Skip to content

Use custom allreduce w/ torch.compile #2185

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

Jump to bottom
Merged
merged 1 commit into from
Nov 25, 2024
Merged

Use custom allreduce w/ torch.compile #2185

merged 1 commit into from
Nov 25, 2024

Conversation

merrymercy
Copy link
Contributor

@merrymercy merrymercy commented Nov 25, 2024 •
edited
Loading

It makes torch.compile with tensor parallelism much faster.

llama 3.1 8b w/ tp = 8

python3 -m sglang.bench_offline_throughput --model meta-llama/Llama-3.1-8B-Instruct --dataset-name random --num-prompt 1 --random-input 1024 --random-output 256 --random-range 1 --tp 8 --enable-torch-compile

before: 315.02 token/s
after: 457.22 token/s

@merrymercy merrymercy merged commit c4336b2 into main Nov 25, 2024
14 of 16 checks passed
@merrymercy merrymercy deleted the pr-fix-torch-compile branch November 25, 2024 22:55
@merrymercy merrymercy mentioned this pull request Nov 26, 2024
Closed
timethink pushed a commit to timethink/sglang that referenced this pull request Mar 9, 2025
@merrymercy
Use custom allreduce w/ torch.compile (sgl-project#2185)
59ae67f
@ZhuJiaqi9905
Copy link
Contributor

Since the setting of tp_group.ca_comm to None has been cancelled here, is it still necessary to backup ca_comm? Could we change the code of patch_model() to the following:

@contextmanager
def patch_model(
    model: torch.nn.Module,
    enable_compile: bool,
    num_tokens: int,
):
    """Patch the model to make it compatible with with torch.compile"""
   

    try:
        if enable_compile:
            _to_torch(model, reverse=False, num_tokens=num_tokens)
            
            # Use custom-allreduce here.
            # We found the custom allreduce is much faster than the built-in allreduce in torch,
            # even with ENABLE_INTRA_NODE_COMM=1.

            yield torch.compile(
                torch.no_grad()(model.forward),
                mode=os.environ.get(
                    "SGLANG_TORCH_COMPILE_MODE", "max-autotune-no-cudagraphs"
                ),
                dynamic=False,
            )
        else:
            yield model.forward
    finally:
        if enable_compile:
            _to_torch(model, reverse=True, num_tokens=num_tokens)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@Ying1123 Ying1123 Awaiting requested review from Ying1123 Ying1123 is a code owner

@hnyls2002 hnyls2002 Awaiting requested review from hnyls2002 hnyls2002 is a code owner

@zhyncs zhyncs Awaiting requested review from zhyncs zhyncs is a code owner

@ispobock ispobock Awaiting requested review from ispobock ispobock is a code owner

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@merrymercy @ZhuJiaqi9905