NVIDIA
Inference failure of TensorRT 10.0.x.x when running my internal model on GPU(T4, A100)
Description
After updating to TensorRT 10.0.1.6, we expected the previously reported issue to be resolved. Unfortunately, not only does the issue persist, but the model’s outputs have deteriorated even further. Specifically, all output values are now nan, making it impossible to use our models. This issue affects both fp16 and fp32 precision settings, rendering the model completely non-functional.
https://github.com/NVIDIA/TensorRT/issues/3292
Environment
TensorRT Version: All version of 10.0.x.x. NGC Container 24.05~24.07.
NVIDIA GPU: T4, A100
NVIDIA Driver Version: 550.90.07
CUDA Version: 12.4
CUDNN Version: x
Operating System:
Container (if so, version): NGC Container from 23.03 and 24.07. https://catalog.ngc.nvidia.com/orgs/nvidia/containers/tensorrt
Operating System:
Python Version (if applicable):
Tensorflow Version (if applicable):
PyTorch Version (if applicable):
Baremetal or Container (if so, version):
Relevant Files
Model link:
Steps To Reproduce
Commands or scripts:
Have you tried the latest release?:
Can this model run on other frameworks? For example run ONNX model with ONNXRuntime (polygraphy run <model.onnx> --onnxrt):
fail on polygraphy
related to https://github.com/NVIDIA/TensorRT/issues/3292
@lix19937 Can I send this to you via email? I prefer not to expose my model publicly. The --verbose option also reveals too much information.
@kimdwkimdw I didn't receive your e-mail.
Please check your spam mail inbox.
I've sent mail via Gmail.
Can you upload the goole drive ?
OK I upload log file to gdrive and shared with your email
@kimdwkimdw So sorry ! Current I has no env, can you upload the full log with trtexec --verbose 2>&1 |tee full_log to google drive to share me ? I will analysis it as soon as possible.
@lix19937 let me know your google address for Google drive. Your email address [email protected] seems doesn't work.
This is same kind of issue from https://github.com/NVIDIA/TensorRT/issues/3292
TensorRT 10.x have significant errors.
cc. @zerollzeng @ttyio
@lix19937 let me know your google address for Google drive. Your email address
[email protected]seems doesn't work.
sent log to email [email protected]
@kimdwkimdw
but the model’s outputs have deteriorated even further. Specifically, all output values are now nan, making it impossible to use our models.
From your logs, it has no valid errors or warnings. Maybe you can use polygraphy, like follow
polygraphy run model.onnx --trt --onnxrt --input-shapes source:[2,160000] wav_lens:[2,1]
to check which layer begin to arise the big nan/diff, check whether a BN after conv, etc. Also you can check the weights max-min range.
Another hand, you can try to use latest version.
@lix19937
Thank you for the suggestion, but I have already tried using polygraphy along with other methods. My question goes back to the root of the issue: why do nan values in the relative difference output from polygraphy start appearing when using TensorRT version 10.x?
I did not encounter this issue with TensorRT 8.5.3, including versions like 23.02 and 23.03, where there were no nan values. However, starting from TensorRT 8.6.1.6, the errors have become more pronounced, and with all versions of TensorRT 10 (e.g., 10.3.0.26, 10.2.0, 10.1.0), the model’s errors seem to overflow dramatically.
In my opinion, form 8.6, tensorrt add more feature, like builder optimization level(the default optimization level is 3. Valid values include integers from 0 to the maximum optimization level 5), and import llm layers fusion(like mha, ln), for normalization layers, particularly if deploying with mixed precision, target the latest ONNX opset that contains the corresponding function ops, for example: opset 17 for LayerNormalization or opset 18 GroupNormalization. Numerical accuracy using function ops is superior to corresponding implementation with primitive ops for normalization layers. And move cuda-x lib depends.
You can try to add --builderOptimizationLevel=5 --noTF32 and adjust the size of --memPoolSize . @kimdwkimdw
I have already tried the suggestions, including setting --builderOptimizationLevel=5 and --noTF32, and updating the ONNX opset from 15 to 17. Currently, proper export for opset 18 is not supported in PyTorch (Ref: PyTorch ONNX Export).
