 qwen35moe ,1 ,prefill ,RMS_NORM+MUL , 8
252 qwen35moe ,1 ,prefill ,ADD+ADD , 2
253 qwen35moe ,1 ,prefill ,GATED_DELTA_NET+CPY , 1
254 qwen35moe ,1 ,decode ,RMS_NORM+MUL , 128
255 qwen35moe ,1 ,decode ,ADD+ADD , 32
256 qwen35moe ,1 ,decode ,GATED_DELTA_NET+CPY , 16
257 qwen3moe ,1 ,prefill ,RMS_NORM+MUL , 9
258 qwen3moe ,1 ,decode ,RMS_NORM+MUL , 144
259 qwen3next ,0 ,prefill ,RMS_NORM+MUL , 8
260 qwen3next ,0 ,prefill ,ADD+ADD , 2
261 qwen3next ,0 ,prefill ,GATED_DELTA_NET+CPY , 1
262 qwen3next ,0 ,decode ,RMS_NORM+MUL , 128
263 qwen3next ,0 ,decode ,ADD+ADD , 32
264 qwen3next ,0 ,decode ,GATED_DELTA_NET+CPY , 16
265 qwen3tts ,0 ,prefill ,RMS_NORM+MUL , 9
266 qwen3tts ,0 ,decode ,RMS_NORM+MUL , 144
267 qwen3vl ,0 ,prefill ,RMS_NORM+MUL , 9
268 qwen3vl ,0 ,decode ,RMS_NORM+MUL , 144
269 qwen3vlmoe ,1 ,prefill ,RMS_NORM+MUL , 9
270 qwen3vlmoe ,1 ,decode ,RMS_NORM+MUL , 144
271 qwen4exp ,0 ,prefill ,RMS_NORM+MUL , 5
272 qwen4exp ,0 ,prefill ,ADD+ADD+ADD , 5
273 qwen4exp ,0 ,prefill ,GATED_DELTA_NET+CPY , 1
274 qwen4exp ,0 ,decode ,RMS_NORM+MUL , 80
275 qwen4exp ,0 ,decode ,ADD+ADD+ADD , 80
276 qwen4exp ,0 ,decode ,GATED_DELTA_NET+CPY , 16
277 refact ,0 ,prefill ,RMS_NORM+MUL , 5
278 refact ,0 ,decode ,RMS_NORM+MUL , 80
279 refact ,0 ,prefill ,RMS_NORM+MUL , 5
280 refact ,0 ,decode ,RMS_NORM+MUL , 80
281 rnd1 ,0 ,prefill ,RMS_NORM+MUL , 9
282 rnd1 ,0 ,decode ,RMS_NORM+MUL , 144
283 seed_oss ,0 ,prefill ,RMS_NORM+MUL , 5
284 seed_oss ,0 ,decode ,RMS_NORM+MUL , 80
285 smallthinker ,0 ,prefill ,RMS_NORM+MUL , 5
286 smallthinker ,0 ,decode ,RMS_NORM+MUL , 80
287 smollm3 ,0 ,prefill ,RMS_NORM+MUL , 5
288 smollm3 ,0 ,decode ,RMS_NORM+MUL , 80
289 stablelm ,0 ,prefill ,NORM+MUL , 4
290 stablelm ,0 ,prefill ,NORM+MUL+ADD , 5
291 stablelm ,0 ,decode ,NORM+MUL , 64
292 stablelm ,0 ,decode ,NORM+MUL+ADD , 80
293 starcoder ,0 ,prefill ,NORM+MUL+ADD , 5
294 starcoder ,0 ,decode ,NORM+MUL+ADD , 80
295 starcoder2 ,0 ,prefill ,NORM+MUL+ADD , 5
296 starcoder2 ,0 ,decode ,NORM+MUL+ADD , 80
297 talkie ,0 ,prefill ,ADD+ADD , 2
298 talkie ,0 ,decode ,ADD+ADD , 32
299 xverse ,0 ,prefill ,RMS_NORM+MUL , 5
300 xverse ,0 ,decode ,RMS_NORM+MUL , 80
-300
View File
@@ -1,300 +0,0 @@
# test-fusion baseline for device MTL
# arch moe mode label count
arcee 0 prefill RMS_NORM+MUL 5
arcee 0 decode RMS_NORM+MUL 80
arctic 0 prefill RMS_NORM+MUL 7
arctic 0 decode RMS_NORM+MUL 112
baichuan 0 prefill RMS_NORM+MUL 5
baichuan 0 decode RMS_NORM+MUL 80
bailingmoe 1 prefill RMS_NORM+MUL 5
bailingmoe 1 prefill ADD+ADD 2
bailingmoe 1 decode RMS_NORM+MUL 80
bailingmoe 1 decode ADD+ADD 32
bailingmoe2 1 prefill RMS_NORM+MUL 9
bailingmoe2 1 prefill ADD+ADD 1
bailingmoe2 1 decode RMS_NORM+MUL 144
bailingmoe2 1 decode ADD+ADD 16
bailingmoe3 1 prefill RMS_NORM+MUL 8
bailingmoe3 1 prefill ADD+ADD 1
bailingmoe3 1 prefill GATED_DELTA_NET+CPY 1
bailingmoe3 1 decode RMS_NORM+MUL 128
bailingmoe3 1 decode ADD+ADD 16
bailingmoe3 1 decode GATED_DELTA_NET+CPY 16
bloom 0 prefill NORM+MUL+ADD 6
bloom 0 decode NORM+MUL+ADD 96
chatglm 0 prefill RMS_NORM+MUL 5
chatglm 0 decode RMS_NORM+MUL 80
codeshell 0 prefill NORM+MUL+ADD 5
codeshell 0 decode NORM+MUL+ADD 80
cogvlm 0 prefill RMS_NORM+MUL 5
cogvlm 0 decode RMS_NORM+MUL 80
command-r 0 prefill NORM+MUL 3
command-r 0 decode NORM+MUL 48
dbrx 0 prefill NORM+MUL 5
dbrx 0 decode NORM+MUL 80
