Renaming quantize to quantize_

Summary:
Addressing feedback for `quantize` API from #391 (comment)

this is an API that changes model inplace, so we want to change the name to reflect that. inplace model quantization is important especially for LLM since it will be hard to load the model to memory. we typically load the model to meta device and then load the quantized weight.

Test Plan:
python test/quantization/test_quant_api.py
python test/integration/test_integration.py

Reviewers:

Subscribers:

Tasks:

Tags:

README.md

@@ -19,8 +19,8 @@ All with no intrusive code changes and minimal accuracy degradation.
 Quantizing your models is a 1 liner that should work on any model with an `nn.Linear` including your favorite HuggingFace model. You can find a more comprehensive usage instructions [here](torchao/quantization/) and a HuggingFace inference example [here](scripts/hf_eval.py)
 
 ```python
-from torchao.quantization.quant_api import quantize, int4_weight_only
-m = quantize(m, int4_weight_only())
+from torchao.quantization.quant_api import quantize_, int4_weight_only
+quantize_(m, int4_weight_only())
 ```
 
 Benchmarks are run on a machine with a single A100 GPU using the script in `_models/llama` which generates text in a latency-optimized way (batchsize=1)
@@ -70,7 +70,7 @@ swap_linear_with_semi_sparse_linear(model, {"seq.0": SemiSparseLinear})
 
 * [MX](torchao/prototype/mx_formats) implementing training and inference support with tensors using the [OCP MX spec](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf) data types, which can be described as groupwise scaled float8/float6/float4/int8, with the scales being constrained to powers of two. This work is prototype as the hardware support is not available yet.
 * [nf4](torchao/dtypes/nf4tensor.py) which was used to [implement QLoRA](https://github.com/pytorch/torchtune/blob/main/docs/source/tutorials/qlora_finetune.rst) one of the most popular finetuning algorithms without writing custom Triton or CUDA code. Accessible talk [here](https://x.com/HamelHusain/status/1800315287574847701)
-* [fp6](torchao/prototype/quant_llm/) for 2x faster inference over fp16 with an easy to use API `quantize(model, fp6_llm_weight_only())`
+* [fp6](torchao/prototype/quant_llm/) for 2x faster inference over fp16 with an easy to use API `quantize_(model, fp6_llm_weight_only())`
 
 ## Composability
 

test/integration/test_integration.py

@@ -23,7 +23,7 @@
  int4_weight_only,
  int8_weight_only,
  int8_dynamic_activation_int8_weight,
- quantize,
+ quantize_,
  _replace_with_custom_fn_if_matches_filter,
 )
 # APIs to be deprecated (used for torch 2.2.2 and 2.3)
@@ -98,21 +98,21 @@
 
 def _int8wo_api(mod):
  if TORCH_VERSION_AFTER_2_4:
- quantize(mod, int8_weight_only(), set_inductor_config=False)
+ quantize_(mod, int8_weight_only(), set_inductor_config=False)
  unwrap_tensor_subclass(mod)
  else:
  change_linear_weights_to_int8_woqtensors(mod)
 
 def _int8da_int8w_api(mod):
  if TORCH_VERSION_AFTER_2_4:
- quantize(mod, int8_dynamic_activation_int8_weight(), set_inductor_config=False)
+ quantize_(mod, int8_dynamic_activation_int8_weight(), set_inductor_config=False)
  unwrap_tensor_subclass(mod)
  else:
  change_linear_weights_to_int8_dqtensors(mod)
 
 def _int4wo_api(mod):
  if TORCH_VERSION_AFTER_2_4:
- quantize(mod, int4_weight_only(), set_inductor_config=False)
+ quantize_(mod, int4_weight_only(), set_inductor_config=False)
  unwrap_tensor_subclass(mod)
  else:
  change_linear_weights_to_int4_woqtensors(mod)
@@ -127,8 +127,8 @@ def _int4wo_api(mod):
 def undo_recommended_configs():
  torch._inductor.config.coordinate_descent_tuning = False
  torch._inductor.config.coordinate_descent_check_all_directions = False
- torch._inductor.config.force_fuse_int_mm_with_mul = False 
- torch._inductor.config.fx_graph_cache = False 
+ torch._inductor.config.force_fuse_int_mm_with_mul = False
+ torch._inductor.config.fx_graph_cache = False
  torch._inductor.config.triton.unique_kernel_names = False
  torch.set_float32_matmul_precision("highest")
 
