Make SmoothQuant more General #2728

namgyu-youn · 2025-08-11T04:52:51Z

Summary

Add SmoothQuantConfig as a base config and SmoothQuantObserver as a smoothing factor computation. Apply corresponding changes in other parts for the SmoothQuant API flows.

Test Plan

Unittest and real run (example.py) using example.py with Llama-2-7b-chat-hf for both quantization and model saving

Future Plan

Build a benchmark within the vLLM ecosystem for AWQ and SmoothQuant. See #2815 for more info

Summary: - Added SmoothQuantConfig as a base config and made corresponding changes in other parts of the flow Test Plan: - Qwen 3-8B with example.py and unittest - Additional test plans requirerd ETC - Fix typo in README.md for SmoothQuant

pytorch-bot · 2025-08-11T04:52:55Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2728

📄 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 · 2025-08-15T17:44:09Z

@jerryzh168 Could you please look into this PR? It was inspired by #2659 (comment) for more generalized SmoothQuant API.

jerryzh168 · 2025-08-15T17:49:18Z

Thanks @namgyu-youn this is a step towards that but not fully general yet, it seems to be a quick change to add it though, commented inline.

also it seems smoothquant is not very popular at the moment: https://huggingface.co/models?search=smoothquant, so I'd like to wait a bit before we invest more effort to it, let me know if you are interested to contribute more to torchao, we have many more higher priority issues that you can help with I think

test/prototype/test_smoothquant.py

torchao/prototype/smoothquant/api.py

namgyu-youn · 2025-08-15T18:57:29Z

Thanks @namgyu-youn this is a step towards that but not fully general yet, it seems to be a quick change to add it though, commented inline.

also it seems smoothquant is not very popular at the moment: https://huggingface.co/models?search=smoothquant, so I'd like to wait a bit before we invest more effort to it, let me know if you are interested to contribute more to torchao, we have many more higher priority issues that you can help with I think

Thanks for the kind info, and I truly love your team's work after reviewing TorchAO: CodeML @ ICML 2025.

The recently updated contribution guide could be a great choice for the next contribution, but personally I prefer the sparsity (pruning) module more. Unfortunately, I heard the main POC (@jcaip) is on vacation, making it hard for me to progress. The following are my recent activities related to the sparsity module:

Since Wanda was already introduced, I recently introduced Wanda++ at feat: RGS for wanda++ #2537.
Computation overhead was missing in your team's workshop (not certain because of my lack of knowledge), and opened issue at Missing benchmark for sparse24_sm90_sparsify overhead #2612
Also interested in Activation compression Accelerate activation sparsity with activation compression #1920, but I have to learn more about it.

If there is no huge progress for the sparsity module, quantization (new APIs or primitive ops) might be a next step. Let me know if there is a good-second-issue about it.

p.s. Could you please check #2644 ? It hasn't merged yet after being approved (no CI broken). Also, #2660 has been waiting for review (I am fine to close this because it is low-priority).

namgyu-youn · 2025-08-16T18:35:40Z

Test result (test_smoothquant.py):

$ python test/prototype/test_smoothquant.py
..............................................
----------------------------------------------------------------------
Ran 46 tests in 15.208s

OK

namgyu-youn · 2025-08-16T18:35:58Z

@jerryzh168 Hi, I am happy to show you more generalized SmoothQuant API by using Quantization API (torchao/quantization/quant_api.py) at ba89d03. Could you review this PR?

jerryzh168 · 2025-08-18T22:50:54Z

test/prototype/test_smoothquant.py

-        "device", ["cpu"] + (["cuda"] if torch.cuda.is_available() else [])
+        "base_config",
+        [
+            int8_dynamic_activation_int8_weight(),


nit: this API is deprecated, use Int8DynamicActivationInt8WeightConfig instead

torchao/prototype/smoothquant/README.md

jerryzh168 · 2025-08-18T22:52:29Z

torchao/prototype/smoothquant/example.py

-        insert_smooth_quant_observer_(model, alpha, quant_mode)
+        # Step 1: Insert observers to find average magnitude and calculate scales
+        config = SmoothQuantConfig(
+            base_config=int8_dynamic_activation_int8_weight(),


can generalize the example API to take quant type configs now, see

ao/torchao/prototype/awq/example.py

Line 307 in 751d7f6

help="Quantization method. Options are either awq-int4wo-<group_size>, or int4wo-<group_size>.",

Thanks, but how about using Int8DynamicActivationInt8WeightConfig as a default in here and devide PR? It might require checking which APIs are compatiable with SmoothQuantConfig, and building unittest.

