Move some quantization ops to pytorch (#77)

Summary:
since executorch is removing dep on torchao and only depend on pytorch for now, we need to move these ops to pytorch so they can have access
we can move back when torchao is in a more mature state

Test Plan:
CI

Reviewers:

Subscribers:

Tasks:

Tags:

Author: jerryzh168

requirements.txt

@@ -1,3 +1,4 @@
 torch
 numpy
 sentencepiece
+packaging

test/quantization/test_quant_api.py

@@ -23,9 +23,9 @@
 from torchao.quantization.quant_api import (
     Quantizer,
     TwoStepQuantizer,
-    Int8DynActInt4WeightGPTQQuantizer,
-    Int8DynActInt4WeightQuantizer,
-    Int8DynActInt4WeightLinear,
+)
+from torchao.quantization.utils import (
+    TORCH_VERSION_AFTER_2_4,
 )
 from pathlib import Path
 from sentencepiece import SentencePieceProcessor
@@ -136,7 +136,11 @@ def test_dynamic_quant_gpu_unified_api_eager_mode_impl(self):
         compiled = m(*example_inputs)
         torch.testing.assert_close(quantized, compiled, atol=0, rtol=0)
 
+    @unittest.skipIf(not TORCH_VERSION_AFTER_2_4, "skipping when torch verion is 2.3 or lower")
     def test_8da4w_quantizer(self):
+        from torchao.quantization.quant_api import Int8DynActInt4WeightQuantizer
+        from torchao.quantization.quant_api import Int8DynActInt4WeightLinear
+
         quantizer = Int8DynActInt4WeightQuantizer(group_size=32)
         m = M().eval()
         example_inputs = m.example_inputs()
@@ -147,6 +151,7 @@ def test_8da4w_quantizer(self):
 
     @unittest.skip("skipping until we get checkpoints for gpt-fast")
     def test_gptq_quantizer(self):
+        from torchao.quantization.quant_api import Int8DynActInt4WeightGPTQQuantizer
         # should be similar to TorchCompileDynamicQuantizer
         precision = torch.bfloat16
         device = "cpu"
@@ -163,8 +168,8 @@ def test_gptq_quantizer(self):
         blocksize = 128
         percdamp = 0.01
         groupsize = 128
-        calibration_tasks = ["hellaswag"]
-        calibration_limit = 200 # 1000
+        calibration_tasks = ["wikitext"]
+        calibration_limit = 5
         calibration_seq_length = 100
         pad_calibration_inputs = False
         quantizer = Int8DynActInt4WeightGPTQQuantizer(