Issue Station
Sign In
Tools hereToolbox

pytorch - 💡(How to fix) Fix cudaErrorLaunchOutOfResources in CTC loss backward on RTX 5090 (sm_120, Blackwell) with CUDA 13.0 [1 comments, 2 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
pytorch/pytorch#179214•Fetched 2026-04-08 02:32:50
View on GitHub
Comments
1
Participants
2
Timeline
19
Reactions
0
Author
Hideki-OoyanagiHideki-Ooyanagi
Participants
eqyeqyHideki-OoyanagiHideki-Ooyanagi
Timeline (top)
mentioned ×7subscribed ×7labeled ×4commented ×1

CTC loss backward raises cudaErrorLaunchOutOfResources on RTX 5090 (Blackwell, sm_120) with CUDA 13.0 when batch size × transcript length exceeds a certain threshold. The error occurs in Variable._execution_engine.run_backward and cannot be caught with CUDA_LAUNCH_BLOCKING=1 at the Python level.

Error Message

torch.AcceleratorError: CUDA error: too many resources requested for launch
Search for `cudaErrorLaunchOutOfResources` in
https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html
for more information.

Root Cause

CTC loss backward raises cudaErrorLaunchOutOfResources on RTX 5090 (Blackwell, sm_120) with CUDA 13.0 when batch size × transcript length exceeds a certain threshold. The error occurs in Variable._execution_engine.run_backward and cannot be caught with CUDA_LAUNCH_BLOCKING=1 at the Python level.

Fix Action

Workaround

Limiting batch size to ≤ 3 avoids the error but severely degrades training throughput.

Code Example

import torch

device = "cuda"
V = 3350
ctc_loss = torch.nn.CTCLoss(blank=0, reduction='none', zero_infinity=True)

# Failing case
B, T_enc, tlen = 8, 250, 65
encoded = torch.randn(B, T_enc, V + 1, device=device, requires_grad=True)
log_probs = encoded.log_softmax(dim=-1)
targets = torch.randint(1, V, (B, tlen), device=device)
encoded_len = torch.full((B,), T_enc, dtype=torch.long, device=device)
target_lengths = torch.full((B,), tlen, dtype=torch.long, device=device)

loss = ctc_loss(log_probs.permute(1, 0, 2), targets, encoded_len, target_lengths)
loss.mean().backward()  # <-- cudaErrorLaunchOutOfResources here

---

torch.AcceleratorError: CUDA error: too many resources requested for launch
Search for `cudaErrorLaunchOutOfResources` in
https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html
for more information.

---

GPU: NVIDIA GeForce RTX 5090
Compute Capability: 12.0
Total Memory: 31.3 GB
Warp size: 32
Multi-processor count: 170
Max threads per multiprocessor: 1536
Regs per multiprocessor: 65536
→ Max registers per thread: 65536 / 1536 = 42.67
Shared memory per block: 49152 bytes

---

Collecting environment information...                                                                 
PyTorch version: 2.9.0a0+50eac811a6.nv25.09                                                           
Is debug build: False                                                                                 
CUDA used to build PyTorch: 13.0                                                                      
ROCM used to build PyTorch: N/A                                                                       
                                                                                                      
OS: Ubuntu 24.04.3 LTS (x86_64)                                                                       
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0                                                    
Clang version: Could not collect                                                                      
CMake version: version 3.31.6                                                                         
Libc version: glibc-2.39                                                                              
                                                                                                      
Python version: 3.12.3 (main, Aug 14 2025, 17:47:21) [GCC 13.3.0] (64-bit runtime)                    
Python platform: Linux-6.14.0-37-generic-x86_64-with-glibc2.39                                        
Is CUDA available: True                                                                               
CUDA runtime version: 13.0.88                                                                         
CUDA_MODULE_LOADING set to: LAZY                                                                      
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 5090                                          
Nvidia driver version: 580.105.08                                                                     
cuDNN version: Probably one of the following:                                                         
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.13.1                                                          
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.13.1                                                      
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.13.1                                                      
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.13.1                                      
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.13.1                                 
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.13.1                                                    
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.13.1                                                
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.13.1                                                      
Is XPU available: False                                                                               
HIP runtime version: N/A                                                                              
MIOpen runtime version: N/A                                                                           
Is XNNPACK available: True                                                                            
Caching allocator config: {'PYTORCH_CUDA_ALLOC_CONF': 'expandable_segments:True,max_split_size_mb:512'
}                                                                                                     
                                                                                                      