@@ -844,7 +844,7 @@ def api(mod):
  kwargs_copy = kwargs.copy()
  kwargs_copy["group_size"] = groupsize
  del kwargs_copy["groupsize"]
- quantize(mod, int4_weight_only(**kwargs_copy))
+ quantize_(mod, int4_weight_only(**kwargs_copy))
  unwrap_tensor_subclass(mod)
  else:
  change_linear_weights_to_int4_woqtensors(mod, **kwargs)
@@ -865,7 +865,7 @@ def test_dynamic_quant(self):
  m = nn.Sequential(nn.Linear(K, N))
 
  y_ref = m(x)
- quantize(m, int8_dynamic_activation_int8_weight())
+ quantize_(m, int8_dynamic_activation_int8_weight())
  y_test = m(x)
 
  sqnr = compute_error(y_ref, y_test)
@@ -1259,7 +1259,7 @@ def test_autoquant_manual(self, device, dtype):
  out3 = mod(example_input)
  sqnr2 = SQNR(out, out3)
  self.assertTrue(sqnr2 >= 30)
- 
+
 
  @parameterized.expand(combine_parameters(COMMON_DEVICE_DTYPE,
  [

test/prototype/test_quant_llm.py

@@ -16,7 +16,7 @@
 )
 from torchao.prototype.quant_llm.quant_llm import _pack_tc_fpx, _pack_tc_fp6
 from torchao.prototype.custom_fp_utils import _f32_to_fpx_unpacked, _fpx_unpacked_to_f32
-from torchao.quantization.quant_api import quantize
+from torchao.quantization.quant_api import quantize_
 
 
 _DEVICES = ["cpu"] + (["cuda"] if torch.cuda.is_available() else [])
@@ -91,7 +91,7 @@ def test_quant_llm_quantize(self, ebits, mbits, bias):
 
  linear = torch.nn.Linear(IC, OC, bias=bias, device=device)
  fpx_linear = copy.deepcopy(linear)
- quantize(fpx_linear, quant_llm_fpx_weight_only(ebits, mbits))
+ quantize_(fpx_linear, quant_llm_fpx_weight_only(ebits, mbits))
 
  x = torch.randn(N, IC, device=device, dtype=torch.half)
  expected = fpx_linear(x)

test/quantization/test_quant_api.py

@@ -31,7 +31,7 @@
  Int8WeightOnlyQuantizedLinearWeight,
  Int4WeightOnlyQuantizedLinearWeight,
 )
-from torchao import quantize
+from torchao import quantize_
 from torchao.quantization.quant_api import (
  _replace_with_custom_fn_if_matches_filter,
  Quantizer,
@@ -89,7 +89,7 @@ def convert(self, model: torch.nn.Module) -> torch.nn.Module:
 
 class TorchCompileDynamicQuantizer(Quantizer):
  def quantize(self, model: torch.nn.Module) -> torch.nn.Module:
- quantize(model, int8_dynamic_activation_int8_weight())
+ quantize_(model, int8_dynamic_activation_int8_weight())
  return model
 