Even more, we can uniform commonly used utils functions in AWQ and SmoothQuant: get_calib_dataset, wiki2_eval, and quantize_and_eval.

torchao/prototype/smoothquant/api.py

torchao/prototype/smoothquant/core.py

jerryzh168 · 2025-08-19T18:08:14Z

torchao/prototype/smoothquant/example.py

+    print(f"time for convert: {time.time() - t0:.02f} seconds")
+
+    # Set up config for loading
+    quant_config.step = SmoothQuantStep.PREPARE_FOR_LOADING


does this work? you can check if it works by the following:

export MODEL=YOUR_SAVED_SMOOTHQUANT_MODEL lm_eval --model hf --model_args pretrained=$MODEL --tasks $TASK --device cuda:0 --batch_size auto --limit 50 # vllm export MODEL=YOUR_SAVED_SMOOTHQUANT_MODEL python benchmarks/benchmark_latency.py --input-len 256 --output-len 256 --model $MODEL

Hoped so because it works similarly to AWQ, but just tested it with the following code for assurance and got the log message:

import tempfile import torch from transformers import AutoModelForCausalLM, AutoTokenizer from torchao.prototype.smoothquant import SmoothQuantConfig from torchao.prototype.smoothquant.core import SmoothQuantStep from torchao.prototype.smoothquant.example import quantize_and_eval from torchao.quantization import quantize_ from torchao.quantization.quant_api import Int8DynamicActivationInt8WeightConfig MODEL_NAME = "microsoft/DialoGPT-small" # Step 1: Create quantized model with tempfile.NamedTemporaryFile(suffix='.pt', delete=False) as f: model_path = f.name quantize_and_eval(MODEL_NAME, 0.5, ['PPL'], 256, 5, 'cuda', torch.float32, False, model_path, None) # Step 2: Test PREPARE_FOR_LOADING tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME) tokenizer.pad_token = tokenizer.eos_token model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch.float32).cuda() quantize_(model, SmoothQuantConfig( base_config=Int8DynamicActivationInt8WeightConfig(), step=SmoothQuantStep.PREPARE_FOR_LOADING, alpha=0.5, )) # Test inference test_input = tokenizer('Hello world', return_tensors='pt').to('cuda') with torch.no_grad(): output = model(**test_input) generated = model.generate(**test_input, max_length=20, do_sample=False) print(f"✓ Inference: {output.logits.shape}") print(f"✓ Generation: {tokenizer.decode(generated[0], skip_special_tokens=True)}")

Loading model on cuda... Time to load model: 1.86 seconds running SmoothQuant prepare and calibrate Repo card metadata block was not found. Setting CardData to empty. Token indices sequence length is longer than the specified maximum sequence length for this model (1443 > 1024). Running this sequence through the model will result in indexing errors time for prepare and calibration: 5.20 seconds running SmoothQuant convert time for convert: 0.04 seconds Saving model to /tmp/tmpqeme5s1r.pt `loss_type=None` was set in the config but it is unrecognised.Using the default loss: `ForCausalLMLoss`. Setting `pad_token_id` to `eos_token_id`:50256 for open-end generation. ✓ Inference: torch.Size([1, 4, 50257]) ✓ Generation: TorchAO TorchAO

For sure, we should benchmark them with your suggestion, but I want to carefully suggest dividing its PR.

