Feat (quantization): torch_function based quantization

src/brevitas/graph/quantize.py

@@ -1,9 +1,14 @@
 # Copyright (C) 2023, Advanced Micro Devices, Inc. All rights reserved.
 # SPDX-License-Identifier: BSD-3-Clause
 
+from typing import Dict
+
+from packaging import version
+import torch
 from torch import nn
 
 from brevitas import config
+from brevitas import torch_version
 from brevitas.core.scaling.standalone import ConstScaling
 from brevitas.core.scaling.standalone import ParameterScaling
 from brevitas.fx.brevitas_tracer import symbolic_trace
@@ -33,6 +38,34 @@
 from brevitas.quant import Uint8ActPerTensorFloatMaxInit
 from brevitas.quant.scaled_int import Int8WeightPerTensorFloat
 
+if torch_version >= version.parse('1.12'):
+    from torch.overrides import TorchFunctionMode
+
+    class functional_quantization_mode(TorchFunctionMode):
+
+        def __init__(self, model: torch.nn.Module, quant_map: Dict, enabled: bool = True):
+            super().__init__()
+            self.quant_map = quant_map
+            self.model = model
+            self.enabled = enabled
+            for stateless_function, stateless_module in quant_map.items():
+                if not hasattr(model, str(stateless_function)):
+                    setattr(model, str(stateless_function), stateless_module())
+
+        def __torch_function__(self, func, types, args=(), kwargs=None):
+            if kwargs is None:
+                kwargs = dict()
+
+            if hasattr(self.model, str(func)) and self.enabled:
+                module = getattr(self.model, str(func))
+                out = module(*args, **kwargs)
+            else:
+                out = func(*args, **kwargs)
+
+            return out
+else:
+    functional_quantization_mode = object()
+
 COMPUTE_LAYER_MAP = {
     nn.AvgPool2d:
         None,

src/brevitas_examples/llm/README.md

@@ -47,8 +47,8 @@ usage: main.py [-h] [--config CONFIG] [--model MODEL] [--seed SEED]
                [--convert-layernorm-to-rmsnorm] [--replace-rmsnorm]
                [--no-quantize] [--no-float16]
                [--scaling-min-val SCALING_MIN_VAL] [--quant-sdpa]
-               [--replace-mha] [--weight-equalization]
-               [--rotation {fx,layerwise,fused_no_fx}]
+               [--functional-sdpa-quant] [--replace-mha]
+               [--weight-equalization] [--rotation {fx,layerwise,fused_no_fx}]
                [--rotation-mode {had,ort}] [--rotation-orphan-sink]
                [--act-equalization {None,layerwise,fx}]
                [--act-equalization-alpha ACT_EQUALIZATION_ALPHA]
@@ -170,6 +170,9 @@ options:
                         fp16 quantization.
   --quant-sdpa          Quantize `F.scaled_dot_product_attention` (default:
                         False)
+  --functional-sdpa-quant
+                        Quantize `F.scaled_dot_product_attention` with
+                        stateless module and torch_function (default: False)
   --replace-mha         Replace HuggingFace Attention with a quantizable
                         version
   --weight-equalization

src/brevitas_examples/llm/config/default_template.yml

@@ -17,6 +17,7 @@ few_shot_tasks:
 - winogrande
 - piqa
 few_shot_zeroshot: false
+functional_sdpa_quant: false
 fuse_sequences: false
 gpfq: false
 gptq: false

src/brevitas_examples/llm/main.py

@@ -2,6 +2,7 @@
 # SPDX-License-Identifier: BSD-3-Clause
 
 import argparse
+from contextlib import nullcontext
 from copy import deepcopy
 import functools
 import sys
@@ -23,8 +24,10 @@
 from brevitas.graph import load_quant_model_mode
 from brevitas.graph.equalize import GraphRotationEqualization
 from brevitas.graph.equalize import LayerwiseActivationRotation
+from brevitas.graph.quantize import functional_quantization_mode
 from brevitas.graph.quantize import layerwise_quantize
 from brevitas.graph.utils import get_module
+from brevitas.nn.quant_sdpa import ScaledDotProductAttention
 from brevitas.utils.python_utils import hooked_on_a_function
 from brevitas_examples.common.accelerate_utils.accelerate import offload_model
 from brevitas_examples.common.accelerate_utils.accelerate import remove_hooks
@@ -132,6 +135,8 @@ def model_export(model, ref_input, args):
 