 class ToyLinearModel(torch.nn.Module):
@@ -152,7 +152,7 @@ class TestQuantFlow(TestCase):
  def test_dynamic_quant_gpu_singleline(self):
  m = ToyLinearModel().eval()
  example_inputs = m.example_inputs()
- m = quantize(m, int8_dynamic_activation_int8_weight())
+ quantize_(m, int8_dynamic_activation_int8_weight())
  quantized = m(*example_inputs)
  # AssertionError: Expecting input to have dtype torch.float32, but got dtype: torch.float64
  # While executing %choose_qparams_tensor_1 : [num_users=2] = call_function[target=torch.ops.quantized_decomposed.choose_qparams.tensor](args = (%arg0_3, -128, 127, 0.000244140625, torch.int8), kwargs = {})
@@ -195,7 +195,7 @@ def test_int8_wo_quant_save_load(self):
  )
  m = ToyLinearModel().eval().cpu()
  def api(model):
- model = quantize(model, int8_weight_only())
+ quantize_(model, int8_weight_only())
  unwrap_tensor_subclass(model)
 
  api(m)
@@ -501,7 +501,7 @@ def test_quantized_tensor_subclass_8da4w(self):
  m = ToyLinearModel().eval()
  m_copy = copy.deepcopy(m)
  example_inputs = m.example_inputs()
- m = quantize(m, int8_dynamic_activation_int4_weight(group_size=group_size))
+ quantize_(m, int8_dynamic_activation_int4_weight(group_size=group_size))
 
  assert isinstance(m.linear1.weight, LinearActQuantizedTensor)
  assert isinstance(m.linear2.weight, LinearActQuantizedTensor)
@@ -530,7 +530,7 @@ def test_quantized_tensor_subclass_int4(self):
  example_inputs = m.example_inputs(dtype=torch.bfloat16, device="cuda")
 
  group_size = 32
- m = quantize(m, int4_weight_only(group_size=group_size))
+ quantize_(m, int4_weight_only(group_size=group_size))
  assert isinstance(m.linear1.weight, AffineQuantizedTensor)
  assert isinstance(m.linear2.weight, AffineQuantizedTensor)
 
@@ -550,7 +550,7 @@ def test_quantized_tensor_subclass_int8_wo(self):
  m_copy = copy.deepcopy(m)
  example_inputs = tuple(map(lambda x: x.to(torch.bfloat16), m.example_inputs()))
 
- m = quantize(m, int8_weight_only())
+ quantize_(m, int8_weight_only())
 
  assert isinstance(m.linear1.weight, AffineQuantizedTensor)
  assert isinstance(m.linear2.weight, AffineQuantizedTensor)
@@ -573,7 +573,7 @@ def test_quantized_tensor_subclass_int8_dyn_quant(self):
  m_copy = copy.deepcopy(m)
  # setting batch_size to 20 to be compatible with the kernel
  example_inputs = m.example_inputs(batch_size=20, dtype=torch.bfloat16, device="cuda")
- m = quantize(m, int8_dynamic_activation_int8_weight())
+ quantize_(m, int8_dynamic_activation_int8_weight())
 
  assert isinstance(m.linear1.weight, LinearActQuantizedTensor)
  assert isinstance(m.linear2.weight, LinearActQuantizedTensor)
@@ -607,7 +607,7 @@ def test_quantized_tensor_subclass_save_load(self):
  m_copy = copy.deepcopy(m)
  example_inputs = m.example_inputs(dtype=torch.bfloat16)
 
- m = quantize(m, int8_weight_only())
+ quantize_(m, int8_weight_only())
  ref = m(*example_inputs)
  with tempfile.NamedTemporaryFile() as f:
  torch.save(m.state_dict(), f)

torchao/__init__.py

@@ -30,14 +30,14 @@
 
 from torchao.quantization import (
  autoquant,
- quantize,
+ quantize_,
 )
 from . import dtypes
 
 __all__ = [
  "dtypes",
  "autoquant",
- "quantize",
+ "quantize_",
 ]
 
 # test-pytorchbot

torchao/_models/llama/eval.py

@@ -13,7 +13,7 @@
 
 )
 from torchao.quantization.quant_api import (
- quantize, int4_weight_only, int8_weight_only, int8_dynamic_activation_int8_weight, unwrap_tensor_subclass
+ quantize_, int4_weight_only, int8_weight_only, int8_dynamic_activation_int8_weight, unwrap_tensor_subclass
 