deci 0 prefill RMS_NORM+MUL 5
deci 0 decode RMS_NORM+MUL 80
deepseek 0 prefill RMS_NORM+MUL 5
deepseek 0 prefill ADD+ADD 1
deepseek 0 decode RMS_NORM+MUL 80
deepseek 0 decode ADD+ADD 16
deepseek2 0 prefill RMS_NORM+MUL 9
deepseek2 0 prefill ADD+ADD 1
deepseek2 0 decode RMS_NORM+MUL 144
deepseek2 0 decode ADD+ADD 16
deepseek32 0 prefill NORM+MUL+ADD 2
deepseek32 0 prefill RMS_NORM+MUL 9
deepseek32 0 prefill ADD+ADD 1
deepseek32 0 decode NORM+MUL+ADD 32
deepseek32 0 decode RMS_NORM+MUL 144
deepseek32 0 decode ADD+ADD 16
deepseek4 0 prefill RMS_NORM+MUL 20
deepseek4 0 decode RMS_NORM+MUL 320
dots1 0 prefill RMS_NORM+MUL 9
dots1 0 prefill ADD+ADD 1
dots1 0 decode RMS_NORM+MUL 144
dots1 0 decode ADD+ADD 16
dream 0 prefill RMS_NORM+MUL 5
dream 0 decode RMS_NORM+MUL 80
ernie4_5-moe 1 prefill RMS_NORM+MUL 5
ernie4_5-moe 1 prefill ADD+ADD 1
ernie4_5-moe 1 decode RMS_NORM+MUL 80
ernie4_5-moe 1 decode ADD+ADD 16
ernie4_5 0 prefill RMS_NORM+MUL 5
ernie4_5 0 decode RMS_NORM+MUL 80
exaone 0 prefill RMS_NORM+MUL 5
exaone 0 decode RMS_NORM+MUL 80
exaone4 0 prefill RMS_NORM+MUL 5
exaone4 0 prefill RMS_NORM+MUL+ADD 4
exaone4 0 decode RMS_NORM+MUL 80
exaone4 0 decode RMS_NORM+MUL+ADD 64
falcon 0 prefill NORM+MUL+ADD 5
falcon 0 prefill ADD+ADD 2
falcon 0 decode NORM+MUL+ADD 80
falcon 0 decode ADD+ADD 32
falcon-h1 0 prefill RMS_NORM+MUL 9
falcon-h1 0 prefill ADD+ADD 2
falcon-h1 0 decode RMS_NORM+MUL 144
falcon-h1 0 decode ADD+ADD 32
gemma 0 prefill RMS_NORM+MUL 5
gemma 0 decode RMS_NORM+MUL 80
gemma2 0 prefill RMS_NORM+MUL 5
gemma2 0 prefill RMS_NORM+MUL+ADD 4
gemma2 0 decode RMS_NORM+MUL 80
gemma2 0 decode RMS_NORM+MUL+ADD 64
glm-dsa 0 prefill NORM+MUL+ADD 2
glm-dsa 0 prefill RMS_NORM+MUL 9
glm-dsa 0 prefill ADD+ADD 1
glm-dsa 0 decode NORM+MUL+ADD 32
glm-dsa 0 decode RMS_NORM+MUL 144
glm-dsa 0 decode ADD+ADD 16
glm4 0 prefill RMS_NORM+MUL 5
glm4 0 prefill RMS_NORM+MUL+ADD 4
glm4 0 decode RMS_NORM+MUL 80
glm4 0 decode RMS_NORM+MUL+ADD 64
glm4moe 1 prefill RMS_NORM+MUL 9
glm4moe 1 prefill ADD+ADD 1
glm4moe 1 decode RMS_NORM+MUL 144
glm4moe 1 decode ADD+ADD 16
gpt-oss 0 prefill RMS_NORM+MUL 5
gpt-oss 0 decode RMS_NORM+MUL 80
gpt2 0 prefill NORM+MUL+ADD 5
gpt2 0 decode NORM+MUL+ADD 80
gptneox 0 prefill NORM+MUL+ADD 5
gptneox 0 decode NORM+MUL+ADD 80
granite 0 prefill RMS_NORM+MUL 5
granite 0 decode RMS_NORM+MUL 80
granite 0 prefill RMS_NORM+MUL 5
granite 0 decode RMS_NORM+MUL 80
granitehybrid 0 prefill RMS_NORM+MUL 6
granitehybrid 0 decode RMS_NORM+MUL 96
granitemoe 1 prefill RMS_NORM+MUL 5
granitemoe 1 decode RMS_NORM+MUL 80
granitemoe 1 prefill RMS_NORM+MUL 5
granitemoe 1 decode RMS_NORM+MUL 80
grok 0 prefill RMS_NORM+MUL 5
grok 0 prefill RMS_NORM+MUL+ADD 4
grok 0 decode RMS_NORM+MUL 80
grok 0 decode RMS_NORM+MUL+ADD 64
grovemoe 1 prefill RMS_NORM+MUL 9
grovemoe 1 prefill ADD+ADD 2
grovemoe 1 decode RMS_NORM+MUL 144
grovemoe 1 decode ADD+ADD 32
hunyuan-dense 0 prefill RMS_NORM+MUL 9
hunyuan-dense 0 decode RMS_NORM+MUL 144
hunyuan-moe 1 prefill RMS_NORM+MUL 9
hunyuan-moe 1 prefill ADD+ADD 2
hunyuan-moe 1 decode RMS_NORM+MUL 144
hunyuan-moe 1 decode ADD+ADD 32
hunyuan_vl 0 prefill RMS_NORM+MUL 9
hunyuan_vl 0 decode RMS_NORM+MUL 144
hy_v3 0 prefill RMS_NORM+MUL 9
hy_v3 0 prefill ADD+ADD 2
hy_v3 0 decode RMS_NORM+MUL 144
hy_v3 0 decode ADD+ADD 32
internlm2 0 prefill RMS_NORM+MUL 5
