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

Fix the bitness of addition/subtraction in quantize op #2234

Merged
merged 2 commits into from
May 1, 2024
Merged

Fix the bitness of addition/subtraction in quantize op #2234

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
77bce32
Fix the bitness of addition/subtraction in quantize op
sdasgup3 Apr 18, 2024
f6fe066
address feedback:I
sdasgup3 Apr 30, 2024
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
19 changes: 12 additions & 7 deletions docs/spec.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6691,13 +6691,18 @@ of the expressed type into corresponding integer values of the storage type
using the zero point and scale associated with the quantized element type.

```python
def quantize(x: Value, type: Type) -> Value:
assert is_float(x) and is_quantized(type)
x_expressed_rounded = round_nearest_even(x / compute_scales(type, type(x)))
x_storage_rounded = convert(x_expressed_rounded, storage_type(type))
x_storage_add = x_storage_rounded + compute_zero_points(type, type(x_storage_rounded))
x_storage = clamp(storage_min(type), x_storage_add, storage_max(type))
return bitcast_convert(x_storage, type)
def quantize(x: Value, result_type: Type) -> Value:
assert is_float(x) and is_quantized(result_type)
zero_points = compute_zero_points(result_type, TensorType(shape(x), storage_type(result_type)))
converted_zero_points = convert(zero_points, expressed_type(result_type))
converted_min = convert(storage_min(result_type), expressed_type(result_type))
converted_max = convert(storage_max(result_type), expressed_type(result_type))

x_scaled = x / compute_scales(result_type, type(x))
x_scaled_add_zp = x_scaled + converted_zero_points
x_clamped = clamp(converted_min, x_scaled_add_zp, converted_max)
x_rounded = round_nearest_even(x_clamped)
return convert(x_rounded, result_type)
```

* `dequantize_op_quantize` is used to specify element-wise computations on
Expand Down
Loading