Introduce SINQ calibration-free quantization algorithm by namgyu-youn · Pull Request #3156 · pytorch/ao

namgyu-youn · 2025-10-11T16:30:57Z

Summary:
Introduce SINQ: Sinkhorn-Normalized Quantization for calibration-free weight quantization (https://arxiv.org/abs/2509.22944).

SINQ uses dual-axis scaling (row + column) vs. HQQ's single-axis approach, achieving 1) 2-3x faster quantization time and 2) better imbalance handling.

(TL;DR) What is SINQ?

Quantized Parameterization

Single-scale (Scales + Shifts)

In normally, weight-only quantization algorithms defined as

$\hat{W} = \vec{s} \odot(Q +\vec{z})$

where w_hat is N × M matrix, \vec{s} is a N × 1 scale factor, Q is a quantized N × M matrix, and \vec{z} is a shift.

Dual-Scacles (SINQ)

Unlike above scales+shift approach, SINQ supply two vectors,

$\hat{W} = \vec{s} \odot Q \odot \vec{t}$

where \vec{s} is a N × 1 vector, \vec{t} is a 1 × M vector and the rest is as above.

2-axis scale factor efficiently collects spatial outlier distribution.

If we don't mind the potential additonal overhead, Q can be updated to Q+\vec{z} with shifting

Representation Space (Matrix Imbalance)

Matrix imbalance (i.e., outlier) is inconvenient to optimize with gradient descent, because of sparse gradients interrupt maximum and minimum operations. HIGGS used rotations (hadamard transform to normalize weight distribution), and AWQ/SmoothQuant used channel-wise scaling to minimizing errors by outliers.

SINQ uses sinkhorn iteration to normalize both row/column std (Algorithm 1).

SINQ: Algorithm 1

Iteratively normalize the standard deviation of the rows and columns of the matrix (weight) to be quantized. Then apply a standard quantization method (e.g., RTN)

Test/Future Plan:
The commented test shows quantization quality and speed comparison with HQQ. Full TorchAO integration with IntxWeightOnlyConfig and Int8DynamicActivationIntxWeightConfig is planned.

Related Issue/PR: #3106

pytorch-bot · 2025-10-11T16:31:00Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3156

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

namgyu-youn · 2025-10-11T16:35:01Z

Small PERF for HQQ vs. SINQ

import time

import torch

from torchao.quantization.quant_primitives import (
    _choose_qparams_and_quantize_affine_hqq,
    _choose_qparams_and_quantize_affine_sinq,
)


def compute_imbalance(W):
    """Compute matrix imbalance as defined in SINQ paper (Eq. 4)"""
    q_max_row = W.std(dim=1).max()
    q_max_col = W.std(dim=0).max()
    q_min_row = W.std(dim=1).min()
    q_min_col = W.std(dim=0).min()

    q_max = max(q_max_row, q_max_col)
    q_min = min(q_min_row, q_min_col)

    imbalance = q_max / max(q_min, 1e-8)
    return imbalance.item()


def compute_metrics(W_orig, W_dq):
    """Compute reconstruction metrics"""
    return {
        "mse": torch.mean((W_orig - W_dq) ** 2).item(),
        "mae": torch.mean(torch.abs(W_orig - W_dq)).item(),
        "rel_error": (torch.norm(W_orig - W_dq) / torch.norm(W_orig)).item(),
    }


def test_quantization_methods(
    shape: tuple[int, int] = (4096, 11008),
    nbits: int = 4,
    group_size: int = 128,
    device: str = "cpu",
):
    """Test and compare HQQ vs SINQ quantization methods."""
    print(f"\nTesting quantization methods on {shape} matrix")
    print(f"Bits: {nbits}, Group size: {group_size}")
    print("=" * 80)

    # Generate test weight matrix
    torch.manual_seed(42)
    W_original = torch.randn(shape, dtype=torch.float32, device=device) * 0.02

    # Add outliers to simulate real LLM weights
    outlier_mask = torch.rand(shape, device=device) < 0.01
    W_original[outlier_mask] *= 5.0

    print("\n Original Matrix Statistics:")
    print(f"  Mean: {W_original.mean():.6f}")
    print(f"  Std: {W_original.std():.6f}")
    print(f"  Min: {W_original.min():.6f}, Max: {W_original.max():.6f}")
    print(f"  Imbalance: {compute_imbalance(W_original):.4f}")