CPU:                                                                                                  
Architecture:                            x86_64                                                       
CPU op-mode(s):                          32-bit, 64-bit                                               
Address sizes:                           48 bits physical, 48 bits virtual                            
Byte Order:                              Little Endian                                                
CPU(s):                                  16                                                           
On-line CPU(s) list:                     0-15                                                         
Vendor ID:                               AuthenticAMD                                                 
Model name:                              AMD Ryzen 7 9700X 8-Core Processor                           
CPU family:                              26                                                           
Model:                                   68                                                           
Thread(s) per core:                      2                                                            
Core(s) per socket:                      8                                                            
Socket(s):                               1                                                            
Stepping:                                0                                                            
Frequency boost:                         enabled                                                      
CPU(s) scaling MHz:                      57%                                                          
CPU max MHz:                             5582.0000                                                    
CPU min MHz:                             600.0000                                                     
BogoMIPS:                                7585.99
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov
 pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_
good amd_lbr_v2 nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor sss
e3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_le
gacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext
 perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmc
all fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap av
x512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_
occup_llc cqm_mbm_total cqm_mbm_local user_shstk avx_vnni avx512_bf16 clzero irperf xsaveerptr rdpru w
bnoinvd cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilt
er pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi
2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid bus_lock_detect movdiri movdir
64b overflow_recov succor smca fsrm avx512_vp2intersect flush_l1d amd_lbr_pmc_freeze
Virtualization:                          AMD-V
L1d cache:                               384 KiB (8 instances)
L1i cache:                               256 KiB (8 instances)
L2 cache:                                8 MiB (8 instances)
L3 cache:                                32 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; IBPB on VMEXIT only
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanit
ization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIB
P always-on; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB on VMEXIT

Versions of relevant libraries:
[pip3] intel-openmp==2021.4.0
[pip3] mkl==2021.1.1
[pip3] mkl-devel==2021.1.1
[pip3] mkl-include==2021.1.1
[pip3] mypy_extensions==1.1.0
[pip3] numpy==1.26.4
[pip3] nv-one-logger-pytorch-lightning-integration==2.3.1
[pip3] nvidia-cuda-nvrtc==13.1.80
[pip3] nvidia-cuda-runtime==13.0.88
[pip3] nvidia-cudnn-frontend==1.16.0
[pip3] nvidia-nvjitlink==13.0.88
[pip3] nvidia-pytriton==0.7.0
[pip3] nvtx==0.2.14
[pip3] onnx==1.18.0
[pip3] onnx_graphsurgeon==0.5.8
[pip3] onnx-ir==0.1.12
[pip3] onnxscript==0.5.7.dev20251209
[pip3] open_clip_torch==3.2.0
[pip3] optree==0.17.0
[pip3] pytorch-lightning==2.6.0
[pip3] tbb==2021.13.1
[pip3] torch==2.9.0a0+50eac811a6.nv25.9
[pip3] torch_tensorrt==2.9.0a0
[pip3] torchao==0.13.0+git
[pip3] torchdiffeq==0.2.5 
[pip3] torchmetrics==1.8.2
[pip3] torchprofile==0.0.4
[pip3] torchsde==0.2.6
[pip3] torchvision==0.24.0a0+98f8b375
[pip3] torchx==0.7.0
[pip3] triton==3.4.0+gitc817b9b6
[pip3] tritonclient==2.51.0
[conda] Could not collect

---

curl -OL https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py
#For security purposes, please check the contents of collect_env.py before running it.
python collect_env.py

---

% Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100 31107  100 31107    0     0  1051k      0 --:--:-- --:--:-- --:--:-- 1084k
Collecting environment information...
PyTorch version: 2.9.0a0+50eac811a6.nv25.09
Is debug build: False
CUDA used to build PyTorch: 13.0
ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.3 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version: Could not collect
CMake version: version 3.31.6
Libc version: glibc-2.39

