mixed-precision quantization milestone1: naive_intNwo + eval/benchmark framework

test/quantization/test_mixed_precision.py

@@ -0,0 +1,32 @@
+import unittest
+
+import torch
+import torch.nn as nn
+from torchao.quantization import quantize_, int8_weight_only, int4_weight_only
+from torchao.quantization.utils import compute_error
+from torchao.quantization.prototype.mixed_precision.scripts.naive_intNwo import intN_weight_only
+
+_CUDA_IS_AVAILABLE = torch.cuda.is_available()
+
+class TestWeightOnlyQuantNaive(unittest.TestCase):
+
+    def test_quantization_intNwo(self):
+        #skip test int4wo for now since it is under development in torchao
+        for quantization_bit in [2, 3, 5, 6, 8]:
+            for symmetric in [False, True]:
+                with self.subTest(quantization_bit=quantization_bit, symmetric=symmetric):
+                    for x_shape in [[64, 32], [80, 80, 80, 32], [16, 64, 32]]:
+                        x = torch.randn(*x_shape, dtype=torch.bfloat16)
+                        m = nn.Sequential(nn.Linear(32, 80)).bfloat16()
+                        y_ref = m(x)
+                        quantize_(m, intN_weight_only(n=quantization_bit, group_size=32, symmetric=symmetric))
+                        y_wo = m(x)
+                        sqnr = compute_error(y_ref, y_wo)
+                        # SQNR_dB can be approximated by 6.02n, where n is the bit width of the quantization
+                        # e.g., we set sqnr threshold = 44 for 8-bit, so that 6.02 * 8= 48.16 fullfills
+                        expected_sqnr_threshold = 44.0 - (8 - quantization_bit) * 6.02
+                        self.assertGreater(sqnr, expected_sqnr_threshold, f"sqnr: {sqnr} is too low")
+
+
+if __name__ == '__main__':
+    unittest.main()

torchao/quantization/prototype/mixed_precision/scripts/__init__.py

@@ -0,0 +1 @@
+from .naive_intNwo import intN_weight_only

torchao/quantization/prototype/mixed_precision/scripts/mp_quant_eval.py

@@ -0,0 +1,95 @@
+import torch
+import torch.nn as nn
+
+from naive_intNwo import intN_weight_only
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from lm_eval.models.huggingface import HFLM
+from lm_eval.evaluator import evaluate
+from lm_eval.tasks import get_task_dict
+
+from torchao.quantization import quantize_, int8_weight_only, int4_weight_only, int8_dynamic_activation_int4_weight
+from torchao._models._eval import TransformerEvalWrapper
+
+from torchao.quantization.quant_primitives import (
+    MappingType,
+    ZeroPointDomain,
+)
+
+from torchao.quantization.quant_api import autoquant
+
+
+torch._inductor.config.force_fuse_int_mm_with_mul = True
+torch._inductor.config.fx_graph_cache = True
+
+
+def run_evaluation(repo_id, tasks, limit, device, precision, quantization, compile, batch_size, max_length, sensi_bit, non_sensi_bit, quant_sym, group_size):
+
+    tokenizer = AutoTokenizer.from_pretrained(repo_id)
+    model = AutoModelForCausalLM.from_pretrained(repo_id).to(device="cpu", dtype=precision)
+
+    if quantization == "autoquant":
+        model = autoquant(model.to(device=device))
+
+    # naive implementation of uniform precision quantization all layers    
+    elif quantization in ["2","3","4","5","6","8"]:
+        quantize_(model.to(device=device), intN_weight_only(n=int(quantization), group_size=group_size, symmetric=quant_sym))
+
+    # mix precision quantization for Llama3    
+    elif quantization == "MP_llama3":
+
+        # filter for sensitive layers (the first 3 and last 2 layers for Llama3)
+        def filter_fn_sen(child: torch.nn.Module, cur_fqn:str) -> bool:
+            return isinstance(child, nn.Linear) and any(skiplayer in cur_fqn for skiplayer in ['.0.', '.1.', '.2.', '.30.', '.31.'])
+
+        # filter for non-sensitive layers (other 27 layers for Llama3)
+        def filter_fn_nonsen(child: torch.nn.Module, cur_fqn:str) -> bool:
+            return isinstance(child, nn.Linear) and not(any(skiplayer in cur_fqn for skiplayer in ['.0.', '.1.', '.2.', '.30.', '.31.']))
+
+        # quantize the sensitive layers
+        if sensi_bit != 16:
+            quantize_(model.to(device=device), intN_weight_only(n=sensi_bit, group_size=group_size, symmetric=quant_sym), filter_fn_sen)
+
+        # quantize the less-sensitive layers
+        if sensi_bit == 4: 
+            quantize_(model, intN_weight_only(n=non_sensi_bit, group_size=group_size, symmetric=quant_sym), filter_fn_nonsen)     
+        else:
+            quantize_(model.to(device=device), intN_weight_only(n=non_sensi_bit, group_size=group_size, symmetric=quant_sym), filter_fn_nonsen)   
+
+    if compile:
+        model = torch.compile(model, mode="max-autotune", fullgraph=True)
+
+    with torch.no_grad():
+
+        result = evaluate(
+            HFLM(
+                pretrained=model,
+                tokenizer=tokenizer, 
+                batch_size=batch_size, 
+                max_length=max_length),
+            get_task_dict(tasks),
+            limit = limit,
+        )
+
+    for task, res in result["results"].items():
+        print(f"{task}: {res}")
+
+
+if __name__ == '__main__':
+    import argparse
+    parser = argparse.ArgumentParser(description='Run HF Model Evaluation')
+    parser.add_argument('--repo_id', type=str, default="checkpoints/meta-llama/Meta-Llama-3-8B", help='Repository ID to download from HF.')
+    parser.add_argument('--tasks', nargs='+', type=str, default=["wikitext"], help='List of lm-eluther tasks to evaluate usage: --tasks task1 task2')
+    parser.add_argument('--limit', type=int, default=None, help='Number of eval samples to evaluate')
+    parser.add_argument('--precision', type=lambda x: getattr(torch, x.split(".")[-1]), default=torch.bfloat16, help='dtype precision to use')
+    parser.add_argument('--device', type=str, default="cuda", help='Device to use for evaluation')
+    parser.add_argument('-q', '--quantization', default = "None", choices = ["2", "3", "4", "5", "6", "8", "MP_llama3", "None"], help='Which quantization technique to apply, choose from ["2", "3", "4", "5", "6", "8"] for uniform quantizatoin, choose "MP_llama3" for mixed-precision for Llama3 and need to set corresponding sensi_bit and non_sensi_bit, choose "None" for no quantization')
+    parser.add_argument('--compile', action='store_true', help='Whether to compile the model.')
+    parser.add_argument('--batch_size', type=int, default=1, help='Batch size to use for evaluation, note int8wo and int4wo work best with small batchsizes, int8dq works better with large batchsizes')
+    parser.add_argument('--max_length', type=int, default=None, help='Length of text to process at one time')
+    parser.add_argument('--sensi_bit', type=int, default=16, choices = [16, 8, 6, 5, 4, 3], help='Bit setting for sensitive layers')
+    parser.add_argument('--non_sensi_bit', type=int, default=8, choices = [8, 6, 5, 4, 3, 2], help='Bit setting for non-sensitive layers')
+    parser.add_argument('--quant_sym', type=bool, default=False, help='Symmetric or asymmetric quantization, asymmetric by default')
+    parser.add_argument('--group_size', type=int, default=32, help='Group size to perform quantization on')
+    args = parser.parse_args()
+    run_evaluation(args.repo_id, args.tasks, args.limit, args.device, args.precision, args.quantization, args.compile, args.batch_size, args.max_length, args.sensi_bit, args.non_sensi_bit, args.quant_sym, args.group_size)

