Support Wan2.2 S2V Quantization by changwangss · Pull Request #1924 · intel/auto-round

changwangss · 2026-06-15T12:14:04Z

Description

Fixed tokenizer save crash in export paths

Issue: exporter called tokenizer.save_pretrained(...) even when tokenizer was a string placeholder ("Placeholder"), causing: AttributeError: 'str' object has no attribute 'save_pretrained'
Fix: added guard in FP8/NVFP exporters to save only when tokenizer has save_pretrained; otherwise skip with a log message.

Fixed model path access in is_mllm_model

Issue: code assumed every model object had name_or_path, but WanModel_S2V does not, causing:
AttributeError: 'WanModel_S2V' object has no attribute 'name_or_path'
Fix: added safe fallback chain (_name_or_path → name_or_path → fallback key) and robust cache-key handling.

  File "/usr/local/lib/python3.12/dist-packages/neural_compressor/torch/algorithms/autoround/autoround.py", line 216, in convert
    rounder = AutoRound(
              ^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/auto_round/autoround.py", line 165, in __new__
    return AutoRoundCompatible(**local_args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/auto_round/compressors/entry.py", line 586, in __new__
    if is_mllm_model(model, platform=platform):
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/auto_round/utils/model.py", line 954, in is_mllm_model
    model_path = model_or_path if isinstance(model_or_path, str) else model_or_path.name_or_path
                                                                      ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/diffusers/models/modeling_utils.py", line 274, in __getattr__
    return super().__getattr__(name)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/nn/modules/module.py", line 1965, in __getattr__
    raise AttributeError(
AttributeError: 'WanModel_S2V' object has no attribute 'name_or_path'

Wan2.2 S2V tokenizer is not from transformers tokenizer, INC pass 'Placeholder' for tokenizer,

  File "/mnt/disk1/wangchang/sd/neural-compressor/examples/pytorch/diffusion_model/diffusers/wan/auto-round/auto_round
/compressors/base.py", line 1580, in quantize_and_save
    model, folders = self.save_quantized(output_dir, inplace=inplace, return_folders=True, **kwargs)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/mnt/disk1/wangchang/sd/neural-compressor/examples/pytorch/diffusion_model/diffusers/wan/auto-round/auto_round
/compressors/mllm_mixin.py", line 122, in save_quantized
    compressed_model = super().save_quantized(
                       ^^^^^^^^^^^^^^^^^^^^^^^
  File "/mnt/disk1/wangchang/sd/neural-compressor/examples/pytorch/diffusion_model/diffusers/wan/auto-round/auto_round
/compressors/base.py", line 1428, in save_quantized
    compressed_model = format.save_quantized(
                       ^^^^^^^^^^^^^^^^^^^^^^
  File "/mnt/disk1/wangchang/sd/neural-compressor/examples/pytorch/diffusion_model/diffusers/wan/auto-round/auto_round
/formats.py", line 1262, in save_quantized
    return self.backend.save_quantized(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/mnt/disk1/wangchang/sd/neural-compressor/examples/pytorch/diffusion_model/diffusers/wan/auto-round/auto_round
/formats.py", line 1293, in save_quantized
    return export_func(
           ^^^^^^^^^^^^
  File "/mnt/disk1/wangchang/sd/neural-compressor/examples/pytorch/diffusion_model/diffusers/wan/auto-round/auto_round
/export/export_to_autoround/export_to_fp8.py", line 284, in save_quantized_as_autoround
    tokenizer.save_pretrained(output_dir)
    ^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'str' object has no attribute 'save_pretrained'

for more information, see https://pre-commit.ci

chensuyue · 2026-06-15T12:47:57Z

/azp run Unit-Test-CUDA-AutoRound

azure-pipelines · 2026-06-15T12:48:06Z

Azure Pipelines successfully started running 1 pipeline(s).

Signed-off-by: changwangss <chang1.wang@intel.com>

for more information, see https://pre-commit.ci

chensuyue · 2026-06-15T17:10:36Z

/azp run Unit-Test-CUDA-AutoRound

azure-pipelines · 2026-06-15T17:10:46Z

Azure Pipelines successfully started running 1 pipeline(s).

n1ck-guo · 2026-06-16T04:52:05Z

LGTM

yiliu30

Agree with heng's comments; others lgtm.