Python version: 3.12.3 (main, Aug 14 2025, 17:47:21) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.14.0-37-generic-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: 13.0.88
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 5090
Nvidia driver version: 580.105.08
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.13.1
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: {'PYTORCH_CUDA_ALLOC_CONF': 'expandable_segments:True,max_split_size_mb:512'
}

CPU:
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           48 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  16
On-line CPU(s) list:                     0-15
Vendor ID:                               AuthenticAMD
Model name:                              AMD Ryzen 7 9700X 8-Core Processor
CPU family:                              26
Model:                                   68
Thread(s) per core:                      2
Core(s) per socket:                      8
Socket(s):                               1
Stepping:                                0
Frequency boost:                         enabled
CPU(s) scaling MHz:                      60%
CPU max MHz:                             5582.0000
CPU min MHz:                             600.0000
BogoMIPS:                                7585.99
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov
 pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_
good amd_lbr_v2 nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor sss
e3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_le
gacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext
 perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmc
all fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap av
x512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_
occup_llc cqm_mbm_total cqm_mbm_local user_shstk avx_vnni avx512_bf16 clzero irperf xsaveerptr rdpru w
bnoinvd cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilt
er pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi
2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid bus_lock_detect movdiri movdir
64b overflow_recov succor smca fsrm avx512_vp2intersect flush_l1d amd_lbr_pmc_freeze
Virtualization:                          AMD-V
L1d cache:                               384 KiB (8 instances)
L1i cache:                               256 KiB (8 instances)
L2 cache:                                8 MiB (8 instances)
L3 cache:                                32 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; IBPB on VMEXIT only
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanit
ization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIB
P always-on; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB on VMEXIT

Versions of relevant libraries:
[pip3] intel-openmp==2021.4.0
[pip3] mkl==2021.1.1
[pip3] mkl-devel==2021.1.1
[pip3] mkl-include==2021.1.1
[pip3] mypy_extensions==1.1.0
[pip3] numpy==1.26.4
[pip3] nv-one-logger-pytorch-lightning-integration==2.3.1
[pip3] nvidia-cuda-nvrtc==13.1.80
[pip3] nvidia-cuda-runtime==13.0.88
[pip3] nvidia-cudnn-frontend==1.16.0
[pip3] nvidia-nvjitlink==13.0.88
[pip3] nvidia-pytriton==0.7.0
[pip3] nvtx==0.2.14
[pip3] onnx==1.18.0
[pip3] onnx_graphsurgeon==0.5.8
[pip3] onnx-ir==0.1.12
[pip3] onnxscript==0.5.7.dev20251209
[pip3] open_clip_torch==3.2.0
[pip3] optree==0.17.0
[pip3] pytorch-lightning==2.6.0
[pip3] tbb==2021.13.1
[pip3] torch==2.9.0a0+50eac811a6.nv25.9
[pip3] torch_tensorrt==2.9.0a0
[pip3] torchao==0.13.0+git
[pip3] torchdiffeq==0.2.5 
[pip3] torchmetrics==1.8.2
[pip3] torchprofile==0.0.4
[pip3] torchsde==0.2.6
[pip3] torchvision==0.24.0a0+98f8b375
[pip3] torchx==0.7.0
[pip3] triton==3.4.0+gitc817b9b6
[pip3] tritonclient==2.51.0
[conda] Could not collect
RAW_BUFFERClick to expand / collapse

🐛 Describe the bug

Description

CTC loss backward raises cudaErrorLaunchOutOfResources on RTX 5090 (Blackwell, sm_120) with CUDA 13.0 when batch size × transcript length exceeds a certain threshold. The error occurs in Variable._execution_engine.run_backward and cannot be caught with CUDA_LAUNCH_BLOCKING=1 at the Python level.

