Skip to content

[Core] Support FP8 weight storage in unquantized linear and embedding layers #35694

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
lucaspirola wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[Core] Support FP8 weight storage in unquantized linear and embedding layers #35694

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c94480d
[Core] Support FP8 weight storage in unquantized linear and embedding…
lucaspirola Mar 2, 2026
8c523de
fixup: use get_default_dtype() instead of hardcoded bfloat16 for FP8 …
lucaspirola Mar 8, 2026
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
8 changes: 6 additions & 2 deletions vllm/model_executor/layers/linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -217,9 +217,13 @@ def apply(
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
weight = layer.weight
# Support FP8 weight storage: cast to compute dtype for GEMM
if weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2):
weight = weight.to(x.dtype)
if vllm_is_batch_invariant() and current_platform.is_cuda_alike():
return linear_batch_invariant(x, layer.weight, bias)
return dispatch_unquantized_gemm()(layer, x, layer.weight, bias)
return linear_batch_invariant(x, weight, bias)
return dispatch_unquantized_gemm()(layer, x, weight, bias)


class LinearBase(PluggableLayer):
Expand Down
20 changes: 18 additions & 2 deletions vllm/model_executor/layers/vocab_parallel_embedding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,10 +66,18 @@ def apply(
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
return dispatch_unquantized_gemm()(layer, x, layer.weight, bias)
weight = layer.weight
# Support FP8 weight storage: cast to compute dtype for GEMM
if weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2):
weight = weight.to(x.dtype)
return dispatch_unquantized_gemm()(layer, x, weight, bias)

def embedding(self, layer: torch.nn.Module, input_: torch.Tensor) -> torch.Tensor:
return F.embedding(input_, layer.weight)
output = F.embedding(input_, layer.weight)
# Support FP8 weight storage: cast to compute dtype after lookup
if output.dtype in (torch.float8_e4m3fn, torch.float8_e5m2):
output = output.to(torch.get_default_dtype())
return output


def pad_vocab_size(vocab_size: int, pad_to: int = DEFAULT_VOCAB_PADDING_SIZE) -> int:
Expand Down Expand Up @@ -429,6 +437,14 @@ def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
shape[output_dim] = self.num_embeddings_per_partition
param.materialize(tuple(shape), dtype=loaded_weight.dtype)

# If loaded weight is FP8, cast parameter to match so FP8 is
# preserved in memory (saves VRAM for e.g. embed_tokens).
if (
loaded_weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2)
and param.data.dtype != loaded_weight.dtype
):
param.data = torch.empty_like(param.data, dtype=loaded_weight.dtype)

# If parameter does not have output dim, then it should
# be copied onto all gpus (e.g. g_idx for act_order gptq).
if output_dim is None:
Expand Down