 
 def validate(args):
+    if args.functional_sdpa_quant:
+        assert args.input_scale_type == 'dynamic' or args.input_bit_width is None, "Functional SDPA Quant requires dynamic activation quantization"
     if args.rotation == 'fx':
         assert args.ln_affine_merge, 'Graph rotation requires to merge LN/RMS norm affine parameters'
         assert args.replace_rmsnorm, 'Graph rotation requires to replace HF RMSNorm with PyTorch ones (torch 2.4+ require)'
@@ -316,7 +321,12 @@ def quantize_llm(args):
         print("Replace `F.scaled_dot_product_attention` with QuantSDPA...")
         model = replace_sdpa_with_quantizable_layers(model)
         print("Replacing done.")
-
+    elif args.functional_sdpa_quant:
+        print("Inserting SDPA quantizable module")
+        model = offload_model(model)
+        with torch.no_grad(), functional_quantization_mode(model, {torch.nn.functional.scaled_dot_product_attention: ScaledDotProductAttention}):
+            model(**calibration_loader[0])
+        remove_hooks(model)
     if args.weight_equalization:
         print("Apply weight equalization...")
         # In case of float16 model, we need to offload to account for missing ops
@@ -421,118 +431,131 @@ def quantize_llm(args):
                 new_funct = functools.partial(update_internal_dict, m)
                 m._hf_hook.post_forward = hooked_on_a_function(m._hf_hook.post_forward, new_funct)
 
-    with torch.no_grad():
-        model(**calibration_loader[0])
+    # If we are doing functional SDPA quantization, we create the correct context manager,
+    # otherwise nullcontext. We would love to avoid the extra indentation level but it doesn't seem easy.
+    if args.functional_sdpa_quant:
+        quantization_cm = functional_quantization_mode(
+            model, {torch.nn.functional.scaled_dot_product_attention: ScaledDotProductAttention})
+    else:
+        quantization_cm = nullcontext()
 
-    # We restore the original behaviour of the post-forward.
-    for k, v in dict_hooks.items():
-        k._hf_hook.post_forward = v
+    with quantization_cm:
+        with torch.no_grad():
+            model(**calibration_loader[0])
 
-    if args.act_calibration and not args.load_checkpoint:
-        print("Apply act calibration...")
-        apply_calibration(model, calibration_loader)
-        print("Act calibration applied.")
+        # We restore the original behaviour of the post-forward.
+        for k, v in dict_hooks.items():
+            k._hf_hook.post_forward = v
 
-    if args.learned_round:
-        print("Applying learned round...")
-        if args.load_checkpoint:
-            iters = 1
-            loader = [calibration_loader[0]]
-        else:
-            iters = args.learned_round_iters
-            loader = calibration_loader
-        remove_hooks(model)
-        apply_learned_round(
-            model,
-            loader,
-            iters=iters,
-            block_name_attribute=args.gpxq_block_name,
-            learn_scale=args.learned_round_scale,
-            scale_optimizer_class='sgd',
-            optimizer_kwargs={'lr': args.learned_round_lr},
-            scale_optimizer_kwargs={
-                'lr': args.learned_round_scale_lr, 'momentum': args.learned_round_scale_momentum},
-            fast_update=args.learned_round_fast_update)
-        print("Learned round applied.")
+        if args.act_calibration and not args.load_checkpoint:
+            print("Apply act calibration...")
+            apply_calibration(model, calibration_loader)
+            print("Act calibration applied.")
 