Despite these adjustments, including experimenting with different precision combinations (fp16, bf16, tf32, fp32 with --noTF32), the output from polygraphy continues to show significant degradation as follows:
[I] Error Metrics: x
[I] Minimum Required Tolerance: elemwise error | [abs=11.761] OR [rel=1393.1] (requirements may be lower if both abs/rel tolerances are set)
[I] Absolute Difference | Stats: mean=1.6023, std-dev=1.614, var=2.605, median=1.0086, min=0 at (1, 234, 1208), max=11.761 at (0, 14, 619), avg-magnitude=1.6023, p90=4.0612, p95=4.0612, p99=6.56
[I] ---- Histogram ----
Bin Range | Num Elems | Visualization
(0 , 1.18) | 8723143 | ########################################
(1.18, 2.35) | 3338792 | ###############
(2.35, 3.53) | 1736355 | #######
(3.53, 4.7 ) | 1026219 | ####
(4.7 , 5.88) | 718419 | ###
(5.88, 7.06) | 308960 | #
(7.06, 8.23) | 65386 |
(8.23, 9.41) | 16262 |
(9.41, 10.6) | 2064 |
(10.6, 11.8) | 400 |
[I] Relative Difference | Stats: mean=0.1035, std-dev=0.84023, var=0.70599, median=0.068093, min=0 at (1, 234, 1208), max=1393.1 at (0, 107, 0), avg-magnitude=0.1035, p90=0.24082, p95=0.24082, p99=0.40837
[I] ---- Histogram ----
Bin Range | Num Elems | Visualization
(0 , 139 ) | 15935951 | ########################################
(139 , 279 ) | 36 |
(279 , 418 ) | 3 |
(418 , 557 ) | 4 |
(557 , 697 ) | 2 |
(697 , 836 ) | 0 |
(836 , 975 ) | 1 |
(975 , 1.11e+03) | 0 |
(1.11e+03, 1.25e+03) | 0 |
(1.25e+03, 1.39e+03) | 3 |
This indicates a severe accuracy issue. Features that were functional in TensorRT 8.5 no longer work correctly in the later versions.
@lix19937
I have already tried the suggestions, including setting
--builderOptimizationLevel=5and--noTF32, and updating the ONNX opset from 15 to 17. Currently, proper export for opset 18 is not supported in PyTorch (Ref: PyTorch ONNX Export).Despite these adjustments, including experimenting with different precision combinations (fp16, bf16, tf32, fp32 with --noTF32), the output from polygraphy continues to show significant degradation as follows:
[I] Error Metrics: x [I] Minimum Required Tolerance: elemwise error | [abs=11.761] OR [rel=1393.1] (requirements may be lower if both abs/rel tolerances are set) [I] Absolute Difference | Stats: mean=1.6023, std-dev=1.614, var=2.605, median=1.0086, min=0 at (1, 234, 1208), max=11.761 at (0, 14, 619), avg-magnitude=1.6023, p90=4.0612, p95=4.0612, p99=6.56 [I] ---- Histogram ---- Bin Range | Num Elems | Visualization (0 , 1.18) | 8723143 | ######################################## (1.18, 2.35) | 3338792 | ############### (2.35, 3.53) | 1736355 | ####### (3.53, 4.7 ) | 1026219 | #### (4.7 , 5.88) | 718419 | ### (5.88, 7.06) | 308960 | # (7.06, 8.23) | 65386 | (8.23, 9.41) | 16262 | (9.41, 10.6) | 2064 | (10.6, 11.8) | 400 | [I] Relative Difference | Stats: mean=0.1035, std-dev=0.84023, var=0.70599, median=0.068093, min=0 at (1, 234, 1208), max=1393.1 at (0, 107, 0), avg-magnitude=0.1035, p90=0.24082, p95=0.24082, p99=0.40837 [I] ---- Histogram ---- Bin Range | Num Elems | Visualization (0 , 139 ) | 15935951 | ######################################## (139 , 279 ) | 36 | (279 , 418 ) | 3 | (418 , 557 ) | 4 | (557 , 697 ) | 2 | (697 , 836 ) | 0 | (836 , 975 ) | 1 | (975 , 1.11e+03) | 0 | (1.11e+03, 1.25e+03) | 0 | (1.25e+03, 1.39e+03) | 3 |This indicates a severe accuracy issue. Features that were functional in TensorRT 8.5 no longer work correctly in the later versions.
can you provide the result of polygraphy compare trt with onnxrt in trt8.5 ?
