Skip to content

[ROCm] Guard group quant RMS norm fusion patterns #30239

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

Merged
yeqcharlotte merged 1 commit into from
Dec 8, 2025
+19 −17
Merged

[ROCm] Guard group quant RMS norm fusion patterns #30239

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
36 changes: 19 additions & 17 deletions vllm/compilation/fusion.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -490,23 +490,25 @@ def __init__(self, config: VllmConfig):
# as the latter is a subset of the former in torch ops
for epsilon in [1e-5, 1e-6]:
# Fuse fused_add_rms_norm + fp8 group quant
FusedAddRMSNormGroupQuantPattern(
epsilon, FP8_DTYPE, group_shape=GroupShape(1, 128)
).register(self.patterns)

# Fuse rms_norm + fp8 group quant
RMSNormGroupQuantPattern(
epsilon, FP8_DTYPE, group_shape=GroupShape(1, 128)
).register(self.patterns)

FusedAddRMSNormGroupQuantPattern(
epsilon, FP8_DTYPE, group_shape=GroupShape(1, 64)
).register(self.patterns)

# Fuse rms_norm + fp8 group quant
RMSNormGroupQuantPattern(
epsilon, FP8_DTYPE, group_shape=GroupShape(1, 64)
).register(self.patterns)
# Only register group quant patterns on CUDA where the C++ op exists
if current_platform.is_cuda():
FusedAddRMSNormGroupQuantPattern(
epsilon, FP8_DTYPE, group_shape=GroupShape(1, 128)
).register(self.patterns)

# Fuse rms_norm + fp8 group quant
RMSNormGroupQuantPattern(
epsilon, FP8_DTYPE, group_shape=GroupShape(1, 128)
).register(self.patterns)

FusedAddRMSNormGroupQuantPattern(
epsilon, FP8_DTYPE, group_shape=GroupShape(1, 64)
).register(self.patterns)

# Fuse rms_norm + fp8 group quant
RMSNormGroupQuantPattern(
epsilon, FP8_DTYPE, group_shape=GroupShape(1, 64)
).register(self.patterns)
Comment on lines +495 to +511
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Dec 8, 2025

Choose a reason for hiding this comment

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

high

While this correctly guards the CUDA-specific patterns, there's an opportunity to reduce code duplication. The logic for group shapes 128 and 64 is identical. You can use a loop to register these patterns, which will make the code more concise and easier to maintain.

                for group_size in [128, 64]:
                    group_shape = GroupShape(1, group_size)
                    FusedAddRMSNormGroupQuantPattern(
                        epsilon, FP8_DTYPE, group_shape=group_shape
                    ).register(self.patterns)

                    # Fuse rms_norm + fp8 group quant
                    RMSNormGroupQuantPattern(
                        epsilon, FP8_DTYPE, group_shape=group_shape
                    ).register(self.patterns)


# Fuse fused_add_rms_norm + static fp8 quant
FusedAddRMSNormStaticQuantPattern(epsilon, FP8_DTYPE).register(
Expand Down
Loading