Skip to content

mx4 quantize: compute per-group scale via exact FP32 bit-construction instead of FP64 exp2 (#6083) #6083

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

Open
qchip wants to merge 1 commit into
base: main
Choose a base branch
from
Open

mx4 quantize: compute per-group scale via exact FP32 bit-construction instead of FP64 exp2 (#6083) #6083

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 10 additions & 1 deletion fbgemm_gpu/fbgemm_gpu/triton/quantize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -224,7 +224,16 @@ def _kernel_quantize_mx4(
group_exp = tl.clamp(group_exp, -127, 125)

# Next we scale A in preparation for quantization.
scale = tl.exp2(group_exp.to(tl.float64)).to(tl.float32)
# Compute scale = 2^group_exp exactly via FP32 bit construction instead of a slow
# per-group FP64 exp2. group_exp is an integer in [-127, 125]; for e >= -126, 2^e is
# a normal fp32 (zero mantissa) so we set the exponent field directly. e == -127 ->
# 2^-127 is subnormal and cannot be built that way, so special-case it.
group_exp_int = group_exp.to(tl.int32)
scale = tl.where(
group_exp_int >= -126,
((group_exp_int + 127) << 23).to(tl.float32, bitcast=True),
2.0**-127,
)
# Apply scale to input. We do this by broadcasting scale.
scaled_a = tl.reshape(a, [GROUP_LOAD, GROUP_SIZE]) / tl.reshape(
scale, [GROUP_LOAD, 1]
Expand Down
Loading