Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add FP16 support to ptq_evaluate.py and update README argument list #1174

Merged
merged 2 commits into from
Feb 7, 2025
Merged

Add FP16 support to ptq_evaluate.py and update README argument list #1174

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
26412f8
Added FP16 support to ptq_evaluate.py and updated README
hkayann Feb 5, 2025
ebe901b
Renamed half to float16 for consistency
hkayann Feb 6, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion .pre-commit-config.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,7 @@
default_language_version:
python: python3

default_stages: [commit]
default_stages: [pre-commit]
hkayann marked this conversation as resolved.
Show resolved Hide resolved

exclude: '(src/brevitas/backport|docs)'

Expand Down
81 changes: 58 additions & 23 deletions src/brevitas_examples/imagenet_classification/ptq/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,7 +71,7 @@ usage: ptq_evaluate.py [-h] --calibration-dir CALIBRATION_DIR --validation-dir
[--batch-size-validation BATCH_SIZE_VALIDATION]
[--export-dir EXPORT_DIR] [--gpu GPU]
[--calibration-samples CALIBRATION_SAMPLES]
[--model-name ARCH] [--dtype {float,bfloat16}]
[--model-name ARCH] [--dtype {float,bfloat16,half}]
[--target-backend {fx,layerwise,flexml}]
[--scale-factor-type {float_scale,po2_scale}]
[--act-bit-width ACT_BIT_WIDTH]
Expand All @@ -82,13 +82,19 @@ usage: ptq_evaluate.py [-h] --calibration-dir CALIBRATION_DIR --validation-dir
[--weight-quant-type {sym,asym}]
[--weight-quant-granularity {per_tensor,per_channel,per_group}]
[--act-quant-granularity {per_tensor,per_group}]
[--weight-quant-calibration-type {stats,mse}]
[--weight-quant-calibration-type {stats,mse,hqo}]
[--act-equalization {fx,layerwise,None}]
[--act-quant-calibration-type {stats,mse}]
[--act-scale-computation-type {static,dynamic}]
[--graph-eq-iterations GRAPH_EQ_ITERATIONS]
[--learned-round {None,linear_round,hard_sigmoid_round,sigmoid_round}]
[--learned-round-block-name LEARNED_ROUND_BLOCK_NAME]
[--learned-round-loss {regularised_mse,mse}]
[--learned-round-mode {layerwise,blockwise}]
[--learned-round-iters LEARNED_ROUND_ITERS]
[--learned-round-lr-scheduler {None,linear}]
[--learned-round-lr LEARNED_ROUND_LR]
[--learned-round-batch-size LEARNED_ROUND_BATCH_SIZE]
[--act-quant-percentile ACT_QUANT_PERCENTILE]
[--export-onnx-qcdq] [--export-torch-qcdq]
[--bias-corr | --no-bias-corr]
Expand All @@ -102,22 +108,24 @@ usage: ptq_evaluate.py [-h] --calibration-dir CALIBRATION_DIR --validation-dir
[--weight-exponent-bit-width WEIGHT_EXPONENT_BIT_WIDTH]
[--act-mantissa-bit-width ACT_MANTISSA_BIT_WIDTH]
[--act-exponent-bit-width ACT_EXPONENT_BIT_WIDTH]
[--accumulator-bit-width ACCUMULATOR_BIT_WIDTH]
[--gpxq-accumulator-bit-width GPXQ_ACCUMULATOR_BIT_WIDTH]
[--gpxq-accumulator-tile-size GPXQ_ACCUMULATOR_TILE_SIZE]
[--onnx-opset-version ONNX_OPSET_VERSION]
[--channel-splitting-ratio CHANNEL_SPLITTING_RATIO]
[--compression-rate COMPRESSION_RATE]
[--gptq | --no-gptq] [--gpfq | --no-gpfq]
[--gpfa2q | --no-gpfa2q]
[--optimizer {adam,sign_sgd}] [--gptq | --no-gptq]
[--gpfq | --no-gpfq]
[--gpxq-act-order | --no-gpxq-act-order]
[--learned-round | --no-learned-round]
[--gptq-use-quant-activations | --no-gptq-use-quant-activations]
[--gpxq-create-weight-orig | --no-gpxq-create-weight-orig]
[--calibrate-bn | --no-calibrate-bn]
[--channel-splitting-split-input | --no-channel-splitting-split-input]
[--merge-bn | --no-merge-bn]
[--uint_sym_act_for_unsigned_values | --no-uint_sym_act_for_unsigned_values]
[--uint-sym-act-for-unsigned-values | --no-uint-sym-act-for-unsigned-values]
[--compile | --no-compile]