-        model = offload_model(model)
+        if args.learned_round:
+            print("Applying learned round...")
+            if args.load_checkpoint:
+                iters = 1
+                loader = [calibration_loader[0]]
+            else:
+                iters = args.learned_round_iters
+                loader = calibration_loader
+            remove_hooks(model)
+            apply_learned_round(
+                model,
+                loader,
+                iters=iters,
+                block_name_attribute=args.gpxq_block_name,
+                learn_scale=args.learned_round_scale,
+                scale_optimizer_class='sgd',
+                optimizer_kwargs={'lr': args.learned_round_lr},
+                scale_optimizer_kwargs={
+                    'lr': args.learned_round_scale_lr,
+                    'momentum': args.learned_round_scale_momentum},
+                fast_update=args.learned_round_fast_update)
+            print("Learned round applied.")
+            model = offload_model(model)
 
-    if args.load_checkpoint:
+        if args.load_checkpoint:
+            remove_hooks(model)
+            with load_quant_model_mode(model):
+                model.load_state_dict(torch.load(args.checkpoint_name, map_location='cpu'))
+            model = offload_model(model)
+
+        if args.gptq and not args.load_checkpoint:
+            print("Applying GPTQ...")
+            apply_gptq(
+                model,
+                calibration_loader,
+                act_order=args.gpxq_act_order,
+                use_quant_activations=args.gpxq_use_quant_activations,
+                create_weight_orig=args.gpxq_create_weight_orig,
+                block_name=args.gpxq_block_name,
+                max_accumulator_bit_width=args.gpxq_max_accumulator_bit_width,
+                max_accumulator_tile_size=args.gpxq_max_accumulator_tile_size)
+            print("GPTQ applied.")
+
+        if args.gpfq and not args.load_checkpoint:
+            print("Applying GPFQ...")
+            apply_gpfq(
+                model,
+                calibration_loader,
+                act_order=args.gpxq_act_order,
+                block_name=args.gpxq_block_name,
+                max_accumulator_bit_width=args.gpxq_max_accumulator_bit_width,
+                max_accumulator_tile_size=args.gpxq_max_accumulator_tile_size)
+            print("GPFQ applied.")
+
+        if args.bias_corr and not args.load_checkpoint:
+            print("Applying bias correction...")
+            apply_bias_correction(model, calibration_loader)
+            print("Bias correction applied.")
+
+            if args.checkpoint_name is not None:
+                print(f"Saving checkpoint to {args.checkpoint_name}")
+                torch.save(model.state_dict(), args.checkpoint_name)
+
+        if args.eval and not args.no_quantize:
+            print("Model eval...")
+            with torch.no_grad(), quant_inference_mode(model):
+                model(**calibration_loader[0])
+                quant_ppl = compute_perplexity(
+                    model, validation_loader, context_length=args.seqlen // 2, tokenizer=tokenizer)
+            print(f"Quantized perplexity ({args.dataset}): {quant_ppl:.3f}")
+
+        if args.few_shot_eval:
+            with torch.no_grad(), quant_inference_mode(model):
+                model(**calibration_loader[0])
+                if args.few_shot_compile:
+                    remove_hooks(model)
+                    model.cuda()
+                    model = torch.compile(model)
+
+                wrapped_model = HFLM(pretrained=model)  # need to wrap for LLM eval
+                results = evaluator.simple_evaluate(
+                    model=wrapped_model,
+                    model_args=None,
+                    tasks=list(args.few_shot_tasks),
+                    device='cuda:0',
+                    limit=args.few_shot_limit,
+                    num_fewshot=0 if args.few_shot_zeroshot else None,
+                    log_samples=False,
+                    batch_size=None,
+                    verbosity="ERROR")
+            results = filter_results(results, args.few_shot_tasks)
+            print("Few shot eval results")
+            print(results)
         remove_hooks(model)
-        with load_quant_model_mode(model):
-            model.load_state_dict(torch.load(args.checkpoint_name, map_location='cpu'))
-        model = offload_model(model)
 