internlm2 0 decode RMS_NORM+MUL 80
jais 0 prefill NORM+MUL+ADD 5
jais 0 decode NORM+MUL+ADD 80
jais2 0 prefill NORM+MUL+ADD 5
jais2 0 decode NORM+MUL+ADD 80
jamba 0 prefill RMS_NORM+MUL 8
jamba 0 decode RMS_NORM+MUL 128
kimi-k3 0 prefill RMS_NORM+MUL 17
kimi-k3 0 prefill GATED_DELTA_NET+CPY 1
kimi-k3 0 decode RMS_NORM+MUL 272
kimi-k3 0 decode GATED_DELTA_NET+CPY 16
kimi-linear 0 prefill RMS_NORM+MUL 7
kimi-linear 0 prefill ADD+ADD 1
kimi-linear 0 prefill GATED_DELTA_NET+CPY 1
kimi-linear 0 decode RMS_NORM+MUL 112
kimi-linear 0 decode ADD+ADD 16
kimi-linear 0 decode GATED_DELTA_NET+CPY 16
lfm2 0 prefill RMS_NORM+MUL 7
lfm2 0 decode RMS_NORM+MUL 112
lfm2moe 1 prefill RMS_NORM+MUL 7
lfm2moe 1 decode RMS_NORM+MUL 112
llada 0 prefill RMS_NORM+MUL 5
llada 0 decode RMS_NORM+MUL 80
llada-moe 1 prefill RMS_NORM+MUL 9
llada-moe 1 decode RMS_NORM+MUL 144
llama 0 prefill RMS_NORM+MUL 5
llama 0 decode RMS_NORM+MUL 80
llama 0 prefill RMS_NORM+MUL 5
llama 0 decode RMS_NORM+MUL 80
llama4 0 prefill RMS_NORM+MUL 9
llama4 0 prefill ADD+ADD 2
llama4 0 decode RMS_NORM+MUL 144
llama4 0 decode ADD+ADD 32
maincoder 0 prefill RMS_NORM+MUL 9
maincoder 0 decode RMS_NORM+MUL 144
mamba 0 prefill RMS_NORM+MUL 3
mamba 0 decode RMS_NORM+MUL 48
mamba2 0 prefill RMS_NORM+MUL 5
mamba2 0 decode RMS_NORM+MUL 80
minicpm 0 prefill RMS_NORM+MUL 5
minicpm 0 decode RMS_NORM+MUL 80
minicpm 0 prefill RMS_NORM+MUL 5
minicpm 0 decode RMS_NORM+MUL 80
minicpm3 0 prefill RMS_NORM+MUL 9
minicpm3 0 decode RMS_NORM+MUL 144
minimax-01 0 prefill RMS_NORM+MUL 6
minimax-01 0 decode RMS_NORM+MUL 96
minimax-m2 0 prefill RMS_NORM+MUL 9
minimax-m2 0 decode RMS_NORM+MUL 144
minimax-m3 0 prefill RMS_NORM+MUL 11
minimax-m3 0 prefill ADD+ADD 1
minimax-m3 0 decode RMS_NORM+MUL 176
minimax-m3 0 decode ADD+ADD 16
mistral3 0 prefill RMS_NORM+MUL 5
mistral3 0 decode RMS_NORM+MUL 80
mistral3 0 prefill RMS_NORM+MUL 5
mistral3 0 decode RMS_NORM+MUL 80
mistral4 0 prefill RMS_NORM+MUL 9
mistral4 0 prefill ADD+ADD 1
mistral4 0 decode RMS_NORM+MUL 144
mistral4 0 decode ADD+ADD 16
mpt 0 prefill NORM+MUL+ADD 5
mpt 0 decode NORM+MUL+ADD 80
nanbeige 0 prefill RMS_NORM+MUL 5
nanbeige 0 decode RMS_NORM+MUL 80
nemotron 0 prefill NORM+MUL+ADD 5
nemotron 0 decode NORM+MUL+ADD 80
nemotron_h 0 prefill RMS_NORM+MUL 5
nemotron_h 0 decode RMS_NORM+MUL 80
nemotron_h_moe 1 prefill RMS_NORM+MUL 5
nemotron_h_moe 1 decode RMS_NORM+MUL 80
olmoe 1 prefill RMS_NORM+MUL 9
olmoe 1 decode RMS_NORM+MUL 144
openelm 0 prefill RMS_NORM+MUL 9
openelm 0 decode RMS_NORM+MUL 144
orion 0 prefill NORM+MUL+ADD 5
orion 0 decode NORM+MUL+ADD 80
paddleocr 0 prefill RMS_NORM+MUL 5
paddleocr 0 decode RMS_NORM+MUL 80
pangu-embedded 0 prefill RMS_NORM+MUL 5
pangu-embedded 0 decode RMS_NORM+MUL 80
phi2 0 prefill NORM+MUL+ADD 3
phi2 0 prefill ADD+ADD 2
phi2 0 decode NORM+MUL+ADD 48
phi2 0 decode ADD+ADD 32
phi3 0 prefill RMS_NORM+MUL 5
phi3 0 decode RMS_NORM+MUL 80
phimoe 1 prefill RMS_NORM+MUL+ADD 5
phimoe 1 decode RMS_NORM+MUL+ADD 80
plamo 0 prefill RMS_NORM+MUL 3
plamo 0 prefill ADD+ADD 2
plamo 0 decode RMS_NORM+MUL 48
plamo 0 decode ADD+ADD 32
plamo2 0 prefill RMS_NORM+MUL 11
plamo2 0 prefill RMS_NORM+MUL+ADD 3
plamo2 0 decode RMS_NORM+MUL 176
plamo2 0 decode RMS_NORM+MUL+ADD 48
pockettts 0 prefill NORM+MUL+ADD 5
pockettts 0 decode NORM+MUL+ADD 80