OK sounds good to divide the PR

jerryzh168 · 2025-08-20T00:13:23Z

torchao/prototype/smoothquant/README.md

-insert_smooth_quant_observer_(model)
-load_smooth_quant_recipe(model, "./smooth_quant_recipe.json")
+
+# Step 3: Convert


I think ideally we can add a tutorial doc for how to save transformer models for vllm/lm-eval as well:

ao/torchao/prototype/awq/example.py

Lines 223 to 265 in 083361b

if quant.startswith("awq-int4wo"):

group_size = int(quant.split("-")[2])

print(f"running {quant} quantization with group size {group_size}")

# TODO: this is temporary, we'll be using Int4WeightOnlyConfig soon

from torchao.quantization import FbgemmConfig

# use_hqq = True

# base_config = Int4WeightOnlyConfig(group_size=group_size, use_hqq=use_hqq)

base_config = FbgemmConfig(

input_dtype=torch.bfloat16,

weight_dtype=torch.int4,

output_dtype=torch.bfloat16,

block_size=[1, group_size],

preshuffle=False,

)

print(f"running {quant} prepare and calibrate")

t0 = time.time()

quant_config = AWQConfig(base_config, step="prepare")

quantize_(

model,

quant_config,

)

from torchao._models._eval import TransformerEvalWrapper

TransformerEvalWrapper(

model=model.to(device),

tokenizer=tokenizer,

max_seq_length=max_seq_length,

device=device,

).run_eval(

tasks=tasks,

limit=calibration_limit,

)

print(f"time for prepare and calibration: {time.time() - t0:.02f} seconds")

print(f"running {quant} convert")

t0 = time.time()

quant_config = AWQConfig(base_config, step="convert")

quantize_(model, quant_config)

print(f"time for convert: {time.time() - t0:.02f} seconds")

quant_config = AWQConfig(base_config, step="prepare_for_loading")

model.config.quantization_config = TorchAoConfig(quant_config)

basically: prepare, convert, and then manually set the config step to "prepare_for_loading"
and then upload the model.

after this the model should be able to be used with vllm and lm-eval.

jerryzh168 · 2025-08-20T00:13:57Z

torchao/prototype/smoothquant/api.py

    if config.set_inductor_config:
        torchao.quantization.utils.recommended_inductor_config_setter()


this should be removed I think

please remove this before landing

Thanks for the reminder. I will remove it.

jerryzh168 · 2025-08-20T00:15:38Z

torchao/prototype/smoothquant/api.py


+    # Get quantization parameters
+    if all(x is not None for x in (config.smoothing_factor, config.wei_scales)):


when is this branch taken? how do people generate these parameters?

Yes we only use the smooth factor in here. wei_scales and act_scales should be totally removed.

jerryzh168 · 2025-08-20T00:17:56Z

torchao/prototype/smoothquant/core.py

        quant_min: Optional[int] = None,
        quant_max: Optional[int] = None,
        eps: Optional[float] = None,


can we remove these args as well, I think these are not needed

yes they can be removed with AffineQuantizedMinMaxObserver, thanks.

jerryzh168 · 2025-08-20T00:18:39Z

torchao/prototype/smoothquant/core.py

        self.act_ic_obs = AffineQuantizedMinMaxObserver(
-            MappingType.SYMMETRIC,
-            torch.int8,
-            PerAxis(-1),
-            eps=eps,
+            MappingType.SYMMETRIC, torch.int8, PerAxis(-1), eps=self.eps
        )
        self.wei_ic_obs = AffineQuantizedMinMaxObserver(
-            MappingType.SYMMETRIC,
-            torch.int8,
-            PerAxis(-1),
-            eps=eps,
+            MappingType.SYMMETRIC, torch.int8, PerAxis(-1), eps=self.eps
        )


are these hardcoded to int8? what if we want to apply them to other types of quantization?

I missed updating it after testing Int8DynamicActivationInt8WeightConfig. It should be fixed.

jerryzh168 · 2025-08-20T00:20:06Z

torchao/prototype/smoothquant/core.py

        wei_min_per_ic = self.wei_ic_obs.min_val
        wei_max_per_ic = self.wei_ic_obs.max_val
        act_min_per_ic = self.act_ic_obs.min_val


so we only need min_val/max_val right? might be easier to just do this ourselves instead of relying on AffineQuantizedMinMaxObserver? we can copy over some of the main logic to record min_val/max_val based on granularity of scale (PerAxis) as well

Oh sorry I misunderstood it. What we only need here are min_val/max_val same as AWQ. The workflow should be updated based on it.