    # ========================================================================
    # HQQ Quantization
    # ========================================================================
    print(f"\n{'─' * 80}")
    print("HQQ (Half-Quadratic Quantization)")
    print(f"{'─' * 80}")

    start_time = time.time()
    W_q_hqq, scale_hqq, zero_hqq, _ = _choose_qparams_and_quantize_affine_hqq(
        tensor=W_original,
        nbits=nbits,
        group_size=group_size,
        optimize=True,
        axis=1,
        device=device,
        raw_output=False,
    )
    hqq_time = time.time() - start_time

    # Dequantize: W = (W_q - zero) * scale
    W_reshaped = W_q_hqq.float().reshape(-1, group_size)
    W_dq_hqq = ((W_reshaped - zero_hqq) * scale_hqq).reshape(shape)
    hqq_metrics = compute_metrics(W_original, W_dq_hqq)

    print(f"  Quantization Time: {hqq_time:.4f}s")
    print(f"  MSE: {hqq_metrics['mse']:.8f}")
    print(f"  MAE: {hqq_metrics['mae']:.8f}")
    print(f"  Relative Error: {hqq_metrics['rel_error']:.6f}")
    print(f"  Dequantized Imbalance: {compute_imbalance(W_dq_hqq):.4f}")

    # ========================================================================
    # SINQ Quantization
    # ========================================================================
    print(f"\n{'─' * 80}")
    print("SINQ (Sinkhorn-Normalized Quantization)")
    print(f"{'─' * 80}")

    start_time = time.time()
    W_q_sinq, scale_row_sinq, zero_sinq, scale_col_sinq, _ = (
        _choose_qparams_and_quantize_affine_sinq(
            tensor=W_original,
            nbits=nbits,
            group_size=group_size,
            device=device,
        )
    )
    sinq_time = time.time() - start_time

    # Dequantize: W = scale_row * (W_q - zero) * scale_col
    W_q_reshaped = W_q_sinq.float().reshape(-1, group_size)
    scale_row_flat = scale_row_sinq.view(-1, 1)  # (262144, 1)
    zero_flat = zero_sinq.view(-1, 1)  # (262144, 1)

    W_dq_sinq = scale_row_flat * (W_q_reshaped - zero_flat)
    W_dq_sinq = W_dq_sinq.reshape(shape) * scale_col_sinq.reshape(1, -1)

    sinq_metrics = compute_metrics(W_original, W_dq_sinq)

    print(f"  Quantization Time: {sinq_time:.4f}s")
    print(f"  MSE: {sinq_metrics['mse']:.8f}")
    print(f"  MAE: {sinq_metrics['mae']:.8f}")
    print(f"  Relative Error: {sinq_metrics['rel_error']:.6f}")
    print(f"  Dequantized Imbalance: {compute_imbalance(W_dq_sinq):.4f}")

    return {
        "hqq": {**hqq_metrics, "time": hqq_time},
        "sinq": {**sinq_metrics, "time": sinq_time},
    }


if __name__ == "__main__":
    test_quantization_methods()

Test result:

Testing quantization methods on (4096, 11008) matrix
Bits: 4, Group size: 128
================================================================================

 Original Matrix Statistics:
  Mean: 0.000003
  Std: 0.022279
  Min: -0.481529, Max: 0.434112
  Imbalance: 1.2571

────────────────────────────────────────────────────────────────────────────────
HQQ (Half-Quadratic Quantization)
────────────────────────────────────────────────────────────────────────────────
  Quantization Time: 7.2840s
  MSE: 0.00754850
  MAE: 0.07414244
  Relative Error: 3.928517
  Dequantized Imbalance: 2.3559

────────────────────────────────────────────────────────────────────────────────
SINQ (Sinkhorn-Normalized Quantization)
────────────────────────────────────────────────────────────────────────────────
  Quantization Time: 2.3842s
  MSE: 0.00000963
  MAE: 0.00247242
  Relative Error: 0.139391
  Dequantized Imbalance: 1.2619

namgyu-youn · 2025-10-30T12:49:10Z

@andrewor14 @metascroy Could you please check if this new quantization algorithm is needed for your teams?

namgyu-youn · 2025-11-03T14:49:07Z

CI failed to build, probably re-run to fix?