qwen 0 prefill RMS_NORM+MUL 5
qwen 0 decode RMS_NORM+MUL 80
qwen2 0 prefill RMS_NORM+MUL 5
qwen2 0 decode RMS_NORM+MUL 80
qwen2moe 1 prefill RMS_NORM+MUL 5
qwen2moe 1 prefill ADD+ADD 2
qwen2moe 1 decode RMS_NORM+MUL 80
qwen2moe 1 decode ADD+ADD 32
qwen2vl 0 prefill RMS_NORM+MUL 5
qwen2vl 0 decode RMS_NORM+MUL 80
qwen3 0 prefill RMS_NORM+MUL 9
qwen3 0 decode RMS_NORM+MUL 144
qwen35 0 prefill RMS_NORM+MUL 8
qwen35 0 prefill GATED_DELTA_NET+CPY 1
qwen35 0 decode RMS_NORM+MUL 128
qwen35 0 decode GATED_DELTA_NET+CPY 16
qwen35moe 1 prefill RMS_NORM+MUL 8
qwen35moe 1 prefill ADD+ADD 2
qwen35moe 1 prefill GATED_DELTA_NET+CPY 1
qwen35moe 1 decode RMS_NORM+MUL 128
qwen35moe 1 decode ADD+ADD 32
qwen35moe 1 decode GATED_DELTA_NET+CPY 16
qwen3moe 1 prefill RMS_NORM+MUL 9
qwen3moe 1 decode RMS_NORM+MUL 144
qwen3next 0 prefill RMS_NORM+MUL 8
qwen3next 0 prefill ADD+ADD 2
qwen3next 0 prefill GATED_DELTA_NET+CPY 1
qwen3next 0 decode RMS_NORM+MUL 128
qwen3next 0 decode ADD+ADD 32
qwen3next 0 decode GATED_DELTA_NET+CPY 16
qwen3tts 0 prefill RMS_NORM+MUL 9
qwen3tts 0 decode RMS_NORM+MUL 144
qwen3vl 0 prefill RMS_NORM+MUL 9
qwen3vl 0 decode RMS_NORM+MUL 144
qwen3vlmoe 1 prefill RMS_NORM+MUL 9
qwen3vlmoe 1 decode RMS_NORM+MUL 144
qwen4exp 0 prefill RMS_NORM+MUL 5
qwen4exp 0 prefill ADD+ADD+ADD 5
qwen4exp 0 prefill GATED_DELTA_NET+CPY 1
qwen4exp 0 decode RMS_NORM+MUL 80
qwen4exp 0 decode ADD+ADD+ADD 80
qwen4exp 0 decode GATED_DELTA_NET+CPY 16
refact 0 prefill RMS_NORM+MUL 5
refact 0 decode RMS_NORM+MUL 80
refact 0 prefill RMS_NORM+MUL 5
refact 0 decode RMS_NORM+MUL 80
rnd1 0 prefill RMS_NORM+MUL 9
rnd1 0 decode RMS_NORM+MUL 144
seed_oss 0 prefill RMS_NORM+MUL 5
seed_oss 0 decode RMS_NORM+MUL 80
smallthinker 0 prefill RMS_NORM+MUL 5
smallthinker 0 decode RMS_NORM+MUL 80
smollm3 0 prefill RMS_NORM+MUL 5
smollm3 0 decode RMS_NORM+MUL 80
stablelm 0 prefill NORM+MUL 4
stablelm 0 prefill NORM+MUL+ADD 5
stablelm 0 decode NORM+MUL 64
stablelm 0 decode NORM+MUL+ADD 80
starcoder 0 prefill NORM+MUL+ADD 5
starcoder 0 decode NORM+MUL+ADD 80
starcoder2 0 prefill NORM+MUL+ADD 5
starcoder2 0 decode NORM+MUL+ADD 80
talkie 0 prefill ADD+ADD 2
talkie 0 decode ADD+ADD 32
xverse 0 prefill RMS_NORM+MUL 5
xverse 0 decode RMS_NORM+MUL 80
1 # test-fusion baseline for device MTL
2 # arch moe mode label count
3 arcee 0 prefill RMS_NORM+MUL 5
4 arcee 0 decode RMS_NORM+MUL 80
5 arctic 0 prefill RMS_NORM+MUL 7
6 arctic 0 decode RMS_NORM+MUL 112
7 baichuan 0 prefill RMS_NORM+MUL 5
8 baichuan 0 decode RMS_NORM+MUL 80
9 bailingmoe 1 prefill RMS_NORM+MUL 5
10 bailingmoe 1 prefill ADD+ADD 2
11 bailingmoe 1 decode RMS_NORM+MUL 80
12 bailingmoe 1 decode ADD+ADD 32
13 bailingmoe2 1 prefill RMS_NORM+MUL 9
14 bailingmoe2 1 prefill ADD+ADD 1
15 bailingmoe2 1 decode RMS_NORM+MUL 144
16 bailingmoe2 1 decode ADD+ADD 16
17 bailingmoe3 1 prefill RMS_NORM+MUL 8
18 bailingmoe3 1 prefill ADD+ADD 1
19 bailingmoe3 1 prefill GATED_DELTA_NET+CPY 1
20 bailingmoe3 1 decode RMS_NORM+MUL 128
21 bailingmoe3 1 decode ADD+ADD 16
22 bailingmoe3 1 decode GATED_DELTA_NET+CPY 16
23 bloom 0 prefill NORM+MUL+ADD 6
24 bloom 0 decode NORM+MUL+ADD 96
25 chatglm 0 prefill RMS_NORM+MUL 5
26 chatglm 0 decode RMS_NORM+MUL 80
27 codeshell 0 prefill NORM+MUL+ADD 5
28 codeshell 0 decode NORM+MUL+ADD 80
29 cogvlm 0 prefill RMS_NORM+MUL 5
30 cogvlm 0 decode RMS_NORM+MUL 80
31 command-r 0 prefill NORM+MUL 3
32 command-r 0 decode NORM+MUL 48
33 dbrx 0 prefill NORM+MUL 5
34 dbrx 0 decode NORM+MUL 80
35 deci 0 prefill RMS_NORM+MUL 5
36 deci 0 decode RMS_NORM+MUL 80
37 deepseek 0 prefill RMS_NORM+MUL 5
38 deepseek 0 prefill ADD+ADD 1
39 deepseek 0 decode RMS_NORM+MUL 80
40 deepseek 0 decode ADD+ADD 16
41 deepseek2 0 prefill RMS_NORM+MUL 9
42 deepseek2 0 prefill ADD+ADD 1
43 deepseek2 0 decode RMS_NORM+MUL 144
44 deepseek2 0 decode ADD+ADD 16
45 deepseek32 0 prefill NORM+MUL+ADD 2
46 deepseek32 0 prefill RMS_NORM+MUL 9
47 deepseek32 0 prefill ADD+ADD 1
48 deepseek32 0 decode NORM+MUL+ADD 32
49 deepseek32 0 decode RMS_NORM+MUL 144
50 deepseek32 0 decode ADD+ADD 16
51 deepseek4 0 prefill RMS_NORM+MUL 20
52 deepseek4 0 decode RMS_NORM+MUL 320
53 dots1 0 prefill RMS_NORM+MUL 9
54 dots1 0 prefill ADD+ADD 1
55 dots1 0 decode RMS_NORM+MUL 144
56 dots1 0 decode ADD+ADD 16
57 dream 0 prefill RMS_NORM+MUL 5
58 dream 0 decode RMS_NORM+MUL 80
59 ernie4_5-moe 1 prefill RMS_NORM+MUL 5
60 ernie4_5-moe 1 prefill ADD+ADD 1
61 ernie4_5-moe 1 decode RMS_NORM+MUL 80
62 ernie4_5-moe 1 decode ADD+ADD 16
63 ernie4_5 0 prefill RMS_NORM+MUL 5
64 ernie4_5 0 decode RMS_NORM+MUL 80
65 exaone 0 prefill RMS_NORM+MUL 5
66 exaone 0 decode RMS_NORM+MUL 80
67 exaone4 0 prefill RMS_NORM+MUL 5
68 exaone4 0 prefill RMS_NORM+MUL+ADD 4
69 exaone4 0 decode RMS_NORM+MUL 80
70 exaone4 0 decode RMS_NORM+MUL+ADD 64
71 falcon 0 prefill NORM+MUL+ADD 5
72 falcon 0 prefill ADD+ADD 2
73 falcon 0 decode NORM+MUL+ADD 80
74 falcon 0 decode ADD+ADD 32
75 falcon-h1 0 prefill RMS_NORM+MUL 9
76 falcon-h1 0 prefill ADD+ADD 2
77 falcon-h1 0 decode RMS_NORM+MUL 144
78 falcon-h1 0 decode ADD+ADD 32
79 gemma 0 prefill RMS_NORM+MUL 5
80 gemma 0 decode RMS_NORM+MUL 80
81 gemma2 0 prefill RMS_NORM+MUL 5
82 gemma2 0 prefill RMS_NORM+MUL+ADD 4
83 gemma2 0 decode RMS_NORM+MUL 80
84 gemma2 0 decode RMS_NORM+MUL+ADD 64
85 glm-dsa 0 prefill NORM+MUL+ADD 2
86 glm-dsa 0 prefill RMS_NORM+MUL 9
87 glm-dsa 0 prefill ADD+ADD 1
88 glm-dsa 0 decode NORM+MUL+ADD 32
89 glm-dsa 0 decode RMS_NORM+MUL 144
90 glm-dsa 0 decode ADD+ADD 16
91 glm4 0 prefill RMS_NORM+MUL 5
92 glm4 0 prefill RMS_NORM+MUL+ADD 4
93 glm4 0 decode RMS_NORM+MUL 80
94 glm4 0 decode RMS_NORM+MUL+ADD 64
95 glm4moe 1 prefill RMS_NORM+MUL 9
96 glm4moe 1 prefill ADD+ADD 1
97 glm4moe 1 decode RMS_NORM+MUL 144
98 glm4moe 1 decode ADD+ADD 16
99 gpt-oss 0 prefill RMS_NORM+MUL 5
100 gpt-oss 0 decode RMS_NORM+MUL 80
101 gpt2 0 prefill NORM+MUL+ADD 5
102 gpt2 0 decode NORM+MUL+ADD 80
103 gptneox 0 prefill NORM+MUL+ADD 5
104 gptneox 0 decode NORM+MUL+ADD 80
105 granite 0 prefill RMS_NORM+MUL 5
106 granite 0 decode RMS_NORM+MUL 80
107 granite 0 prefill RMS_NORM+MUL 5
108 granite 0 decode RMS_NORM+MUL 80
109 granitehybrid 0 prefill RMS_NORM+MUL 6
110 granitehybrid 0 decode RMS_NORM+MUL 96
111 granitemoe 1 prefill RMS_NORM+MUL 5
112 granitemoe 1 decode RMS_NORM+MUL 80
113 granitemoe 1 prefill RMS_NORM+MUL 5
114 granitemoe 1 decode RMS_NORM+MUL 80
115 grok 0 prefill RMS_NORM+MUL 5
116 grok 0 prefill RMS_NORM+MUL+ADD 4
117 grok 0 decode RMS_NORM+MUL 80
118 grok 0 decode RMS_NORM+MUL+ADD 64
119 grovemoe 1 prefill RMS_NORM+MUL 9
120 grovemoe 1 prefill ADD+ADD 2
121 grovemoe 1 decode RMS_NORM+MUL 144
122 grovemoe 1 decode ADD+ADD 32
123 hunyuan-dense 0 prefill RMS_NORM+MUL 9