jerryzh168 · 2025-08-20T00:20:56Z

torchao/prototype/smoothquant/example.py

-    print("Loading dataset")
-    t0 = time.time()
+# TODO: Uniform this with torchao/prototype/awq/example.py and expand more tasks
+def benchmark(model, tokenizer, max_seq_length=512, tasks=["PPL"], device="cuda"):


I was hoping to remove these in the future and just rely on vllm for performance evaluation and lm-eval for model quality evaluation

Filed an issue at #2815 for testing more quantization API and benchmarks. And these task sound good to me; let me engage it after this PR.

jerryzh168 · 2025-08-20T22:35:33Z

torchao/prototype/smoothquant/api.py

+    )
+    weight = observed_linear.weight * smoothing_factor
+
+    # Create new linear layer
    linear = torch.nn.Linear(


nit: one trick we can have when creating these linear weights is to create them in meta device: https://github.com/vllm-project/vllm/blob/c86af22f31838ee654c856279ac5110ae3fdb2cc/vllm/model_executor/layers/quantization/torchao.py#L159 to save memory I think, similar for awq

Thanks for teaching me the Meta device. LGTM for memory saving when transformations are done before loading the actual data.

jerryzh168 · 2025-08-20T22:37:01Z

torchao/prototype/smoothquant/core.py



 class SmoothQuantObserver(torch.nn.Module):
    def __init__(
        self,
        weight: torch.Tensor,
+        bias: Optional[torch.Tensor],
+        base_config: AOBaseConfig,


actually seems like base_config is not used here?

Since the observer only computes smoothing factor, base_config can be removed, thanks.

jerryzh168 · 2025-08-20T22:39:15Z

torchao/prototype/smoothquant/api.py

        set_inductor_config: if True, adjusts `torchinductor` settings to recommended values.
    """

+    base_config: AOBaseConfig
+    step: SmoothQuantStep
+    alpha: Optional[float] = 0.5
    smoothing_factor: Optional[torch.Tensor] = None


should this be removed as well? seems like not used?

Removing it looks good because it can be computed without initialization, thanks.

jerryzh168 · 2025-08-20T22:39:28Z

torchao/prototype/smoothquant/api.py

    smoothing_factor: Optional[torch.Tensor] = None
-    act_scales: Optional[torch.Tensor] = None
-    wei_scales: Optional[torch.Tensor] = None
    set_inductor_config: bool = True


also this flag, I don't think we need this

jerryzh168 · 2025-08-20T22:40:33Z

torchao/prototype/smoothquant/api.py

+    # Get quantization parameters
+    smoothing_factor = (
+        config.smoothing_factor
+        if config.smoothing_factor is not None


probably just remove this arg, I don't see when we'll use it

jerryzh168 · 2025-08-20T22:41:15Z

torchao/prototype/smoothquant/core.py

        self.obs = obs

    def forward(self, input: torch.Tensor):
        input = self.obs(input)
-        output = F.linear(input, self.weight, self.bias)
-        return output
+        return F.linear(input, self.weight, self.bias)

    @classmethod
    def from_float(cls, float_linear: torch.nn.Linear, obs: SmoothQuantObserver):
        observed_linear = cls(


also here, can use https://github.com/vllm-project/vllm/blob/c86af22f31838ee654c856279ac5110ae3fdb2cc/vllm/model_executor/layers/quantization/torchao.py#L159 to save memoery

Thanks for teaching me meta devices again!

jerryzh168 · 2025-08-21T01:11:18Z

torchao/prototype/smoothquant/README.md

@@ -68,29 +64,15 @@ Running the example with `torch.compile` on a NVIDIA A10G GPU.
 Perplexity
 | Quant Method | alpha=0.25 | alpha=0.5 | alpha=0.75 | alpha=None* |
 |-|-|-|-|-|
-| Dynamic | 8.1872 | 7.4257 | 7.2518 | 7.5509 |
-| Static | 43.8051 | 11.2984 | 7.5791 | 19.5050 |
+| SmoothQuant | - | - | - | - |


I think we'd need to run some initial testing to make sure the refactor works? otherwise we might be landing non-working code

