From 94555437e277949e6b3be9e97317df2a5e046d95 Mon Sep 17 00:00:00 2001
From: HDCharles <39544797+HDCharles@users.noreply.github.com>
Date: Thu, 20 Mar 2025 06:23:14 -0400
Subject: [PATCH] Disable inductor config setter by default (#36608)

* Disable inductor config setter by default

This is hard to debug and should be off by default

* remove default settings in autoquant too

* Add info to torchao.md about recommended settings

* satisfying Ruff format

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

---------

Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>
---
 docs/source/en/quantization/torchao.md           | 3 +++
 src/transformers/quantizers/quantizer_torchao.py | 7 +++++--
 2 files changed, 8 insertions(+), 2 deletions(-)

diff --git a/docs/source/en/quantization/torchao.md b/docs/source/en/quantization/torchao.md
index 31e2d4f020..42231d759c 100644
--- a/docs/source/en/quantization/torchao.md
+++ b/docs/source/en/quantization/torchao.md
@@ -150,6 +150,9 @@ output = bf16_model.generate(**input_ids, max_new_tokens=10, cache_implementatio
 print("bf16 model:", benchmark_fn(bf16_model.generate, **input_ids, max_new_tokens=MAX_NEW_TOKENS, cache_implementation="static"))
 ```
 
+> [!TIP]
+> For best performance, you can use recommended settings by calling `torchao.quantization.utils.recommended_inductor_config_setter()`
+
 </hfoption>
 <hfoption id="automatic">
 
diff --git a/src/transformers/quantizers/quantizer_torchao.py b/src/transformers/quantizers/quantizer_torchao.py
index 1dddaf19d4..d15f09be8e 100644
--- a/src/transformers/quantizers/quantizer_torchao.py
+++ b/src/transformers/quantizers/quantizer_torchao.py
@@ -236,7 +236,7 @@ class TorchAoHfQuantizer(HfQuantizer):
         else:
             assert isinstance(self.quantization_config, TorchAoConfig)
             module._parameters[tensor_name] = torch.nn.Parameter(param_value).to(device=target_device)
-            quantize_(module, self.quantization_config.get_quantize_config())
+            quantize_(module, self.quantization_config.get_apply_tensor_subclass(), set_inductor_config=False)
 
     def _process_model_after_weight_loading(self, model, **kwargs):
         """No process required for torchao quantized model"""
@@ -246,7 +246,10 @@ class TorchAoHfQuantizer(HfQuantizer):
 
             model = torch.compile(model, mode="max-autotune")
             model = autoquant(
-                model, qtensor_class_list=ALL_AUTOQUANT_CLASS_LIST, **self.quantization_config.quant_type_kwargs
+                model,
+                qtensor_class_list=ALL_AUTOQUANT_CLASS_LIST,
+                set_inductor_config=False,
+                **self.quantization_config.quant_type_kwargs,
             )
             return model
         return