  + pip install .
  Processing /pytorch/ao
    Installing build dependencies ... 25l-� �\� �|� �done
  25h  Getting requirements to build wheel ... 25l-� �error
    error: subprocess-exited-with-error
    
    × Getting requirements to build wheel did not run successfully.
    │ exit code: 1
    ╰─> [15 lines of output]
        Traceback (most recent call last):
          File "/opt/conda/envs/venv/lib/python3.10/site-packages/pip/_vendor/pyproject_hooks/_in_process/_in_process.py", line 389, in <module>
            main()
          File "/opt/conda/envs/venv/lib/python3.10/site-packages/pip/_vendor/pyproject_hooks/_in_process/_in_process.py", line 373, in main
            json_out["return_val"] = hook(**hook_input["kwargs"])
          File "/opt/conda/envs/venv/lib/python3.10/site-packages/pip/_vendor/pyproject_hooks/_in_process/_in_process.py", line 143, in get_requires_for_build_wheel
            return hook(config_settings)
          File "/tmp/pip-build-env-chxekihp/overlay/lib/python3.10/site-packages/setuptools/build_meta.py", line 331, in get_requires_for_build_wheel
            return self._get_build_requires(config_settings, requirements=[])
          File "/tmp/pip-build-env-chxekihp/overlay/lib/python3.10/site-packages/setuptools/build_meta.py", line 301, in _get_build_requires
            self.run_setup()
          File "/tmp/pip-build-env-chxekihp/overlay/lib/python3.10/site-packages/setuptools/build_meta.py", line 317, in run_setup
            exec(code, locals())
          File "<string>", line 101, in <module>
        ModuleNotFoundError: No module named 'torch'
        [end of output]
    
    note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed to build 'file:///pytorch/ao' when getting requirements to build wheel
      main()
    File "/home/ec2-user/actions-runner/_work/ao/ao/test-infra/.github/scripts/run_with_env_secrets.py", line 98, in main
      run_cmd_or_die(f"docker exec -t {container_name} /exec")
    File "/home/ec2-user/actions-runner/_work/ao/ao/test-infra/.github/scripts/run_with_env_secrets.py", line 39, in run_cmd_or_die
      raise RuntimeError(f"Command {cmd} failed with exit code {exit_code}")
  RuntimeError: Command docker exec -t af025e4a848b71038da4722d85d198a1d3c8ab42a5d3e6acd3175647c6e718b3 /exec failed with exit code 1
  25h
  Error: Process completed with exit code 1.

namgyu-youn · 2025-11-16T14:49:16Z

@pytorchbot label "topic: new feature"

metascroy · 2025-11-18T05:44:35Z

@andrewor14 @metascroy Could you please check if this new quantization algorithm is needed for your teams?

Thanks @namgyu-youn! Does this algorithm support scale-only quantization, or only scale + zero point? Do you have any data on LLM benchmarks from lm-eval using this algorithm?

namgyu-youn · 2025-11-18T06:41:24Z

@metascroy

Does this algorithm support scale-only quantization, or only scale + zero point?

SINQ both supports symmetric(2-scale only; equation.2) and asymmetric (2-scale + zero point; equation.3). Figure 4 show symmetric losses of small accuracy but efficient for dequantization ops. In my opinion, we can go with symmetric first because it is much easier to implement and less overhead.

Do you have any data on LLM benchmarks from lm-eval using this algorithm?

Before the start, for api (Int4WeightOnlyConfig(version=2)) side, we can add to following, similar to HQQ:

ao/torchao/quantization/quant_api.py

Lines 1153 to 1160 in 1b80794

    
           if int4_choose_qparams_algorithm == Int4ChooseQParamsAlgorithm.HQQ: 
        
               assert int4_packing_format in [ 
        
                   Int4PackingFormat.TILE_PACKED_TO_4D, 
        
                   Int4PackingFormat.OPAQUE, 
        
               ], ( 
        
                   f"Int4ChooseQParamsAlgorithm.HQQ is not supported by packing format {int4_packing_format}, " 
        
                   f"it's only supported by Int4PackingFormat.TILE_PACKED_TO_4D and Int4PackingFormat.OPAQUE currently" 
        
               )

The benchmark (lm-eval) side, TorchAO have multiple implementations, and integration is in progress (#3289) .

.github/scripts/torchao_model_releases/quantize_and_upload.py: lm-eval

This implementation uses wrapper (TransformerEvalWrapper) to run benchmark in lm-eval. Because we have already implemented AWQ+HQQ (#3106 (comment)), we can add options for SINQ I think.

torchao/_models/llama/generate.py: not lm-eval (optional)

ao/torchao/_models/llama/generate.py

Lines 433 to 436 in 1b80794

    
           quantize_( 
        
               model, 
        
               Int4WeightOnlyConfig(group_size=group_size, use_hqq=use_hqq, version=1), 
        
           )

This implementation is not using lm-eval, but testing HQQ using version 1. We can add SINQ with version 1 if needed.

metascroy · 2025-11-18T18:25:11Z

@metascroy

Does this algorithm support scale-only quantization, or only scale + zero point?

