Add parq utility to create an optimizer (#3165)

* Add parq util

* up

Author: metascroy

torchao/prototype/parq/api.py

@@ -0,0 +1,145 @@
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Tuple, Type
+
+import torch
+
+from torchao.prototype.parq.optim import QuantOptimizer
+from torchao.prototype.parq.quant import (
+    Quantizer,
+    StretchedUnifTorchaoQuantizer,
+    UnifTorchaoQuantizer,
+)
+
+
+@dataclass(frozen=True, slots=True)
+class QuantConfig:
+    bitwidth: int
+    group_size: Optional[int] = None
+    quantizer: Optional[Quantizer] = None
+
+    def __post_init__(self):
+        if self.bitwidth < 2:
+            raise ValueError("bitwidth must be >= 2")
+        if self.group_size is not None and self.group_size <= 0:
+            raise ValueError("group_size must be positive")
+
+        if self.quantizer is None:
+            if self.bitwidth in [2, 3]:
+                q = StretchedUnifTorchaoQuantizer(b=self.bitwidth)
+            else:
+                q = UnifTorchaoQuantizer()
+            object.__setattr__(self, "quantizer", q)
+
+
+def create_param_groups_and_group_quantizer_map(
+    model: torch.nn.Module,
+    quant_configs_and_filter_fns: List[
+        Tuple[QuantConfig, Callable[[torch.nn.Module, str], bool]]
+    ],
+):
+    param_groups = []
+    group_quantizer_map = {}
+    for idx, (config, _) in enumerate(quant_configs_and_filter_fns):
+        params_quant = []
+        param_group = {
+            "params": params_quant,
+            "quant_bits": config.bitwidth,
+        }
+        if config.group_size is not None:
+            param_group["quant_block_size"] = config.group_size
+        param_group["_quantizer"] = config.quantizer
+        param_groups.append(param_group)
+
+    # Non-quantized group at end so that index in param_groups
+    # is the index in the subset of quantized param groups, which is
+    # used in defining group_quantizer_map
+    params_no_quant = []
+    param_groups.append({"params": params_no_quant, "weight_decay": 0.0})
+
+    seen_data_ptrs = {}
+    for param_name, param in model.named_parameters():
+        module_name, _, param_basename = param_name.rpartition(".")
+        owning_module = model.get_submodule(module_name) if module_name else model
+
+        data_ptr = param.data_ptr()
+        if data_ptr in seen_data_ptrs:
+            print(
+                f"Not considering {param} because it shares a data_ptr with {seen_data_ptrs[data_ptr]}, which was previously considered"
+            )
+            continue
+        seen_data_ptrs[data_ptr] = param_name
+
+        print(
+            "param_name",
+            param_name,
+            "module_type",
+            type(owning_module),
+            "matching_config:",
+            end="",
+        )
+        matching_config = None
+        for idx, (config, filter_fn) in enumerate(quant_configs_and_filter_fns):
+            if filter_fn(owning_module, param_name):
+                param_groups[idx]["params"].append(param)
+                if matching_config is None:
+                    matching_config = config
+                    print(f"{config.bitwidth},{config.group_size}")
+                else:
+                    raise ValueError(
+                        f"Found multiple matching configs for {param_name}. Previous match={matching_config}, new match={config}."
+                    )
+
+        # If no match, add to no-quant group at last idx
+        if matching_config is None:
+            print("NONE")
+            param_groups[-1]["params"].append(param)
+
+    # Filter out empty param groups
+    param_groups = [pg for pg in param_groups if len(pg["params"]) > 0]
+
+    # After filter define group_quantizer_map
+    # The index in group_quantizer_map must correspond to index in
+    # quantized params
+    group_quantizer_map = {}
+    for idx, param_group in enumerate(param_groups):
+        if "_quantizer" in param_group:
+            group_quantizer_map[idx] = param_group.pop("_quantizer")
+
+    expected_n_params = sum(1 for p in model.parameters())
+    n_found_params = sum(len(pg["params"]) for pg in param_groups)
+    assert n_found_params == expected_n_params, (
+        f"{n_found_params} != {expected_n_params=}"
+    )
+
+    return param_groups, group_quantizer_map
+
+
+from torchao.prototype.parq import ProxHardQuant
+
+
+def create_optimizer(
+    model: torch.nn.Module,
+    quant_configs_and_filter_fns: List[
+        Tuple[QuantConfig, Callable[[torch.nn.Module, str], bool]]
+    ],
+    base_optimizer_cls: Type[torch.optim.Optimizer],
+    base_optimizer_kwargs: Dict[str, Any],
+    *,
+    warmup_steps: int = 0,
+    quant_period: int = 1,
+    quant_per_channel: bool = True,
+):
+    param_groups, group_quantizer_map = create_param_groups_and_group_quantizer_map(
+        model, quant_configs_and_filter_fns
+    )
+    base_optimizer = base_optimizer_cls(param_groups, **base_optimizer_kwargs)
+    optimizer = QuantOptimizer(
+        base_optimizer,
+        quantizer=UnifTorchaoQuantizer(),
+        prox_map=ProxHardQuant(),
+        warmup_steps=warmup_steps,
+        quant_period=quant_period,
+        quant_per_channel=quant_per_channel,
+        group_quantizer_map=group_quantizer_map,
+    )
+    return optimizer