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(l4): fix fp8 logic on l4 #2277

Merged
merged 3 commits into from
Jul 23, 2024
Merged

fix(l4): fix fp8 logic on l4 #2277

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
3d0c7b8
fix(l4): fix fp8 logic on l4
OlivierDehaene Jul 22, 2024
473f968
also quant weights with single scale
OlivierDehaene Jul 22, 2024
025f80d
use marlin even on 89
OlivierDehaene Jul 23, 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
17 changes: 11 additions & 6 deletions server/text_generation_server/layers/fp8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,8 +32,8 @@ def get_fp8_linear() -> torch.nn.Module:
"""

if SYSTEM == "cuda":
major, minor = torch.cuda.get_device_capability()
if major == 8 and minor < 9:
major, _ = torch.cuda.get_device_capability()
if major == 8:
from text_generation_server.layers.marlin import GPTQMarlinFP8Linear

return GPTQMarlinFP8Linear
Expand All @@ -42,8 +42,10 @@ def get_fp8_linear() -> torch.nn.Module:
return Fp8Linear


def fp8_quantize(weight, scale_upper_bound=None, qdtype=torch.float8_e4m3fn):
if FBGEMM_DYN_AVAILABLE:
def fp8_quantize(
weight, scale_upper_bound=None, qdtype=torch.float8_e4m3fn, scalar=False
):
if FBGEMM_DYN_AVAILABLE and not scalar:
qweight, scale = torch.ops.fbgemm.quantize_fp8_per_row(
weight, bs=None, scale_ub=scale_upper_bound, output_dtype=qdtype
)
Expand Down Expand Up @@ -186,6 +188,9 @@ def __init__(
dtype,
) -> None:
super().__init__()
if FBGEMM_MM_AVAILABLE:
log_once(logger.info, "Using FBGEMM fp8 optimized kernels")

self.dtype = dtype
self.qweight = qweight
self.scale = scale
Expand All @@ -201,7 +206,7 @@ def __init__(

@classmethod
def from_unquant(cls, weight, bias, dtype):
qweight, scale = fp8_quantize(weight)
qweight, scale = fp8_quantize(weight, scalar=not FBGEMM_MM_AVAILABLE)
return cls(
qweight=qweight, scale=scale, scale_upper_bound=None, bias=bias, dtype=dtype
)
Expand Down Expand Up @@ -232,7 +237,7 @@ def forward(self, input: torch.Tensor) -> torch.Tensor:
)
return y.to(self.dtype)

qinput, scale = fp8_quantize(input)
qinput, scale = fp8_quantize(input, scalar=True)
output, _ = torch._scaled_mm(
qinput,
self.qweight.t(),
Expand Down
Loading