PyTorch ImageNet PTQ Validation

optional arguments:
options:
-h, --help show this help message and exit
--calibration-dir CALIBRATION_DIR
Path to folder containing Imagenet calibration folder
Expand Down Expand Up @@ -159,8 +167,9 @@ optional arguments:
vgg16 | vgg16_bn | vgg19 | vgg19_bn | vit_b_16 |
vit_b_32 | vit_h_14 | vit_l_16 | vit_l_32 |
wide_resnet101_2 | wide_resnet50_2 (default: resnet18)
--dtype {float,bfloat16}
Data type to use
--dtype {float,bfloat16,half}
Data type to use (float for FP32, bfloat16, or half
for FP16)
--target-backend {fx,layerwise,flexml}
Backend to target for quantization (default: fx)
--scale-factor-type {float_scale,po2_scale}
Expand All @@ -182,7 +191,7 @@ optional arguments:
Weight quantization type (default: per_tensor)
--act-quant-granularity {per_tensor,per_group}
Activation quantization type (default: per_tensor)
--weight-quant-calibration-type {stats,mse}
--weight-quant-calibration-type {stats,mse,hqo}
Weight quantization calibration type (default: stats)
--act-equalization {fx,layerwise,None}
Activation equalization type (default: None)
Expand All @@ -195,11 +204,25 @@ optional arguments:
--graph-eq-iterations GRAPH_EQ_ITERATIONS
Numbers of iterations for graph equalization (default:
20)
--learned-round {None,linear_round,hard_sigmoid_round,sigmoid_round}
Learned round type (default: None)
--learned-round-block-name LEARNED_ROUND_BLOCK_NAME
Block name for learned round. It works only if FX is
not needed (default: layer\d+)
--learned-round-loss {regularised_mse,mse}
Learned round type (default: none)
--learned-round-mode {layerwise,blockwise}
Learned round mode (default: none)
--learned-round-iters LEARNED_ROUND_ITERS
Numbers of iterations for learned round for each layer
(default: 1000)
--learned-round-lr-scheduler {None,linear}
Learning rate scheduler for learned round (default:
None)
--learned-round-lr LEARNED_ROUND_LR
Learning rate for learned round (default: 1e-3)
--learned-round-batch-size LEARNED_ROUND_BATCH_SIZE
Learning rate for learned round (default: 1)
--act-quant-percentile ACT_QUANT_PERCENTILE
Percentile to use for stats of activation quantization
(default: 99.999)
Expand Down Expand Up @@ -243,26 +266,36 @@ optional arguments:
--act-exponent-bit-width ACT_EXPONENT_BIT_WIDTH
Exponent bit width used with float quantization for
activations (default: 3)
--accumulator-bit-width ACCUMULATOR_BIT_WIDTH
Accumulator Bit Width for GPFA2Q (default: None)
--gpxq-accumulator-bit-width GPXQ_ACCUMULATOR_BIT_WIDTH
Accumulator Bit Width for GPxQ (default: None)
--gpxq-accumulator-tile-size GPXQ_ACCUMULATOR_TILE_SIZE
Accumulator tile size for GPxQ (default: None)
--onnx-opset-version ONNX_OPSET_VERSION
ONNX opset version
--channel-splitting-ratio CHANNEL_SPLITTING_RATIO
Split Ratio for Channel Splitting. When set to 0.0,
Channel Splitting will not be applied. (default: 0.0)
--compression-rate COMPRESSION_RATE
Specify compression rate < 1.0 for random projection.
Default is 0.0 and does not use RP.
--optimizer {adam,sign_sgd}
Optimizer to use with learnable rounding (default:
adam)
--gptq Enable GPTQ (default: disabled)
--no-gptq Disable GPTQ (default: disabled)
--gpfq Enable GPFQ (default: disabled)
--no-gpfq Disable GPFQ (default: disabled)
--gpfa2q Enable GPFA2Q (default: disabled)
--no-gpfa2q Disable GPFA2Q (default: disabled)
--gpxq-act-order Enable GPxQ Act order heuristic (default: disabled)