SINQ both supports symmetric(2-scale only; equation.2) and asymmetric (2-scale + zero point; equation.3). Figure 4 show symmetric losses of small accuracy but efficient for dequantization ops. In my opinion, we can go with symmetric first because it is much easier to implement and less overhead.

Do you have any data on LLM benchmarks from lm-eval using this algorithm?

Before the start, for api (Int4WeightOnlyConfig(version=2)) side, we can add to following, similar to HQQ:

ao/torchao/quantization/quant_api.py

Lines 1153 to 1160 in 1b80794

if int4_choose_qparams_algorithm == Int4ChooseQParamsAlgorithm.HQQ:

assert int4_packing_format in [

Int4PackingFormat.TILE_PACKED_TO_4D,

Int4PackingFormat.OPAQUE,

], (

f"Int4ChooseQParamsAlgorithm.HQQ is not supported by packing format {int4_packing_format}, "

f"it's only supported by Int4PackingFormat.TILE_PACKED_TO_4D and Int4PackingFormat.OPAQUE currently"

)

The benchmark (lm-eval) side, TorchAO have multiple implementations, and integration is in progress (#3289) .

.github/scripts/torchao_model_releases/quantize_and_upload.py: lm-eval

This implementation uses wrapper (TransformerEvalWrapper) to run benchmark in lm-eval. Because we have already implemented AWQ+HQQ (#3106 (comment)), we can add options for SINQ I think.

torchao/_models/llama/generate.py: not lm-eval (optional)

ao/torchao/_models/llama/generate.py

Lines 433 to 436 in 1b80794

quantize_(

model,

Int4WeightOnlyConfig(group_size=group_size, use_hqq=use_hqq, version=1),

)

This implementation is not using lm-eval, but testing HQQ using version 1. We can add SINQ with version 1 if needed.

Great! When you do integrations, if you consider using adding support to intx_choose_qparams_algorithm (see 01849b2 for example of adding HQQ there), we could potentially use this algorithm in ExecuTorch as a default, depending on how model level benchmarks look.

For this PR, can you add some unit test for your change here: https://github.com/pytorch/ao/blob/main/test/quantization/test_quant_primitives.py

namgyu-youn · 2025-11-19T03:50:56Z

Great! When you do integrations, if you consider using adding support to intx_choose_qparams_algorithm (see 01849b2 for example of adding HQQ there), we could potentially use this algorithm in ExecuTorch as a default, depending on how model level benchmarks look.

Int4ChooseQParamsAlgorithm was in my mind, but IntxChooseQParamsAlgorithm looks better I feel. I will move into it; hope to introduce this algorithm to ExecuTorch!

For this PR, can you add some unit test for your change here: https://github.com/pytorch/ao/blob/main/test/quantization/test_quant_primitives.py

Added unit test, could you look into it? btw, it seems there is only E2E test for HQQ (no unit test), probably small unit test needed?

metascroy · 2025-11-19T18:06:40Z

        self.assertEqual(new_scale5.shape, torch.Size([3, 2, 8]))
        self.assertEqual(new_scale5.unique(dim=-1).shape, torch.Size([3, 2, 2]))

+    @unittest.skipIf(not torch.cuda.is_available(), "SINQ requires CUDA")


I guess this is fine, but does it require cuda? I didn't see anything cuda specific?

Updated device to CPU; no CUDA requirement for unit test I feel.

metascroy · 2025-11-19T18:07:50Z

+        reconstructed = (
+            qdata_reshaped * scale_row_expanded * scale_col_expanded
+        ).reshape(input.shape)
+        self.assertFalse(torch.isnan(reconstructed).any())


This will almost surely fail a lint check. Did you run ruff format and ruff check?

Fixed using pre-commit run; sorry I forgot pre-commit installation in laptop.

metascroy · 2025-11-19T18:28:06Z

+        nbits: Number of quantization bits (default: 4)
+        group_size: Quantization group size (default: 64)
+        niter: Number of Sinkhorn iterations (default: 20)
+        compute_dtype: Target compute dtype (default: torch.float16)


It does look like compute_dtype is being used?