This is the omitted section and will be updated after the refactor is finished. I will only add int8 dynamic in this PR, but am definitely interested in expanding to vLLM benchmarks with more quantization APIs.

do you plan to update this before this PR is merged or later? I feel we should update this before the PR can be merged to make sure this is working

performance benchmark can come later, but we should have accuracy test to make sure smoothquant implementation is correct I think

Sorry I didn't make it clear; it will be updated in this PR after refactoring is finished. Actually, I have been checking its performance using TinyLlama/TinyLlama-1.1B-Chat-v1.0 for each commit. The architecture is quite different, so I didn't mention them, although the accuracy (perplexity) was affordable. Following is one of the experiment results:

Similar to AWQ, it will be updated using the Llama-2-7b-chat-hf model. Please feel free to direct commit about it because my setup is quite different from your team's setup (1xA100 80GB SXM4 instance).

OK please request review when you are ready

namgyu-youn · 2025-08-23T06:58:07Z

torchao/prototype/smoothquant/README.md


-Note*: Conventional quantization without SmoothQuant
+Evaluation perplexity numbers were calculated using the script in `smoothquant/example.py`. For Llama-2-7b-chat-hf, performance benchmarks were calculated using the `torchao/_models/llama/generate.py` script and run on a 1xA100 48GB PCIe interface.


@jerryzh168 Since my setup (1xA100 48GB PCIe) is not the same as your team (1xA100 80GB SXM4 instance), the result can be quite different. Please feel free to update it if needed.

that's fine I feel, we are mainly interested in the perplexity changes with and without AWQ I think, could you show that? you can use https://github.com/pytorch/ao/blob/main/torchao/_models/llama/eval.py to get the perplexity I think, or try to add a non-awq version in example.py

Make SmoothQuant more General

c482371

Summary: - Added SmoothQuantConfig as a base config and made corresponding changes in other parts of the flow Test Plan: - Qwen 3-8B with example.py and unittest - Additional test plans requirerd ETC - Fix typo in README.md for SmoothQuant