--no-gpxq-act-order Disable GPxQ Act order heuristic (default: disabled)
--learned-round Enable Learned round (default: disabled)
--no-learned-round Disable Learned round (default: disabled)
--gptq-use-quant-activations
Enable Use quant activations for GPTQ (default:
disabled)
--no-gptq-use-quant-activations
Disable Use quant activations for GPTQ (default:
disabled)
--gpxq-create-weight-orig
Enable Maintain original weights for non-quant forward
pass (default: disabled)
--no-gpxq-create-weight-orig
Disable Maintain original weights for non-quant
forward pass (default: disabled)
--calibrate-bn Enable Calibrate BN (default: disabled)
--no-calibrate-bn Disable Calibrate BN (default: disabled)
--channel-splitting-split-input
Expand All @@ -275,12 +308,14 @@ optional arguments:
(default: enabled)
--no-merge-bn Disable Merge BN layers before quantizing the model
(default: enabled)
--uint_sym_act_for_unsigned_values
--uint-sym-act-for-unsigned-values
Enable Use unsigned act quant when possible (default:
enabled)
--no-uint_sym_act_for_unsigned_values
--no-uint-sym-act-for-unsigned-values
Disable Use unsigned act quant when possible (default:
enabled)
--compile Enable Use torch.compile (default: disabled)
--no-compile Disable Use torch.compile (default: disabled)
```

The script requires to specify the calibration folder (`--calibration-dir`), from which the calibration samples will be taken (configurable with the `--calibration-samples` argument), and a validation folder (`--validation-dir`).
Expand Down
5 changes: 4 additions & 1 deletion src/brevitas_examples/imagenet_classification/ptq/ptq_evaluate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,7 +87,10 @@ def validate_args(args):
choices=model_names,
help='model architecture: ' + ' | '.join(model_names) + ' (default: resnet18)')
parser.add_argument(
'--dtype', default='float', choices=['float', 'bfloat16'], help='Data type to use')
'--dtype',
default='float',
choices=['float', 'bfloat16', 'half'],
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think half and float16 are equivalent, and for consistency reasons I think I prefer float16.

If I am missing something, let me know please.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, they are functionally equivalent. My initial thought was that using half could help prevent typos, as the only difference is the letter b.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Now, I am trying to save custom-bit-width models, for example FP16 with mantissa 9 bits, exponent 6 bits etc, but seems like not possible given the available PyTorch dtypes.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Unfortunately I'm not sure I can fully help with the second issue, since we mostly focus on minifloat quantization with 8 bits of fewer.

In the meantime, would you mind changing half to float16? I understand the potential for typos but I still prefer trying to be consistent across the codebase, and we never (or rarely) use half instead of float16.

Thanks!

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I can simulate for now, so I have a working workaround which is something. I have done all the requested changes as requested. Many thanks again.

help='Data type to use (float for FP32, bfloat16, or half for FP16)')
parser.add_argument(
'--target-backend',
default='fx',
Expand Down