Updated to use, it is designed for scale factor dtype

metascroy · 2025-11-19T18:47:56Z

+    group_size: int = 64,
+    niter: int = 20,
+    compute_dtype: torch.dtype = torch.float16,
+    device: str = "cuda",


It does not look like device is used? Do we need device, or can we just assume whatever device tensor is on?

metascroy · 2025-11-19T18:48:08Z

+    Args:
+        tensor: Input weight tensor
+        nbits: Number of quantization bits (default: 4)
+        group_size: Quantization group size (default: 64)


Is group_size here for both row/col?

Yes group_size is applied to both. We can try distinct ranges, but it isn't addressed in the paper I remember.

metascroy · 2025-11-19T18:49:54Z

Other than comments I left, I think it looks good. Let's see what CI says

metascroy · 2025-11-19T18:50:34Z

Great! When you do integrations, if you consider using adding support to intx_choose_qparams_algorithm (see 01849b2 for example of adding HQQ there), we could potentially use this algorithm in ExecuTorch as a default, depending on how model level benchmarks look.

Int4ChooseQParamsAlgorithm was in my mind, but IntxChooseQParamsAlgorithm looks better I feel. I will move into it; hope to introduce this algorithm to ExecuTorch!

For this PR, can you add some unit test for your change here: https://github.com/pytorch/ao/blob/main/test/quantization/test_quant_primitives.py

Added unit test, could you look into it? btw, it seems there is only E2E test for HQQ (no unit test), probably small unit test needed?

Yeah, it looks like HQQ doesn't have a unit test, only E2E test.

metascroy · 2025-11-19T20:27:36Z

@namgyu-youn it looks like the new unit test you added is failing CI:

=========================== short test summary info ============================
FAILED test/quantization/test_quant_primitives.py::TestQuantPrimitives::test_choose_qparams_and_quantize_scale_only_sinq - AssertionError: torch.float16 != torch.int8
==== 1 failed, 635 passed, 6797 skipped, 266 warnings in 594.01s (0:09:54) =====
    main()
  File "/home/ec2-user/actions-runner/_work/ao/ao/test-infra/.github/scripts/run_with_env_secrets.py", line 98, in main
    run_cmd_or_die(f"docker exec -t {container_name} /exec")
  File "/home/ec2-user/actions-runner/_work/ao/ao/test-infra/.github/scripts/run_with_env_secrets.py", line 39, in run_cmd_or_die
    raise RuntimeError(f"Command {cmd} failed with exit code {exit_code}")
RuntimeError: Command docker exec -t 92eda2ebb7097ce0dd8ae2ac96db659ec5144f4b3e338ff4f4475c926addc842 /exec failed with exit code 1

namgyu-youn · 2025-11-22T14:01:55Z

@namgyu-youn it looks like the new unit test you added is failing CI:

=========================== short test summary info ============================
FAILED test/quantization/test_quant_primitives.py::TestQuantPrimitives::test_choose_qparams_and_quantize_scale_only_sinq - AssertionError: torch.float16 != torch.int8
==== 1 failed, 635 passed, 6797 skipped, 266 warnings in 594.01s (0:09:54) =====
    main()
  File "/home/ec2-user/actions-runner/_work/ao/ao/test-infra/.github/scripts/run_with_env_secrets.py", line 98, in main
    run_cmd_or_die(f"docker exec -t {container_name} /exec")
  File "/home/ec2-user/actions-runner/_work/ao/ao/test-infra/.github/scripts/run_with_env_secrets.py", line 39, in run_cmd_or_die
    raise RuntimeError(f"Command {cmd} failed with exit code {exit_code}")
RuntimeError: Command docker exec -t 92eda2ebb7097ce0dd8ae2ac96db659ec5144f4b3e338ff4f4475c926addc842 /exec failed with exit code 1

Fixed; please take a look at it.

metascroy · 2025-11-24T16:59:51Z

Great! When you do integrations, if you consider using adding support to intx_choose_qparams_algorithm (see 01849b2 for example of adding HQQ there), we could potentially use this algorithm in ExecuTorch as a default, depending on how model level benchmarks look.

Int4ChooseQParamsAlgorithm was in my mind, but IntxChooseQParamsAlgorithm looks better I feel. I will move into it; hope to introduce this algorithm to ExecuTorch!

For this PR, can you add some unit test for your change here: https://github.com/pytorch/ao/blob/main/test/quantization/test_quant_primitives.py

Added unit test, could you look into it? btw, it seems there is only E2E test for HQQ (no unit test), probably small unit test needed?