Environment

  • GPU: NVIDIA GeForce RTX 5090
  • Compute Capability: sm_120 (Blackwell)
  • Driver Version: 580.105.08
  • CUDA Version: 13.0
  • PyTorch Version: 2.9.0a0+50eac811a6.nv25.09
  • Python Version: 3.12
  • OS: Ubuntu (Docker container)

Reproduction

import torch

device = "cuda"
V = 3350
ctc_loss = torch.nn.CTCLoss(blank=0, reduction='none', zero_infinity=True)

# Failing case
B, T_enc, tlen = 8, 250, 65
encoded = torch.randn(B, T_enc, V + 1, device=device, requires_grad=True)
log_probs = encoded.log_softmax(dim=-1)
targets = torch.randint(1, V, (B, tlen), device=device)
encoded_len = torch.full((B,), T_enc, dtype=torch.long, device=device)
target_lengths = torch.full((B,), tlen, dtype=torch.long, device=device)

loss = ctc_loss(log_probs.permute(1, 0, 2), targets, encoded_len, target_lengths)
loss.mean().backward()  # <-- cudaErrorLaunchOutOfResources here

Threshold observed

Through systematic testing, the following threshold was identified:

Batch size (B)Max transcript length (OK)Min transcript length (FAIL)
1–3173+ ✅(not found)
4–7120 ✅130 ❌
≥ 860 ✅65 ❌

The threshold appears to correspond to next_power_of_2(2 × transcript_len + 1):

  • transcript ≤ 120 → 241 → 256 threads ✅
  • transcript ≥ 128 → 257 → 512 threads ❌ (exceeds register limit on sm_120)

Key observations

  • The error occurs only during backward, not forward
  • encoder backward alone passes without error
  • CTC loss backward is the failing component
  • Reproducible with both bf16 and fp32 precision
  • Reproducible with both single GPU and multi-GPU (DDP)
  • Batch size alone does not trigger the error; transcript length is the primary factor

Error message

torch.AcceleratorError: CUDA error: too many resources requested for launch
Search for `cudaErrorLaunchOutOfResources` in
https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html
for more information.

GPU resource limits

GPU: NVIDIA GeForce RTX 5090
Compute Capability: 12.0
Total Memory: 31.3 GB
Warp size: 32
Multi-processor count: 170
Max threads per multiprocessor: 1536
Regs per multiprocessor: 65536
→ Max registers per thread: 65536 / 1536 = 42.67
Shared memory per block: 49152 bytes

Expected behavior

CTC loss backward should run successfully on sm_120 (Blackwell) with the same configurations that work on previous GPU architectures.

Workaround

Limiting batch size to ≤ 3 avoids the error but severely degrades training throughput.

Versions

curl -sL https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py | python

Result

Collecting environment information...                                                                 
PyTorch version: 2.9.0a0+50eac811a6.nv25.09                                                           
Is debug build: False                                                                                 
CUDA used to build PyTorch: 13.0                                                                      
ROCM used to build PyTorch: N/A                                                                       
                                                                                                      
OS: Ubuntu 24.04.3 LTS (x86_64)                                                                       
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0                                                    
Clang version: Could not collect                                                                      
CMake version: version 3.31.6                                                                         
Libc version: glibc-2.39                                                                              
                                                                                                      
Python version: 3.12.3 (main, Aug 14 2025, 17:47:21) [GCC 13.3.0] (64-bit runtime)                    
Python platform: Linux-6.14.0-37-generic-x86_64-with-glibc2.39                                        
Is CUDA available: True                                                                               
CUDA runtime version: 13.0.88                                                                         
CUDA_MODULE_LOADING set to: LAZY                                                                      
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 5090                                          
Nvidia driver version: 580.105.08                                                                     
cuDNN version: Probably one of the following:                                                         
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.13.1                                                          
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.13.1                                                      
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.13.1                                                      
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.13.1                                      
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.13.1                                 
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.13.1                                                    
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.13.1                                                
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.13.1                                                      
Is XPU available: False                                                                               
HIP runtime version: N/A                                                                              
MIOpen runtime version: N/A                                                                           
Is XNNPACK available: True                                                                            
Caching allocator config: {'PYTORCH_CUDA_ALLOC_CONF': 'expandable_segments:True,max_split_size_mb:512'
}                                                                                                     
                                                                                                      