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 Aug 11, 2025

refactor: use predefined ToyLinearModel

e16edc2

namgyu-youn marked this pull request as draft August 12, 2025 06:53

namgyu-youn added 2 commits August 12, 2025 16:51

fix incorrect parameters

5ec0dcf

add type hint for dataclass

2475ad1

namgyu-youn marked this pull request as ready for review August 12, 2025 08:01

Merge branch 'main' into refactor-smoothquant

ccb7b84

jerryzh168 approved these changes Aug 15, 2025

View reviewed changes

jerryzh168 reviewed Aug 15, 2025

View reviewed changes

test/prototype/test_smoothquant.py Outdated Show resolved Hide resolved

jerryzh168 reviewed Aug 15, 2025

View reviewed changes

torchao/prototype/smoothquant/api.py Outdated Show resolved Hide resolved

jerryzh168 self-requested a review August 15, 2025 17:53

namgyu-youn marked this pull request as draft August 16, 2025 15:27

use Quantization API for more generalized SmoothQuant API

ba89d03

namgyu-youn marked this pull request as ready for review August 16, 2025 18:31

jerryzh168 reviewed Aug 18, 2025

View reviewed changes

torchao/prototype/smoothquant/README.md Outdated Show resolved Hide resolved

jerryzh168 reviewed Aug 18, 2025

View reviewed changes

torchao/prototype/smoothquant/api.py Show resolved Hide resolved

namgyu-youn added 2 commits August 19, 2025 18:54

add PREPARE_FOR_LOADING mode for loading quantized weight

a6df6af

update example and doc for updated SmoothQuant API

0fc6539

namgyu-youn requested a review from jerryzh168 August 19, 2025 11:03

jerryzh168 reviewed Aug 19, 2025

View reviewed changes

torchao/prototype/smoothquant/core.py Outdated Show resolved Hide resolved

jerryzh168 reviewed Aug 19, 2025

View reviewed changes

jerryzh168 reviewed Aug 20, 2025

View reviewed changes

namgyu-youn mentioned this pull request Aug 20, 2025

Benchmark AWQ and SmoothQuant within vLLM ecosystem #2815

Open

remove overused/misunderstood parameters

46d5d31

namgyu-youn requested a review from jerryzh168 August 20, 2025 06:57

jerryzh168 reviewed Aug 20, 2025

View reviewed changes

jerryzh168 reviewed Aug 21, 2025

View reviewed changes

remove unused variable from SmoothQuant

fc8ae4d

namgyu-youn requested a review from jerryzh168 August 21, 2025 15:17

update SmoothQuant docs for user guide

4f7def9

namgyu-youn commented Aug 23, 2025

View reviewed changes

jerryzh168 mentioned this pull request Aug 26, 2025

Implement an AWQ algorithm with dynamic activation quantization for ExecuTorch #2388

Open

	if quant.startswith("awq-int4wo"):
	group_size = int(quant.split("-")[2])
	print(f"running {quant} quantization with group size {group_size}")
	# TODO: this is temporary, we'll be using Int4WeightOnlyConfig soon
	from torchao.quantization import FbgemmConfig

	# use_hqq = True
	# base_config = Int4WeightOnlyConfig(group_size=group_size, use_hqq=use_hqq)
	base_config = FbgemmConfig(
	input_dtype=torch.bfloat16,
	weight_dtype=torch.int4,
	output_dtype=torch.bfloat16,
	block_size=[1, group_size],
	preshuffle=False,
	)
	print(f"running {quant} prepare and calibrate")
	t0 = time.time()
	quant_config = AWQConfig(base_config, step="prepare")

	quantize_(
	model,
	quant_config,
	)
	from torchao._models._eval import TransformerEvalWrapper

	TransformerEvalWrapper(
	model=model.to(device),
	tokenizer=tokenizer,
	max_seq_length=max_seq_length,
	device=device,
	).run_eval(
	tasks=tasks,
	limit=calibration_limit,
	)

	print(f"time for prepare and calibration: {time.time() - t0:.02f} seconds")
	print(f"running {quant} convert")
	t0 = time.time()
	quant_config = AWQConfig(base_config, step="convert")
	quantize_(model, quant_config)
	print(f"time for convert: {time.time() - t0:.02f} seconds")
	quant_config = AWQConfig(base_config, step="prepare_for_loading")
	model.config.quantization_config = TorchAoConfig(quant_config)

		if config.set_inductor_config:
		torchao.quantization.utils.recommended_inductor_config_setter()


		# Get quantization parameters
		if all(x is not None for x in (config.smoothing_factor, config.wei_scales)):


		Note*: Conventional quantization without SmoothQuant
		Evaluation perplexity numbers were calculated using the script in `smoothquant/example.py`. For Llama-2-7b-chat-hf, performance benchmarks were calculated using the `torchao/_models/llama/generate.py` script and run on a 1xA100 48GB PCIe interface.

Make SmoothQuant more General #2728

Are you sure you want to change the base?

Make SmoothQuant more General #2728

Conversation

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

Uh oh!

Summary

Test Plan

Future Plan

Uh oh!

pytorch-bot bot commented Aug 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/2728

Uh oh!

namgyu-youn commented Aug 15, 2025

Uh oh!

jerryzh168 commented Aug 15, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

namgyu-youn commented Aug 15, 2025

Uh oh!

namgyu-youn commented Aug 16, 2025

Uh oh!

namgyu-youn commented Aug 16, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

namgyu-youn Aug 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!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

jerryzh168 Aug 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!

namgyu-youn Aug 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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

namgyu-youn Aug 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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

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

namgyu-youn commented Aug 11, 2025 •

edited

Loading

pytorch-bot bot commented Aug 11, 2025 •

edited

Loading

jerryzh168 commented Aug 15, 2025 •

edited

Loading

namgyu-youn Aug 19, 2025 •

edited

Loading

jerryzh168 Aug 19, 2025 •

edited

Loading

namgyu-youn Aug 19, 2025 •

edited

Loading

namgyu-youn Aug 20, 2025 •

edited

Loading

namgyu-youn Aug 20, 2025 •

edited

Loading

namgyu-youn Aug 21, 2025 •

edited

Loading

jerryzh168 Aug 21, 2025 •

edited

Loading

namgyu-youn Aug 22, 2025 •

edited

Loading