-    if args.gptq and not args.load_checkpoint:
-        print("Applying GPTQ...")
-        apply_gptq(
-            model,
-            calibration_loader,
-            act_order=args.gpxq_act_order,
-            use_quant_activations=args.gpxq_use_quant_activations,
-            create_weight_orig=args.gpxq_create_weight_orig,
-            block_name=args.gpxq_block_name,
-            max_accumulator_bit_width=args.gpxq_max_accumulator_bit_width,
-            max_accumulator_tile_size=args.gpxq_max_accumulator_tile_size)
-        print("GPTQ applied.")
-
-    if args.gpfq and not args.load_checkpoint:
-        print("Applying GPFQ...")
-        apply_gpfq(
-            model,
-            calibration_loader,
-            act_order=args.gpxq_act_order,
-            block_name=args.gpxq_block_name,
-            max_accumulator_bit_width=args.gpxq_max_accumulator_bit_width,
-            max_accumulator_tile_size=args.gpxq_max_accumulator_tile_size)
-        print("GPFQ applied.")
-
-    if args.bias_corr and not args.load_checkpoint:
-        print("Applying bias correction...")
-        apply_bias_correction(model, calibration_loader)
-        print("Bias correction applied.")
-
-    if args.eval and not args.no_quantize:
-        print("Model eval...")
-        with torch.no_grad(), quant_inference_mode(model):
-            model(**calibration_loader[0])
-            quant_ppl = compute_perplexity(
-                model, validation_loader, context_length=args.seqlen // 2, tokenizer=tokenizer)
-        print(f"Quantized perplexity ({args.dataset}): {quant_ppl:.3f}")
+        if args.checkpoint_name is not None and not args.load_checkpoint:
+            print(f"Saving checkpoint to {args.checkpoint_name}")
+            torch.save(model.state_dict(), args.checkpoint_name)
 
-    if args.few_shot_eval:
-        with torch.no_grad(), quant_inference_mode(model):
-            model(**calibration_loader[0])
-            if args.few_shot_compile:
-                remove_hooks(model)
-                model.cuda()
-                model = torch.compile(model)
-
-            wrapped_model = HFLM(pretrained=model)  # need to wrap for LLM eval
-            results = evaluator.simple_evaluate(
-                model=wrapped_model,
-                model_args=None,
-                tasks=list(args.few_shot_tasks),
-                device='cuda:0',
-                limit=args.few_shot_limit,
-                num_fewshot=0 if args.few_shot_zeroshot else None,
-                log_samples=False,
-                batch_size=None,
-                verbosity="ERROR")
-        results = filter_results(results, args.few_shot_tasks)
-        print("Few shot eval results")
-        print(results)
-    remove_hooks(model)
-
-    if args.checkpoint_name is not None and not args.load_checkpoint:
-        print(f"Saving checkpoint to {args.checkpoint_name}")
-        torch.save(model.state_dict(), args.checkpoint_name)
-
-    if args.export_target:
-        print(f"Export to {args.export_target}")
-        # Currently we always export on CPU with a float32 container to avoid float16 CPU errors
-        model = model.to(dtype=torch.float32)
-        model_export(model, calibration_loader[0], args)
+        if args.export_target:
+            print(f"Export to {args.export_target}")
+            # Currently we always export on CPU with a float32 container to avoid float16 CPU errors
+            model = model.to(dtype=torch.float32)
+            model_export(model, calibration_loader[0], args)
 
     return float_ppl, quant_ppl, model
 
@@ -762,6 +785,12 @@ def parse_args(args, override_defaults={}):
         '--quant-sdpa',
         action='store_true',
         help='Quantize `F.scaled_dot_product_attention` (default: %(default)s)')
+    parser.add_argument(
+        '--functional-sdpa-quant',
+        action='store_true',
+        help=
+        'Quantize `F.scaled_dot_product_attention` with stateless module and torch_function (default: %(default)s)'
+    )
     parser.add_argument(
         '--replace-mha',
         action='store_true',