Please support yolo26 - yolo26n-pose fails with a stack trace at model_deploy.py

[juodumas]

Versions:

• container docker.io/sophgo/tpuc_dev:v3.4-20250411
• built https://github.com/sophgo/tpu-mlir from git at 3c489cf
• with python packages ultralytics==8.4.30 onnxruntime==1.22.1 onnx==1.19.1 onnxslim==0.1.90

Problems:

• yolo26n-pose fails at model_deploy.py (see logs below)
• also tried yolo26n and it failed at run_calibration.py (did not attach logs because I would like to try running yolo26n-pose on Milk-V Duo 256M)

The error from model_deploy:

$ model_deploy.py --mlir model_yolo26n-pose/yolo26n-pose.mlir --chip cv181x --quantize INT8 --calibration_table model_yolo26n-pose/yolo26n-pose_cali_table --model model_yolo26n-pose/yolo26n-pose-cv181x-int8.bmodel --fuse_preprocess --aligned_input

Entering FusePreprocessPass.
WARNING: qscale > 1,  = 9.999990e-01
Not support now.
UNREACHABLE executed at /w/tpu-mlir/lib/Conversion/TopToTpu/CV18xx/CompareConst.cpp:32!
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.      Program arguments: tpuc-opt model_yolo26n-pose/yolo26n-pose.mlir --init "--processor-assign=chip=cv181x mode=INT8 num_device=1 num_core=1 addr_mode=auto high_precision=False" "--import-calibration-table=file=model_yolo26n-pose/yolo26n-pose_cali_table asymmetric=False" --processor-top-optimize "--fuse-prepr
ocess=mode=INT8 customization_format=RGB_PLANAR align=True" "--convert-top-to-tpu=weightFileName=yolo26n-pose_cv181x_int8_sym_tpu_weights.npz asymmetric=False doWinograd=False q_group_size=0 q_symmetric=False matmul_perchannel=False gelu_mode=normal" --canonicalize --weight-fold --deinit --mlir-print-debuginfo -o yolo26n-pose_cv181x_int8_sym_tpu.mlir
 #0 0x0000559086914ce7 (/w/tpu-mlir/install/bin/tpuc-opt+0x91ece7)
 #1 0x0000559086912a0e (/w/tpu-mlir/install/bin/tpuc-opt+0x91ca0e)
 #2 0x000055908691566a (/w/tpu-mlir/install/bin/tpuc-opt+0x91f66a)
 #3 0x00007fa0c4c3f520 (/lib/x86_64-linux-gnu/libc.so.6+0x42520)
 #4 0x00007fa0c4c939fc pthread_kill (/lib/x86_64-linux-gnu/libc.so.6+0x969fc)
 #5 0x00007fa0c4c3f476 gsignal (/lib/x86_64-linux-gnu/libc.so.6+0x42476)
 #6 0x00007fa0c4c257f3 abort (/lib/x86_64-linux-gnu/libc.so.6+0x287f3)
 #7 0x0000559086912831 (/w/tpu-mlir/install/bin/tpuc-opt+0x91c831)
 #8 0x0000559086d58633 (/w/tpu-mlir/install/bin/tpuc-opt+0xd62633)
 #9 0x0000559086b20406 (/w/tpu-mlir/install/bin/tpuc-opt+0xb2a406)
#10 0x0000559086b1fba4 (/w/tpu-mlir/install/bin/tpuc-opt+0xb29ba4)
#11 0x0000559088200e67 (/w/tpu-mlir/install/bin/tpuc-opt+0x220ae67)
#12 0x00005590881fd77f (/w/tpu-mlir/install/bin/tpuc-opt+0x220777f)
#13 0x00005590881c697c (/w/tpu-mlir/install/bin/tpuc-opt+0x21d097c)
#14 0x00005590881c37cc (/w/tpu-mlir/install/bin/tpuc-opt+0x21cd7cc)
#15 0x0000559086a713d4 (/w/tpu-mlir/install/bin/tpuc-opt+0xa7b3d4)
#16 0x000055908822a154 (/w/tpu-mlir/install/bin/tpuc-opt+0x2234154)
#17 0x000055908822a781 (/w/tpu-mlir/install/bin/tpuc-opt+0x2234781)
#18 0x000055908822cc28 (/w/tpu-mlir/install/bin/tpuc-opt+0x2236c28)
#19 0x000055908690636b (/w/tpu-mlir/install/bin/tpuc-opt+0x91036b)
#20 0x0000559086905734 (/w/tpu-mlir/install/bin/tpuc-opt+0x90f734)
#21 0x0000559088447a28 (/w/tpu-mlir/install/bin/tpuc-opt+0x2451a28)
#22 0x00005590868ffa3a (/w/tpu-mlir/install/bin/tpuc-opt+0x909a3a)
#23 0x00005590868fff04 (/w/tpu-mlir/install/bin/tpuc-opt+0x909f04)
#24 0x00005590868fe94a (/w/tpu-mlir/install/bin/tpuc-opt+0x90894a)
#25 0x00007fa0c4c26d90 (/lib/x86_64-linux-gnu/libc.so.6+0x29d90)
#26 0x00007fa0c4c26e40 __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x29e40)
#27 0x00005590868fdd55 (/w/tpu-mlir/install/bin/tpuc-opt+0x907d55)
Aborted