CPU:                                                                                                  
Architecture:                            x86_64                                                       
CPU op-mode(s):                          32-bit, 64-bit                                               
Address sizes:                           48 bits physical, 48 bits virtual                            
Byte Order:                              Little Endian                                                
CPU(s):                                  16                                                           
On-line CPU(s) list:                     0-15                                                         
Vendor ID:                               AuthenticAMD                                                 
Model name:                              AMD Ryzen 7 9700X 8-Core Processor                           
CPU family:                              26                                                           
Model:                                   68                                                           
Thread(s) per core:                      2                                                            
Core(s) per socket:                      8                                                            
Socket(s):                               1                                                            
Stepping:                                0                                                            
Frequency boost:                         enabled                                                      
CPU(s) scaling MHz:                      57%                                                          
CPU max MHz:                             5582.0000                                                    
CPU min MHz:                             600.0000                                                     
BogoMIPS:                                7585.99
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov
 pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_
good amd_lbr_v2 nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor sss
e3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_le
gacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext
 perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmc
all fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap av
x512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_
occup_llc cqm_mbm_total cqm_mbm_local user_shstk avx_vnni avx512_bf16 clzero irperf xsaveerptr rdpru w
bnoinvd cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilt
er pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi
2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid bus_lock_detect movdiri movdir
64b overflow_recov succor smca fsrm avx512_vp2intersect flush_l1d amd_lbr_pmc_freeze
Virtualization:                          AMD-V
L1d cache:                               384 KiB (8 instances)
L1i cache:                               256 KiB (8 instances)
L2 cache:                                8 MiB (8 instances)
L3 cache:                                32 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; IBPB on VMEXIT only
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanit
ization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIB
P always-on; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB on VMEXIT

Versions of relevant libraries:
[pip3] intel-openmp==2021.4.0
[pip3] mkl==2021.1.1
[pip3] mkl-devel==2021.1.1
[pip3] mkl-include==2021.1.1
[pip3] mypy_extensions==1.1.0
[pip3] numpy==1.26.4
[pip3] nv-one-logger-pytorch-lightning-integration==2.3.1
[pip3] nvidia-cuda-nvrtc==13.1.80
[pip3] nvidia-cuda-runtime==13.0.88
[pip3] nvidia-cudnn-frontend==1.16.0
[pip3] nvidia-nvjitlink==13.0.88
[pip3] nvidia-pytriton==0.7.0
[pip3] nvtx==0.2.14
[pip3] onnx==1.18.0
[pip3] onnx_graphsurgeon==0.5.8
[pip3] onnx-ir==0.1.12
[pip3] onnxscript==0.5.7.dev20251209
[pip3] open_clip_torch==3.2.0
[pip3] optree==0.17.0
[pip3] pytorch-lightning==2.6.0
[pip3] tbb==2021.13.1
[pip3] torch==2.9.0a0+50eac811a6.nv25.9
[pip3] torch_tensorrt==2.9.0a0
[pip3] torchao==0.13.0+git
[pip3] torchdiffeq==0.2.5 
[pip3] torchmetrics==1.8.2
[pip3] torchprofile==0.0.4
[pip3] torchsde==0.2.6
[pip3] torchvision==0.24.0a0+98f8b375
[pip3] torchx==0.7.0
[pip3] triton==3.4.0+gitc817b9b6
[pip3] tritonclient==2.51.0
[conda] Could not collect
curl -OL https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py
#For security purposes, please check the contents of collect_env.py before running it.
python collect_env.py

Result

% Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                Dload  Upload   Total   Spent    Left  Speed
100 31107  100 31107    0     0  1051k      0 --:--:-- --:--:-- --:--:-- 1084k
Collecting environment information...
PyTorch version: 2.9.0a0+50eac811a6.nv25.09
Is debug build: False
CUDA used to build PyTorch: 13.0
ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.3 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version: Could not collect
CMake version: version 3.31.6
Libc version: glibc-2.39

Python version: 3.12.3 (main, Aug 14 2025, 17:47:21) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.14.0-37-generic-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: 13.0.88
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 5090
Nvidia driver version: 580.105.08
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.13.1
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.13.1
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: {'PYTORCH_CUDA_ALLOC_CONF': 'expandable_segments:True,max_split_size_mb:512'
}