torchao/quantization/prototype/mixed_precision/scripts/naive_intNwo.py

@@ -0,0 +1,60 @@
+import torch
+
+from torchao.quantization.quant_primitives import (
+    MappingType,
+    ZeroPointDomain,
+)
+
+from torchao.quantization import int8_weight_only, int4_weight_only
+from torchao.quantization.quant_api import _get_linear_subclass_inserter
+
+def intN_weight_only(group_size=32, n=8, symmetric=False):
+    '''
+        Apply int N-bit weight only quantization to a linear layer.
+        Args:
+            `groupsize`: parameter for quantization, controls the granularity of quantization, smaller size is more fine grained, choices are [512, 256, 128, 64, 32]
+            `n`: number of bits to quantize to, choices are [8, 6, 5, 4, 3, 2]
+        Usage:
+            from torchao.quantization import quantize_
+            quantize_(model, intN_weight_only(n=your_bit_choice, group_size=group_size), optional_filter_func_for_desired_layers_to_quantize)
+    '''
+    # for asymmetric quantization
+    def apply_intN_weight_only_quant_asym(weight):
+        # avoid circular dependency
+        from torchao.dtypes import to_affine_quantized
+        mapping_type = MappingType.ASYMMETRIC
+        block_size = (1, group_size)
+        target_dtype = torch.uint8
+        quant_min = 0
+        quant_max = 2**n-1
+        eps = 1e-6
+        preserve_zero = True
+        zero_point_dtype = torch.int64
+        zero_point_domain = ZeroPointDomain.INT
+        return to_affine_quantized(weight, mapping_type, block_size, target_dtype, quant_min, quant_max, eps, zero_point_dtype=zero_point_dtype)#, preserve_zero=preserve_zero,zero_point_domain=zero_point_domain)
+
+    # for symmetric quantization
+    def apply_intN_weight_only_quant_sym(weight):
+        # avoid circular dependency
+        from torchao.dtypes import to_affine_quantized
+        mapping_type = MappingType.SYMMETRIC
+        block_size = (1, group_size)
+        target_dtype = torch.int8
+        eps = 1e-6
+        zero_point_dtype = torch.int64
+        return to_affine_quantized(weight, mapping_type, block_size, target_dtype, eps=eps, zero_point_dtype=zero_point_dtype)
+
+    try:
+        assert n in [8, 6, 5, 4, 3, 2], "n must be one of [8, 6, 5, 4, 3, 2]"
+        if n == 8:
+            return int8_weight_only()
+        elif n == 4:
+            return int4_weight_only(group_size=group_size)
+        else:
+            if symmetric:
+                return _get_linear_subclass_inserter(apply_intN_weight_only_quant_sym)
+            else:
+                return _get_linear_subclass_inserter(apply_intN_weight_only_quant_asym)
+    except Exception as e:
+        raise  
+