Signed-off-by: changwangss <chang1.wang@intel.com>

changwangss · 2026-06-16T09:09:12Z

Good point, thanks. I’ve extracted the path resolution into get_model_name_or_path and updated is_mllm_model to use it consistently. This makes the code cleaner and avoids duplicated fallback logic. @n1ck-guo @yiliu30

changwangss · 2026-06-16T09:56:24Z

CI update: Unit-Test-AutoRound parts 3/4/5 failed during Env Setup (tests did not start).

part3: exited with code 137 (likely runner resource/OOM kill)
part5: exited with code 137 (likely runner resource/OOM kill)
part4: exited with code 1 during setup
Given failures happened before UT execution and two jobs show 137, this looks infrastructure/setup-related rather than a code regression in this PR.
Could we please rerun these jobs? @XuehaoSun @chensuyue

changwangss · 2026-06-16T11:58:12Z

/azp run Unit-Test-AutoRound

azure-pipelines · 2026-06-16T11:58:24Z

Azure Pipelines successfully started running 1 pipeline(s).

chensuyue · 2026-06-16T12:36:38Z

/azp run Unit-Test-CUDA-AutoRound

azure-pipelines · 2026-06-16T12:36:48Z

Azure Pipelines successfully started running 1 pipeline(s).

changwangss · 2026-06-17T02:43:55Z

/azp run Unit-Test-CUDA-AutoRound

azure-pipelines · 2026-06-17T02:44:05Z

Azure Pipelines successfully started running 1 pipeline(s).

chensuyue · 2026-06-17T06:10:18Z

/azp run Unit-Test-CUDA-AutoRound

azure-pipelines · 2026-06-17T06:10:29Z

Azure Pipelines successfully started running 1 pipeline(s).

wenhuach21 · 2026-06-25T01:43:37Z

        tokenizer.save_pretrained(output_dir)
+    else:
+        print(f"Skip tokenizer save, tokenizer={tokenizer}")


for wan s2v, there is no tokenizer, https://huggingface.co/Wan-AI/Wan2.2-S2V-14B. general, if the tokenizer is not from transformers, it doesn't attribute the "save_pretrained".
Add a reminder to distinguish it from the standard process.

I mean, why not use logger here

Agreed, makes sense to use logger here. I'll update it. Thanks!

get model path for Wan S2V

2754589

changwangss requested review from n1ck-guo and yiliu30 June 15, 2026 12:14

[pre-commit.ci] auto fixes from pre-commit.com hooks

f7e9fa8

for more information, see https://pre-commit.ci

improve tokenizer for Wan 2.2 S2V

c3a1fa0

Signed-off-by: changwangss <chang1.wang@intel.com>

changwangss changed the title ~~Get correct model path for Wan2.2 S2V~~ Support Wan2.2 S2V Quantization Jun 15, 2026

changwangss and others added 2 commits June 15, 2026 16:35

git correct original key

5d07261

Signed-off-by: changwangss <chang1.wang@intel.com>

[pre-commit.ci] auto fixes from pre-commit.com hooks

23867d4

for more information, see https://pre-commit.ci

n1ck-guo reviewed Jun 16, 2026

View reviewed changes

Comment thread auto_round/utils/model.py Outdated

n1ck-guo approved these changes Jun 16, 2026

View reviewed changes

chensuyue added this to the 0.14.0 milestone Jun 16, 2026

yiliu30 approved these changes Jun 16, 2026

View reviewed changes

changwangss added 2 commits June 16, 2026 08:50

add function for get_model_name_or_path

61b6461

Signed-off-by: changwangss <chang1.wang@intel.com>

add optional

d0189b6

Signed-off-by: changwangss <chang1.wang@intel.com>

Merge branch 'main' into changwangss-patch-1

d25924b

Merge branch 'main' into changwangss-patch-1

a04241a

XuehaoSun merged commit ca37879 into main Jun 18, 2026
33 of 35 checks passed

XuehaoSun deleted the changwangss-patch-1 branch June 18, 2026 05:20

wenhuach21 reviewed Jun 25, 2026

View reviewed changes

Uh oh!

Conversation

changwangss commented Jun 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Uh oh!

chensuyue commented Jun 15, 2026

Uh oh!

azure-pipelines Bot commented Jun 15, 2026

Uh oh!

chensuyue commented Jun 15, 2026

Uh oh!

azure-pipelines Bot commented Jun 15, 2026

Uh oh!

Uh oh!

n1ck-guo commented Jun 16, 2026

Uh oh!

yiliu30 left a comment

Choose a reason for hiding this comment

Uh oh!

changwangss commented Jun 16, 2026

Uh oh!

changwangss commented Jun 16, 2026

Uh oh!

changwangss commented Jun 16, 2026

Uh oh!

azure-pipelines Bot commented Jun 16, 2026

Uh oh!

chensuyue commented Jun 16, 2026

Uh oh!

azure-pipelines Bot commented Jun 16, 2026

Uh oh!

changwangss commented Jun 17, 2026

Uh oh!

azure-pipelines Bot commented Jun 17, 2026

Uh oh!

chensuyue commented Jun 17, 2026

Uh oh!

azure-pipelines Bot commented Jun 17, 2026

Uh oh!

Uh oh!

wenhuach21 Jun 25, 2026

Choose a reason for hiding this comment

Uh oh!

changwangss Jun 25, 2026 • 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!

wenhuach21 Jun 25, 2026

Choose a reason for hiding this comment

Uh oh!

changwangss Jun 25, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

changwangss commented Jun 15, 2026 •

edited

Loading

changwangss Jun 25, 2026 •

edited

Loading