Yeah, it looks like HQQ doesn't have a unit test, only E2E test.

Re-running CI

metascroy · 2025-11-25T18:16:42Z

LGTM! Looking forward to the follow up PR for E2E workflow integration :)

I think the CI failures are not related.

* feat: SINQ quantization algorithm * update sinq algorithm * update sinq ops and add unit test * update device to cpu in SINQ test * fix scale dtype, device * update device to direct override * add qmin, qmax args similar to HQQ

meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Oct 11, 2025

namgyu-youn changed the title ~~Introduce SINQ quantization algorithm~~ Introduce SINQ weight-only quantization algorithm Oct 17, 2025

namgyu-youn changed the title ~~Introduce SINQ weight-only quantization algorithm~~ Introduce SINQ quantization algorithm Oct 20, 2025

pytorch-bot Bot added the topic: new feature Use this tag if this PR adds a new feature label Nov 16, 2025

namgyu-youn changed the title ~~Introduce SINQ quantization algorithm~~ Introduce SINQ calibration-free quantization algorithm Nov 18, 2025

metascroy reviewed Nov 19, 2025

View reviewed changes

namgyu-youn requested a review from metascroy November 19, 2025 18:23

metascroy reviewed Nov 19, 2025

View reviewed changes

namgyu-youn requested a review from metascroy November 20, 2025 02:57

namgyu-youn added 6 commits November 22, 2025 22:57

feat: SINQ quantization algorithm

711bf76

update sinq algorithm

6d4c81c

update sinq ops and add unit test

4070892

update device to cpu in SINQ test

a08dc8b

fix scale dtype, device

b9147d9

update device to direct override

0d36e77

namgyu-youn force-pushed the int4-sinq branch from e0f9c78 to 0d36e77 Compare November 22, 2025 13:59

add qmin, qmax args similar to HQQ

b60fd17

metascroy approved these changes Nov 25, 2025

View reviewed changes

metascroy merged commit 1cfd3b5 into pytorch:main Nov 25, 2025
16 of 19 checks passed

namgyu-youn deleted the int4-sinq branch November 25, 2025 19:06

namgyu-youn mentioned this pull request Nov 30, 2025

Add unit test for HQQ #3403

Open

Conversation

namgyu-youn commented Oct 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Quantized Parameterization

Single-scale (Scales + Shifts)

Dual-Scacles (SINQ)

Representation Space (Matrix Imbalance)

SINQ: Algorithm 1

Uh oh!

pytorch-bot Bot commented Oct 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3156

Uh oh!

namgyu-youn commented Oct 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

namgyu-youn commented Oct 30, 2025

Uh oh!

namgyu-youn commented Nov 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

namgyu-youn commented Nov 16, 2025

Uh oh!

metascroy commented Nov 18, 2025

Uh oh!

namgyu-youn commented Nov 18, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

metascroy commented Nov 18, 2025

Uh oh!

namgyu-youn commented Nov 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

metascroy Nov 19, 2025

Choose a reason for hiding this comment

Uh oh!

namgyu-youn Nov 19, 2025

Choose a reason for hiding this comment

Uh oh!

metascroy Nov 19, 2025

Choose a reason for hiding this comment

Uh oh!

namgyu-youn Nov 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

metascroy Nov 19, 2025

Choose a reason for hiding this comment

Uh oh!

namgyu-youn Nov 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

metascroy Nov 19, 2025

Choose a reason for hiding this comment

Uh oh!

metascroy Nov 19, 2025

Choose a reason for hiding this comment

Uh oh!

namgyu-youn Nov 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

metascroy commented Nov 19, 2025

Uh oh!

metascroy commented Nov 19, 2025

Uh oh!

metascroy commented Nov 19, 2025

Uh oh!

namgyu-youn commented Nov 22, 2025

Uh oh!

metascroy commented Nov 24, 2025

Uh oh!

metascroy commented Nov 25, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

namgyu-youn commented Oct 11, 2025 •

edited

Loading

pytorch-bot Bot commented Oct 11, 2025 •

edited

Loading

namgyu-youn commented Oct 11, 2025 •

edited

Loading

namgyu-youn commented Nov 3, 2025 •

edited

Loading

namgyu-youn commented Nov 18, 2025 •

edited

Loading

namgyu-youn commented Nov 19, 2025 •

edited

Loading

namgyu-youn Nov 19, 2025 •

edited

Loading

namgyu-youn Nov 20, 2025 •

edited

Loading

namgyu-youn Nov 20, 2025 •

edited

Loading