Here are two version of polygraphy compare trt with onnxrt between 8.5 and TensorRT 10.3.0.26
trt8.5 with fp32
[I] trt-runner-N0-08/14/24-04:33:32
---- Inference Input(s) ----
{source [dtype=float32, shape=(1, 160000)],
wav_lens [dtype=float32, shape=(1, 1)]}
[I] trt-runner-N0-08/14/24-04:33:32
---- Inference Output(s) ----
{x [dtype=float32, shape=(1, 999, 2000)]}
[38;5;10m[I] trt-runner-N0-08/14/24-04:33:32 | Completed 1 iteration(s) in 45.12 ms | Average inference time: 45.12 ms.[0m
[38;5;14m[I] Accuracy Comparison | onnxrt-runner-N0-08/14/24-04:33:32 vs. trt-runner-N0-08/14/24-04:33:32[0m
[38;5;14m[I] Comparing Output: 'x' (dtype=float32, shape=(1, 999, 2000)) with 'x' (dtype=float32, shape=(1, 999, 2000))[0m
[I] Tolerance: [abs=1e-05, rel=1e-05] | Checking elemwise error
[I] onnxrt-runner-N0-08/14/24-04:33:32: x | Stats: mean=-15.392, std-dev=3.0801, var=9.487, median=-15.933, min=-27.929 at (0, 489, 2), max=-0.0010415 at (0, 0, 4), avg-magnitude=15.392
[I] ---- Histogram ----
Bin Range | Num Elems | Visualization
(-27.9 , -25.1 ) | 1126 |
(-25.1 , -22.3 ) | 5617 |
(-22.3 , -19.6 ) | 107791 | ######
(-19.6 , -16.8 ) | 679836 | ########################################
(-16.8 , -14 ) | 585686 | ##################################
(-14 , -11.2 ) | 384150 | ######################
(-11.2 , -8.38 ) | 209469 | ############
(-8.38 , -5.59 ) | 22838 | #
(-5.59 , -2.79 ) | 479 |
(-2.79 , -0.00104) | 1008 |
[I] trt-runner-N0-08/14/24-04:33:32: x | Stats: mean=-15.392, std-dev=3.0801, var=9.487, median=-15.933, min=-27.929 at (0, 489, 2), max=-0.0010418 at (0, 0, 4), avg-magnitude=15.392
[I] ---- Histogram ----
Bin Range | Num Elems | Visualization
(-27.9 , -25.1 ) | 1126 |
(-25.1 , -22.3 ) | 5617 |
(-22.3 , -19.6 ) | 107793 | ######
(-19.6 , -16.8 ) | 679836 | ########################################
(-16.8 , -14 ) | 585686 | ##################################
(-14 , -11.2 ) | 384146 | ######################
(-11.2 , -8.38 ) | 209471 | ############
(-8.38 , -5.59 ) | 22838 | #
(-5.59 , -2.79 ) | 479 |
(-2.79 , -0.00104) | 1008 |
[I] Error Metrics: x
[I] Minimum Required Tolerance: elemwise error | [abs=0.00096989] OR [rel=0.00058633] (requirements may be lower if both abs/rel tolerances are set)
[I] Absolute Difference | Stats: mean=2.6239e-05, std-dev=3.3076e-05, var=1.094e-09, median=1.7166e-05, min=0 at (0, 1, 150), max=0.00096989 at (0, 489, 4), avg-magnitude=2.6239e-05
[I] ---- Histogram ----
Bin Range | Num Elems | Visualization
(0 , 9.7e-05 ) | 1945723 | ########################################
(9.7e-05 , 0.000194) | 42371 |
(0.000194, 0.000291) | 4457 |
(0.000291, 0.000388) | 2794 |
(0.000388, 0.000485) | 1888 |
(0.000485, 0.000582) | 648 |
(0.000582, 0.000679) | 71 |
(0.000679, 0.000776) | 31 |
(0.000776, 0.000873) | 14 |
(0.000873, 0.00097 ) | 3 |
[I] Relative Difference | Stats: mean=1.8316e-06, std-dev=2.9922e-06, var=8.9532e-12, median=1.1515e-06, min=0 at (0, 1, 150), max=0.00058633 at (0, 493, 0), avg-magnitude=1.8316e-06
[I] ---- Histogram ----
Bin Range | Num Elems | Visualization
(0 , 5.86e-05) | 1997412 | ########################################
(5.86e-05, 0.000117) | 477 |
(0.000117, 0.000176) | 91 |
(0.000176, 0.000235) | 14 |
(0.000235, 0.000293) | 2 |
(0.000293, 0.000352) | 2 |
(0.000352, 0.00041 ) | 0 |