CPU:
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           48 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  16
On-line CPU(s) list:                     0-15
Vendor ID:                               AuthenticAMD
Model name:                              AMD Ryzen 7 9700X 8-Core Processor
CPU family:                              26
Model:                                   68
Thread(s) per core:                      2
Core(s) per socket:                      8
Socket(s):                               1
Stepping:                                0
Frequency boost:                         enabled
CPU(s) scaling MHz:                      60%
CPU max MHz:                             5582.0000
CPU min MHz:                             600.0000
BogoMIPS:                                7585.99
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov
pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_
good amd_lbr_v2 nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor sss
e3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_le
gacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext
perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmc
all fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap av
x512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_
occup_llc cqm_mbm_total cqm_mbm_local user_shstk avx_vnni avx512_bf16 clzero irperf xsaveerptr rdpru w
bnoinvd cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilt
er pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi
2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid bus_lock_detect movdiri movdir
64b overflow_recov succor smca fsrm avx512_vp2intersect flush_l1d amd_lbr_pmc_freeze
Virtualization:                          AMD-V
L1d cache:                               384 KiB (8 instances)
L1i cache:                               256 KiB (8 instances)
L2 cache:                                8 MiB (8 instances)
L3 cache:                                32 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; IBPB on VMEXIT only
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanit
ization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIB
P always-on; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Mitigation; IBPB on VMEXIT

Versions of relevant libraries:
[pip3] intel-openmp==2021.4.0
[pip3] mkl==2021.1.1
[pip3] mkl-devel==2021.1.1
[pip3] mkl-include==2021.1.1
[pip3] mypy_extensions==1.1.0
[pip3] numpy==1.26.4
[pip3] nv-one-logger-pytorch-lightning-integration==2.3.1
[pip3] nvidia-cuda-nvrtc==13.1.80
[pip3] nvidia-cuda-runtime==13.0.88
[pip3] nvidia-cudnn-frontend==1.16.0
[pip3] nvidia-nvjitlink==13.0.88
[pip3] nvidia-pytriton==0.7.0
[pip3] nvtx==0.2.14
[pip3] onnx==1.18.0
[pip3] onnx_graphsurgeon==0.5.8
[pip3] onnx-ir==0.1.12
[pip3] onnxscript==0.5.7.dev20251209
[pip3] open_clip_torch==3.2.0
[pip3] optree==0.17.0
[pip3] pytorch-lightning==2.6.0
[pip3] tbb==2021.13.1
[pip3] torch==2.9.0a0+50eac811a6.nv25.9
[pip3] torch_tensorrt==2.9.0a0
[pip3] torchao==0.13.0+git
[pip3] torchdiffeq==0.2.5 
[pip3] torchmetrics==1.8.2
[pip3] torchprofile==0.0.4
[pip3] torchsde==0.2.6
[pip3] torchvision==0.24.0a0+98f8b375
[pip3] torchx==0.7.0
[pip3] triton==3.4.0+gitc817b9b6
[pip3] tritonclient==2.51.0
[conda] Could not collect

extent analysis

TL;DR

Limit the batch size to ≤ 3 or reduce the transcript length to avoid exceeding the GPU's register limit and triggering the cudaErrorLaunchOutOfResources error.

Guidance

  • Identify the threshold for batch size and transcript length that triggers the error, as it seems to be related to the next_power_of_2(2 × transcript_len + 1) calculation.
  • Consider reducing the transcript length or batch size to stay within the GPU's register limit.
  • Verify that the error only occurs during the backward pass of the CTC loss calculation, as this might help in isolating the issue.
  • Check if using a different GPU architecture or updating the CUDA version could potentially resolve the issue.

Example

No specific code example is provided, as the issue seems to be related to the GPU's hardware limitations rather than a software bug. However, the user can try modifying the batch size or transcript length in the provided reproduction code to see if it resolves the issue.

Notes

The error is specific to the NVIDIA GeForce RTX 5090 GPU with CUDA 13.0, and it's not clear if this issue affects other GPU models or CUDA versions. The user may need to experiment with different batch sizes and transcript lengths to find a suitable workaround.

Recommendation

Apply a workaround by limiting the batch size to ≤ 3 or reducing the transcript length, as this seems to be the most straightforward way to avoid the cudaErrorLaunchOutOfResources error. Upgrading to a different GPU architecture or updating the CUDA version might also be considered, but this would require further testing to confirm.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

FAQ

Expected behavior

CTC loss backward should run successfully on sm_120 (Blackwell) with the same configurations that work on previous GPU architectures.

#api#memory management#API rate limit#retriever error#indexing error

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING