Fast-ACVNet
A stereo network by Gangwei Xu et al. that uses attention-selected disparities and a MobileNetV2 backbone, in a base and a plus version.
stable Fork of gangweix/Fast-ACVNet@61d5375 · backends: portable, kernels
- Paper Accurate and Efficient Stereo Matching via Attention Concatenation Volume TPAMI 2023 · first released 23 Sept 2022
Paper 2022 Added in 0.1.0 Evaluated with 0.2.0
Use it
pip install fastforks Lean install with only what this model needs: pip install fastforks-core
from fastforks import stereo
model = stereo.load("fast-acvnet-generalization") # kernels on a supported GPU, else portable PyTorch
result = model(left, right)
result.disparity
result.backend # what ran: mode, device, dtype, kernels fastforks bench fast-acvnet-generalization Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.
Speed and memory
Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 101 | 0.60× | 1,784 MiB/1,878 MiB | – | – |
| Original fp16torch.autocast(fp16) around the forward | fp16 | 60.6 | – | 1,421 MiB/1,506 MiB | – | – |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portablemeasured on fast-acvnet-sceneflow (same architecture, other checkpoint) | fp16 | 54.7 | 1.1× | 653 MiB/756 MiB | 0.3 | 0.06 |
| Kernelsmeasured on fast-acvnet-sceneflow (same architecture, other checkpoint) | fp16 | 38.1 | 1.6× | 654 MiB/716 MiB | 0.6 | 0.04 |
| Portable + torch.compile | 36.4 | 1.7× | 549 MiB/678 MiB | 42 (build) | – | |
| TensorRT fp16 | fp16 | 36.1 | 1.7× | 492 MiBenginenot comparable | 68 (build) | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-01 to 2026-11-16. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two
times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.
Other devices
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14 | Original | fp32 | 1334 | 4,975 MiB process | matches | 994x1482 |
| Portable | fp16 | 478 | 2,737 MiB process | matches | 994x1482 | |
| Portable | fp32 | 944 | 3,646 MiB process | matches | 994x1482 | |
| Kernels | fp16 | not implemented | – | 994x1482 | ||
| NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 75.7 | 662 MiB/756 MiB | – | 994x1482 |
| Kernels | fp16 | 51.7 | 738 MiB/916 MiB | – | 994x1482 | |
| NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 25.3 | 662 MiB/756 MiB | – | 994x1482 |
| Kernels | fp16 | 19.2 | 738 MiB/916 MiB | – | 994x1482 | |
| NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 54.3 | 662 MiB/756 MiB | – | 994x1482 |
| Kernels | fp16 | 43.6 | 738 MiB/856 MiB | – | 994x1482 | |
| NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 31.7 | 662 MiB/756 MiB | – | 994x1482 |
| Kernels | fp16 | 21.5 | 738 MiB/956 MiB | – | 994x1482 | |
| NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08 | Portable | fp16 | 15.4 | 662 MiB/736 MiB | – | 994x1482 |
| Kernels | fp16 | 9.33 | 709 MiB/794 MiB | – | 994x1482 | |
CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.
Licence and weights
Commercial use allowed
- Code
- MIT LICENSE
- Weights
- MIT (assumed from code)(not stated by the authors; assumed to follow the code) source
- Training data
-
- Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa) research-only, non-commercial
- weights licence not stated by the original authors; assumed to be the code licence
Our reading of the licences, not legal advice.
Weights
Fast-ACVNet/generalization.ckpt 12.6 MB· hosted on Google Drivefastforks downloads and checks these files on first use (fastforks download fast-acvnet-generalization).
Accuracy
The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.
ETH3D two-view train
bad > 1 px % / EPE
| Run | dtype | bad px % ↓ | EPE ↓ | Δ bad px % vs fp32 original | pairs | ms / pair |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 6.50 | 0.775 | reference | 27 | 29.2 |
| Original fp16 | fp16 | 6.51 | 0.778 | −0.01 | 27 | 34.6 |
| Portable fp32 | fp32 | 6.50 | 0.776 | ±0.00 | 27 | 25.0 |
| Portable | fp16 | 6.50 | 0.776 | ±0.00 | 27 | 18.6 |
| Kernels | fp16 | 6.51 | 0.776 | −0.01 | 27 | 13.3 |
| Original .half() | fp16 | fails: RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same | ||||
Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16, original .half(): RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-05; portable fp32, portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.
Middlebury 2014 trainingH
bad > 2 px % / EPE
| Run | dtype | bad px % ↓ | EPE ↓ | Δ bad px % vs fp32 original | pairs | ms / pair |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 18.57 | 4.203 | reference | 15 | 90.2 |
| Original fp16 | fp16 | 18.62 | 4.216 | −0.05 | 15 | 75.8 |
| Portable fp32 | fp32 | 18.57 | 4.202 | ±0.00 | 15 | 90.5 |
| Portable | fp16 | 18.61 | 4.224 | −0.04 | 15 | 59.5 |
| Kernels | fp16 | 18.60 | 4.220 | −0.03 | 15 | 35.7 |
| Original .half() | fp16 | fails: RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same | ||||
Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16, original .half(): RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-05; portable fp32, portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.
Closeness to the original
Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 px and p99 <= 0.25 px.
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.002 px | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.02 px | 0.00092 px |
| GPU default dtypes, p99 |diff| | ≤ 0.25 px | 0.00938 px |
Recorded for fast-acvnet-generalization on RTX 5060 Ti, torch 2.14.1+cu130, 2026-10-31, with these kernels: cudnn_hacccutlass_convigev
Details
Fast-ACVNet (Xu et al. 2022): MobileNetV2 features; attention weights from a correlation volume (1/8 group-wise, propagated to 1/4; Fast-ACVNet+: 1/4) select the top-24 disparities of a concatenation volume, then a 3-D hourglass, top-2 soft-argmax and superpixel upsampling (max disparity 192).
- Input / output
- Input: a rectified left / right image pair. Output: a disparity map (pixels, left view).
- Input size
- any H x W >= 32, padded to a multiple of 32; tested up to 3000x4000 (5,077 MiB allocated)
- Options
cuda_graphs=true
- Outputs
disparity- Checkpoint
plus=false- CUDA kernels
igevIGEV stereo: convolution epilogues, correlation lookups, GRU gatescutlass_convCUTLASS convolutions with the recurrent update (GRU) fused incudnn_haccconvolutions with fp16 accumulation (cuDNN)
- Apple (MPS)
- fp32, fp16
Limitations
- Kernels at 994 x 1482 reserve 960 MB vs 756 MB portable (peak allocated 664 vs 651 MB; original 1441 MB), 34 % faster; the difference is cached allocator blocks of the CUDA graph / cuDNN workspaces.
Full reference, generated from the repository
A stereo network by Gangwei Xu et al. that uses attention-selected disparities and a MobileNetV2 backbone, in a base and a plus version.
Commercial use: yes.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
fast-acvnet-generalization | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 101 60.6 54.7 38.1 36.4 36.1 | 1784 / 1878 1421 / 1506 653 / 756 654 / 716 549 / 678 573 / 624 |
fast-acvnet-sceneflow | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 101 60.6 54.7 38.1 36.2 36.7 | 1784 / 1878 1421 / 1506 653 / 756 654 / 716 584 / 698 563 / 576 |
fast-acvnet-kitti12 | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 101 60.6 54.7 38.1 36.2 36.2 | 1784 / 1878 1421 / 1506 653 / 756 654 / 716 549 / 678 573 / 624 |
fast-acvnet-kitti15 | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 101 60.6 54.7 38.1 36.5 36.2 | 1784 / 1878 1421 / 1506 653 / 756 654 / 716 549 / 678 573 / 624 |
fast-acvnet-plus-generalization | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 138 80.9 70.5 46.5 42.0 40.8 | 1861 / 1978 1441 / 1522 651 / 756 651 / 716 549 / 830 1127 / 1178 |
fast-acvnet-plus-sceneflow | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 138 80.9 70.5 46.5 41.9 40.9 | 1861 / 1978 1441 / 1522 651 / 756 651 / 716 549 / 830 1127 / 1178 |
fast-acvnet-plus-kitti12 | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 138 80.9 70.6 46.5 42.1 40.6 | 1861 / 1978 1441 / 1522 651 / 756 651 / 716 549 / 830 1127 / 1178 |
fast-acvnet-plus-kitti15 | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 138 80.9 70.5 46.5 42.1 40.7 | 1861 / 1978 1441 / 1522 651 / 756 651 / 716 549 / 830 1127 / 1178 |
Details
- Original code: https://github.com/gangweix/Fast-ACVNet at
61d5375e1816 - Code licence: MIT (permissive); files:
LICENSE - Weights licence: MIT (assumed from code)
- Licence evidence: https://github.com/gangweix/Fast-ACVNet/blob/61d5375e1816f124f4417cea506eeaddac1504e4/LICENSE.md
- Training data: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa)
fast-acvnet-kitti12: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2012)fast-acvnet-kitti15: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2015)fast-acvnet-plus-kitti12: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2012)fast-acvnet-plus-kitti15: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015 (SceneFlow, then fine-tuned on KITTI 2015)
- Options:
cuda_graphs= True (speed only) - Kernel libraries:
igev,cutlass_conv,cudnn_hacc - Variants:
fast-acvnet-generalization,fast-acvnet-sceneflow,fast-acvnet-kitti12,fast-acvnet-kitti15,fast-acvnet-plus-generalization,fast-acvnet-plus-sceneflow,fast-acvnet-plus-kitti12,fast-acvnet-plus-kitti15 - Outputs: disparity
- Input size: any H x W >= 32, padded to a multiple of 32
- Tested up to (portable, fp16, RTX 5060 Ti 16 GB):
fast-acvnet-generalization3000x4000 (5.0 GiB peak allocated, 7.9 reserved);fast-acvnet-sceneflow3000x4000 (5.0 GiB peak allocated, 7.9 reserved);fast-acvnet-kitti123000x4000 (5.0 GiB peak allocated, 7.9 reserved);fast-acvnet-kitti153000x4000 (5.0 GiB peak allocated, 7.9 reserved);fast-acvnet-plus-generalization3000x4000 (4.9 GiB peak allocated, 7.9 reserved);fast-acvnet-plus-sceneflow3000x4000 (4.9 GiB peak allocated, 7.9 reserved);fast-acvnet-plus-kitti123000x4000 (4.9 GiB peak allocated, 7.9 reserved);fast-acvnet-plus-kitti153000x4000 (4.9 GiB peak allocated, 7.9 reserved) fast-acvnet-generalization994x1482, first call: portable + torch.compile 42 s (cold Inductor cache), TensorRT fp16 68 s (ONNX + engine).fast-acvnet-sceneflow994x1482, first call: portable + torch.compile 42 s (cold Inductor cache), TensorRT fp16 71 s (ONNX + engine).fast-acvnet-kitti12994x1482, first call: portable + torch.compile 42 s (cold Inductor cache), TensorRT fp16 69 s (ONNX + engine).fast-acvnet-kitti15994x1482, first call: portable + torch.compile 42 s (cold Inductor cache), TensorRT fp16 70 s (ONNX + engine).fast-acvnet-plus-generalization994x1482, first call: portable + torch.compile 41 s (cold Inductor cache), TensorRT fp16 74 s (ONNX + engine).fast-acvnet-plus-sceneflow994x1482, first call: portable + torch.compile 41 s (cold Inductor cache), TensorRT fp16 73 s (ONNX + engine).fast-acvnet-plus-kitti12994x1482, first call: portable + torch.compile 42 s (cold Inductor cache), TensorRT fp16 72 s (ONNX + engine).fast-acvnet-plus-kitti15994x1482, first call: portable + torch.compile 40 s (cold Inductor cache), TensorRT fp16 74 s (ONNX + engine).fast-acvnet-generalizationoriginal 994x1482: torch.autocast(fp16) around the forwardfast-acvnet-generalizationportable 994x1482, kernels 994x1482: measured on fast-acvnet-sceneflow (same architecture, other checkpoint)fast-acvnet-scenefloworiginal (fp32) 994x1482: timed with fast-acvnet-generalization (same network, other weights)fast-acvnet-scenefloworiginal 994x1482: timed with fast-acvnet-generalization (same network, other weights); torch.autocast(fp16) around the forwardfast-acvnet-kitti12original (fp32) 994x1482: timed with fast-acvnet-generalization (same network, other weights)fast-acvnet-kitti12original 994x1482: timed with fast-acvnet-generalization (same network, other weights); torch.autocast(fp16) around the forwardfast-acvnet-kitti12portable 994x1482, kernels 994x1482: measured on fast-acvnet-sceneflow (same architecture, other checkpoint)fast-acvnet-kitti15original (fp32) 994x1482: timed with fast-acvnet-generalization (same network, other weights)fast-acvnet-kitti15original 994x1482: timed with fast-acvnet-generalization (same network, other weights); torch.autocast(fp16) around the forwardfast-acvnet-kitti15portable 994x1482, kernels 994x1482: measured on fast-acvnet-sceneflow (same architecture, other checkpoint)fast-acvnet-plus-generalizationoriginal 994x1482: torch.autocast(fp16) around the forwardfast-acvnet-plus-generalizationportable 994x1482, kernels 994x1482: measured on fast-acvnet-plus-sceneflow (same architecture, other checkpoint)fast-acvnet-plus-scenefloworiginal (fp32) 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights)fast-acvnet-plus-scenefloworiginal 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights); torch.autocast(fp16) around the forwardfast-acvnet-plus-kitti12original (fp32) 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights)fast-acvnet-plus-kitti12original 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights); torch.autocast(fp16) around the forwardfast-acvnet-plus-kitti12portable 994x1482, kernels 994x1482: measured on fast-acvnet-plus-sceneflow (same architecture, other checkpoint)fast-acvnet-plus-kitti15original (fp32) 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights)fast-acvnet-plus-kitti15original 994x1482: timed with fast-acvnet-plus-generalization (same network, other weights); torch.autocast(fp16) around the forwardfast-acvnet-plus-kitti15portable 994x1482, kernels 994x1482: measured on fast-acvnet-plus-sceneflow (same architecture, other checkpoint)fast-acvnet-generalizationload / first call: original 0.24 s / 1.732 s (375x1242), portable 0.27 s / 0.787 s (375x1242), kernels 0.59 s / 0.625 s (375x1242)fast-acvnet-sceneflowload / first call: original 0.24 s / 1.732 s (375x1242), portable 0.27 s / 0.76 s (375x1242), kernels 0.57 s / 0.627 s (375x1242)fast-acvnet-kitti12load / first call: original 0.24 s / 1.732 s (375x1242), portable 0.27 s / 0.756 s (375x1242), kernels 0.68 s / 0.619 s (375x1242)fast-acvnet-kitti15load / first call: original 0.24 s / 1.732 s (375x1242), portable 0.18 s / 0.832 s (375x1242), kernels 0.57 s / 0.618 s (375x1242)fast-acvnet-plus-generalizationload / first call: original 0.23 s / 0.037 s (375x1242), portable 0.27 s / 0.709 s (375x1242), kernels 0.56 s / 0.554 s (375x1242)fast-acvnet-plus-sceneflowload / first call: original 0.23 s / 0.037 s (375x1242), portable 0.27 s / 0.712 s (375x1242), kernels 0.57 s / 0.553 s (375x1242)fast-acvnet-plus-kitti12load / first call: original 0.23 s / 0.037 s (375x1242), portable 0.27 s / 0.698 s (375x1242), kernels 0.57 s / 0.553 s (375x1242)fast-acvnet-plus-kitti15load / first call: original 0.23 s / 0.037 s (375x1242), portable 0.27 s / 0.692 s (375x1242), kernels 0.66 s / 0.553 s (375x1242)- Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5
Measured
Original = the original code named under Details, same checkpoint and protocol.
| variant | dataset | original fp32 | original fp16 | original half | portable | portable fp32 | kernels | note |
|---|---|---|---|---|---|---|---|---|
fast-acvnet-generalization | ETH3D two-view train, bad > 1 px % / EPE | 6.50 / 0.775 | 6.51 / 0.778 | fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same) | 6.50 / 0.776 | 6.50 / 0.776 | 6.51 / 0.776 | |
fast-acvnet-generalization | Middlebury 2014 trainingH, bad > 2 px % / EPE | 18.57 / 4.203 | 18.62 / 4.216 | fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same) | 18.61 / 4.224 | 18.57 / 4.202 | 18.60 / 4.220 | |
fast-acvnet-sceneflow | ETH3D two-view train, bad > 1 px % / EPE | 8.40 / 0.654 | 8.44 / 0.659 | fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same) | 8.40 / 0.657 | 8.41 / 0.655 | 8.43 / 0.657 | |
fast-acvnet-sceneflow | Middlebury 2014 trainingH, bad > 2 px % / EPE | 25.53 / 7.667 | 25.55 / 7.686 | fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same) | 25.53 / 7.676 | 25.53 / 7.670 | 25.55 / 7.690 | |
fast-acvnet-kitti12 | ETH3D two-view train, bad > 1 px % / EPE | 19.07 / 0.825 | 19.06 / 0.825 | fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same) | 19.06 / 0.826 | 19.06 / 0.825 | 19.06 / 0.825 | |
fast-acvnet-kitti12 | Middlebury 2014 trainingH, bad > 2 px % / EPE | 28.20 / 6.892 | 28.23 / 6.888 | fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same) | 28.21 / 6.880 | 28.21 / 6.887 | 28.21 / 6.886 | |
fast-acvnet-kitti15 | ETH3D two-view train, bad > 1 px % / EPE | 17.58 / 0.774 | 17.58 / 0.772 | fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same) | 17.61 / 0.772 | 17.58 / 0.773 | 17.59 / 0.772 | |
fast-acvnet-kitti15 | Middlebury 2014 trainingH, bad > 2 px % / EPE | 28.67 / 6.920 | 28.71 / 6.918 | fails (RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor) should be the same) | 28.67 / 6.918 | 28.68 / 6.919 | 28.69 / 6.928 | |
fast-acvnet-plus-generalization | ETH3D two-view train, bad > 1 px % / EPE | 6.19 / 0.638 | 6.19 / 0.640 | fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward)) | 6.23 / 0.640 | 6.20 / 0.640 | 6.20 / 0.637 | |
fast-acvnet-plus-generalization | Middlebury 2014 trainingH, bad > 2 px % / EPE | 18.68 / 4.225 | 18.70 / 4.227 | fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward)) | 18.69 / 4.229 | 18.68 / 4.227 | 18.69 / 4.229 | |
fast-acvnet-plus-sceneflow | ETH3D two-view train, bad > 1 px % / EPE | 12.64 / 3.063 | 12.61 / 3.061 | fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward)) | 12.61 / 3.059 | 12.64 / 3.071 | 12.62 / 3.065 | |
fast-acvnet-plus-sceneflow | Middlebury 2014 trainingH, bad > 2 px % / EPE | 33.29 / 9.716 | 33.28 / 9.701 | fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward)) | 33.27 / 9.689 | 33.30 / 9.720 | 33.29 / 9.705 | |
fast-acvnet-plus-kitti12 | ETH3D two-view train, bad > 1 px % / EPE | 21.63 / 0.821 | 21.64 / 0.821 | fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward)) | 21.61 / 0.821 | 21.64 / 0.821 | 21.62 / 0.821 | |
fast-acvnet-plus-kitti12 | Middlebury 2014 trainingH, bad > 2 px % / EPE | 29.87 / 7.098 | 29.88 / 7.104 | fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward)) | 29.87 / 7.103 | 29.88 / 7.097 | 29.87 / 7.105 | |
fast-acvnet-plus-kitti15 | ETH3D two-view train, bad > 1 px % / EPE | 21.62 / 0.822 | 21.63 / 0.822 | fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward)) | 21.61 / 0.822 | 21.63 / 0.822 | 21.63 / 0.822 | |
fast-acvnet-plus-kitti15 | Middlebury 2014 trainingH, bad > 2 px % / EPE | 28.09 / 6.656 | 28.10 / 6.664 | fails (RuntimeError: Expected tensor for argument #1 'input' to have the same type as tensor for argument #2 'grid'; but type torch.cuda.HalfTensor does not equal torch.cuda.FloatTensor (while checking arguments for cudnn_grid_sampler_forward)) | 28.09 / 6.663 | 28.10 / 6.657 | 28.09 / 6.662 |
Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.
Parity
CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 px and p99 <= 0.25 px.
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
fast-acvnet-generalization | 0.00092 / 0.00938 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
fast-acvnet-sceneflow | 0.00216 / 0.03439 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
fast-acvnet-kitti12 | 0.00112 / 0.00401 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
fast-acvnet-kitti15 | 0.00147 / 0.00436 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
fast-acvnet-plus-generalization | 0.00114 / 0.01085 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
fast-acvnet-plus-sceneflow | 0.01838 / 0.06381 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
fast-acvnet-plus-kitti12 | 0.00068 / 0.00258 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
fast-acvnet-plus-kitti15 | 0.00125 / 0.01454 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, igev@f075827d68c6:sm_120 |
Limitations
- Kernels at 994 x 1482 reserve 960 MB vs 756 MB portable (peak allocated 664 vs 651 MB; original 1441 MB), 34 % faster; the difference is cached allocator blocks of the CUDA graph / cuDNN workspaces.
Example output
One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).
Input: Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), Scharstein et al., Middlebury College (research dataset, no licence stated).