 )
 from torchao._models._eval import TransformerEvalWrapper, InputRecorder
@@ -60,13 +60,13 @@ def run_evaluation(
 
  if quantization:
  if "int8wo" in quantization:
- quantize(model, int8_weight_only())
+ quantize_(model, int8_weight_only())
  if "int8dq" in quantization:
- quantize(model, int8_dynamic_activation_int8_weight())
+ quantize_(model, int8_dynamic_activation_int8_weight())
  if "int4wo" in quantization and not "gptq" in quantization:
  groupsize=int(quantization.split("-")[-1])
  assert groupsize in [32,64,128,256], f"int4wo groupsize needs to be one of [32,64,128,256] but got {groupsize}"
- quantize(model.to(device), int4_weight_only(group_size=groupsize))
+ quantize_(model.to(device), int4_weight_only(group_size=groupsize))
  if "int4wo" in quantization and "gptq" in quantization:
  groupsize=int(quantization.split("-")[-2])
  assert groupsize in [32,64,128,256], f"int4wo groupsize needs to be one of [32,64,128,256] but got {groupsize}"
@@ -94,8 +94,8 @@ def run_evaluation(
  model = torch.compile(model, mode="max-autotune", fullgraph=True)
  with torch.no_grad():
  TransformerEvalWrapper(
- model=model.to(device), 
- tokenizer=tokenizer, 
+ model=model.to(device),
+ tokenizer=tokenizer,
  max_seq_length=max_length,
  input_prep_func=prepare_inputs_for_model,
  device=device,
@@ -122,16 +122,16 @@ def run_evaluation(
 
  args = parser.parse_args()
  run_evaluation(
- args.checkpoint_path, 
- args.tasks, 
- args.limit, 
- args.device, 
- args.precision, 
- args.quantization, 
- args.compile, 
- args.max_length, 
- args.calibration_tasks, 
- args.calibration_limit, 
- args.calibration_seq_length, 
- args.pad_calibration_inputs, 
+ args.checkpoint_path,
+ args.tasks,
+ args.limit,
+ args.device,
+ args.precision,
+ args.quantization,
+ args.compile,
+ args.max_length,
+ args.calibration_tasks,
+ args.calibration_limit,
+ args.calibration_seq_length,
+ args.pad_calibration_inputs,
  )

torchao/_models/llama/generate.py

@@ -100,7 +100,7 @@ def generate(
  T_new = T + max_new_tokens
  seq = torch.empty(T_new, dtype=prompt.dtype, device=device)
  seq[:T] = prompt.view(-1)
- 
+
  # setup model cache
  max_seq_length = min(T_new, model.config.block_size) if not interactive else 350
  with torch.device(device):
@@ -158,7 +158,7 @@ def main(
  """
 
  torchao.quantization.utils.recommended_inductor_config_setter()
- 
+
  assert checkpoint_path.is_file(), checkpoint_path
  tokenizer_path = checkpoint_path.parent / "tokenizer.model"
  assert tokenizer_path.is_file(), str(tokenizer_path)
@@ -180,11 +180,11 @@ def main(
  prompt_length = encoded.size(0)
 
  torch.manual_seed(1234)
- 
+
 
  if quantization:
  from torchao.quantization.quant_api import (
- quantize,
+ quantize_,
  int8_weight_only,
  int8_dynamic_activation_int8_weight,
  int4_weight_only,
@@ -193,13 +193,13 @@ def main(
  )
 
  if "int8wo" in quantization:
- quantize(model, int8_weight_only())
+ quantize_(model, int8_weight_only())
  if "int8dq" in quantization:
- quantize(model, int8_dynamic_activation_int8_weight())
+ quantize_(model, int8_dynamic_activation_int8_weight())
  if "int4wo" in quantization:
  groupsize=int(quantization.split("-")[-1])
  assert groupsize in [32,64,128,256], f"int4wo groupsize needs to be one of [32,64,128,256] but got {groupsize}"
- quantize(model, int4_weight_only(group_size=groupsize))
+ quantize_(model, int4_weight_only(group_size=groupsize))
  if "autoquant" == quantization:
  model = autoquant(model, manual=True)
 