Partial run logs below

1. onnx → mlir finishes, but needs --tolerance 0.94,0.66 and also logs WARNING : onnxsim opt failed.

$ model_transform.py --model_def model_yolo26n-pose/yolo26n-pose.onnx --model_name yolo26n-pose --mlir model_yolo26n-pose/yolo26n-pose.mlir --input_shapes '[[1,3,640,640]]' --mean 0,0,0 --scale 0.0039216,0.0039216,0.0039216 --pixel_format rgb --add_postprocess yolov26 --test_input images/dog.jpg --test_result model_yol
o26n-pose/yolo26n-pose_top_outputs.npz --tolerance 0.94,0.66                                                                                                                                                                                                                                                                    
2026/03/28 02:51:02 - INFO : TPU-MLIR v1.0.0.dev-3c489cf-20260328                                                                                                                                                                                                                                                               
2026/03/28 02:51:02 - INFO :                                                                                                                                                                                                                                                                                                    
         _____________________________________________________                                                                                                                                                                                                                                                                  
        | preprocess:                                           |                                                                                                                                                                                                                                                               
        |   (x - mean) * scale                                  |                                                                                                                                                                                                                                                               
        '-------------------------------------------------------'                                                                                                                                                                                                                                                               
  config Preprocess args :                                                                                                                                                                                                                                                                                                      
        resize_dims           : same to net input dims                                                                                                                                                                                                                                                                          
        keep_aspect_ratio     : False                                                                                                                                                                                                                                                                                           
        keep_ratio_mode       : letterbox                                                                                                                                                                                                                                                                                       
        pad_value             : 0                                                                                                                                                                                                                                                                                               
        pad_type              : center                                                                                                                                                                                                                                                                                          
        --------------------------                                                                                                                                                                                                                                                                                              
        mean                  : [0.0, 0.0, 0.0]                                                                                                                                                                                                                                                                                 
        scale                 : [0.0039216, 0.0039216, 0.0039216]                                                                                                                                                                                                                                                               
        --------------------------                                                                                                                                                                                                                                                                                              
        pixel_format          : rgb                                                                                                                            
        channel_format        : nchw                                                                                                                           
        yuv_type              :                                                                                                                                
                                                                                                                                                               
2026/03/28 02:51:03 - INFO : Input_shape assigned                                                                                                              
2026/03/28 02:51:03 - INFO : ConstantFolding finished                                                                                                          
2026/03/28 02:51:03 - INFO : skip_fuse_bn:False                                                                                                                
2026/03/28 02:51:03 - WARNING : onnxsim opt failed.                                                                                                            
2026/03/28 02:51:03 - INFO : Onnxsim opt finished                                                                                                              
2026/03/28 02:51:04 - INFO : ConstantFolding finished                                                                                                          
2026/03/28 02:51:04 - INFO : Save mlir file: model_yolo26n-pose/yolo26n-pose_origin.mlir                                                                       
[Running]: tpuc-opt model_yolo26n-pose/yolo26n-pose_origin.mlir --struct-optimize --shape-infer --add-postprocess="type=yolov26" --canonicalize --extra-optimize -o model_yolo26n-pose/yolo26n-pose.mlir
[Success]: tpuc-opt model_yolo26n-pose/yolo26n-pose_origin.mlir --struct-optimize --shape-infer --add-postprocess="type=yolov26" --canonicalize --extra-optimize -o model_yolo26n-pose/yolo26n-pose.mlir
2026/03/28 02:51:04 - INFO : Mlir file generated:model_yolo26n-pose/yolo26n-pose.mlir                                                                          
2026/03/28 02:51:04 - INFO :                                                                                                                                   
  load_config Preprocess args :                                                                                                                                
        resize_dims           : [640, 640]                                                                                                                     
        keep_aspect_ratio     : False                                                                                                                          
        keep_ratio_mode       : letterbox                                                                                                                      
        pad_value             : 0                                                                                                                              
        pad_type              : center                                                                                                                         
        input_dims            : [640, 640]                                                                                                                     
        --------------------------                                                                                                                             
        mean                  : [0.0, 0.0, 0.0]                                                                                                                
        scale                 : [0.0039216, 0.0039216, 0.0039216]                                                                                              
        --------------------------                                                                                                                             
        pixel_format          : rgb                                                                                                                            
        channel_format        : nchw                                                                                                                           
        yuv_type              :                                                                                                                                
                                                                                                                                                                                                                                                                                                                                