124 hunyuan-dense 0 decode RMS_NORM+MUL 144
125 hunyuan-moe 1 prefill RMS_NORM+MUL 9
126 hunyuan-moe 1 prefill ADD+ADD 2
127 hunyuan-moe 1 decode RMS_NORM+MUL 144
128 hunyuan-moe 1 decode ADD+ADD 32
129 hunyuan_vl 0 prefill RMS_NORM+MUL 9
130 hunyuan_vl 0 decode RMS_NORM+MUL 144
131 hy_v3 0 prefill RMS_NORM+MUL 9
132 hy_v3 0 prefill ADD+ADD 2
133 hy_v3 0 decode RMS_NORM+MUL 144
134 hy_v3 0 decode ADD+ADD 32
135 internlm2 0 prefill RMS_NORM+MUL 5
136 internlm2 0 decode RMS_NORM+MUL 80
137 jais 0 prefill NORM+MUL+ADD 5
138 jais 0 decode NORM+MUL+ADD 80
139 jais2 0 prefill NORM+MUL+ADD 5
140 jais2 0 decode NORM+MUL+ADD 80
141 jamba 0 prefill RMS_NORM+MUL 8
142 jamba 0 decode RMS_NORM+MUL 128
143 kimi-k3 0 prefill RMS_NORM+MUL 17
144 kimi-k3 0 prefill GATED_DELTA_NET+CPY 1
145 kimi-k3 0 decode RMS_NORM+MUL 272
146 kimi-k3 0 decode GATED_DELTA_NET+CPY 16
147 kimi-linear 0 prefill RMS_NORM+MUL 7
148 kimi-linear 0 prefill ADD+ADD 1
149 kimi-linear 0 prefill GATED_DELTA_NET+CPY 1
150 kimi-linear 0 decode RMS_NORM+MUL 112
151 kimi-linear 0 decode ADD+ADD 16
152 kimi-linear 0 decode GATED_DELTA_NET+CPY 16
153 lfm2 0 prefill RMS_NORM+MUL 7
154 lfm2 0 decode RMS_NORM+MUL 112
155 lfm2moe 1 prefill RMS_NORM+MUL 7
156 lfm2moe 1 decode RMS_NORM+MUL 112
157 llada 0 prefill RMS_NORM+MUL 5
158 llada 0 decode RMS_NORM+MUL 80
159 llada-moe 1 prefill RMS_NORM+MUL 9
160 llada-moe 1 decode RMS_NORM+MUL 144
161 llama 0 prefill RMS_NORM+MUL 5
162 llama 0 decode RMS_NORM+MUL 80
163 llama 0 prefill RMS_NORM+MUL 5
164 llama 0 decode RMS_NORM+MUL 80
165 llama4 0 prefill RMS_NORM+MUL 9
166 llama4 0 prefill ADD+ADD 2
167 llama4 0 decode RMS_NORM+MUL 144
168 llama4 0 decode ADD+ADD 32
169 maincoder 0 prefill RMS_NORM+MUL 9
170 maincoder 0 decode RMS_NORM+MUL 144
171 mamba 0 prefill RMS_NORM+MUL 3
172 mamba 0 decode RMS_NORM+MUL 48
173 mamba2 0 prefill RMS_NORM+MUL 5
174 mamba2 0 decode RMS_NORM+MUL 80
175 minicpm 0 prefill RMS_NORM+MUL 5
176 minicpm 0 decode RMS_NORM+MUL 80
177 minicpm 0 prefill RMS_NORM+MUL 5
178 minicpm 0 decode RMS_NORM+MUL 80
179 minicpm3 0 prefill RMS_NORM+MUL 9
180 minicpm3 0 decode RMS_NORM+MUL 144
181 minimax-01 0 prefill RMS_NORM+MUL 6
182 minimax-01 0 decode RMS_NORM+MUL 96
183 minimax-m2 0 prefill RMS_NORM+MUL 9
184 minimax-m2 0 decode RMS_NORM+MUL 144
185 minimax-m3 0 prefill RMS_NORM+MUL 11
186 minimax-m3 0 prefill ADD+ADD 1
187 minimax-m3 0 decode RMS_NORM+MUL 176
188 minimax-m3 0 decode ADD+ADD 16
189 mistral3 0 prefill RMS_NORM+MUL 5
190 mistral3 0 decode RMS_NORM+MUL 80
191 mistral3 0 prefill RMS_NORM+MUL 5
192 mistral3 0 decode RMS_NORM+MUL 80
193 mistral4 0 prefill RMS_NORM+MUL 9
194 mistral4 0 prefill ADD+ADD 1
195 mistral4 0 decode RMS_NORM+MUL 144
196 mistral4 0 decode ADD+ADD 16
197 mpt 0 prefill NORM+MUL+ADD 5
198 mpt 0 decode NORM+MUL+ADD 80
199 nanbeige 0 prefill RMS_NORM+MUL 5
200 nanbeige 0 decode RMS_NORM+MUL 80
201 nemotron 0 prefill NORM+MUL+ADD 5
202 nemotron 0 decode NORM+MUL+ADD 80
203 nemotron_h 0 prefill RMS_NORM+MUL 5
204 nemotron_h 0 decode RMS_NORM+MUL 80
205 nemotron_h_moe 1 prefill RMS_NORM+MUL 5
206 nemotron_h_moe 1 decode RMS_NORM+MUL 80
207 olmoe 1 prefill RMS_NORM+MUL 9
208 olmoe 1 decode RMS_NORM+MUL 144
209 openelm 0 prefill RMS_NORM+MUL 9
210 openelm 0 decode RMS_NORM+MUL 144
211 orion 0 prefill NORM+MUL+ADD 5