torchao/prototype/quant_llm/README.md

@@ -5,15 +5,15 @@ This is a FP16 x FPx mixed matmul kernel optimized for io bound workloads per [F
 ## Usage
 
 ```python
-from torchao.quantization.quant_api import quantize
+from torchao.quantization.quant_api import quantize_
 from torchao.prototype.quant_llm import fp6_llm_weight_only, quant_llm_fpx_weight_only
 
 model = ...
 model.half() # not necessary, but recommeneded to maintain accuracy
-quantize(model, fp6_llm_weight_only()) # convert nn.Lineaer.weight to FP6 E3M2 in-place
+quantize_(model, fp6_llm_weight_only()) # convert nn.Lineaer.weight to FP6 E3M2 in-place
 
 # for generic FPx EyMz where x = 1 + y + z
-# quantize(model, quant_llm_fpx_weight_only(2, 2)) # use FP5 E2M2 instead
+# quantize_(model, quant_llm_fpx_weight_only(2, 2)) # use FP5 E2M2 instead
 
 # fully compatible with torch.compile()
 model.compile(mode="max-autotune", fullgraph=True)

torchao/quantization/README.md

@@ -74,7 +74,7 @@ from torchao.quantization.quant_primitives import MappingType, ZeroPointDomain
 from torchao.dtypes import to_affine_quantized
 import copy
 from torchao.quantization.quant_api import (
- quantize,
+ quantize_,
  int4_weight_only,
 )
 
@@ -101,7 +101,7 @@ m_bf16 = torch.compile(m_bf16, mode='max-autotune')
 # apply int4 weight only quant (compatible with tinygemm int4 weight only quant mm kernel in torchao)
 group_size = 32
 # only works for torch 2.4+
-m = quantize(m, int4_weight_only(group_size=group_size))
+quantize_(m, int4_weight_only(group_size=group_size))
 
 # temporary workaround for tensor subclass + torch.compile
 from torchao.utils import unwrap_tensor_subclass
@@ -168,7 +168,7 @@ torch._inductor.config.force_fuse_int_mm_with_mul = True
 
 # for torch 2.4+
 from torchao.quantization import quantize, int8_dynamic_activation_int8_weight
-quantize(model, int8_dynamic_activation_int8_weight())
+quantize_(model, int8_dynamic_activation_int8_weight())
 
 # for torch 2.2.2 and 2.3
 from torchao.quantization.quant_api import change_linear_weights_to_int8_dqtensors
@@ -180,7 +180,7 @@ change_linear_weights_to_int8_dqtensors(model)
 ```python
 # for torch 2.4+
 from torchao.quantization import quantize, int8_weight_only
-quantize(model, int8_weight_only())
+quantize_(model, int8_weight_only())
 
 # for torch 2.2.2 and 2.3
 from torchao.quantization.quant_api import change_linear_weights_to_int8_woqtensors
@@ -195,7 +195,7 @@ This technique works best when the torch._inductor.config.use_mixed_mm option is
 ```python
 # for torch 2.4+
 from torchao.quantization import quantize, int4_weight_only
-quantize(model, int4_weight_only())
+quantize_(model, int4_weight_only())
 
 # for torch 2.2.2 and 2.3
 from torchao.quantization.quant_api import change_linear_weights_to_int4_woqtensors

torchao/quantization/__init__.py

@@ -29,7 +29,7 @@
  "quantize_affine",
  "dequantize_affine",
  "choose_qprams_affine",
- "quantize",
+ "quantize_",
  "int8_dynamic_activation_int4_weight",
  "int8_dynamic_activation_int8_weight",
  "int4_weight_only",