[CMD]: model_runner.py --input yolo26n-pose_in_f32.npz --model model_yolo26n-pose/yolo26n-pose.onnx --output yolo26n-pose_ref_outputs.npz                       
2026/03/28 02:51:04 - INFO : Saving yolo26n-pose_ref_outputs.npz                
[CMD]: model_runner.py --input yolo26n-pose_in_f32.npz --model model_yolo26n-pose/yolo26n-pose.mlir --output model_yolo26n-pose/yolo26n-pose_top_outputs.npz    
[##################################################] 100%                       
2026/03/28 02:51:05 - INFO : Saving model_yolo26n-pose/yolo26n-pose_top_outputs.npz                                                                             
[Running]: npz_tool.py compare model_yolo26n-pose/yolo26n-pose_top_outputs.npz yolo26n-pose_ref_outputs.npz --tolerance 0.94,0.66 --except - -vv
<...>
3 equal, 24 close, 313 similar
<...>
min_similiarity = (0.9447258114814758, 0.6628196046536068, 1.6779352724552155)

2. calibration finishes with warnings

$ run_calibration.py model_yolo26n-pose/yolo26n-pose.mlir --dataset tpu-mlir/regression/dataset/COCO2017 --input_num 100 -o model_yolo26n-pose/yolo26n-pose_cali_table                                                                                                                                                          
TPU-MLIR v1.0.0.dev-3c489cf-20260328                                                                                                                                                                                                                                                                                            
2026/03/28 03:00:17 - INFO :                                                                                                                                                                                                                                                                                                    
  load_config Preprocess args :                                                                                                                                                                                                                                                                                                 
        resize_dims           : [640, 640]                                                                                                                                                                                                                                                                                      
        keep_aspect_ratio     : False                                                                                                                                                                                                                                                                                           
        keep_ratio_mode       : letterbox                                                                                                                                                                                                                                                                                       
        pad_value             : 0                                                                                                                                                                                                                                                                                               
        pad_type              : center                                                                                                                                                                                                                                                                                          
        input_dims            : [640, 640]                                                                                                                                                                                                                                                                                      
        --------------------------                                                                                                                                                                                                                                                                                              
        mean                  : [0.0, 0.0, 0.0]                                                                                                                                                                                                                                                                                 
        scale                 : [0.0039216, 0.0039216, 0.0039216]                                                                                                                                                                                                                                                               
        --------------------------                                                                                                                                                                                                                                                                                              
        pixel_format          : rgb                                                                                                                                                                                                                                                                                             
        channel_format        : nchw                                                                                                                           
        yuv_type              :                                                
                                       
input_num = 100, ref = 100                                                     
real input_num = 100
<...>
activation_collect_and_calc_th for sample: 99: 100%|█████████████████████████████████████████████████████████████████████████| 100/100 [00:43<00:00,  2.31it/s]                                                                                                                                                                 
WARNING: layer /model.23/Mod_output_0_Mod is all zeros. Please check the input data correctness.                                                                                                                                                                                                                                
WARNING: layer /model.23/Mod_output_0_Mod_Mul is all zeros. Please check the input data correctness.                                                                                                                                                                                                                            
WARNING: layer /model.23/Unsqueeze_2_output_0_Unsqueeze is all zeros. Please check the input data correctness.                                                                                                                                                                                                                  
WARNING: layer /model.23/Cast_2_output_0_Cast is all zeros. Please check the input data correctness.                                                                                                                                                                                                                            
[2048] threshold: /model.23/Split_output_2_Split: 100%|█████████████████████████████████████████████████████████████████████| 340/340 [00:00<00:00, 903.73it/s]                                                                                                                                                                 
UserWarning: Warning: op /model.23/Mod_output_0_Mod not in thresholds_map, set to 1.0                                                                                                                                                                                                                                           
UserWarning: Warning: op /model.23/Mod_output_0_Mod_Mul not in thresholds_map, set to 1.0                                                                                                                                                                                                                                       
UserWarning: Warning: op /model.23/Unsqueeze_2_output_0_Unsqueeze not in thresholds_map, set to 1.0                                                                                                                                                                                                                             
UserWarning: Warning: op /model.23/Cast_2_output_0_Cast not in thresholds_map, set to 1.0                                                                                                                                                                                                                                       
WARNING: layer /model.23/Mod_output_0_Mod threshold is zero. Please check the input data correctness.                                                                                                                                                                                                                           
WARNING: layer /model.23/Mod_output_0_Mod_Mul threshold is zero. Please check the input data correctness.                                                                                                                                                                                                                       
WARNING: layer /model.23/Unsqueeze_2_output_0_Unsqueeze threshold is zero. Please check the input data correctness.                                                                                                                                                                                                             
WARNING: layer /model.23/Cast_2_output_0_Cast threshold is zero. Please check the input data correctness.                                                                                                                                                                                                                       
tune op: yolo_post: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 343/343 [01:00<00:00,  5.69it/s]                                                                                                                                                                 
auto tune end, run time:60.6066997051239                                                                                                                                                                                                                                                                                        
WARNING: layer /model.23/Mod_output_0_Mod threshold is zero. Please check the input data correctness.                                                                                                                                                                                                                           
WARNING: layer /model.23/Mod_output_0_Mod_Mul threshold is zero. Please check the input data correctness.                                                                                                                                                                                                                       
WARNING: layer /model.23/Unsqueeze_2_output_0_Unsqueeze threshold is zero. Please check the input data correctness.                                                                                                                                                                                                             
WARNING: layer /model.23/Cast_2_output_0_Cast threshold is zero. Please check the input data correctness.