(0.00041 , 0.000469) | 1 |
(0.000469, 0.000528) | 0 |
(0.000528, 0.000586) | 1 |
[38;5;9m[E] FAILED | Output: 'x' | Difference exceeds tolerance (rel=1e-05, abs=1e-05)[0m
[38;5;9m[E] FAILED | Mismatched outputs: ['x'][0m
trt10.x with fp32
[I] trt-runner-N0-08/14/24-04:04:19
---- Inference Input(s) ----
{source [dtype=float32, shape=(1, 160000)],
wav_lens [dtype=float32, shape=(1, 1)]}
[I] trt-runner-N0-08/14/24-04:04:19
---- Inference Output(s) ----
{x [dtype=float32, shape=(1, 999, 2000)]}
[38;5;10m[I] trt-runner-N0-08/14/24-04:04:19 | Completed 1 iteration(s) in 39.22 ms | Average inference time: 39.22 ms.[0m
[38;5;14m[I] Accuracy Comparison | onnxrt-runner-N0-08/14/24-04:04:19 vs. trt-runner-N0-08/14/24-04:04:19[0m
[38;5;14m[I] Comparing Output: 'x' (dtype=float32, shape=(1, 999, 2000)) with 'x' (dtype=float32, shape=(1, 999, 2000))[0m
[I] Tolerance: [abs=1e-05, rel=1e-05] | Checking elemwise error
[I] onnxrt-runner-N0-08/14/24-04:04:19: x | Stats: mean=-15.392, std-dev=3.0801, var=9.487, median=-15.933, min=-27.929 at (0, 489, 2), max=-0.0010415 at (0, 0, 4), avg-magnitude=15.392
[I] ---- Histogram ----
Bin Range | Num Elems | Visualization
(-27.9 , -25.1 ) | 1126 |
(-25.1 , -22.3 ) | 5617 |
(-22.3 , -19.6 ) | 107791 | ######
(-19.6 , -16.8 ) | 679836 | ########################################
(-16.8 , -14 ) | 585686 | ##################################
(-14 , -11.2 ) | 384150 | ######################
(-11.2 , -8.38 ) | 209469 | ############
(-8.38 , -5.59 ) | 22838 | #
(-5.59 , -2.79 ) | 479 |
(-2.79 , -0.00104) | 1008 |
[I] trt-runner-N0-08/14/24-04:04:19: x | Stats: mean=nan, std-dev=nan, var=nan, median=nan, min=nan at (0, 0, 0), max=nan at (0, 0, 0), avg-magnitude=nan
[I] None
[I] Error Metrics: x
[I] Minimum Required Tolerance: elemwise error | [abs=nan] OR [rel=nan] (requirements may be lower if both abs/rel tolerances are set)
[I] Absolute Difference | Stats: mean=nan, std-dev=nan, var=nan, median=nan, min=nan at (0, 0, 0), max=nan at (0, 0, 0), avg-magnitude=nan
[I]
[I] Relative Difference | Stats: mean=nan, std-dev=nan, var=nan, median=nan, min=nan at (0, 0, 0), max=nan at (0, 0, 0), avg-magnitude=nan
[I]
[38;5;9m[E] FAILED | Output: 'x' | Difference exceeds tolerance (rel=1e-05, abs=1e-05)[0m
[38;5;9m[E] FAILED | Mismatched outputs: ['x'][0m
[38;5;9m[E] Accuracy Summary | onnxrt-runner-N0-08/14/24-04:04:19 vs. trt-runner-N0-08/14/24-04:04:19 | Passed: 0/1 iterations | Pass Rate: 0.0%[0m
[38;5;9m[E] FAILED | Runtime: 32.87
@moraxu I've shared google drive link with [email protected].
Checkout 'raw_model' folder. There is 2024.model.onnx model file
And I've shared it with [email protected] last year.
@kimdwkimdw Can you please also try the latest TRT 10.5? There are some known accuracy issues that have been fixed in 10.5.
@yuanyao-nv
Currently, all of our tests are being conducted using the following official TensorRT container, but it appears that version 10.5 is not yet included in the 24.09 release: TensorRT 10.4.0.26 in NGC 24.09 - https://catalog.ngc.nvidia.com/orgs/nvidia/containers/tensorrt/tags
Once an update is available, I will test with version 10.5.
@yuanyao-nv "NGC 24.10 with TensorRT 10.5.0.18" updated recently. I've tested all of tests with TRT 10.5.
Nothing improved.