212 orion 0 decode NORM+MUL+ADD 80
213 paddleocr 0 prefill RMS_NORM+MUL 5
214 paddleocr 0 decode RMS_NORM+MUL 80
215 pangu-embedded 0 prefill RMS_NORM+MUL 5
216 pangu-embedded 0 decode RMS_NORM+MUL 80
217 phi2 0 prefill NORM+MUL+ADD 3
218 phi2 0 prefill ADD+ADD 2
219 phi2 0 decode NORM+MUL+ADD 48
220 phi2 0 decode ADD+ADD 32
221 phi3 0 prefill RMS_NORM+MUL 5
222 phi3 0 decode RMS_NORM+MUL 80
223 phimoe 1 prefill RMS_NORM+MUL+ADD 5
224 phimoe 1 decode RMS_NORM+MUL+ADD 80
225 plamo 0 prefill RMS_NORM+MUL 3
226 plamo 0 prefill ADD+ADD 2
227 plamo 0 decode RMS_NORM+MUL 48
228 plamo 0 decode ADD+ADD 32
229 plamo2 0 prefill RMS_NORM+MUL 11
230 plamo2 0 prefill RMS_NORM+MUL+ADD 3
231 plamo2 0 decode RMS_NORM+MUL 176
232 plamo2 0 decode RMS_NORM+MUL+ADD 48
233 pockettts 0 prefill NORM+MUL+ADD 5
234 pockettts 0 decode NORM+MUL+ADD 80
235 qwen 0 prefill RMS_NORM+MUL 5
236 qwen 0 decode RMS_NORM+MUL 80
237 qwen2 0 prefill RMS_NORM+MUL 5
238 qwen2 0 decode RMS_NORM+MUL 80
239 qwen2moe 1 prefill RMS_NORM+MUL 5
240 qwen2moe 1 prefill ADD+ADD 2
241 qwen2moe 1 decode RMS_NORM+MUL 80
242 qwen2moe 1 decode ADD+ADD 32
243 qwen2vl 0 prefill RMS_NORM+MUL 5
244 qwen2vl 0 decode RMS_NORM+MUL 80
245 qwen3 0 prefill RMS_NORM+MUL 9
246 qwen3 0 decode RMS_NORM+MUL 144
247 qwen35 0 prefill RMS_NORM+MUL 8
248 qwen35 0 prefill GATED_DELTA_NET+CPY 1
249 qwen35 0 decode RMS_NORM+MUL 128
250 qwen35 0 decode GATED_DELTA_NET+CPY 16
251 qwen35moe 1 prefill RMS_NORM+MUL 8
252 qwen35moe 1 prefill ADD+ADD 2
253 qwen35moe 1 prefill GATED_DELTA_NET+CPY 1
254 qwen35moe 1 decode RMS_NORM+MUL 128
255 qwen35moe 1 decode ADD+ADD 32
256 qwen35moe 1 decode GATED_DELTA_NET+CPY 16
257 qwen3moe 1 prefill RMS_NORM+MUL 9
258 qwen3moe 1 decode RMS_NORM+MUL 144
259 qwen3next 0 prefill RMS_NORM+MUL 8
260 qwen3next 0 prefill ADD+ADD 2
261 qwen3next 0 prefill GATED_DELTA_NET+CPY 1
262 qwen3next 0 decode RMS_NORM+MUL 128
263 qwen3next 0 decode ADD+ADD 32
264 qwen3next 0 decode GATED_DELTA_NET+CPY 16
265 qwen3tts 0 prefill RMS_NORM+MUL 9
266 qwen3tts 0 decode RMS_NORM+MUL 144
267 qwen3vl 0 prefill RMS_NORM+MUL 9
268 qwen3vl 0 decode RMS_NORM+MUL 144
269 qwen3vlmoe 1 prefill RMS_NORM+MUL 9
270 qwen3vlmoe 1 decode RMS_NORM+MUL 144
271 qwen4exp 0 prefill RMS_NORM+MUL 5
272 qwen4exp 0 prefill ADD+ADD+ADD 5
273 qwen4exp 0 prefill GATED_DELTA_NET+CPY 1
274 qwen4exp 0 decode RMS_NORM+MUL 80
275 qwen4exp 0 decode ADD+ADD+ADD 80
276 qwen4exp 0 decode GATED_DELTA_NET+CPY 16
277 refact 0 prefill RMS_NORM+MUL 5
278 refact 0 decode RMS_NORM+MUL 80
279 refact 0 prefill RMS_NORM+MUL 5
280 refact 0 decode RMS_NORM+MUL 80
281 rnd1 0 prefill RMS_NORM+MUL 9
282 rnd1 0 decode RMS_NORM+MUL 144
283 seed_oss 0 prefill RMS_NORM+MUL 5
284 seed_oss 0 decode RMS_NORM+MUL 80
285 smallthinker 0 prefill RMS_NORM+MUL 5
286 smallthinker 0 decode RMS_NORM+MUL 80
287 smollm3 0 prefill RMS_NORM+MUL 5
288 smollm3 0 decode RMS_NORM+MUL 80
289 stablelm 0 prefill NORM+MUL 4
290 stablelm 0 prefill NORM+MUL+ADD 5
291 stablelm 0 decode NORM+MUL 64
292 stablelm 0 decode NORM+MUL+ADD 80
293 starcoder 0 prefill NORM+MUL+ADD 5
294 starcoder 0 decode NORM+MUL+ADD 80
295 starcoder2 0 prefill NORM+MUL+ADD 5
296 starcoder2 0 decode NORM+MUL+ADD 80
297 talkie 0 prefill ADD+ADD 2
298 talkie 0 decode ADD+ADD 32
299 xverse 0 prefill RMS_NORM+MUL 5
300 xverse 0 decode RMS_NORM+MUL 80
+16 -16
View File
@@ -6,14 +6,14 @@
// - the NMSE between the fused and unfused logits
// - the NMSE between the device and a CPU reference
//
// the per-fusion-type counters are compared against a per-device baseline file (TSV) so a
// the per-fusion-type counters are compared against a per-device baseline file (CSV) so a
// fusion pattern that silently stops matching (or fires when it should not) is caught as a
// regression.
//
// usage:
// test-fusion --models DIR --device MTL0 --record baseline.tsv # generate a baseline
// test-fusion --models DIR --device MTL0 --check baseline.tsv # validate against it
// test-fusion --model FILE --device MTL0 --check baseline.tsv # validate a single model
// test-fusion --models DIR --device MTL0 --record baseline.csv # generate a baseline
// test-fusion --models DIR --device MTL0 --check baseline.csv # validate against it
// test-fusion --model FILE --device MTL0 --check baseline.csv # validate a single model
#include "common.h"
#include "log.h"
@@ -84,7 +84,7 @@ static std::vector<llama_token> get_tokens(const uint32_t n_tokens, const uint32
return ret;
}
// trim leading/trailing whitespace (used when parsing padded TSV columns)
// trim leading/trailing whitespace (used when parsing padded CSV columns)
static std::string trim(const std::string & s) {
const size_t b = s.find_first_not_of(" \t\r\n");
if (b == std::string::npos) {
@@ -215,8 +215,8 @@ static void usage(const char * argv0) {
printf(" --models DIR run over all .gguf models in a directory\n");
printf(" --model FILE run over a single model file (mutually exclusive with --models)\n");
printf(" --device NAME device to run on (e.g. MTL0, CPU)\n");
printf(" --record TSV write the golden baseline\n");
printf(" --check TSV validate the counters against a baseline (default)\n");
printf(" --record CSV write the golden baseline\n");
printf(" --check CSV validate the counters against a baseline (default)\n");
printf(" -h, --help show this message and exit\n");
}
@@ -343,7 +343,7 @@ int main(int argc, char ** argv) {
}
std::vector<std::string> cols;
size_t pos = 0;
while ((pos = line.find('\t')) != std::string::npos) {
while ((pos = line.find(',')) != std::string::npos) {
cols.push_back(trim(line.substr(0, pos)));
line.erase(0, pos + 1);
}
@@ -471,10 +471,10 @@ int main(int argc, char ** argv) {
if (!record_path.empty()) {
os << "# test-fusion baseline for device " << base_name << "\n";
os << "# " << std::left
<< std::setw(18) << "arch" << '\t'
<< std::setw(4) << "moe" << '\t'
<< std::setw(8) << "mode" << '\t'
<< std::setw(28) << "label" << '\t'
<< std::setw(18) << "arch" << ','
<< std::setw(4) << "moe" << ','
<< std::setw(8) << "mode" << ','
<< std::setw(28) << "label" << ','
<< std::right << std::setw(7) << "count" << '\n';
}
@@ -494,10 +494,10 @@ int main(int argc, char ** argv) {
(unsigned long long) r.expected, r.nmse_fus, r.nmse_dev, status);
if (!record_path.empty()) {
os << std::left
<< std::setw(20) << r.arch << '\t'
<< std::setw(4) << (r.moe ? "1" : "0") << '\t'
<< std::setw(8) << r.mode << '\t'
<< std::setw(28) << r.label << '\t'
<< std::setw(20) << r.arch << ','
<< std::setw(4) << (r.moe ? "1" : "0") << ','
<< std::setw(8) << r.mode << ','
<< std::setw(28) << r.label << ','
<< std::right << std::setw(7) << r.count_fused << '\n';
}
}