3. mlir → bmodel fails

$ model_deploy.py --mlir model_yolo26n-pose/yolo26n-pose.mlir --chip cv181x --quantize INT8 --calibration_table model_yolo26n-pose/yolo26n-pose_cali_table --model model_yolo26n-pose/yolo26n-pose-cv181x-int8.bmodel --fuse_preprocess --aligned_input
2026/03/28 03:02:02 - INFO : TPU-MLIR v1.0.0.dev-3c489cf-20260328                                                                                                                                                                                                                                                               
2026/03/28 03:02:02 - INFO :                                                                                                                                                                                                                                                                                                    
  load_config Preprocess args :                                                                                                                                                                                                                                                                                                 
        resize_dims           : [640, 640]                                                                                                                                                                                                                                                                                      
        keep_aspect_ratio     : False                                                                                                                                                                                                                                                                                           
        keep_ratio_mode       : letterbox                                                                                                                                                                                                                                                                                       
        pad_value             : 0                                                                                                                                                                                                                                                                                               
        pad_type              : center                                                                                                                                                                                                                                                                                          
        input_dims            : [640, 640]                                                                                                                                                                                                                                                                                      
        --------------------------                                                                                                                                                                                                                                                                                              
        mean                  : [0.0, 0.0, 0.0]                                                                                                                                                                                                                                                                                 
        scale                 : [0.0039216, 0.0039216, 0.0039216]                                                                                                                                                                                                                                                               
        --------------------------                                                                                                                                                                                                                                                                                              
        pixel_format          : rgb                                             
        channel_format        : nchw                                            
        yuv_type              :                                                 

[Success] Create config file 'yolo26n-pose_cv181x_int8.layer_group_config.json'.
Content:                                                                        
{                                                                               
  "shape_secs_search_strategy": 0,                                              
  "structure_detect_opt": true,                                                 
  "sc_method_configs": [                                                        
    {                                                                           
      "sc_method": "sc_method_quick_search",                                    
      "MAX_TRY_NUM": 20                                                         
    },                                                                          
    {                                                                           
      "sc_method": "sc_method_search_better_v1",                                
      "NSECS_SEARCH_RECORD_THRESHOLD": 3,                                       
      "CSECS_SEARCH_RECORD_THRESHOLD": 3,                                       
      "DSECS_SEARCH_RECORD_THRESHOLD": 3,                                       
      "HSECS_SEARCH_RECORD_THRESHOLD": 3,                                       
      "WSECS_SEARCH_RECORD_THRESHOLD": 3                                        
    },                                                                          
    {                                                                           
      "sc_method": "sc_method_search_better_v2",                                
      "MAX_NSECS": 32,                                                          
      "MAX_CSECS": 32,                                                          
      "MAX_DSECS": 32,                                                          
      "MAX_HSECS": 32,                                                          
      "MAX_WSECS": 32,                                                          
      "NSECS_SEARCH_RECORD_THRESHOLD": 2,                                       
      "CSECS_SEARCH_RECORD_THRESHOLD": 2,                                       
      "DSECS_SEARCH_RECORD_THRESHOLD": 2,                                       
      "HSECS_SEARCH_RECORD_THRESHOLD": 2,                                       
      "WSECS_SEARCH_RECORD_THRESHOLD": 2                                        
    }                                                                           
  ]                                                                             
}                                                                               
[Running]: tpuc-opt model_yolo26n-pose/yolo26n-pose.mlir --processor-assign="chip=cv181x mode=INT8 num_device=1 num_core=1 addr_mode=auto high_precision=False" --import-calibration-table="file=model_yolo26n-pose/yolo26n-pose_cali_table asymmetric=False" --processor-top-optimize --fuse-preprocess="mode=INT8 customizatio
n_format=RGB_PLANAR align=True" --convert-top-to-tpu="weightFileName=yolo26n-pose_cv181x_int8_sym_tpu_weights.npz asymmetric=False doWinograd=False q_group_size=0 q_symmetric=False matmul_perchannel=False gelu_mode=normal" --canonicalize --weight-fold -o yolo26n-pose_cv181x_int8_sym_tpu.mlir
module @"yolo26n-pose" attributes {module.FLOPs = 7757193000 : i64, module.addr_mode = "basic", module.chip = "cv181x", module.cores = 1 : i64, module.devices = 1 : i64, module.high_precision = false, module.mode = "INT8", module.platform = "ONNX", module.postprocess = "yolov26", module.state = "TOP_CALIBRATED", module
.top_run_mode = "STATIC", module.weight_file = "yolo26n-pose_top_f32_all_weight.npz"} {                                                                         
  func.func @main(%arg0: tensor<1x3x640x640xf32> loc(unknown)) -> tensor<1x1x200x7x!quant.calibrated<f32<-679.54791260000002:679.54791260000002>>> {            
    %0 = "top.None"() : () -> none loc(unknown)                                 
    %1 = "top.Input"(%arg0) {channel_format = "nchw", do_preprocess = true, keep_aspect_ratio = false, keep_ratio_mode = "letterbox", mean = [0.000000e+00, 0.000000e+00, 0.000000e+00], pad_type = "center", pad_value = 0 : i64, pixel_format = "rgb", resize_dims = [640, 640], scale = [0.0039216000586748123, 0.00392160005
86748123, 0.0039216000586748123], yuv_type = ""} : (tensor<1x3x640x640xf32>) -> tensor<1x3x640x640x!quant.calibrated<f32<-1.000008:1.000008>>> loc("images")    
    %2 = "top.Weight"() : () -> tensor<16x3x3x3xf32> loc("model.0.conv.weight")
    %3 = "top.Weight"() : () -> tensor<16xf32> loc("model.0.conv.bias")         
    %4 = "top.Conv"(%1, %2, %3) {dilations = [1, 1], do_relu = false, dynweight_reorderd = false, group = 1 : i64, kernel_shape = [3, 3], pads = [1, 1, 1, 1], relu_limit = -1.000000e+00 : f64, strides = [2, 2], weight_is_coeff = 1 : i64} : (tensor<1x3x640x640x!quant.calibrated<f32<-1.000008:1.000008>>>, tensor<16x3x3x3
xf32>, tensor<16xf32>) -> tensor<1x16x320x320x!quant.calibrated<f32<-46.822227499999997:46.822227499999997>>> loc("/model.0/conv/Conv_output_0_Conv")           
    %5 = "top.SiLU"(%4) : (tensor<1x16x320x320x!quant.calibrated<f32<-46.822227499999997:46.822227499999997>>>) -> tensor<1x16x320x320x!quant.calibrated<f32<-39.618843099999999:39.618843099999999>>> loc("/model.0/act/Mul_output_0_Mul")
    %6 = "top.Weight"() : () -> tensor<32x16x3x3xf32> loc("model.1.conv.weight")
    %7 = "top.Weight"() : () -> tensor<32xf32> loc("model.1.conv.bias")         
    %8 = "top.Conv"(%5, %6, %7) {dilations = [1, 1], do_relu = false, dynweight_reorderd = false, group = 1 : i64, kernel_shape = [3, 3], pads = [1, 1, 1, 1], relu_limit = -1.000000e+00 : f64, strides = [2, 2], weight_is_coeff = 1 : i64} : (tensor<1x16x320x320x!quant.calibrated<f32<-39.618843099999999:39.61884309999999
9>>>, tensor<32x16x3x3xf32>, tensor<32xf32>) -> tensor<1x32x160x160x!quant.calibrated<f32<-68.7999954:68.7999954>>> loc("/model.1/conv/Conv_output_0_Conv")     
    %9 = "top.SiLU"(%8) : (tensor<1x32x160x160x!quant.calibrated<f32<-68.7999954:68.7999954>>>) -> tensor<1x32x160x160x!quant.calibrated<f32<-47.066119899999997:47.066119899999997>>> loc("/model.1/act/Mul_output_0_Mul")
    %10 = "top.Weight"() : () -> tensor<32x32x1x1xf32> loc("model.2.cv1.conv.weight")                                                                           
    %11 = "top.Weight"() : () -> tensor<32xf32> loc("model.2.cv1.conv.bias")    
    %12 = "top.Conv"(%9, %10, %11) {dilations = [1, 1], do_relu = false, dynweight_reorderd = false, group = 1 : i64, kernel_shape = [1, 1], pads = [0, 0, 0, 0], relu_limit = -1.000000e+00 : f64, strides = [1, 1], weight_is_coeff = 1 : i64} : (tensor<1x32x160x160x!quant.calibrated<f32<-47.066119899999997:47.06611989999
9997>>>, tensor<32x32x1x1xf32>, tensor<32xf32>) -> tensor<1x32x160x160x!quant.calibrated<f32<-55.433734899999997:55.433734899999997>>> loc("/model.2/cv1/conv/Conv_output_0_Conv")
    %13 = "top.SiLU"(%12) : (tensor<1x32x160x160x!quant.calibrated<f32<-55.433734899999997:55.433734899999997>>>) -> tensor<1x32x160x160x!quant.calibrated<f32<-39.016002700000001:39.016002700000001>>> loc("/model.2/cv1/act/Mul_output_0_Mul")
    %14 = "top.Slice"(%13, %0, %0, %0) {axes = [], ends = [9223372036854775807, 16, 9223372036854775807, 9223372036854775807], hasparamConvert_axes = [1], offset = [0, 0, 0, 0], steps = [1, 1, 1, 1]} : (tensor<1x32x160x160x!quant.calibrated<f32<-39.016002700000001:39.016002700000001>>>, none, none, none) -> tensor<1x16
x160x160x!quant.calibrated<f32<-39.016002700000001:39.016002700000001>>> loc("/model.2/Split_output_0_Split")
<...>
    %563 = "top.Reshape"(%562) {flatten_start_dim = -1 : i64} : (tensor<1x300x!quant.calibrated<f32<-1.000000e-05:1.000000e-05>>>) -> tensor<1x300x1x!quant.calibrated<f32<-1.000000e-05:1.000000e-05>>> loc("/model.23/Unsqueeze_2_output_0_Unsqueeze")
    %564 = "top.Gather"(%555, %indices_1) {axis = 0 : si32, is_lora = false, is_scalar = false, keepdims = true} : (tensor<300x1x!quant.calibrated<f32<-8383.0830077999999:8383.0830077999999>>>, tensor<1x300x!quant.calibrated<f32<-2.990000e+02:2.990000e+02>>>) -> tensor<1x300x1x!quant.calibrated<f32<-8383.0830077999999:8383.0830077999999>>> loc("/model.23/Reshape_12_output_0_Gather")
    %565 = "top.Cast"(%563) {round_mode = "HalfAwayFromZero", to = "F32"} : (tensor<1x300x1x!quant.calibrated<f32<-1.000000e-05:1.000000e-05>>>) -> tensor<1x300x1x!quant.calibrated<f32<-1.000000e-05:1.000000e-05>>> loc("/model.23/Cast_2_output_0_Cast")
    %566 = "top.Tile"(%564) {tile = [1, 1, 4]} : (tensor<1x300x1x!quant.calibrated<f32<-8383.0830077999999:8383.0830077999999>>>) -> tensor<1x300x4x!quant.calibrated<f32<-8383.0830077999999:8383.0830077999999>>> loc("/model.23/Tile_1_output_0_Tile")
    %567 = "top.Tile"(%564) {tile = [1, 1, 51]} : (tensor<1x300x1x!quant.calibrated<f32<-8383.0830077999999:8383.0830077999999>>>) -> tensor<1x300x51x!quant.calibrated<f32<-8383.0830077999999:8383.0830077999999>>> loc("/model.23/Tile_2_output_0_Tile")
    %568 = "top.GatherElements"(%550, %566) {axis = 1 : i64} : (tensor<1x8400x4x!quant.calibrated<f32<-762.5643311:762.5643311>>>, tensor<1x300x4x!quant.calibrated<f32<-8383.0830077999999:8383.0830077999999>>>) -> tensor<1x300x4x!quant.calibrated<f32<-646.18310550000001:646.18310550000001>>> loc("/model.23/GatherElements_1_output_0_GatherElements")
    %569 = "top.GatherElements"(%552, %567) {axis = 1 : i64} : (tensor<1x8400x51x!quant.calibrated<f32<-640.83679199999995:640.83679199999995>>>, tensor<1x300x51x!quant.calibrated<f32<-8383.0830077999999:8383.0830077999999>>>) -> tensor<1x300x51x!quant.calibrated<f32<-679.57073969999999:679.57073969999999>>> loc("/model.23/GatherElements_2_output_0_GatherElements")
    %570 = "top.Concat"(%568, %558, %565, %569) {axis = 2 : si32, do_relu = false, only_merge = false, relu_limit = -1.000000e+00 : f64, round_mode = "HalfAwayFromZero"} : (tensor<1x300x4x!quant.calibrated<f32<-646.18310550000001:646.18310550000001>>>, tensor<1x300x1x!quant.calibrated<f32<-0.94782730000000004:0.94782730000000004>>>, tensor<1x300x1x!quant.calibrated<f32<-1.000000e-05:1.000000e-05>>>, tensor<1x300x51x!quant.calibrated<f32<-679.57073969999999:679.57073969999999>>>) -> tensor<1x300x57x!quant.calibrated<f32<-724.69607980000001:724.69607980000001>>> loc("output0_Concat")
    %571 = "top.YoloDetection"(%570) {agnostic_nms = false, anchors = [0.000000e+00], class_num = 80 : i64, keep_topk = 200 : i64, net_input_h = 640 : i64, net_input_w = 640 : i64, nms_threshold = 5.000000e-01 : f64, num_boxes = 3 : i64, obj_threshold = 5.000000e-01 : f64, version = "yolov26"} : (tensor<1x300x57x!quant.calibrated<f32<-724.69607980000001:724.69607980000001>>>) -> tensor<1x1x200x7x!quant.calibrated<f32<-679.54791260000002:679.54791260000002>>> loc("yolo_post")
    return %571 : tensor<1x1x200x7x!quant.calibrated<f32<-679.54791260000002:679.54791260000002>>> loc(unknown)
  } loc(unknown)
} loc(unknown)
Entering FusePreprocessPass.
WARNING: qscale > 1,  = 9.999990e-01
Not support now.
UNREACHABLE executed at /w/tpu-mlir/lib/Conversion/TopToTpu/CV18xx/CompareConst.cpp:32!
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.      Program arguments: tpuc-opt model_yolo26n-pose/yolo26n-pose.mlir --init "--processor-assign=chip=cv181x mode=INT8 num_device=1 num_core=1 addr_mode=auto high_precision=False" "--import-calibration-table=file=model_yolo26n-pose/yolo26n-pose_cali_table asymmetric=False" --processor-top-optimize "--fuse-prepr
ocess=mode=INT8 customization_format=RGB_PLANAR align=True" "--convert-top-to-tpu=weightFileName=yolo26n-pose_cv181x_int8_sym_tpu_weights.npz asymmetric=False doWinograd=False q_group_size=0 q_symmetric=False matmul_perchannel=False gelu_mode=normal" --canonicalize --weight-fold --deinit --mlir-print-debuginfo -o yolo26n-pose_cv181x_int8_sym_tpu.mlir
 #0 0x0000559086914ce7 (/w/tpu-mlir/install/bin/tpuc-opt+0x91ece7)
 #1 0x0000559086912a0e (/w/tpu-mlir/install/bin/tpuc-opt+0x91ca0e)
 #2 0x000055908691566a (/w/tpu-mlir/install/bin/tpuc-opt+0x91f66a)
 #3 0x00007fa0c4c3f520 (/lib/x86_64-linux-gnu/libc.so.6+0x42520)
 #4 0x00007fa0c4c939fc pthread_kill (/lib/x86_64-linux-gnu/libc.so.6+0x969fc)
 #5 0x00007fa0c4c3f476 gsignal (/lib/x86_64-linux-gnu/libc.so.6+0x42476)
 #6 0x00007fa0c4c257f3 abort (/lib/x86_64-linux-gnu/libc.so.6+0x287f3)
 #7 0x0000559086912831 (/w/tpu-mlir/install/bin/tpuc-opt+0x91c831)
 #8 0x0000559086d58633 (/w/tpu-mlir/install/bin/tpuc-opt+0xd62633)
 #9 0x0000559086b20406 (/w/tpu-mlir/install/bin/tpuc-opt+0xb2a406)
#10 0x0000559086b1fba4 (/w/tpu-mlir/install/bin/tpuc-opt+0xb29ba4)
#11 0x0000559088200e67 (/w/tpu-mlir/install/bin/tpuc-opt+0x220ae67)
#12 0x00005590881fd77f (/w/tpu-mlir/install/bin/tpuc-opt+0x220777f)
#13 0x00005590881c697c (/w/tpu-mlir/install/bin/tpuc-opt+0x21d097c)
#14 0x00005590881c37cc (/w/tpu-mlir/install/bin/tpuc-opt+0x21cd7cc)
#15 0x0000559086a713d4 (/w/tpu-mlir/install/bin/tpuc-opt+0xa7b3d4)
#16 0x000055908822a154 (/w/tpu-mlir/install/bin/tpuc-opt+0x2234154)
#17 0x000055908822a781 (/w/tpu-mlir/install/bin/tpuc-opt+0x2234781)
#18 0x000055908822cc28 (/w/tpu-mlir/install/bin/tpuc-opt+0x2236c28)
#19 0x000055908690636b (/w/tpu-mlir/install/bin/tpuc-opt+0x91036b)
#20 0x0000559086905734 (/w/tpu-mlir/install/bin/tpuc-opt+0x90f734)
#21 0x0000559088447a28 (/w/tpu-mlir/install/bin/tpuc-opt+0x2451a28)
#22 0x00005590868ffa3a (/w/tpu-mlir/install/bin/tpuc-opt+0x909a3a)
#23 0x00005590868fff04 (/w/tpu-mlir/install/bin/tpuc-opt+0x909f04)
#24 0x00005590868fe94a (/w/tpu-mlir/install/bin/tpuc-opt+0x90894a)
#25 0x00007fa0c4c26d90 (/lib/x86_64-linux-gnu/libc.so.6+0x29d90)
#26 0x00007fa0c4c26e40 __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x29e40)
#27 0x00005590868fdd55 (/w/tpu-mlir/install/bin/tpuc-opt+0x907d55)
Aborted
Traceback (most recent call last):
  File "/w/tpu-mlir/python/tools/model_deploy.py", line 670, in <module>
    lowering_patterns = tool.lowering()
  File "/w/tpu-mlir/python/tools/model_deploy.py", line 247, in lowering
    patterns = mlir_lowering(self.mlir_file,
  File "/w/tpu-mlir/install/python/utils/mlir_shell.py", line 642, in mlir_lowering
    _os_system(cmd, mute=mute, log_level=log_level)
  File "/w/tpu-mlir/install/python/utils/mlir_shell.py", line 495, in _os_system
    raise RuntimeError("[!Error]: {}".format(cmd_str))
RuntimeError: [!Error]: tpuc-opt model_yolo26n-pose/yolo26n-pose.mlir --processor-assign="chip=cv181x mode=INT8 num_device=1 num_core=1 addr_mode=auto high_precision=False" --import-calibration-table="file=model_yolo26n-pose/yolo26n-pose_cali_table asymmetric=False" --processor-top-optimize --fuse-preprocess="mode=INT8 customization_format=RGB_PLANAR align=True" --convert-top-to-tpu="weightFileName=yolo26n-pose_cv181x_int8_sym_tpu_weights.npz asymmetric=False doWinograd=False q_group_size=0 q_symmetric=False matmul_perchannel=False gelu_mode=normal" --canonicalize --weight-fold -o yolo26n-pose_cv181x_int8_sym_tpu.mlir

[hans-bbt]

How did you solve the problem of the missing mod operator? Thank you.

[vvwcap]

Has this problem been resolved? I'm having the same issue; it doesn't support YOLOv26-Pose model conversion.

[vvwcap]

Has this problem been resolved? I'm having the same issue; it doesn't support YOLOv26-Pose model conversion.

mlir → bmodel fails

[hans-bbt]

Has this problem been resolved? I'm having the same issue; it doesn't support YOLOv26-Pose model conversion.

mlir → bmodel fails

model_transform.py
--model_name yolov26s
--model_def ../best26.onnx
--input_shapes [[1,3,640,640]]
--mean 0.0,0.0,0.0
--scale 0.0039216,0.0039216,0.0039216
--keep_aspect_ratio
--tolerance 0.8,0.5
--pixel_format rgb
--test_input ../image/people.jpg
--test_result yolov26s_top_outputs.npz
--mlir yolo26s.mlir

model_deploy.py
--mlir yolo26s.mlir
--quantize F16
--processor bm1684x
--tolerance 0.7,0.3
--test_input yolov26s_in_f32.npz
--test_reference yolov26s_top_outputs.npz
--model yolov26s_1684x_f16.bmodel

use this command can be convert

[vvwcap]

Has this problem been resolved? I'm having the same issue; it doesn't support YOLOv26-Pose model conversion.

mlir → bmodel fails

model_transform.py --model_name yolov26s --model_def ../best26.onnx --input_shapes [[1,3,640,640]] --mean 0.0,0.0,0.0 --scale 0.0039216,0.0039216,0.0039216 --keep_aspect_ratio --tolerance 0.8,0.5 --pixel_format rgb --test_input ../image/people.jpg --test_result yolov26s_top_outputs.npz --mlir yolo26s.mlir

model_deploy.py --mlir yolo26s.mlir --quantize F16 --processor bm1684x --tolerance 0.7,0.3 --test_input yolov26s_in_f32.npz --test_reference yolov26s_top_outputs.npz --model yolov26s_1684x_f16.bmodel

use this command can be convert

It's not the YOLOv26 detection or segmentation model, but its pose model. When I convert YOLOv26 to pose, I get the same error as the original poster (3. mlir → bmodel fails).